Beruflich Dokumente
Kultur Dokumente
Editors
Jan Kster
Henk van Riemsdijk
Mouton de Gruyter
Berlin New York
by
Carmen Dobrovie-Sorin
Mouton de Gruyter
Berlin New York
1994
Data
Dobrovie-Sorin, Carmen,
The Syntax of Romanian : comparative studies in Romance /
Carmen Dobrovie-Sorin,
p.
cm. (Studies in generative grammar ; 40)
Originally presented as the author's thesis (These d'Etat).
Includes bibliographical references and index.
ISBN 3-11-013541-8 (cloth)
1. Romanian language Syntax. 2. Romanian language
Verb. 3. Romanian languageGrammar, Generative.
4. Romanian languageGrammar, ComparativeRomance.
5. Romance languages Grammar, Comparative Romanian.
6. Romance languages Syntax. 7. Romance languagesVerb.
8. Romance languages Grammar, Generative.
I. Title. II. Series.
PC725.D63
1993
459'.5 dc20
93-5554
CIP
Data
Dobrovie-Sorin, Carmen:
The syntax of Romanian : comparative studies in romance / by Carmen
Dobrovie-Sorin. Berlin ; New York : Mouton de Gruyter, 1993
(Studies in generative grammar ; 40)
ISBN 3-11-013541-8
NE: GT
Ce serait entreprendre le rcit d'un cauchemar que de vous raconter par le menu
l'histoire de mes relations avec cet idiome d'emprunt [le frangais], avec tous ces
mots pensis et repenss, affinis, subtils jusqu' l'inexistence, courbis sous les
exactions de la nuance, inexpressifs pour avoir tout exprim, effrayants de
prdcision, chargis de fatigue et de pudeur, discrets jusque dans la vulgarity [...]
plus aucune trace de terre, de sang, d'me en eux. Une syntaxe d'une raideur,
d'une dignity cadavirique les enserre et leur assigne une place d'ou Dieu meme ne
pourrait les ddloger. [...] sans quoi jamais je n'eusse abandonnd la notre [le
roumain], dont il m'arrive de regretter l'odeur de fraicheur et de pourriture, le
mlange de soleil et de bouse, la laideur nostalgique, le superbe d6braillement
. M. Cioran, Histoire et Utopie
Preface
This book started out as a translation of my Thfese d'Etat, but it has become a
quite different piece of work. Chapters 1 through 4, as well as chapter 7 are entirely
new. The central ideas of chapters 5 and 6 go back to my dissertation, but have
been revised in essential ways. Chapter 6 is reprinted from Linguistic Inquiry
(1990, 3) without any change in content; because of technical difficulties, the
numbering of examples differs from the version in Linguistic Inquiry, but
everything else is identical.
During the four years I spent writing this book, I greatly benefited from
opportunities to present my ideas to other people. Preliminary versions of the whole
book were presented in various talks in Paris, in a research course at Stony Brook
(1992) and in talks given at the University of Bucarest and the Romanian Academy
of Sciences. These visits were made possible by financial support from the CNRS.
During my stays in Bucarest (1990-1992), Stony Brook and (1991-1992),
discussions with Noam Chomsky, Pusi Cornilescu, Dan Finer, Irene Heim, Jim
Higginbotham, Peter Ludlow, Richard Larson, Eric Reuland and Anna Szabolcsi,
were especially fruitful. The central idea of Chapter 7 was first presented at the LF
Conference in Tilburg (1988), and in a more refined form in Le Deuxifeme Congrs
des Langues Romanes (Groningen 1992), GLOW 1992 (Lisbon), and talks given
in MIT, Stony Brook and New Jersey (1992). Parts of chapter 5 were presented in
Going Romance (Utrecht 1991) and Linguistica Romena Oggi (Venice 1992).
Chapters 1 and 2 were presented at CUNY. I am extremely grateful to the
audiences at these events, notably Guglielmo Cinque, Giuliana Giusti, Ruohmei
Hsieh, Richie Kayne, Tony Kroch, Pino Longobardi, Ken Safir and Eriko SatoZhu for challenging questions and insightful criticism.
Several people gave me helpful comments on preliminary drafts of the
manuscript. Among them, I am especially indebted to Anna Cardinaletti, Dan Finer,
Daniele Godard, Sandu Grosu, Maria Teresa Guasti, Jacqueline Guiron, Sarah
Kennelly, Richard Larson, Peter Ludlow, Lea Nash, Hans Obenauer, Georges
Rebuschi, Henk van Riemsdijk, Maria-Luisa Rivero, Isabelle Simatos and Elisabeth
Villalta, Anne Zribi-Hertz. The acknowledgments for Chapter 6 are reproduced
with it. Special thanks go to Antoine Culioli, Richard Kayne and Jean Claude
Milner for having shown me how to become a linguist and to Irene Heim and Anna
Szabolcsi for their invaluable help with Chapter 7 .
The manuscript was more or less completed in September 1990, and circulated
in preliminary versions since then. In preparing the final text, I received editorial
assistance from Mouton de Gruyter Publishers, and technical assistance from the
staff of the Linguistics Department of the University of Paris 7. Finally, I would
like to thank Sarah Kennelly and Kristin Stromberg for having tried to improve the
style of my English and Elisabeth Villalta for having compiled the Index.
Table of Contents
IX
Table of Contents
Preface
Introduction
1. Auxiliaries and the structure of IP constituents
VII
XIII
1
2
5
6
12
14
15
15
15
18
20
23
24
24
26
27
29
31
33
35
36
37
38
39
40
40
41
45
47
49
49
50
53
55
57
59
61
2. 2. Romanian clitics
2. 2. 1. Deriving linear order from hierarchical structure
2. 2. 2. A definition of clitics
2. 3. The rule of V-preposing and clitic Merging
2. 3. 1. Adverbial clitics and V-preposing
2.3.2. Move I lands in Comp
2. 3. 3. Negation
2. 3. 4. Proclisis, enclisis and Merging
Conclusions
2. 4. Move I-to-C (V-preposing) in auxiliary structures
2. 4. 1. The distribution of clitics in auxiliary structures
2. 4. 2. Auxiliary inversion and endoclitic pronouns
Conclusions .
Appendix: A diachronic note: early Romance inverted conjugations
62
63
65
66
66
67
68
70
72
72
73
78
79
80
82
82
82
87
91
93
93
98
104
106
107
109
112
112
114
115
118
120
121
122
126
126
128
Introduction
5. 1. Passives with (in)transitives
5. 1. 1. The data
128
129
129
Table of Contents XI
145
153
154
167
168
169
172
175
175
175
179
183
185
187
193
196
6.
6.
6.
6.
6.
197
197
202
204
132
137
140
210
218
218
220
223
223
224
229
234
236
236
239
240
240
242
242
243
244
245
248
249
250
251
254
257
259
260
260
261
262
263
264
264
265
265
266
268
271
272
273
273
274
274
275
276
Bibliography
Index of names
Index of subjects
278
290
293
Introduction XIII
Introduction
The chapters of this book are the best answers I have so far found to a number of
intriguing questions, whose common format is: why is a given phenomenon found
in Romanian but not in (the) other Romance languages? This type of question has
become central in generative grammar only recently, after a major theoretical shift
took place, which led from rule-based models towards what is currently called the
"principles and parameters" framework adopted in the Pisa lectures.
This conceptual move has fundamentally modified our hypotheses concerning
the common properties of natural languages. Within the older system, the various
grammars of natural languages were viewed as sets of rules, completely distinct
from each other, developed through language acquisition. Questions of comparative
syntax could not be asked because the assumption was that any given set of rules
(and in particular a given natural language) was different from any other set of
rules: universal grammar was essentially a specification of the general format of
these rules. Within the present framework, the theory consists of general principles
which are supposed to be common across natural languages, and therefore they can
be tested on any particular natural language.
This view immediately raises the question of crosslinguistic variation. Why do
natural languages, viewed now as instantiations of general principles, appear to be
so different from one another? A plausible hypothesis is that languages differ from
each other by a restricted number of primitive differences which, given the
deductive power of the theory, have wide-ranging, apparently unconnected
consequences. The primitive differences are conceived of as different values of
open "parameters". Thus, the existence of parameters is an empirical hypothesis
inherent to the principles framework; parameters bridge the gap between the
general, supposedly universal, principles of language and the quite wide range of
observable crosslinguistic variation: Universal Grammar would contain, along with
general principles, a set of parameters, whose values (in general restricted to two,
positive and negative) are set differently across languages. One important goal of
current research is to determine the relevant parameters, through a careful
investigation of comparative data. Set against this background, the detailed study of
any particular language sheds light on the analysis of the others, and may
sometimes lead to revisions of the principles themselves.
The label "principles-and-parameters" thus appears to designate two distinct
things, on the one hand a highly articulated theory (a system of principles organized
in distinct modules: government, binding, X'-theory, etc.) and on the other hand an
important empirical hypothesis concerning language variation, and a research
program associated with it. Part of this research program is the formulation of the
theory of parameters itself: (a) what the general format of a parameter should be;
(b) on what kind of elements, structural configurations or relations it should bear;
(c) what levels or modules of the grammar are in principle open to parametric
variation; (d) how many values a parameter may take, etc. Tentative answers can be
found in the literature, but I do not think there is any emerging agreement on any of
these fundamental issues. This means that what makes it possible for generative
grammar to deal with comparative syntax is not the "parametric", but the
"principles" framework.
Introduction XV
One type of parameter that has been frequently proposed recently relies on
abstract features (I use the term "abstract feature" to refer to pairs of opposite
features that do not correspond to any overt, phonological difference) such as [+/strong] or [+/- pron] Infi, [+/- referential] or [+/- argumental] pronominal clitics,
etc. According to Aoun (1981, 1985) languages may differ from one another with
respect to the type of clitics that they present: four types can be defined on the basis
of features such as [+/-referential] and [+/- Case]; according to Cinque (1988), the
reflexive clitic se/si may be characterized as [+/- argumental], and Romance
languages differ from one another by the particular choice they make with respect
to these features. One problem with this kind of proposal is the fact that features
such as [+/- referential] or [+/- argumental] do not - and should not be allowed to count among the primitive notions of our theory; such features are quite often mere
abbreviations for structural observations. Thus, the label "[-referential] clitics" used
by Aoun designates clitics that are allowed to double R-expressions (and variables
in particular). This is a direct empirical observation concerning syntactic
configurations, and as far as I can see there is no evidence regarding the nature of
the clitic itself.2 Parameters stated in terms of abstract features of lexical items may
thus obscure structural differences. Analyses relying on abstract features are
currently presented as an "explanation" for a number of observable phenomena; it
seems to me that instead of an explanation we simply give an abstract name to the
observed phenomena, without any further understanding. In Chapters 5 and 6 of
this book it will be shown that for a number of cases not only do abstract features
fail to explain an observed phenomenon, but moreover they postulate a nonexistent
difference between languages or between two elements in a given language: it is I
think undebatable that, contrary to Aoun (1981, 1985), there is no difference in the
clitic systems of Romanian and (River Plate) Spanish.3
The use of "referential" features calls for further comment. I do not think that a
given linguistic element can be said to be either "referential" or "non referential"
independently of the syntactic context in which it appears. Therefore referential
properties should not be directly introduced as features characterizing linguistic
elements, but rather read off the syntactic representation. An adequate theory of
reference must provide (a) a formal procedure that freely assigns indices
(independently of referential properties); (b) well-formedness conditions on
configurations of indices; (c) interpretive procedures by which given configurations
of indices are associated with referential properties (for a specific proposal see
Chapter 5). Given the chain formalism, it is reasonable to assume that referential
properties are only relevant at the level of the chain, and not at the level of the
elements of the chain. Thus, even if we could use the features [+/-referential], we
still would not be able to use it for clitics, but only for clitic chains.
2
It is well-known that different distributional patterns of clitics do not necessarily correlate with
different types of clitics. According to "Kayne's generalization" - for which Jaeggli (1982) has
provided a principled explanation - it is indeed possible to maintain that the possibility of clitic
doubling is not due to the nature of the clitics, but rather to the existence of prepositional
Accusatives. In Dobrovie-Sorin (1990) reprinted here as Chapter 6,1 have shown that Romanian
clitics behave exactly as Spanish clitics regarding Jaeggli's generalization that variables (more
precisely variables bound by "weak" wh-phrases) cannot be clitic-doubled.
3This does not mean that Aoun's parametrization of clitic systems may not prove correct for other
languages.
One important disadvantage related to the use of abstract features is that it tends
to increase the number of cases of synonymy and homonymy in the syntax and the
morphology of natural languages, which decreases explanatory power (explanatory
accounts of linguistic phenomena aim at reducing synonymy and homonymy).
To make the point clear let me take an example. It is known that "passive
meaning" can be expressed in various ways, among which are middle/passive
morphemes (see Greek or Turkish), reflexive clitics (see se/si in Romance
languages) or copula verbs followed by past participles. All these constructions
have in common the absorption of the external th-role (the absorption of objective
Case is much less clear), and we may try to derive this characteristic from the
properties of the syntactic configurations that display it. Jaeggli's (1986a) proposal
(taken up recently by Baker, Johnson and Roberts (1989)) is that the the passive
suffix -en absorbs the external th-role (and also Objective case in certain
languages). The problem with this analysis is the existence of past participles which
do not trigger the passive phenomenon (see the present perfect/ past perfect
paradigms). One must then stipulate the existence of two types of homonymous
past participles, "passive" and "non passive" (cf. Baker, Johnson and Roberts
(1989)).4
This kind of analysis can of course apply to passive reflexives, and this has
indeed been proposed by Belletti (1982): their passive meaning would also be due
to a "passivizing" morpheme, the clitic se, which would be characterized in exactly
the same way as the past participle morpheme. Put together, Jaeggli's and Belletti's
hypotheses lead to the curious conclusion that the "passive" past participle
morpheme and "passive" se are synonyms 5 , and that the "passive" se and the
"reflexive" se are homonyms.
By avoiding the use of abstract features we may provide an analysis within
which semantics is interpretive, that is, it can be read off structures which have
been generated by a truly autonomous syntax, one which makes no reference to the
"passive" meaning of a given element. The syntactic representation should take into
account the distinct overt properties of the constructions under discussion: the verb
BE (we use capitals to refer to English be, but also to French etre, Romanian a fi,
etc.) for copula passives, and se for "reflexive" passives. None of these elements
should be stipulated as being a passivizer. We should also try to understand why se
(see also Greek morphological passives) is compatible with both the passive and
die reflexive meaning, unlike copula passives.
Cinque (1988) has observed an important difference concerning the distribution
of Italian si in [-finite] clauses, between on the one hand (in)transitives and on the
other hand unaccusatives, passives and copula constructions. This data suggests
that what was currently supposed to be just one item, namely the Italian
"impersonal" ("nominative", "subject") si, must in fact be analyzed as two items,
^There is in fact a difference between the two -en morphemes, namely the fact that only "passive"
-en necessarily takes AGR features, thus qualifying as an element (cf. Jaeggli (1986a: 592, fn
6)). However, it is not clear at all that this difference is due to intrinsic features, or rather to the
fact that in passives, as opposed to present perfect, -en appears in a copula construction. Note also
that in French and Italian the non-passive past participle morpheme takes AGR features in a
number of cases.
5
Note that according to J. C. Milner (class lectures) complete synonymy is probably non existent
in general; morphological synonymy is even more questionable than lexical synonymy.
Introduction
XVII
and Cinque proposes to distinguish them by the feature [+argument] versus [argument]. This analysis is difficult to accept, in so far as it postulates perfect
homonymy: two elements that present exactly the same overt properties are
supposed to be characterized by distinct "content" properties (i.e., by distinct
features). It is in fact possible to show that the distinction between +arg and -arg si
is an artefact. The important empirical data that this distinction is supposed to
capture can be treated in terms of the difference between the anaphoric "object" si
and the "subject" (or "nominative") si (see Chapter 5 ). The latter distinction is not a
case of homonymy, because Case constitutes a formal property, which can be used
to distinguish between linguistic elements: Nominative si and Accusative si are two
formally distinct entities, and not two homonyms. The question of why such a
difference should exist for phonologically identical elements is another problem
(diachronic reanalysis may be invoked, see Naro (1976)).
Let us now consider overt features. Notions such as "clitic" or "affix" are
morphophonological features that characterize certain lexical items as opposed to
others. Features of this kind provide important descriptive information, which may
help us in characterizing crosslinguistic variation. It is obvious that languages differ
from each other with respect to the presence of clitic elements: Romance languages,
as opposed to English, present pronominal clitics; Slavonic languages, as opposed
to Romance languages (with the exception of Romanian which in this respect
behaves as a Slavonic language) present verbal auxiliaries which have clitic status.
The presence/absence of (pronominal and/or verbal) clitics is a real crosslinguistic
difference, but a quite "superficial" one, which probably does not deserve the status
of "parametrical" option: if we have a good definition of clitics we may easily
classify languages with respect to the presence of these elements, but this
classification (which can be established independently of our theory of principles)
does not really help us understand crosslinguistic variation. We must therefore
pursue our investigation and try to see whether the existence of clitics can be
shown to follow from other parametrical choices. It seems that the distribution of
clitics is related to functional categories in general, and to Infi (or Tense) in
particular. It is thus, in principle, possible to account for the presence or absence of
clitics on the basis of properties of Tense, which are closely related to the structure
of IP constituents. It would then be interesting to derive the existence of clitic
elements as a consequence of a particular choice concerning IP structure (S-V-O,
S-O-V or V-S-O). This is the line of investigation that I will try to pursue in
Chapter 1: if I am correct, the fact that Romanian verbal auxiliaries are clitics is
related to the fact that this language is V-initial. More precisely, NP subjects are
generated as sisters to VP and V raises to Infi, bypassing the intervening subject;
the subject stays in its base-position - compare S-V-0 languages, which according
to Sportiche (1988) would be characterized similarly, but would furthermore
involve obligatory subject-raising to (Spec,I').
The recent research concerning IP constituents is characterized by the
proliferation of syntactic positions: not only is each inflectional morpheme
generated in a distinct syntactic position (hence syntactic nodes such as Tense,
Agreement, Aspect, Negation, etc.), but moreover each of these morphemes is
considered to be the head of a functional projection; hence AGRP, NegP, AspP,
etc., each of which takes its own Spec, or subject position. The stand that I have
taken in this book is more conservative with respect to two assumptions: (a) I take
AGR to be a functional head of "nominal" type, which could only project a DP-like
Introduction
XIX
The idea that intransitives may vary crosslinguistically with respect to their
Case-assigning properties has been suggested in order to account for a well-known
fact: intransitive verbs are allowed to passivize in some languages (see German,
Norwegian, Arabic, etc. and marginally French), as opposed to others (Romance
languages other than French, English, etc.) which do not. The problem is that by
parametrizing (in)transitivity we get rid of the puzzle rather than solve iL The same
remark applies to the other possible alternative, the parametrization of the Case
absorption properties of passive morphemes (see Baker (1988); Roberts (1985,
1987), among others). In Chapter 5, I argue instead that the passivizability of
intransitives need not be parametrized; the above mentioned crosslinguistic
differences are due to the existence of various types of impersonal constructions
(see il impersonals in French, inverted subjects in "pro-drop" Romance languages,
and es/impersonal "pro"constructions in German).
The proposed analysis of passive intransitives opens the way towards a better
understanding of the Romance construction that involves intransitive verbs used
with the reflexive pronoun se/si. According to a generalization unanimously
maintained by both traditional and generative scholars (see in particular Belletti
(1982)), "pro-drop" Romance languages present a subject/Nominative se/si, which
shows up, among other constructions, in se Vjntransitive sequences. In Chapter 5,
it is shown that this generalization should be revised. Although it is a "pro-drop"
language, Romanian does not have at its disposal "Nominative" se. The matter is
complicated by the fact that Romanian does present the construction se
^intransitive (se doarme 'se sleeps': "one sleeps", "they sleep"), which will be
analyzed as relying on middle/passivizing se.
The foregoing discussion anticipates arguments that show that certain proposed
parameters concerning the Lexicon are not needed; other parameters capture real
differences between linguistic elements (see the clitic status of pronouns and
auxiliaries in certain languages) but these differences can be shown to derive from
structural properties of syntactic constituents. Certain crosslinguistic differences
concerning the distribution of comparable elements in various languages should not
be interpreted as the consequence of some abstract difference between the elements
themselves, but rather as the consequence of the different structural properties of
their syntactic environment: V-initial IP constituents, various types of impersonals,
or the existence versus the absence of "null operator" configurations (Romanian
does not present i/iai-relatives, clefts, Topicalizations, etc). In Chapter 6, I argue
that the lack of "null operator" configurations explains why Romanian presents
systematic contrasts between "strong" and "weak" (in the sense of Milsark) whquantifiers. I believe that this result is correct, but it is clearly not the end of the
story: a "parameter" that distinguishes between languages with and without "null
operators" has no real theoretical status; it is at most a descriptive generalization
made possible by our theory. We must now try to understand better the notion of
"null operator" itself (or rather the type of licensing mechanism of those empty
"variables" which are not bound by lexical quantifiers). We must then try to
establish a relation between this type of licensing mechanism and other
crosslinguistic differences. A possible line of inquiry would be to try to establish a
relation between word-order typology (SVO, SOV, VSO, etc.), "clitic doubling
typology" (Romanian and Spanish, as opposed to French and Italian, allow
(in)direct objects to be doubled by a clitic pronoun) and the typology of whmovement. The "null operator option" is thus clearly not a "parameter" if by that
notion we mean a primitive difference between languages. The term "parameter" is,
however, currently used in a less strict way, to designate one difference that
accounts for a cluster of differences between languages. Such a parametrical
variation may then be shown to follow as a consequence of another, more abstract
parametrical option.
By restricting the parametrization of the Lexicon we reduce quite drastically the
number of possible parameters, because the possible variations in syntactic
structures are much fewer than the possible variations in the Lexicon. This book
eliminates a number of unjustified parameters concerning the Lexicon, and does not
propose any real "parameter", in the sense of primitive difference between
languages. My main aim has been to establish correlations between clusters of covarying crosslinguistic differences; I hope that further investigation will lead us
from these "intermediate/mid-way parameters" to the discovery of some primitive
differences between languages.
This constrained view of parametrization has led me to propose certain
theoretical refinements, which bear on the projection principle, on the theory of whmovement and quantification, and on the identification of empty categories and
pronouns. I shall very briefly summarize the main conclusions and refer the reader
to the various relevant chapters.
If we do not want to introduce unjustified parameters concerning either
intransitivity or passivization, the possibility of passive intransitives must no longer
be thought of as a "marked" characteristic of certain languages, but as a virtual
possibility of any language. In Chapter 5 evidence will be provided in favour of the
idea that the representation of passive intransitives necessarily involves the
presence of an empty category in the object position (which, depending on the type
of impersonal construction, may form a th-chain by itself, or belong to the same
chain as the subject position). Active intransitives, on the other hand, may appear
with overt cognate objects, but no empty category should be postulated when such
objects do not show up. This means that the projection principle is not alone
responsible for the positions projected in the syntax: syntactic configurations,
passives in particular, may force the instantiation of certain positions; as a result,
"implicit/prototypical" or empty "cognate" objects may surface in the syntax in
passive configurations, but in active structures they are saturated in the Lexicon.
The analysis of Romanian wA-structures proposed in Chapters 6 and 7 leads to
certain revisions of the GB theory of quantification, which appear to be
independently motivated by the behaviour of dislocated quantified expressions and
of quantified expressions in situ. My proposals concern the formal mechanisms on
which quantification structures rely. How do we define a quantified NP as opposed
to a "referential" NP? Are quantificational relations established between quantified
NPs and variables, or between nominal determiners and variables? In addition, I
propose new accounts for certain well-known phenomena, such as the systematic
interpretive contrasts between the specific versus nonspecific readings of
indefinites and other quantified NPs, and "weak" islands. My results depart from
the current view of quantification initiated by May (1977), but converge with work
by Cinque (1990), Pesetsky (1987), Williams (1986) and others.
Romanian auxiliary constructions show properties that set this language apart
among the other Romance languages. The main aim of this chapter is to establish
the correct underlying configuration of Romanian auxiliary constructions. I will
then try to show that a correlation exists between types of auxiliaries and types of
IP constituents. If this correlation is correct, the type of auxiliaries found in
Romanian constitutes an indirect argument in favour of the idea that Romanian IP
constituents are VC+ty-initial.1
Let us start by defining the notion of auxiliary:
(A)
The statement in (A)a defines the thematic properties of those auxiliaries that are
sometimes referred to as "functional"/"weak" (see Lema and Rivero (1989, 1990)),
"aspectual" (Chomsky (1986)b), or "temporal" (Bennis and Hoekstra (1988),
Gueron and Hoekstra (1988)). The term "auxiliary" will be used here in this
restrictive sense. The statement in (A)b is the structural characteristic of auxiliaries.
It is however well-known that this structural property can be found in other
elements, such as English modals and Infi, which differ from auxiliaries with
respect to (A)a: Infi is assumed to th-mark VP (see Chomsky (1986)b), and
English modals L-mark VP.
The structural definition of auxiliaries (and modals) goes back to Chomsky (1955):
auxiliary verbs are not sisters to V itself, but rather sisters to VP; 2 in other words,
auxiliaries are not inside, but outside the minimal VP constituent which dominates
the lexical verb. It is currently assumed (see in particular Akmajian, Steele and
Wasow (1979), Gueron and Hoekstra (1988) and Rouveret (1987)), that the
definition in (A)b is an empirical generalization that subsumes (besides English
modals) aspectual auxiliaries in Germanic and Romance languages (see have, the
French etre and avoir, etc.); it does not cover the passive auxiliary, nor English
causatives.
It will be shown (Section 1.1.) that Romanian auxiliaries cannot be analyzed in
terms of (A)b; they appear to be subsumed by the definition in (B):
(B)
The characteristic stated in (B) should not be understood as meaning that Romanian
auxiliaries are comparable to the French aller 'go' in the colloquial future
'(Spec, IP) is an A'-position, which can be occupied by subjects and non-subjects alike; (Spec,
VP) is the -position of subjects.
2
For other hypotheses concerning auxiliaries see Ross (1969), Emonds (1978) and Akmajian,
Steele and Wasow (1979).
construction, which from the structural point of view behaves as a raising verb.
Romanian auxiliaries will be shown to be different from raising verbs. Romanian
raising structures involve two IP constituents and two subject positions, whereas
the auxiliaries defined in (B) can be shown to appear in configurations in which
only one Infi element is active and only one subject position is projected (note that
by these properties the Romanian auxiliaries defined in (B) resemble the structural
auxiliaries defined in (A)b).
Romanian presents a modal verb, a putea 'may, can', which shows the structural
properties of Romance and Germanic auxiliaries (see (A)b):
(C)
An understanding of why (B) and (C) should hold relies on the "Tense filter", a
licensing requirement to which both auxiliaries and lexical verbs are subject.
1 . 1 . Auxiliary constructions
The auxiliaries to be discussed in this chapter are those found in the present perfect
("perfectul compus"), future and conditional paradigms. The passive auxiliary be
will not be considered, because it shows the morpho-syntactic behaviour of lexical
verbs.
Romanian and English have in common the use of auxiliary verbs that encode
grammatical information such as "future" and "conditional"; as in English, these
auxiliaries are followed by a bare infinitive, i.e., an infinitive devoid of the particle
a. For ease of reference I have glossed a as to, although the two particles do not
represent the same syntactic category (see Chapter 3).
(1)
It is not obvious that the Romanian examples in (1) should be analyzed on a par
with their English glosses. We might as well think that (l)a-b are comparable to the
French "futur proche", which is currently attributed a biclausal structure:
(2)
The main argument in favour of a biclausal structure comes from the distribution of
clitics. Assuming that clitics attach necessarily to an Infi node (see Chapter 2), the
position of the clitic in (2) indicates that the lower verb is governed by an Infi node;
hence an embedded IP constituent must be assumed in (2). This means that the verb
aller has the structural properties of raising verbs, not those of auxiliaries (see the
definition in (A)b); thus, to classify aller as an auxiliary is to take into account its
semantics (future meaning) rather than its structural properties (for further
discussion of the difference between structural auxiliaries and raising verbs see
Section 1.2.).
Returning to the Romanian examples in (1), clitics necessarily attach to the
higher verb,3 and this may be taken to indicate that there is no lower Infi node:
(3)
a. Mana il va invita
sigur.
sigur
Furthermore, Romanian future and conditional auxiliaries are not used as lexical
verbs, as opposed to aller in French. Both the Romanian future auxiliary voi/vei/..
and the verb a vrea 'want' stem from the same etymon, uolo. However, in modern
Romanian the paradigm of voi is distinct from the paradigm of a vrea 'to want' (see
(4) and (5)); (4) is also distinct from the paradigm of the lexical verb a voi 'to want'
(see (5)b), felt as an archaic variant of a vrea::4
(4)
eu voi / tu vei/el
va/noi
vorpleca
[future]
'to want!
The morphological differences between the paradigms (4) and (5) indicate clearly
that in modern Romanian the future auxiliary is an element whose use is restricted
to the future periphrastic construction. Likewise, the paradigm in (6)a, used in the
"perfectul compus" is distinct from the conjugated forms of a avea 'to have' (see
(6)b), although they are historically related:
(6)
a. am/ai/a/am/ap/au plecat
(I) have/ (you) have/ (he) has ... left
b. am/ai/are/avem/avepi/au
["perfectul compus"]
[a avea'to have']
Compare the French avoir or English have, which function both as auxiliaries and
as lexical verbs.
The case of the Romanian conditional is even clearer. The inflected auxiliary
element is a/ai/ar..., whose distribution is restricted to the conditional forms. It no
longer bears any relation to a avea 'to have', to which it is historically related:
(7)
[conditional]
The feminine clitic is a notable exception to this generalization, which will lead us to abandon
the hypothesis envisaged here (see Section 1.2.3. below).
4
The etymology of voi is not completely clear; it may go back to Latin uolo, or be of Slavonic
origin.
The verbs that fall under (8) are not all of the same type. English modals present
certain well-known formal characteristics that distinguish them from lexical verbs:
(a) their position with respect to the negation particle and adverbs; (b) participation
in subject-Aux inversion; (c) the lack of Agreement inflections and (d) the lack of
non-finite forms. This irregular morpho-syntactic behaviour has been captured by
assuming that English modals are not generated under V, but under a specific
position, whose label has changed from Aux (Chomsky (1957, 1965)) to Infi
(Chomsky (1981)):
(9)
a.
I'.
I
Aux
VP^
V
NP 0
The label Aux in (9) should not be mistaken for the node Aux in Chomsky (1965),
which has survived as Infi in the current theory (Chomsky (1981)). Infi is a more
inclusive category than the former Aux; the two notions are nevertheless essentially
comparable since they both have the status of a syntactic category. The Aux in (9)
is simply a label for a given lexical class, whose elements are defined in (A).
Unlike English modals, aspectual auxiliaries (see English have, French avoir and
ere, etc.) show the morpho-syntactic behaviour characteristic of lexical verbs5 and
are therefore assumed to be generated outside Infi, under a V node (for ease of
reference the VPs headed by an auxiliary will be notated AuxP; the reader should
recall that this label does not mean that we assume the existence of a distinct
syntactic category) that takes a VP complement: at S-structure French auxiliary
verbs move to Infi (as French verbs normally do, see Emonds (1978) and Pollock
(1989)):
(9)
b.
I'
NP,
What is the correct underlying structure of Romanian auxiliary structures? Are they
of the type given in either (9)a or (9)b?
5
Note however that non-finite auxiliaries behave differently from non-finite lexical verbs, cf.
Pollock (1989).
1. 1. 1. Auxiliaries
and Tense
features
Consider (10)-(11) and compare the corresponding English glosses with the French
and English paradigms in (12)-(13):
(10)
euam/tu
ai/el
a/noi
am/voi
ap/ei
au
mincat
mincasem
/ mlncasefi
/ mincase
/ mincasem
/ mincasefi
mincaser
a. tu as mang6
/tu
es
parti
a. j'avais
mang6/j'6tais
parti
a. eu voi
fi
mincat
I will BE eaten
will have eaten'
b . eu a fi
mincat
I would BE eaten
would have eaten'
These examples are not built with the perfect auxiliary am/ai/a..., used in (10), but
with the invariable auxiliary fi 'be',7 which appears with any kind of verb
^This means that auxiliaries can take Tense inflections only if they are generated outside the Infi
position, as in (9)b. The English modals should and would constitute a potential problem for
this generalization: they bear the Tense inflection -ed, but their distributional properties indicate
that they are generated under Infi. To solve this problem one may assume that should and would
are listed in the Lexicon as such and directly inserted under Infi (rather than deriving from the
raising of shall/will to an Infi position dominating -ed).
^Besides (14)a-b, fi is also used in perfect infinitivals and subjunctives:
(i)
a fi mincat
a.
Nuplecmiine.
(I) not leave tomorrow
b. *Plec nu miine.
c. Je ne pars pas demain.
I NE leave NOT tomorrow
In (15)c pas necessarily shows up after the verb; assuming that pas is generated in
front of VP, its S-structure position indicates that the verb has raised over pas
(compare French infinitives, in which the verb does not raise: ne pas partir/*ne
partirpas 'not pas leave'). The data in (15)a-b does not show that V-raising does not
apply in Romanian; they simply indicate that the negative particle nu is basegenerated in front of Infi, and not between Infi and VP. 9 Consider also the
'to have eaten'
a$ vrea sS fi mincat
would like s have eaten.'
8
See Emonds (1978), Pollock (1989), and the references cited there.
9
S e e Zanuttini (1989) who gives evidence in favour of the existence of two distinct types of
negative elements: (a) nu in Romanian, ne in French, or non in Italian would be functional heads
that take IP complements; (b) pas is an adverbial element, generated in front of VP.
(ii)
NegP
Neg
IP
At this point let me specify the analysis of V-to-I movement to be adopted below.
Following Roberts (1985) and Rizzi and Roberts (1989), I assume that the landing
site of V-to-I movement is created at D-structure, as a function of the lexical
properties of Infi: Infi morphologically subcategorizes for a V position (for
morphological subcategorization see Lieber (1980)), to which the verb raises.
Under this analysis V-to-I movement satisfies a strong version of the structurepreservation constraint, one which rules out the adjunction option: this means that
V cannot adjoin to the element dominated by Infi. If that element does not present a
morphologically subcategorized position, V cannot raise to Infi, but stays under
VP. I will assume that Head to Head movement in general is structure-preserving: a
head can substitute either in an empty head position (the standard case is I-to-C
movement, but one can imagine that V-to-I movement can also apply in case Infi is
empty) or in a morphologically subcategorized position (V-to-I movement). The
adjunction possibility is ruled out for Head to Head movement.10
In line with current research, I assume that functional categories are heads that
project functional projections (see Pollock (1989)); however, I do not think that all
functional projections have subject positions. Even if we do not have at our
disposal an explicit theory of subject positions, I would like to preserve the natural
idea that sentential constituents have just one subject position. In order to achieve
this, we would have to investigate the relation between Tense and the subject NP,
and the relations between the various functional categories: Tense, AGR and Neg
appear to cluster together, which may be taken to indicate that they are subject to an
incorporation process (see also Chapters 2 and 3) giving rise to an X constituent
labelled Infi: [i n fl Neg-V+Tense+AGR] ; correlatively, the projection of this
element would be IP, 11 and the subject would occupy the (Spec, IP) position.
lORizzi and Roberts (1989) note that cliticization may constitute an exception to this strong
version of structure preservation (see Chapter 2).
l^This analysis of functional projections could be theoretically supported by Grimshaw's (1991)
theory of "extended projection", under which NegP, TenseP, AgrP are all extended projections of
the verb.
A short note is in order here, concerning the position and status of AGR
inflections, which in Romanian can combine both with auxiliaries and with Tense
inflections (bound morphemes): in (4), (6) and (7) the auxiliary is inflected for
AGR and the lexical verb takes on the bare infinitive or the past participle form; in
(18) the lexical verb carries both Tense and AGR inflections:
(18) eu plec-a-m/tu plec-a-i/el plec-a- / noi plec-a-m / voi plec-a-p
I leave-imperf-lst / you leave-imperf-2nd / he leave-imperf-3rd ...
If we assume that each inflectional morpheme is generated in a distinct syntactic
position, we may add an AGR position to the structure in (17). However,
Romanian presents no clear evidence in favour of the idea that AGRP and TenseP
are two distinct maximal projections (see Pollock (1989)). We may even question
the idea that AGR is a syntactic head that projects a functional projection; AGR
may instead be viewed as an affix that is nominal in nature, on a par with
pronominal clitics. I take AGR to adjoin to Tense. As discussed above, adjunction
is not allowed for Head to Head movement, but base-generated adjunction is
allowed. Thus we obtain a structure in which AGR is higher than Infi (see also
Belletti (1990), as opposed to Pollock (1989)), which accounts for the surface word
order of AGR and Tense morphemes. (17) thus becomes (19), in which the Tense
node dominates either auxiliary verbs or Tense inflections.12 The latter present a
morphologically subcategorized position to which the verb raises:
AGR
Tense
Aux
^past/present J
Consider now the distribution of adverbs and floating quantifiers:
(20) a. (i)
12
Since Chomsky (1955) it has been assumed that the same position may host both bound and
free morphemes (the English Infi node may dominate modals, auxiliary do, -edand -s).
Example (21)a is ungrammatical because V-raising did not apply; compare (21)b,
obtained by V-raising.
Let us now examine the contrast between (20)c and (21)c. In the French example
the auxiliary has raised by V-to-I movement; the past participle itself is left in situ,
under VP: 14
(21') c. les eleves
Uux-TenseOnt]
A u
x P t A u x [VP
To account for the word order in (20)b we must assume that in Romanian the
lexical verb itself is subject to V-raising. Example (20)c is ungrammatical because
V-raising did not apply. The contrast between (20)b and (20)c thus indicates that in
Romanian V-raising necessarily applies to lexical verbs, even in the presence of
auxiliary verbs:
(20') b. Oelevii mei [Tensevor] [yscriei] [yp toji [yp t; poezie]]
As indicated by the diamond, the structure in (20')b is illegitimate, because V
movement violates the structure preservation constraint: there is no available head
position to which the verb may raise. The fact that (20')b is illicit accounts for a
reasonably well-established generalization, according to which the presence of an
13
The same conclusion concerning V-raising can be reached if we assume that floating quantifiers
are generated under the (Spec, NP) node of NP subjects, which would be themselves generated
inside or adjoined to VP (see Sportiche (1988)).
14
This is probably an oversimplification (see Appendix); we may assume that the past participle
morpheme is generated under an Infi node which governs VP (see Kayne (1987); Baker, Johnson
and Roberts (1989)).
The ungrammatically of (24)c indicates that the verb cannot raise over
monosyllabic adverbs. This does not necessarily mean that V-raising does not
apply. We could assume that the adverb is generated adjoined to the position to
which the verb raises (see below). Or alternatively we may suppose that the adverb
is base-generated in an adjunction position to the verb (see Rivero (1988a; to
appear)); the constituent Adv-V, where Adv belongs to the class of monosyllabic
adverbs enumerated above, would be subject to all the rules that affect verbs, V
raising out of VP in particular. Example (24)b could thus be represented as either
(24')b or (24")b:
(24') b. aIt dat nu [aux^J [mai [yp spunej [yp ti prostii]
tZ
These two structures have in common the idea that the monosyllabic adverbs that
show up in (24) are clitic elements (on the definition of clitics see Chapter 2). This
hypothesis is strongly supported by their distribution relative to the distribution of
pronominal clitics:
(25) a. Nu-1 mai intreb nimic.
(I) not-him longer ask about anything
b. II mai /i examinez din cind in cind.
(I) him still/ also examine from time to time
c. *Mai/i il examinez din cind in cind 16
(I) still/ also him examine from time to time
Note also the example in (26), where phonological endocliticization of mai can be
observed:
(26) nemaiplecind Ion de acas
not-again-leaving John from home
To summarize, the rule of V-raising appears obligatorily to apply in Romanian,
even in the presence of an auxiliary. This fact cannot be captured under the
hypothesis that Romanian auxiliary structures are of type (A)b: the only
representation that could be proposed was that in (20')b, which was shown to be
^Sequences in which clitic adverbs precede pronominal clitics do appear in certain non-standard
idio-/dialects: ?? mai il intreb j/ eu ceva '[I] still him ask me too something'; they seem to
improve with the negation: ? nu mai il intreb nimic. These examples might be obtained at
surface structure, by some morpheme reordering. The improvement observed in negative contexts
may be due to the fact that nu mai is reinterpreted as numai 'only', which is not a clitic adverb,
and as such occupies a different position.
a. * Va/arIon/elpleca miine?
will/ would John/ he leave tomorrow?
b. * A Ion/ el plecat ieri?
has John/ he left yesterday?
The type of movement that is illicit in (27)b is quite general across languages.
Following den Besten (1977, 1983), it is currently assumed that (28)a-c are
particular instantiations of the rule known as "V second" in Germanic languages:
auxiliaries (and lexical verbs) move out of the Infi position to which they raise at Sstructure, over the NP subject, towards a sentence-initial position, presumably
Comp:
(28)
a. Ist er gekommen?
has he come?
b. Kommt er?
comes he?
c. Will/ would John/ he leave tomorrow?
According to den Besten (1977,1983), Kayne (1984, Chapter 10), and Roberts and
Rizzi (1989), the same rule may be assumed to underlie the French examples in
(29):
(29)
a. L'as-tu mang6?
it have you eaten?
b. Le verra-t-il?
him seefuture he
'Will he see him?'
17
This problem may apparently be solved if we adopt the hypothesis proposed by Pollock
(1989), according to which the Infi node is split into two distinct functional heads, AGR and
Tense (see V. Motapanyane (1989)): AGR would host auxiliaries, and Tense would host lexical
verbs. This hypothesis immediately raises the question why in the other Romance languages
AGR and Tense cannot dominate auxiliaries and verbs respectively. Other technical questions
concern the conditions on verb/ auxiliary movement: is it possible to assume that the auxiliary,
which presumably governs VP, raises to AGR skipping Tense (in violation of Head to Head
movement), thus leaving Tense available as a landing site for V?
13
The impossibility of (30)b is due to the fact that in English, lexical verbs do not
raise to Infi, which is a necessary step (due to the condition on Head to Head
movement, see Travis (1984) and Baker (1988)) for the verb to reach Comp.
Let us now come back to Romanian: if a rule such as V-second exists in this
language, it applies either to inflected verbs (see (31 )a) or to Aux V sequences (see
(31)b), and this cannot easily be accounted for by the general principles of
language. The example in (31)a seems to be analogous to (29)b and (28)b in French
and German respectively, but (31)b does not have any counterpart in any of these
two languages:
(31) a. Pleac Ion miine ?
leaves John tomorrow
b. Vapleca Ion miine?^
will leave John tomorrow
'Will John leave tomorrow?'
Our task will be to account for (31)b and for the ungrammaticality of (27)a. It is
interesting to note that Romance languages other than Romanian show the
impossibility illustrated in (27)a. But they nonetheless differ from Romanian in that
they have other constructions obtained by a rule of Aux raising to Comp; the
relevant examples and analysis are due to Rizzi (1982):
(32) a. Questa commissione ritiene [aver loro sempre ottemperato agli obblighi
previsti dalla legge].
this commission thinks [have they always accomplished the requirements
imposed by the law]
'This commission thinks that they have always ...'
b. Avendo Mario accettato di aiutarci, potremo risolvere ilproblema.
having Mario accepted to help, we could solve the problem
Romanian does not present any construction of this type, in which the subject NP
would intervene between the auxiliary and the lexical verb.
l^Note that the order V-S-0 is possible, and probably preferred to V-O-S in questions,
independently of whether V is a simple verb or a AuxV sequence: crezi c va citi Ion cartea asta?
'do you think that will read John this book'.
This is what we might expect, if we assume (a) that adverbs and floating
quantifiers are generated in a pre-VP position and (b) that the subjunctive verb
obligatorily raises to Infi. We must of course also assume that adverbs are not
allowed to be generated in front of the upper VP. It may be reasonable to think that
this restriction is due to the auxiliary nature of the verb; adverbs modify lexical
verbs only.
What is not expected is the ungrammaticality of (35)a-b:
(35) a. *Are Ion s einte la pian?
has John s play the piano
b. *Aveam s-1 conduc la gar.20
[I] had s -him take to the station
(35)a shows that the auxiliary is not able to raise above the NP subject (a
grammatical question would be are s einte Ion la pian? 'has s play John (at) the
19-rhe complete paradigm is: (eu) am s plec / (tu) ai s pleci / (el) are s piece / (noi) avem s
plecm/ (voi) avefi s pleca(i / (ei) au s piece; but the plural forms tend to become less
productive, and are currently replaced by invariable : noi s plecm/ voi s pleca(i/ei s
piece.
20 T he star indicates that the purely future meaning is excluded; examples of this form are
nonetheless acceptable, but they become synonymous with the examples with a urma 'to follow,
to go to'.
15
piano) and (35)b indicates that the auxiliary cannot take Tense inflections. How can
we explain (35)a-b? Is it possible to show that these two impossibilities are
correlated? Why do mono- and bi-clausal auxiliary constructions behave alike?
Conclusions
To summarize, the following characteristics of Romanian auxiliaries have been
illustrated: (a) they cannot combine with Tense affixes, nor can they be embedded
under Tense auxiliaries; b) Romanian lexical verbs necessarily raise out of VP even
if an Aux is present; c) Aux is not subject to V-second (Romanian presents no rule
comparable to either the English subject-Aux inversion, or the French subject-clitic
inversion, or the Aux-to-Comp rule characteristic of Italian (and Portuguese, see
Raposo (1987)); in descriptive terms, Romanian presents no context in which the
subject NP can intervene between an auxiliary and the lexical verb. 21 Another
important fact about Romanian is that the three properties listed above define not
only monoclausal, but also biclausal auxiliary constructions.
The characteristics given in (b) and (c) could be captured by an adjacency
requirement: Romanian auxiliaries are necessarily adjacent to the verb, or to clitic
elements attached to the verb. This suggests that Romanian auxiliaries are clitic
elements, which is indeed correct (see Chapter 2). But this does not conclude the
analysis. We must still provide an adequate representation. It is not clear what kind
of position should be posited for clitic auxiliaries: the idea that they are under Infi,
supported by the property stated in (a), may seem appealing, but this is clearly not
sufficient, since English modals are generated under Infi, but do not qualify as
clitics.
What is then the underlying representation of Romanian auxiliary constructions?
No answer could so far be found to this question. This failure strongly suggests
that we must abandon the null hypothesis according to which Romanian auxiliaries
would be of the type defined in (A)b, found in Romance and Germanic languages.
21
T h e characteristics in a) and c) distinguish Romanian from French, Italian, Spanish and
Portuguese; b) seems to be instantiated in other pro-drop languages such as Italian and Spanish
(see Belletti (1990)). Note, however, that in Italian past participles may, but do not necessarily,
raise to Aux, as indicated by the distribution of floating quantifiers (which can intervene between
the auxiliary and the lexical verb); in Romanian on the other hand past participles necessarily raise
out of VP.
(36)
IP
V-Infl
VP
NP S
V'
tv
j
IP
CI
IP
V-Ir
VP
tv
NP<
In (36) no (Spec, IP) node has been inserted: the one in the upper clause is an A'position, while the one in the lower clause, besides being an A'-position, cannot
even be instantiated. The latter impossibility characterizes not only the subjunctives
that appear in biclausal tenses, but also those that appear with raising and control
configurations (see Chapter 3). I assume that clitics adjoin to a projection of Infi
(see Chapter 2). 22
Plausible as it may seem, the structure in (36) should be ruled out, since it
incorrectly predicts examples such as (35)a, repeated here, to be grammatical:
(35)
a.
IP
V-Infl
VP
*Are
Ion
[s ein
einte la pian]
Has
John
s plaj
play the piano
'Will John play the piano?'"
The ungrammatically of (35)a may stem from the fact that the auxiliary has moved
to Infi. Romanian auxiliaries would then be characterized as not being allowed to
raise to Infi. This would explain not only the ungrammatically of (35)a, but also
that of (35)b, if we assume that the only way for the verb to get Tense inflections is
by raising to Infi:
22
The adjunction of an X element to an XP constituent is ruled out by the current theory. It may
however be argued that adjunction to Tense is legitimate, due to its defective nature (see
Chomsky (1986)b on the defective character of IP).
(35)
Since the auxiliary cannot move to Infi, this node is not only empty, but altogether
absent, because Infi cannot surface separated from V. Since Infi is absent,
Nominative case cannot be assigned in the higher clause. In other words, the
auxiliary structure in (33) is "biclausal" in the sense that the auxiliary takes a CP
complement, but unlike standard biclausal structures, the "main clause" of this
configuration lacks both an Infi node and an NP subject position. The only element
of the higher clause is the auxiliary itself, hence the configuration in (37):
(37)
AuxP
Aux
CP/IP
are
[s vin Ion]
The problem is that normally AuxP does not count as a sentential constituent. It
seems plausible to assume that the auxiliary is reanalyzed as being adjoined to
CP/IP. The following conventions are a possible technical implementation of this
idea:
(38) Functional Coindexation: Coindex adjacent functional X categories.
(39) Functional Adjunction: Adjoin X categories to the YP functional projection
with which they are coindexed.
Given (38) and (39), the configuration in (37) becomes (40), given in its developed
form:
(40)
CP/IP
Aux;
CP/IP
Clj IP
V-Inflj
VP
I \
NP V'
tv
NP
splec/pleci/piece...
s leave- lp / leave-2p,..
'I'll leave, etc., I am going to leave.'
It is clear that aller does not fall under the sirucfura/definition of auxiliaries given in
(A)b: in (42) aller does not take a VP, but an IP constituent. This future
construction could instead be analyzed in terms of structural raising. But aller
nonetheless falls under the thematic definition of auxiliaries (see (A)a), which
might explain why raising is obligatory (see *il va que Jean le voit 'it goes that John
sees him').
The Romanian structure in (40) differs from (42) insofar as one single IP
constituent is activated. Compare the following raising constructions: 2 4
23
x h i s seems consistent with the hypothesis suggested in Section 1.1.2. that AGR is an X
nominal element that does not project a maximal category, but adjoins to a projection of Infi.
24 0 n raising (and control) in s subjunctives see Chapter 4.
19
25
Note that concerning the position of adverbs and floating quantifiers, raising verbs behave on a
par with auxiliaries (compare (34)a-b):
(i)
Ion urmeaz (*adesea) s einte (adesea) la pian.
John follows (*often) s play (often) piano
'John is going to often play piano.'
(ii)
Copiii urmeaz (*to(i) s einte (to(i) la pian.
the children follow (*all) s play (all) the piano
The ungrammaticality of (i) is due to the fact that raising verbs do not allow for adverb
modification. The impossibility of (ii) seems to indicate that quantifiers can only be stranded in
the -position in which they have been generated (see Sportiche's analysis).
a. *am il rugat
b. *a$ilruga
Let us then come back to the unifying hypothesis, according to which the auxiliaries
in (45)a-c take a CP/IP complement. The absence of a could be due to the rule of
Move I-to-C, which moves the inflected verb to Comp and thereby makes it
impossible for a to show up. 26
Let us then assume that (45)a-b are represented by an abstract configuration of
type (40), and by the obligatory movement of the lower (V-)Infl to Comp. The
structure of (45)a-b would then be (45'):
26
S e e den Besten's (1977, 1983) analysis of V second; according to Roberts's (1991) updating of
den Besten's proposal, Move I-to-C is a case of substitution into an empty head position.
(45') b.
CP/IP
Aux
CP/IP
C
IP
V-lnfl
VP
I \
NP V'
Is
tv NP;
Jt
The idea that Move I to C applies as shown in (45')b seems inadequate, because in
general this rule does not apply in embedded clauses. This problem is only
apparent, because it is natural to assume that Move I to C is free to apply if the
higher verb subcategorizes for a projection headed by Infi. And this might be the
case for auxiliaries.
It is however not completely clear why I-to-C movement is obligatory in (45):
due to the rule of Restructuring Incorporation, to be proposed in Chapter 2, the
sequence a V+Infl counts as [Infia V+], and this should suffice to satisfy the
requirement imposed on auxiliaries. Note that this analysis must indeed be assumed
for biclausal auxiliaries: in am s plec the auxiliary am takes as a complement CP/IP
headed by [inf V+Infl]. The fact that a infinitivals do not show up with
auxiliaries may be related to the fact that (full) infinitivals are generally rare in
auxiliary structures, possibly because of their "intrinsic" nominal character (see
Kayne (1982)): in English the particle to is absent with modals 27 and in French the
infinitival became incorporated into future and conditional paradigms: manger-ai(s),
manger-a(i)s, etc. (see Lema and Rivero (1989)). The obligatory I-to-C movement
that applies in (45) could be a means to turn an infinitival into a verbal projection. 28
Let us now consider the position of pronominal clitics, and assume, as we did in
(40), that they occupy the IP-adjunction position:
2 7 W e may of course stipulate that modals and to occupy the same position (see, however, the
problems that this raises concerning the position of not), but we may wonder why this should be
so.
2 8 Note
that we must also assume I-to-C for participles (this is true in Romanian not only for past
participles (see (48)a), but also for present participles (see Chapter 2)) - the reason in this case
would be that the verb has to move as high up as it can, i.e., to C, if no lexical complementizer
is there.
(47)
CP/IP
Aux CP/IP
C
IP
CI;
IP
l \\
V-Infl VP
NP V'
Jt
As discussed in Chapter 2, Move I to C may either take pronominal clitics along 29
or strand them. 30 The first option is clearly not correct here, because examples (46)
are ungrammatical. The second option on the other hand, directly accounts for the
position of the singular feminine clitic 'her':
(48)
a.
Amrugat-o.
ruga-o.
a. L-am
rugat..
L-aruga.
29
23
verbal form hosts both and another clitic, clitic splitting arises, one clitic
preceding the verb, and the other following it:
(50)
*Pot vedea-o.
[I] can see-her
*Pot vedea.
[I] can her see
pot vedea.34
[I] her can see
In modern Romanian the verb a putea 'can, may' takes either a s subjunctive or a
bare infinitive as a complement. But in the latter case the distribution of clitics
clearly indicates that the rule of I-to-C does not apply (see (52) and compare with
(48)b). Example (53) on the other hand shows that the clitic cannot attach on the left
of the embedded verb either, the only possibility being attachment on the left of the
auxiliary, as in (54). The distribution in (53)-(54) can be understood if we assume
a) that at S-structure Romanian clitics cannot adjoin to V, but only to Infi and b)
that the bare infinitive in these examples presents no Infi node to which the clitic
could attach; hence the clitic is forced to adjoin to the first available Infi, the one in
the main clause. What is then the categorial status of the bare infinitive in (52)-(54)?
It may be VP (see the structure in (55)). Since no Infi node is present, the verb
cannot raise to Infi, but must stay under VP:
to the relation between the complementizer s and nu (see (i)), and therefore can only be imputed
to the presence of the auxiliary: these elements cannot govern a NegP constituent. Finally, (iv) is
grammatical, as opposed to (ii), because here the CPAP complement headed by s has the status
of an IP constituent (due to the presence of the auxiliary, which can only adjoin to a constituent
headed by the (inflected) verb. Thus, s subjunctives, which are systematically ambiguous
between CP and IP assume one or the other categorial identity, depending on the context. If the
various elements of the context impose contradictory requirements on s subjunctives, we end up
with an ungrammatical sequence (see (iii), where the subjunctive should be analyzed as CP due to
the position of the negation, but as IP, due to the presence of the auxiliary).
^ N o t e that the example in (54) can be said to be characterized by clitic climbing, because the Aposition identified by the clitic is inside the VP governed by pot, note however that no clitic A'position can be assumed out of which the clitic would have raised. In other words, (54) is
characterized by a direct relation between the clitic and the -position that it identifies.
(55)
25
IP
CI
IP
l \
V-Infl VP
NPS
IX-VP
L\.- v
NP
Oj
pot
NPo
vedea t;
Unlike auxiliaries, a putea is able to raise to Infi. The two manifestations of this
movement can indeed be observed; (56)a shows that the subject of the main clause
can intervene between a putea and the bare infinitive (compare with future and
conditional structures); (56)b shows that a putea combines with Tense morphemes,
like lexical verbs, and unlike auxiliaries:
(56) a. Poate Ion veni miine?
can John come tomorrow
b. Ion putuse veni.
John can-past perfect come
'John had been able to come.'
The structure in (55) thus appears to be supported both by the distribution of clitics
and by the syntactic properties of a putea. This verb is comparable to its
counterparts in French and other Romance languages, insofar as it presents two
distinct interpretations, currently labelled "epistemic" and "root", which are
respectively illustrated below. Note that in (57) the subject occupies th (Spec, IP)
position, after having passed through the [Spec, VP] position of poate, as shown in
(55):
(57) a. Ion poate ajunge dinlr-o clip in alta.
'John may arrive from one moment to the other.'
b. Ion poate citi 100 pagini intr-o or.
'John is able to read 100 pages in an hour.'
Under the epistemic meaning a putea does not impose selectional restrictions on its
subject; it is therefore reasonable to assume that in this case it is a sort of raising
verb: the embedded subject raises to the main subject position, leaving behind an
NP trace. Under the root meaning, a putea does impose selectional restrictions on
its subject, and therefore it is reasonable to assume that the example in (57)b is a
control configuration, with a PRO in the subject position of the lower VP. Since
PRO is ungoverned (by virtue of the PRO theorem, see Chomsky (1981)), we
must assume that "root" a putea does not govern into its VP complement, but we
will not investigate the technical details of this proposal here. The exact analysis of
these examples is only partially relevant here. The data in (52)-(54) and (56)a-b
indicates that the bare infinitive governed by a putea 'may, can' is a VP constituent,
as opposed to the bare infinitives governed by auxiliary verbs, which are CP
constituents in which I-to-C movement applies; Romanian bare infinitives thus
appear to be of different categorial types. Note however that a putea may take a CP
complement: in older stages of Rumanian a infinitives were possible (cf (58)a), and
in both those stages and the modern language s subjunctives are frequent, in free
variation with bare infinitives:
(58) a. Ion poate a cinta dac vrea.
John can to sing if he wants
b. Ion poate s einte dac vrea.
John can s sing-subj if he wants
The question is then why, given the possibility of (58), I-to-C is not allowed to
apply in the embedded CP, and in particular in (58)a. It is reasonable to assume that
this impossibility, which is clearly indicated by the distribution of clitics in (56)a-c),
is again due to the fact that a putea is not an auxiliary. More precisely, I-to-C
movement can apply in embedded clauses only if the main verb subcategorizes for a
V projection (on a similar suggestion for the French clitic subject inversion see
Rizzi and Roberts (1989)), and this is not the case for a putea.
1. 3. 2. Bare innitives and clitic adverbs
Consider next the distribution of clitic adverbs:
(59) a. (nu) (mai) poate (*mai) scrie
[he] (not) (again) can (*again) write
b. (prea) poate (*prea) spune ce vrea
[he] (too much) can (*too much) say what [he] wants
(60) a. ar mai citi-o / va mai citi-o/a mai citit-o
[he] would again read-it / will again read-it / has again read-it
b. ar gi scrie / va scrie / a sens
[he] would even write / will even write / has even written
(61) a. nemaivzindu-1
notagainseeing-him
b. Mai las-1 in pace.
again leave-him in peace
The data in (60)a-b indicates that in simple tenses clitic adverbs necessarily show
up in front of the inflected verb, but after pronominal clitics. In compound tenses
they always appear between auxiliaries and the lexical verb. It has been suggested
that two possible analyses could be adopted: (a) the adverb could be generated in a
position which precedes the position to which the lexical verb raises; (b) the adverb
is directly base generated on the verb. The data in (59) quite clearly indicates that
the second hypothesis is incorrect: were the adverb attached to the verb itself, we
could not understand why these examples are incorrect. This impossibility can be
captured if we assume that these adverbs necessarily attach to an Infi node, which is
absent in (59). The examples in (60) indicate that these adverbs cliticize on Infi, and
therefore on the verb raised to Infi, and are taken along by the rule of I-to-C
movement, which leaves pronominal clitics behind (see (60)a). Note that a
comparable sequence, Adv-V-cl, can be observed in (61), which is also obtained
by the rule of I-to-C.
To summarize, the proposed analysis accounts for the following cluster of
differences between Romanian auxiliaries and a putea 'can, may':
(62)
a.
b.
c.
d.
pot
Vinfinitive
Aux
a. miratu-m-am
/miratu-te-ai...
mira-m-a
wonder-me-would [I]
'I'd wonder'
c.
pleca-voi
leave-shall [I]
shall leave'
The syntactic difference between (63) and their non-inverted counterparts in (64)
correlates with semantic or illocutionary differences, which will not be discussed
here:35
(64)
a. m-am mirat/te-ai
mkat...
mira
voipleca
35
S e e Rivero (1988a; to appear) for a suggestion concerning the triggering elements of the V
movement responsible for the inverted conjugations.
Old Spanish in particular.36 The survival of this construction (labelled "Long Head
Movement" in Lema and Rivero (1989, 1990); Rivera (to appear)) in Romanian
may be related to the existence of parallel constructions in the other Balkan
languages and surrounding areas (Albanian, Bulgarian, Czech, Serbo-Croation,
Slovene, Slovak, Polish).
I will not try to compare the Romanian data with the parallel constructions found
in other languages. My purpose will be to discuss the theoretical problems raised
by this construction. It will be shown that a potential ECP violation is circumvented
due to the particular type of auxiliary constructions that was argued for in previous
sections. By contrast, auxiliary inversion would lead to an ECP violation in the
other modern Romance languages, due to the structural properties of their
auxiliaries. This theory is more constrained than Lema and Rivera's and as such
seems preferable, provided that crosslinguistic evidence supports it. As usual, this
is extremely difficult to evaluate at this stage of the investigation, but I would like to
stress how important a careful analysis of the data is.
The type of structure that one has to assign a given sequence of elements is an
extremely complicated issue, and very subtle arguments have to be taken into
account in order to choose between competing alternatives. A good case in point is
the analysis of auxiliary constructions in Romanian itself, for which the intuitively
most obvious hypothesis, unanimously adopted by previous research (including
Dobrovie-Sorin (1987)) was discarded in the previous section. Lema and Rivero
assume that the auxiliary constructions found in the various languages that present
auxiliary inversion are of the same fundamental type as that found in the wellstudied Romance and Germanic languages, where the auxiliary takes a VP
complement. A closer investigation of this area may lead to a revision of this
hypothesis.
It is also not clear that all the auxiliary inversions that can be observed are of the
same type. We owe to Lema and Rivero (1989, 1990) a careful distinction between
VP-movement and the auxiliary inversion illustrated in (63), which is to be
analyzed as head movement of the verb across the intervening auxiliary. There
seems to exist at least a third construction - labelled Participle-fronting by Lema and
Rivero (1990) - which resembles Platzack's (1987) Stylistic Fronting in Modern
Icelandic. A more precise typology of head movement is clearly needed. The
prediction that I would like to make is that the type of head movement to which the
Romanian paradigm in (63) belongs can only be found in languages which present
the Romanian type of auxiliary constructions.
3The examples in (i)-(ii) are taken from Lema and Rivero (1989); EP and OS stand respectively
for European Portuguese and Old Spanish:
(i)
DIR- se-ia
um povo predestinado.
(EP)
TELL-SE+imp-had a people predestined
'One would say it is a predestined people.'
(ii)
Si yo vivo, DOBLAR vos he la soldada
(OS)
If 1 live, DOUBLE you I-have the wages
'If I live, I will double your pay'.
29
CP/IP
V-I
Aux
CP/IP
C
tv-i
IP.
I
[vp...tv]
tV-I
[vp...tv.]
vazut-am
Seen-have-lst
It is clear that the movement rule illustrated in (65) violates the Head Movement
Constraint (see Travis (1984) and Baker (1988)), which forbids head movement
over an intervening head: (V+)Infl is an X category that bypasses another X
category, the auxiliary. Chomsky (1986)b has pointed out that the Head Movement
Constraint need not be stated in the grammar as such, but derives as a consequence
from the ECP. Since the Head Movement Constraint per se is not part of the
grammar, it may be violated, provided the ECP is obeyed. This predicted
dissociation between the Head Movement Constraint and the ECP is precisely what
we observe in configurations like (65). Since the construction is grammatical, the
ECP is presumably obeyed, a conclusion forced upon any GB analysis of the data
(see Ouhalla (1988), Lema and Rivero (1989, 1990), Roberts (1991)). It is also
commonly assumed that the relevant ECP requirement in the case of head
movement is antecedent government:
(66) properly governs iff (th-governs or/and) antecedent-governs .
Given on the one hand the disjunctive formulation of the ECP, and on the other
hand the empirical necessity to require antecedent-government of Head movement,
Chomsky (1986b: 68-78) has suggested that the clause (b) be suppressed from the
Auxiliaries are "defective" verbs that select but do not th-mark VP.
31
Since they do not th-mark VP, auxiliaries do not Tense-mark VP either. It is only
via their relation to Infi (e.g., by raising to Infi) that auxiliaries are able to th-mark,
or more precisely to Tense-mark their VP complement.37
Having outlined these current assumptions, let us now turn to the configuration
in (65), and see in which way it satisfies the ECP; (73) is the bracketed
representation of the tree in (65):
(73)
V+I
[CP/IP Aux
[VP
The verb moved to Infi L-marks the VP that it has raised out of; therefore VP is no
longer a barrier, and does not transmit barrierhood to IP. There are at least two
reasons that may be invoked for the non-barrierhood of CP: its head is empty (it is
filled only as a result of I-to-C movement); note, furthermore, that even in case the
complementizer a were present, it would count as adjoined to IP (see Chapter 3),
and thus no CP barrierhood would arise. Finally, the last step is also legitimate,
because the auxiliary itself adjoins to CP/IP. Since no barrier intervenes between
the various steps of the V raising in (73), antecedent government holds, as required
by the ECP.
Let me stress that the proposed explanation concerning the way in which the
ECP is satisfied does not introduce any ad hoc assumption: what has been used is
the current idea that V movement out of VP is licit only if (on the first step) the
verb raises to Infi, from where V is able to L-mark the VP out of which it has
raised, thus voiding its barrierhood. The subsequent steps are legitimate due to the
adjunction configuration, which does not create barrierhood. This straightforward
account of inverted conjugations is made possible by the particular configuration of
auxiliary constructions proposed in previous sections on independent grounds.
1. 4. 2. On the absence of auxiliary inversion
Consider for comparison the configuration characteristic of the (A)b type auxiliaries
found in modem Romance languages other than Romanian; the structure given in
(74) is currently assumed for the perfect auxiliaries avoir/etre:
(74)
7 Here I differ slightly from Chomsky (1986b: 73), who assumes that by raising to Infi, an
aspectual becomes able to L-mark VP. It is difficult to see how we can make L-marking come
into play, since neither Infi nor the aspectual element are L-markers.
38 It
is however reasonable to assume that participial VPs are governed by an Infi node (see Kayne
(1987)), which has not been inserted in (43); if this hypothesis is correct, we must explain why,
in Romance languages other than Romanian, participial Infi is not an adequate host for clitics (be
they pronouns or auxiliaries). One may suggest that Infi can be a host for clitics only if it triggers
V-to-I movement, but not if it is subject to Affix-hopping. If this is correct, we would have to
show that in Romance languages other than Romanian, participial Infi is subject to Affixhopping.
The auxiliary in (74) does not adjoin to a projection of Infi, but instead it moves
(just like any other verb) to Infi.
According to (72) the auxiliary does not th-mark VP; a fortiori it does not Lmark VP, which is thereby a barrier. Given the configuration in (74), this barrier
cannot be voided by V raising, because there is no available landing site from
which the raised V could L-mark the VP. The V raising in (74) is thus blocked, as
indicated by the diamond, by the barrierhood of VP, and redundantly by the
barrierhood of AuxP and IP.
According to this analysis the absence of inverted conjugations in modern
French or Italian is due to the type of auxiliary constructions that they present. In
these configurations V raising is blocked by the ECP, under the current definitions
of the relevant notions.
It is interesting to compare the account proposed here with Lema and Rivera
(1989, 1990), who consider - modulo slight variations from one paper to another that inverted conjugations are permitted in configurations such as (74). The ECP
would be obeyed due to the following assumptions which would have to be added
to the current theory :
(75)
These assumptions differ from the current theory presented above: (75)a
generalizes to auxiliaries a property that Chomsky assumes for Infi exclusively
(recall that Tense-marking is simply a label for "th-marking by Infi"); (75)b is a
weakening of the L-marking requirement for non-barriers; (75)c may be viewed as
an extension of the class of defective blocking categories: besides IP we would also
have a certain type of AuxPs.
The first objection to this account is, of course, its stipulative and costly
character. The second problem concerns the account of the (crosslinguistic)
difference between those auxiliary constructions that allow and those that do not
allow auxiliary inversion. According to Lema and Rivera, the possibility of inverted
conjugations depends on the auxiliary qualifying as a "weak"/"functional" auxiliary
(as opposed to "strong"/"lexical" auxiliaries); it is due to the presence of such an
auxiliary that (75) can apply and thereby the ECP can be met.
I agree with Lema and Rivera that functional auxiliaries (defined as in (72)), as
opposed to lexical auxiliaries or modals (see below), allow the verb to bypass them.
However, I do not believe that this is a sufficient condition. A case in point is the
Spanish aspect auxiliary haber (the case of Italian and French may be slightly
different, given the existence of the avoir/etre alternation), which qualifies as a
functional auxiliary by the definition in (72). In order to account for the
impossibility of auxiliary inversion, Lema and Rivero (1989, 1990) have to
stipulate that haber is not a functional auxiliary, but it is difficult to see by which
definition of functional auxiliaries haber would count as lexical and the Romanian
am/ai... as functional. One may wonder whether independent evidence exists, other
than the possibility of the inverted conjugations itself, to determine whether a given
auxiliary is functional or not; otherwise the term "functional auxiliary" is simply a
It
It
The conditions on Merging stated in (77) are met in (78): the auxiliary and the verb
raised to Infi bear the same functional index (see the mechanism of Functional
Coindexation defined in (38) and the raised V is adjacent to the auxiliary.
Granting that this account is correct, and that Merging obligatorily applies in
inverted conjugations (in order to avoid a potential ECP violation), let us see
whether type (A)b auxiliaries allow Merging. In (74), repeated here, the auxiliary
and the verb do not bear the same functional index because the verb has not raised
39
This differs from Roberts (to appear) who assumes an A/A' distinction between X positions,
and uses it in order to void the effects of minimality in Long Head Movement.
Aux-I[ A u x P t A u x [ V p..V..]
splec
spleci/
s mai
ascult
The particle is obviously a clitic: it appears in a fixed position, it cannot move, nor
appear in isolation. These characteristics correlate with the fact that Functional
Coindexation may apply (see (38)): o, s and the Infi head to which the verb has
raised are adjacent functional elements. Hence (79'):
(79')
CP/IP
Aiixj
CP/IP
Pronj
IP
i n f i i ^ ^ VP
Adv
V-Infl NP
NP;
s
4
il
mai
ascult
This paradigm is in the process of replacing the one discussed in Section 1.2.1, made up of the
verb a avea 'to have' (inflected for AGR) followed by a subjunctive.
35
Note now that the examples in (79) do not present inverted forms:
(80) a. *plec-o-s
b. *spleci-o
(80)a can be ruled out if we assume that I-to-C cannot bypass two elements in one
step;41 the ungrammaticality of (80)b on the other hand is due to the fact that I-to-C
can move V-I but not s V-I.
To sum up, the rule of I-to-C that bypasses auxiliaries has been treated as an
instance of the rule Move Infl-to-Comp. The ECP is met due to the particular type
of auxiliary constructions characteristic of Romanian. Among the other Romance
languages, auxiliary inversion is also permitted in European Portuguese and in Old
Spanish, for future and conditional auxiliaries (but not for the perfect auxiliary).
According to the theory proposed here, this possibility indicates that these
auxiliaries present the structural properties of Romanian auxiliaries. This may
explain why it is precisely future and conditional auxiliaries that have become fully
bound morphemes in modern Spanish; compare perfect auxiliaries, which used to
function as free morphemes and have preserved that status.
1. 4. 4. Modals
The proposed analysis also accounts for the impossibility of inverted conjugations
with a putea 'can, may', for which the structure in (81) has been proposed on
independent grounds (see (55) and the generalizations in (62)):
(81)
IP
pot
tp0t
ascult
In (81) vedea 'see' cannot prepose over pot, because there are no intermediary
landing sites for the verb to raise stepwise, thus enabling it to void the various
intervening barriers (the two VPs and IP):
Two pronominal clitics can be bypassed by the verb, because strings of pronominal clitics
probably occupy one single clitic position, see the discussion of examples (24) in Chapter 2.
(82)
,
I
Note that this account is exactly the same as that proposed for the ungrammaticality
of inverted conjugations with type (A)b functional auxiliaries (see (74)).
The proposed analysis thus accounts for the contrast between Romanian
auxiliaries and a putea as to the possibility of inverted conjugations. This contrast
can now be added (see (62)e) to the list given in (62)a-d, of the other differences
between the two constructions:
(62) pot Vinfinitive
Aux V
a.
b.
c.
d.
e.
31
(87)
EPV
NPS
X
V-Infl
VP
i ^ v
tv
(88)
NPr
m
l
V-Infl
VP
NP S
V'
tV
NP 0
In (87) and (88) Sportiche's Vmax has been replaced by VP, and correlatively VP
by V'; the internal subject stands under (Spec, VP). These alterations are needed if
we want to assume a consistent X'-theory of the type proposed in Chomsky
(1986)b).
Let us now consider the (Spec, IP) position, which clearly has a different status
in (87) and (88). In (87) (Spec, IP) is necessarily present, because the subject
cannot survive in its base position (presumably because it cannot be assigned Case
there, but the precise reason is of no consequence here); in this case (Spec, IP) is an
-position, the canonical structural position of the subject.
Compare (88): the (Spec, IP) position has not been represented here, because it
is an optional A'-position (the site characteristic of topicalizations and left
42
This hypothesis can be found (with certain technical variations, which concern (a) the categorial
status of the node that dominates [NP VP] and (b) the languages for which this hypothesis holds)
in Kuroda (1986); Contreras (1987); Kitagawa (1986); Fukui and Speas (1986); Speas (1986);
Sportiche (1988); Zagona (1988); and Dobrovie-Sorin (1987).
'
Note that a somewhat more complicated version of (89)b could be assumed, with a
specifier position for AuxP, which would constitute an intermediary step for the
NP movement from the Spec position of VP to the Spec position of IP:
(89') b
itl
The difference between (89)b and (89')b is irrelevant here, because what interests
us is the way in which auxiliary constructions satisfy the Tense filter, and I assume
that NP-traces are not visible for the relations on which Tense-indexing depends.
Let us now turn to V-initial languages, characterized by the structure in (88): in
such a language the NP subject does not raise to (Spec, IP) but stays in its base
generated position, as in (90):
(90)
The Tense filter is the verbal counterpart of the Case filter, the visibility condition
on nominals. The two ways in which the Tense filter can be satisfied, namely by
incorporation or by government are parallel to the two ways in which nominals can
satisfy the Case filter: clitic pronominals incoiporate into the head bearing the Case
feature, whereas full NPs are assigned a Case feature under government by the
head bearing the Case feature (see Baker (1988); Baker, Johnson and Roberts
(1989)).
Both auxiliaries and lexical verbs must satisfy the Tense filter. Consider first
(89)a-b: in (89)a the auxiliary is Tense-indexed because it is base-generated in the
Infi position; in (89)b the auxiliary satisfies the Tense-filter by incorporating to Infi
(see (91)b(i)), via V-to-I movement. As to the lexical verb itself, it is licensed via
Tense-marking (see (91)b(ii)): in both (89)a and (89)b the auxiliary raised to Infi
th-governs V'; the sisterhood condition is satisfied if we assume that neither NPtraces nor V-traces are visible for Tense-indexing.
Let me stress that it is not the auxiliary itself that Tense-marks VP (and indirectly
its V head); as stated in (72), auxiliaries do not th-mark (and therefore they do not
Tense-mark) their VP complement. It is only Infi that is a Tense-marker: in (89)a-b
the auxiliary Tense-marks only because it occupies the Infi position.43
Turning now to (90), it is easy to see that this configuration violates the Tense
filter: the NP subject intervenes between (the auxiliary raised to) Infi and V', in
violation of the sisterhood condition in (91)b (ii) 4 4 The problem cannot be solved
by inserting the auxiliary lower, i.e., inside VP, just above V', as in (92). Since it is
only by raising to Infi that the auxiliary can license the lexical verb, the sisterhood
condition is still violated after Aux-to-I raising:45
(92)
To sum up, a correlation appears to hold between IP types and types of auxiliaries:
because of the Tense-filter, Aux VP configurations are ruled out in V-initial
langugages characterized by (88). If this correlation is correct, a tentative answer to
(83) could be that Romanian IP constituents are of the type in (88).
1. 5. 3. The licensing of verbs embedded under modals
Consider next the configuration in (93), which was shown to characterize a putea
'may, can':
(93)
From the structural point of view (93) is identical to the configuration characteristic
of type (A)b auxiliaries, and it has just been shown that this configuration is
43 A
41
cartea,...
ceva.
lecfia.
lecfia.
46
,..46
This ungrammaticality is very clear in contemporary standard Romanian. Note, however, that
examples of this type can be found in the spoken language, and are also found in the literature of
the beginning of the century (see Bredemeier 1976, pp 56-58). The data is more complicated for
c. Ion ar mai
scrie.
IP
I
^VP
mai
NPS
fi
V-I
V
V
NP 0
tv
^^Note however that within this analysis, the position of 'her' is somewhat surprising in (i)(iii):
(i)
inainte de a fi trimis-o mama,...
before of a have sent-it mother,...
'before mother had sent-it,...'
(ii) Nu cred s fi spus-o Ion.
[I do] not believe s i be said John
don't believe John to have said such a thing.'
(iii) inainte de a trimite-o mama
before of to send-it mother
This problem arises not only for two-auxiliary constructions, but also for one-auxiliary structures
(see (iii)), and therefore it is not our analysis of fi which is at stake, but rather our analysis of
cliticization in auxiliary structures in general. In neither a infinitives or s subjunctives does (V)Infl to Comp apply, and this is the only syntactic rule that gives rise to postverbal clitics (see
Chapter 2). The difficulty raised by (i)-(iii) is less important than it appears at first sight; the
position of illustrated in (i)-(iii) is in fact not obligatory, only optional:
(iv) Inainte de a fi trimis mama,...
(v) nu cred s-o fi spus Ion
(vi) Inainte de a trimite mama
Given (iv)-(vi), we are not obliged to assume that the position of in (i)-(iii) is the one that this
clitic occupies in conditional and "perfectul compus" configurations, since in all these structures
is necessarily postverbal. We may instead assume that the syntactic position of is the one in
(iv)-(vi), which is consistent with our analysis of infinitives and subjunctives, and with our
hypothesis concerning '; the postverbal position in (i)-(iii) would be the result of a phonological
rule. Note that this rule is sensitive to the presence of AGR features on the subjunctive. Thus, (i)(vi) show that is compatible with either the pre-auxiliary or the postverbal position in perfect
subjunctives and both present and perfect infinitives. But present subjunctives, which are inflected
for AGR, as opposed to perfect subjunctives, do not allow postverbal o.
(vii) *s vd-o
s [I] see-her
4 9 T h e meaning of this impossible example (see the English gloss) can be conveyed by an
embedded subjunctive: Ion poate s fi ajuns ieri, where fi occupies the same position as in (96)b.
(101) C P / I P ^
Auxi
CP/IP
I \
AUX2 CP/IP
I
C
\
IP
I
V- Infi
...
In order to rule (101) out we may assume that adjunction to CP/IP is not recursive.
Therefore, a language that has at its disposal only auxiliaries that adjoin to CP/IP
cannot present AuxP constituents embedded under another Aux. Compare
auxiliaries that govern VP complements, for example, in Romance languages other
than Romanian: they allow embedded auxiliaries.
To sum up, we have provided the following answers to the questions posed at
the beginning of this section: (a) am/ai/a... cannot be used in (96) a-b, because this
auxiliary cannot move to Infi; it cannot be used in (96)c-d either, because it cannot
be embedded under another auxiliary; (b) conversely, fi is impossible in the present
perfect, because fi is part of a larger morpheme that encodes the information
"perfect infinitive"; this also explains (c) why fi is invariable. The necessary
alternation between two perfect auxiliaries (one for the present perfect/"perfectul
compus", and the other for the paradigms in (96)) 50 is thus the only possibility
available for a language that presents only auxiliaries that adjoin to CPAP. Note that
the term "auxiliary", which we have used for convenience, is inadequate for fi,
which is just part of a bound morpheme, and therefore does not fall under the
definition of auxiliaries given in (B).
Conclusions
It has been shown that type (A)b auxiliaries cannot exist in V-initial languages;
insofar as these languages present any auxiliaries at all, these are type (B)
auxiliaries. Given these correlations, the fact that Romanian presents type (B)
auxiliaries and no (A)b auxiliary may be used as indirect evidence in favour of the
idea that Romanian IP constituents are V-initial.51
The structure proposed here for Romanian auxiliaries has in common with type
(A)b auxiliaries the hypothesis that auxiliaries are external to the VP with which
they are associated (this idea goes back to Chomsky (1955) and has been
reformulated by Jackendoff (1972), Emonds (1976), Culicover (1976), Akmajian
and Wasow (1975), Akmajian, Steele and Wasow (1979)). We have reached an
extreme variant of this hypothesis, which brings to mind Ross's (1970) hypothesis
of auxiliaries as main verbs: Romanian auxiliaries are sisters of CP/IP constituents.
50
This analysis does not explain why it is am/ai/a.. which adjoins to CP, and fi which adjoins
to Infi and not vice versa. We leave this open.
51
Note that the reverse does not hold: a language may present auxiliary clitics even if it is not a
V-initial language. In this case auxiliary clitics could adjoin to IP.
\
V
V
laves
I
6te
avons
Another possibility would be to assume that V' is a flat constituent, i.e., the lexical
verb and the preceding auxiliaries would be sisters to each other.
A hypothesis along these lines seems at first sight adequate for Romanian: it
directly captures the strong coherence that exists between auxiliaries and lexical
verbs; it could in fact be claimed that the evidence in favour of generalized Vraising on the one hand and in favour of the absence of subject-Aux inversion on
the other constitutes evidence in favour of the constituenthood of the Aux-V string.
Such a hypothesis is close to the current analysis of Romanian auxiliaries adopted
by the Romanian structuralist school; see in particular Gu{u-Romalo (1962) and
Avram (1988), who reach the conclusion that Romanian "auxiliaries are not verbs,
but grammatical morphemes". The VP-external analysis presented above seems at
first sight incompatible with the "affixal" (more precisely "clitic") status of
Romanian auxiliaries. In Chapter 2 it will be shown that the proposed configuration
does in fact account for the clitic status of Romanian auxiliaries. It is also important
to note that even if Romanian auxiliaries were affixes (which is, as a matter of fact
incorrect), we would not be obliged to assume that at D-structure they form a
constituent with the verb: under the current analysis of IP constituents, inflectional
affixes are not base generated on the verb, but under Infi, which takes VP as a
complement. This is however a theory-internal assumption.
Let me then stress again that the foregoing analysis, relying on the VP-external
hypothesis, is empirically superior to the competing V-complex analysis by
correctly accounting for the following properties: (a) the affix-like behaviour of
Romanian auxiliaries can also be observed with the future paradigms which are
clearly "biclausal", in the sense explicated in Section 1.2.1. It is clear that am s plec
/ai s pleci 'have-1st sing that leave-1st sing/ have that 2nd sing that leave2nd sing' are frozen sequences, on a par with the other Aux-V sequences, but in
this case it is clear that am/ai/are
cannot be analyzed as being intrinsically either
affixes or clitics (the clitic/non-clitic status of am/ai/a depends on the configurations
in which it appears). Moreover, the AGR features on the lexical verb, as well as the
presence of s and the distribution of clitics clearly indicate that these configurations
present an embedded CP/IP constituent; (b) the distribution of the feminine
pronominal clitic o, and that of clitic adverbs (see Chapter 2); (c) the existence of
inverted conjugations such as V Aux ty (see Section 1.4. above).
47
In the next chapter it will be shown that the analysis proposed here is compatible
with the clitic status of Romanian auxiliaries.
Appendix
This remark concerns participial constructions in S-V-0 languages. It has been
recently proposed (see Kayne (1987); Baker, Johnson and Roberts (1989)) that the
participial inflection -en in English (see -6 in French) is generated under an Infi
node which takes VP as a complement. Therefore, we end up with biclausal
configurations for passives and present perfect constructions: the auxiliary verbs
have and be move to the upper Infi node, and the lexical verb is related to its own
Infi node,52 as in (103).
(103)
IP
I
NPS
AuxP
i \
NPS Aux'
Aux
IP53
I
VP
NPs
v:
NPn
The question is then in what way the lexical verb takes the participial inflection; this
can be done either by Affix Hopping or by V-to-I movement. The former option is
the only one available in English; both options may be postulated for French. There
is however one argument in favour of the idea that the participial inflection is
subject to Affix Hopping rather than a triggerer of V-to-I: pronominal clitics cannot
adjoin to the past participle itself, but only to the auxiliary verb (je l'ai mang6 avec
plaisir "I it-have eaten with pleasure'; compare *j'ai le mang6avecplaisir). The GB
52niis hypothesis is interesting insofar as it leads to a very neat system: every verb is generated
inside a VP,which is governed by an Infi node (this generalization could be stated as a visibility
principle on verbs, but note that such a principle would be stronger than the one proposed in
Roberts (1985)). No VP constituent is directly dominated by another verb, auxiliary or not.
English modals do not constitute an exception: since they are generated under Infi, they are
allowed to take VP complements. Compare have, which is generated outside Infi, and
correlatively takes an IP complement.
53 I t is not clear that a (Spec, IP) node is projected here. If it is, the subject NP goes through this
position on its ascendent movement, but cannot stay there because of the condition on auxiliaries.
theory of clitics is not sufficiently clear, but there is much convergent evidence in
favour of the idea that clitics adjoin to Infi (see Chapter 2). This condition must be
strengthened if we want to capture the distribution of clitics in French auxiliary
constructions: Infi can be a host for clitics only if Infi triggers V-raising. This
hypothesis captures correctly the Romanian data: in this language the participial
inflection does trigger V-raising, and correlatively it serves as a host for clitics.
Note also that the absence of clitics in English might be related to the fact that in this
language V-to-I is limited to auxiliaries.
These remarks are only a suggestion, and do not entitle us totally to exclude the
idea that in French, V-to-I may apply towards the past participle inflection. It thus
appears that the three properties of Romanian auxiliaries presented in Section 1.1.
do not necessarily cluster together: the raising of the lexical verb out of VP may be
allowed in French, 54 without any constraint on the auxiliary itself, which is free to
move to the upper Infi node (hence subject-auxiliary inversion, and Tense
inflections on the auxiliary). In Romanian on the other hand, the auxiliary cannot
move to Infi (hence no subject-auxiliary inversion, 55 and no Tense inflection on
auxiliaries).
certain pro-drop Romance languages, such as Spanish, this rule seems to be obligatory: thus,
in Spanish, neither adverbs nor floating quantifiers are allowed to intervene between auxiliaries
and lexical verbs.
5%
our hypothesis concerning the sentence structure of Romanian is correct, the example in (31)a
is not necessarily obtained by V-second: V-to-Infl raising suffices to explain the order V NP S ,
since the subject occupies a sister position to VP, that the verb necessarily bypasses on its way to
Infi; V-second, i.e., "Move (V-)Infl to Comp" may be at work in (31)a, but if it is, it applies
vacuously (if we leave aside [Spec, IP], which is an A'-position, Comp is adjacent to Infi. The
example in (31)b is also directly obtained in the representation in (45')b, where V raises to Infi
bypassing NPS
precisely, pre- and post-verbal clitics),1 which will be shown to be related to each
other via a rule that preposes the verb.
The main theoretical problem raised by clitic elements is the relation between
syntax and the lexicon. Any analysis of clitics must reconcile their morphophonological properties (the fact that the clitic and its host make up a very coherent
element, which looks much like a word) with their syntactic characteristics
(pronominal clitics, for instance, satisfy subcategorization requirements of verbs).
Two main distinct approaches have been adopted: (a) according to a syntactic
theory of clitics (Kayne (1975)), the clitic is generated by the rules of syntax in a
standard syntactic position from which it is moved via an obligatory rule of clitic
placement and adjoined to V; (b) according to the morphological theory of clitics
(see Borer (1984)), the complex [clitic + verb] is a lexical item formed by the word
formation rules of the Lexicon and directly inserted into the syntactic structure (the
phrasemarker obtained by the application of base rules). The difference between the
two approaches can be viewed as relying on a derivational as opposed to a
representational framework: thus the S-structure configurations are indeed identical
(at least in the rough presentation given here) in Kayne's and Borer's analyses:
(1)
VP
I
\
V
NP
I
I
cli+V
ej
The two analyses differ with respect to the postulated D-structure: for Kayne
(1975) the position occupied by ei is occupied by the clitic at D-structure, whereas
for Borer the relation between the clitic and ei is purely representational. The
representational as opposed to the derivational approach correlates here with a
difference between a morphological as opposed to a syntactic approach of
cliticization because of the particular type of S-structure configuration, namely a
complex X element - the sequence [ycl+V] - made up of two distinct components.
Underlying Borer's analysis is the strong Lexicalist Hypothesis, according to which
X constitutents can only be formed in the Lexicon; this view is compatible solely
with the representational approach.
2. 1. 1. Clitic Placement as adjunction to Infi
This strong version of the Lexicalist Hypothesis is not unanimously accepted,
however, consider for instance the current analysis of verbal inflection: die verb and
its agreement+tense affixes form an X element which is assumed to be derived via
a syntactic rule, V-to-Infl raising (the verbal root is generated under V, the
agreement and tense morphemes under Infi - or, according to Pollock (1989), under
two distinct functional heads). Given this current analysis, it is plausible to assume
an analogous rule for clitic placement. To be more precise we can refer to Roberts's
(1991) typology of incorporation processes, which distinguishes three types of
Jean le connait
John him knows
'John knows him'
For a detailed discussion of the differences between clitics and affixes see Zwicky and Pullum
(1982).
le
connsSt
Consider now the order V-cl that characterizes positive imperatives in all Romance
languages, as well as Italian infinitivals, Romanian gerunds and certain
subjunctives, etc.; I give French examples, but the same data can be found in Italian,
Spanish, Romanian:
(4)
a. aide-le
b. *n'aide-le pas
c. nel'aidepas.
The first problem to be solved is whether the relative position of the clitic and the
verb that can be observed in (4)a - as opposed to (2) - is due to a rule of Vpreposing or to a rule of clitic postposing.3 Since downgrading rules are marginal,
if not altogether absent, it is reasonable to assume the first alternative (which fares
better on many other grounds, which I will not review here). This hypothesis,
which goes back to den Besten (1977, 1983), can be found in Rivero (1988a; to
appear), Kayne (1991), Roberts (1991): the clitic inversion characteristic of
Romance imperatives (see (4)a) is due to the rule of Move (V-)Infl to Comp4 (this
is an updated restatement of den Besten's formulation, which takes into account the
current hypothesis concerning the structure of IP constituents, and the Head to
Head Constraint).
Given the structure in (3), the sequence V-cl characteristic of (4)a would be
derived as in (5):
3For still another possibility see Rouveret (1989), who assumes that enclitic pronouns are not
transformationally related to proclitic pronouns.
4
F o r Kayne (1991) the landing site of the preposed verb may be other than Comp. This
assumption is meant to capture the fact that in Italian infinitives postverbal clitics coexist with a
lexical element that is probably dominated by Comp.
(5)
53
CP
IP
-I
clj V-I
f t
VP
V
NP
tv
ej
regarde le
The problem with this analysis is that it relies on an "excorporation" process: in
order to bypass the clitic, the sequence V-I has to move out of an incorporated
sequence, cl-V-I. Within Baker's (1988) theory, excorporation is banned by a
lexicalist principle which rules out words that contain traces (due to incorporation
the sequence cl-V-I is a word; by moving V-I out of it we would obtain another
word which contains a trace: cl-tv-l) Another way of banning excorporation is by
invoking an ECP violation (see Roberts (1991), among others): the trace of the
moved inflected verb is not antecedent governed because of the intervening clitic.
2. 1.2. Clitic Placement as adjunction to IP
It is, of course, possible to devise various assumptions and mechanisms which will
have the effect of allowing excorporation for incorporation by adjunction and ruling
it out only for incorporation by substitution into a morphologically subcategorized
head (see in particular Roberts (1991)). But this kind of theory will not really solve
the excorporation problem, which is, in fact, one aspect of a more general problem:
by applying Move (V)-I we are bound to move the whole sequence [i n fl cl V-Infl]
because the only visible I node is the highest one (by the A-over-A principle, or
whatever restatement of it that we may want to adopt). In order to leave behind the
clitic, we must stipulate that Move I applies to a lower I node, the one that
dominates V-I.
In what follows, I will propose a theory of cliticization which does not rely on
excorporation. 5 My main hypothesis will be that cliticization relies on IPadjunction:
5
Another proposal that avoids excorporation can be found in Kayne (1991), who assumes that Vcl sequences rely on configurations of the type in (i):
(i)
...V ... Cl+I ,..[ V p[ve]....]
(i) is like (5) insofar as the clitic has adjoined to I; (i) differs from (5) in that the verb does not
move first to I and then further on, but directly to the upper position, which according to Kayne
is an I-adjoined position. According to this analysis, the V movement that gives rise to
postverbal clitics does not obey the structure preservation condition, which is maintained under
my own analysis.
(6)
CP
I
IP.
I
CL
cl;
A
IP
I
-I
V-I
VP.
I
V
NP
I
tv
ej
_J
regarde
le
Move I is free to bypass the clitic because the node Infi does not dominate the clitic;
no excorporation is needed.
The hypothesis that pronominal cliticization involves adjunction to IP rather than
to I (I argue below that adverbial clitics are the only Romanian clitics that adjoin to I
itself) may be derived independently as a consequence of a well-formedness
condition on clitic chains: a pronominal clitic must c-command its trace, and an
element adjoined to Infi does not c-command (if we define c-command in terms of
branching nodes, see Reinhart (1976)) any position inside VP.6 The C-command
problem can be solved if we assume that pronominal clitics necessarily adjoin
higher, to IP, as shown in (7). Pronominal clitics cannot adjoin higher than IP, e.g.,
to NegP or CP, probably because Neg and Comp would block the antecedent
government of the clitic trace by the clitic.
6\Ve may of course assume that the Infi node that dominates the suing cl-I does not branch, but
this is a mere stipulation. This solution used to be commonly adopted for a similar problem
posed by the idea that the clitic attaches to V:
(i)
W
V NP
/ I I
cl V t c ,
In (i) the node V dominating the sequence cl V has to be considered as non-branching for the
clitic to be allowed to c-command its trace.
a. Ne la venra-t-il pas?
not her will-see-he not?
'Won't he see her?'
b. N'est-il pas parti trop tt?
Not-is-he not left too soon?
'Didn't he leave too soon?'
According to den Besten (1977, 1983), Kayne (1983,1984) and Rizzi and Roberts
(1989), examples (8)a-b are derived from (9):
(9)
a. II ne la verra pas.
b. II n'est pas parti trop tt.
The sequences set in bold characters move leftward, by (residual) V-second. But
given the IP-adjunction hypothesis proposed above, these sequences are not
dominated by Infl, and therefore should not be accessible to Move Infi (since they
do not even form a constituent they should not be accessible to any movement rule):
(10)
CP
IP
Spec
(NegP)
Pronj IP
Infl VP
V-Infl
a. ne la verra
b. n'est
il
il
tNeg-Pron-V-Infl
tNeg-V-Infl
pas
pas parti
Note that the problem just mentioned arises independently of the IP-adjunction
hypothesis, in relation to the negative clitic. It is indeed currently assumed (see in
particular Zanuttini (1989)) that the negative element ne heads a NegP projection
is well-known that in certain languages clitics do not cluster around Infi but rather around
Comp. This suggests that (15) should be stated not in terms of Infi but more generally, in terms
of functional heads. It would be interesting to find out whether the variation concerning the host
of clitics (Infi versus Comp) correlates with other cross-linguistic differences. NP-internal (or
rather DetP-intemal) clitics, which are not under investigation here, are covered by the suggested
extension of (15) because Determiners are functional heads that govern NPs (see Abney (1987)).
0 regardej le ej
A
regardej-le); e k e;
A ;
(Move I-to-C)
(Merging)
Move I + Merging
(V-preposing).
The impossibility of the verb moving over the negative clitic may be due to the fact that the
negative clitic cannot incorporate to the verb, and this may in turn be due to the fact that Neg is a
head which projects a functional category, unlike pronominal clitics; an S-structure constraint may
also be invoked, which requires that Neg should C-command Tense at S-structure (see Laka
(1989)).
violates the HMC (because the X position occupied by the clitic is skipped by Vpreposing), but obeys the ECP, due to clitic merging.
2.1.5. The adjacency condition on Merging
In order to further specify the constraints that govern V-preposing and clitic
merging, let us assume that a lexical element X intervenes between Comp and IP;
after V-preposing to Comp we end up with (20), which is illicit because the
intervening clitic blocks the antecedent government of the verb trace:
(20)
This violation can be avoided if the clitic is allowed to move and adjoin to V-I
bypassing X. I shall assume that an adjacency condition holds on Merging, which
rules this derivation out:
(21)
regarde-le
(23)
[ C ompregarde r le k ] [ I P t k U]
b. trimite-mi-o
send-meDat-herAccus
What is the derivation of these examples: did the V bypass both clitics in one step,
or did it first land in some intermediary position? The second hypothesis is
probably incorrect because there is no intermediary landing position for the verb.
Following the first hypothesis, on the other hand, we must understand what makes
it possible for both clitics to incorporate: because of the adjacency requirement the
second clitic cannot merge with the inflected verb. This problem is solved if we
assume that strings of pronominal clitics are flat structures, dominated by a single
node that adjoins to IP:
da-i-1
give-him-it
Given the configuration in (24'), V-preposing bypasses just one clitic position, and
clitic merging of the two-pronominal string i-1 'himDat-himAccus' applies under
adjacency.
The hypothesis that sequences of pronominal clitics are flat is corroborated by
the various idiosyncracies that characterize the order of pronominal clitics: in (25)a
the accusative clitic precedes the dative clitic, and in (25)b the reverse order is
observed:
(25) a. Je le lui dirai.
I it him sayfut
b. Je te le dirai.
I you it sayfut
Insofar as we cannot find any syntactic explanation for this peculiar distribution, we
may assume that the strings of pronominal clitics are not hierarchichally structured.
(Restructuring Incorporation)
(Merging)
Besides the different positions they occupy with respect to the verb, the two types
of clitics present a number of other correlated differences, which are directly
explained by the account proposed here: (a) unlike restructuring clitics, merged
clitics cannot be stranded by subsequent applications of Move Infi; (b) coordinating
merged clitics is impossible, whereas coordinating preverbal clitics is marginally
acceptable; (c) in Section 2.3.4. below it will be shown that Merging has an
important effect on the phonological shape of Romanian clitics.
10
For the existence of a Morphological Component that relates PF to Phonology see Halle
(1989).
11
The latter possibility does, however, arise in cases where the main verb subcategorizes for a
Infl-headed complement (on this proposal see also Rizzi and Roberts (1989)), and this is the case
with Romanian auxiliaries (see Chapter 1 and Section 2.4. below).
2. 2. Romanian clitics
The clitic system of Romanian is richer than that of the other Romance languages:
besides pronominal clitics, this language has at its disposal adverbial clitics (other
than the negative adverb, which is found in all Romance languages) and clitic
conjunctions (see a for infinitives and s for subjunctives, discussed in Chapter 3);
more importantly, Romanian auxiliaries qualify as verbal clitics.
All these elements share a peculiar distribution, namely they cluster around the
inflected verb. The ordered sequences given in (27)-(28) are the two maximal
possible strings in Romanian; in case one or several elements are not instantiated,
the relative order of the other elements is not affected:
(27)
a.
b.
(28)
a.
b.
mai fi vzut
thatsubj not
her again be seen
'that [I] should not have seen her again'
a
nu
mai fi
to
not her again be
'not to have seen her again'
vzut
seen
The order in (27) characterizes simple tenses. Note that the only clitic
complementizers of Romanian are s and a for subjunctives and infinitives
respectively. In Chapter 3, an explanation will be proposed for the fact that the
complementizer that heads Romanian indicatives is not a clitic.
12
For the sake of exhaustiveness I have inserted Aux2, but in Section 1.6. it has been shown that
fi does not behave as a clitic, but rather as part of a discontinous morpheme, directly generated
under Infi. The distribution of fi, which was discussed in detail in Section 1.6., will be left aside
here.
The sequence in (28) differs from (27) in two ways, which will be shown to be
correlated: Auxi appears and the feminine singular clitic 'her' is postverbal. All
the other clitics show up in the pie-auxiliary position.
At this point in our discussion, the sequences in (27) and (28) seem to be
disconnected from each other; it will be shown that they rely on the same
underlying configuration, their differences being due to the Move-Infl rule that
applies in (28), but not in (27).
2.2. 1. Deriving linear order from hierarchical structure
It is easy to see that the very strict linear order in (27) need not be stipulated (as in
template morphology), but is due to general principles and well-formedness
conditions of the grammar. Thus, the relative position of Neg, Infi and Comp is
presumably due to the government requirements that hold between these functional
categories and their projections: in quite a number of languages, the Neg head
governs an IP constituent; NegP itself is governed by Comp (see Zanuttini (1989)).
Hence (29):
(29) CP
C NegP
Neg
IP
c. *pot il vedea
[I] can him see
il pot vedea
Example (30)a shows that the position of adverbial clitics is between pronominal
clitics and the verb. Given the hypothesis that clitic pronouns adjoin to IP, the order
in (30)a is compatible with two possible positions for the adverb: adjunction to Infi
or adjunction to V. The ungrammaticality of (30)b allows us to choose among these
possibilities; if adverbial clitics were allowed to adjoin directly to the verb (as
proposed by Rivero (1988a, to appear)) nothing could rule (30)b out. If instead we
assume that adverbial clitics attach obligatorily to Infi, the ungrammaticality of
(30)b may be due to the fact, which is independently motivated (see Section 1.4 and
below), that the verb a putea 'may, can' subcategorizes for a VP complement: the
infinitive in (30)b-c is truly "bare", in the sense that it does not raise to Infi, but
stays inside VP: in the absence of Infi, the adverbial clitic, just as the pronominal
clitic (see (30)c)), cannot show up attached to the lexical verb. The
ungrammaticality of (30)b thus forces us to assume that V cannot be a host for
adverbial clitics. Let me stress that adverbial clitics are the only clitics that could be
assumed to be directly adjoined to V for the following reasons: (a) they are closest
to the verb (whenever a pronominal clitic is present the order Pron Adv V i s
obligatory and Adv Pron V precluded); (b) they are taken along under Move I-to-C
(see the discussion in section 2.3.1. below). We then reach the conclusion that no
clitic adjoins to V.
On the basis of the preceding remarks, we are led to assume that the linear order
given in (27) relies on the hierarchical configuration in (31); I have left aside Aux2,
which has been discussed in Chapter 1:
(31)
CP
l \
C NegP
I
\
Neg IP
Pron;
IP
Infi
Adv
VP
V-Infl NP
V'
tv
NP
In sum, the very strict linear order in (27) need not be stipulated; it derives from the
hierarchical structure directly obtained by the base rules of syntax and by the
adjunction of pronominal and adverbial clitics to IP and I, respectively. In the next
section, it will be shown that the highly structured configuration in (31) allows us
to account for the various reorderings of clitic clusters.
The present analysis may be compared with lexico-morphological approaches,
which could only stipulate distinct slots for each clitic (as one would do in template
morphology); besides its non-explanatory character, such an approach makes the
false prediction that clitic clusters cannot be reordered by the rules of syntax (under
the hypothesis that words are not accessible to syntactic rules). These shortcomings
would also arise within a syntactic theory of clitics, according to which clitic
clusters would be flat constituents generated under a CL node adjoined to Infi in the
syntax:
(32)
I
CL
cl
I
cl
...
Such a structure would not account for the asymmetries discussed below between
Romanian adverbial and pronominal clitics (the order V-cl is possible, in certain
well-defined contexts, for pronominal clitics, but not for adverbial clitics); nor could
we assign an adequate structure to Romanian auxiliary structures. What we need is
a syntactic theory which assigns a hierarchical structure to clitic clusters (but recall
that there is a flat sequence inside the clitic cluster, namely the sequence of
pronominal clitics; see the discussion of examples (24)).
The hierarchical structure in (31) is subject to Restructuring Incorporation and
Merging, under the general conditions proposed in previous sections.
2. 2. 2. A definition of clitics
To conclude this presentation of Romanian clitics, let me try to define the notion of
syntactic clitic (to be kept distinct from phonological clitics), by trying to establish
the common property of clitic elements belonging to distinct syntactic categories
such as N, Adv, V:
(33)
appears to exist between the case depicted in (a) and the other two: functional
categories such as Comp or Neg are clitics but they are not inserted by Clitic
Placement in an adjunction position; rather they are heads that take complements.
Therefore, adjunction (to either IP or Infi) cannot be viewed as the defining
property of clitics. Note further that syntactic clitics do not necessarily present nonclitic counterparts (the only clitics that present them are pronominal clitics). Also, it
is not the syntactic position perse that defines clitichood: clitic conjunctions occupy
the position which is occupied by non-clitic conjunctions; the clitic status of certain
conjunctions is due to their being subject to Functional Coindexation/Restructuring
Incorporation (due to the fact that they are adjacent to the inflected verb, unlike nonclitic conjunctions, which are separated from Infi by (Spec, IP)).
a. spune-mi
tell-me
b. d-mi-1
give-me-it
(36)
CP
Iv
C (*NegP)
(*Neg)
IP,
I
Prori;
IP
Infi
VP
(Adv) V-Infl
(34) a. spune
tell
b. d
give
mi
me
mi-1
me-it
mi
me
ty-Infl
ty-Infl
tAdv V--Infl
CP
C
(*NegP)
(*Neg)
IP.
I
Prorij
IP
Infi
VP
(Adv) V-Infl
a. s
se
intimple
b. indmple
se
tv_Infl
l^Rivero's (1988a, to appear) rule of V movement differs from mine only insofar as it presents an
intermediary step, a position labelled Infi. Note that Rivero's Infi slot, which dominates s,
corresponds to my Comp. The difference is not only terminological: Rivero postulates two
positions, Comp and Infi, whereas I postulate just one position; in my analysis V lands in the
position of s in (37)b; for Rivero it goes through this position, but does not land in it. It is not
clear what kind of empirical evidence could help distinguish between these two hypotheses, since
both approaches predict that elements such as s are bound to be absent in V-preposing
constructions (here because this is the landing site of the rule, according to Rivero because the
same position is an intermediary slot).
(39)
69
a. trimitej-l ej acas15
send-him home
b. nu-ltnmite
not-him send
c. *nu trimitei-1 ei
d. *tiimitei nu-lei
(39)c is ungrammatical because between the negative adverb and the clitic there is
no head position for the verb to land on (see (36)); (39)d is ungrammatical because
an intervening head, the negative clitic, has been skipped, which leads to an ECP
violation: the negative particle is a head which blocks antecedent government of the
V trace by the raised verb. Recall that Functional Coindexation must be assumed to
apply to Neg V+I sequences, in order to explain why V second can move not just
V+I, but the whole sequence Neg V+I. The blocking effect of Neg (39)d clearly
indicates that an intervening X blocks antecedent government of the trace of a
moved Y element, even if X and Y are functionally coindexed. Merging is the
only escape strategy; the ungrammaticality of (39)d indicates that Neg cannot merge
with the preposed verb, probably because it must preserve its status of head of
NegP. The brackets around NegP and Neg in (36) and (38) are intended to indicate
that the negative head is necessarily absent whenever V-preposing applies.
Consider next the example in (40):
(40)
Negsinduj-1 ey acas...
not-finding-him home
The peculiarity of this example resides in the fact that the negative adverb (note that
the phonological form of this element is different from that of the negative element
nu) is taken along by V-preposing; 16 antecedent government is therefore not
blocked, and the example is grammatical. The exact derivation of (40) is not
relevant; we may assume either of the following possibilities: (a) ne is generated
attached to Infi itself, on a par with the other clitic adverbs (mai 'still, again', cam
'quite a bit'; $i 'also, too', etc.); the Move Infi rule applies to the string dominated by
the highest Infi node, thus taking ne along, just like the other clitic adverbs (see
nemaigsindu-1 acas 'no-longer-finding-him home'); (b) ne is generated in the
same position as nu, i.e., under Neg, a head that subcategorizes for a projection of
^Romanian imperatives preceded by a conjunction allow the clitic to precede them (as notated by
the dash, in (i) phonological cliticization may apply between the clitic and the conjunction; this is
a general property of Romanian clitics, see Section 2.3.4.); the clitic may also occupy the
standard postverbal position:
(i)
ia-1 }i-l du
ia-1 du-I
'take-him and-him take away'
(ii)
ia-ne i ne du
ia-ne i du-ne
'take-us and-us take away'
assume that this rule ends up in Comp, just as in the general case (see Romance imperatives).
Compare Rivero (1988a, to appear), who assumes that in Romanian gerunds the verb ends up in a
position labelled Infi (which is distinct from AGR/Tense), the host of the invariable particles of
subjunctives and infinitives (see footnote 14 above).
enclisis
and
Merging
The clitic pronouns in (41) differ from their non-clitic counterparts both by their
phonological shape and by their distribution: 18 non-clitic object pronouns (pe mine
/ tine / el / ea 'pe me / you / him'; mie, (ie, lui 'me-Dat, you-Dat') occur in the
environments characteristic of NPs, as opposed to clitics, which necessarily precede
the verb:
(41)
a. Ion m / t e / f l / o / n e / v / i i / l e apreciaz
Indeed, it should be noted that Romanian does not present diphthongs of the type
a, which would have to be postulated by any analysis that does not insert a word
limit between the clitic and the verb in mapreciaz (see (41)a). Compare the clitics
in (43), which occupy the preverbal position, like those in (41), but differ from
them by their phonological shape (compare -1, -i, m-, -p to fl, ii, m, ip), which
signals phonological cliticization. Thus we may say that the clitics in (41) are
syntactic clitics, whereas those in (43) are both syntactic (since they occupy the
same position with respect to the verb) and phonological:
(43)
a. (i)
agteapt
mama.
Mana-1 /Maria-i
ateapt de or.
(iii)
Nu-I /n-o
/nu-i
a$tept
M-agteapt
te-agteapt...
scrie
Ion.
Maria-mi
des.
Nu-mi/nu-{i
/nu-i place
marea.
mi-aduce
p-aduce...
a.
Nutiec/co/ciia$teaptmama.
[he] does not know that him/ that her/ that them waits for mother
Maria il ateapt.
agteapt...
Ion.
mother does not know that meDat/ that youQat/ that himDat writes John
Maria imi scrie
des.
aduce...
19 If
this adjacency requirement is strict, i.e., if intervening empty categories are disallowed, on a
par with lexical elements, phonological cliticization may be used as a test for the distribution of
empty categories. In particular, (43)a would indicate that no empty subject is present between the
conjunction and the cl V sequence.
20we may also recall cases such as I'll go, where 7/ is a phonological, but not a syntactic clitic (it
occupies the same syntactic position as shall/will, its non-clitic counterparts).
Consider next the following examples, which are characterized by two correlated
properties: the clitics occupy the postverbal position, and they show obligatory
phonological cliticization (as indicated by the ungrammaticality of (46), with a
purely syntactic, non-phonological clitic):
(45) a. apnndu-m /-te /-1..
defending-me / -you / -him
b. apr-m/-te / -1..
defend-me/ defend-you[rself]...
(46) a. *apMnd(u) il
b. *aprfl
An obvious coiTelation can then be established between the obligatory character of
phonological cliticization in (45) - as opposed to its optional nature in (43) - and the
application of V-preposing. It seems natural to assume that the obligatory
phonological cliticization in (45) is an effect of Merging, which has been shown to
be itself triggered by V-preposing. Compare Restructuring Incorporation, which
allows, but does not force, phonological cliticization (phonological cliticization is
optional in (43)).
Conclusions
To sum up, the data examined in this section supports the theory outlined
previously. Moreover, we now have at our disposal a number of important tests,
which may help us infer the underlying structure in certain cases which would be
compatible with more than one analysis:
(a) Clitics cannot adjoin to V; they can only adjoin to Infi or to a Spec-less IP.
(b) If V-preposing does not apply, the relative order between pronominal and
adverbial clitics is Pron Adv V-I. Whenever a different order appears, we may infer
that a rule of V-preposing has applied.
(c) V-preposing strands pronouns, but never strands adverbs.
(d) V-preposing triggers obligatory phonological cliticization. In all the other
environments phonological cliticization is optional.
2. 4. Move I-to-C (V-preposing) in auxiliary structures
I now turn to the initial motivation of the foregoing investigation of cliticization,
which was the attempt to establish the underlying structure of Romanian auxiliary
constructions. It will be shown that the tests established above converge in
indicating that in Romanian auxiliary constructions the lexical verb is subject to Vpreposing; more precisely, the lexical verb raises to Infi and (V+)Infl raises to
Comp. This hypothesis is central to the analysis proposed in Chapter 1, which is
thereby strongly supported.
73
a 1-am rugat
[I] him-have asked
b l-a$ruga
[I] him-would ask
c *amil/o rugat
[I] have him/ her asked
d *ail / ruga
[I] would him/ her ask
(48) a amrugat-o
[I] have asked-her
b a ruga-o21
[I] would ask-her
(49) a il agtept /l-atept
[I] him wait
b il am in buzunar/ *l-am in buzunar
[I] it have in the pocket
c *il am rugat
[] him have asked
As observed in Chapter 1, the fact that in auxiliary constructions 'her' is
postverbal whereas all the other clitics show up before the auxiliary (see the data in
(47)a-b vs (48)), cannot be related to a syntactic peculiarity but most probably
involves phonological processes. It is indeed difficult to understand why is
necessarily postverbal on the basis of its [+fem] feature; its phonological
"weakness" is instead likely to be relevant. Let us then assume that the two
positions in (47) and (48) are not S-structure, but rather Surface Structure
positions; there is only one S-structure position, correlated with phonological
processes that account for the existence of two Surface Structure positions. Now
we have to decide whether the S-structure position is that of (47)a-b or that of (48).
The first possibility is to assume that the pre-auxiliary position is the "basic"
syntactic position, by assimilating "pre-auxiliary" with "pre-verbal".
The future construction is characterized by the same structure as (48)a-b, but the distribution of
is less clear-cut; both the pre-auxiliary and the pre-verbal positions are allowed:
(i)
voi ajuta-o
[I] will help-her
(ii)
voi ajuta
The crucial point is that the postverbal position is allowed; compare simple tenses, where can
only be in the pre-verbal position.
blUuxpAuxVP]]
CP/IP
I
\
Aux CP/IP
I
\
C
MP
Aux
V-I
ty.j
tv
NP0
In what follows it will be shown that this structure allows us to account for the
distribution of the clitics shown in (47)-(48) by assuming the analysis of
cliticization proposed above for non auxiliary constructions. I will then return to the
alternative analysis in (50), and point out its shortcomings.
According to an important generalization obtained in previous sections, adverbial
clitics adjoin to Infi and pronominal clitics to IP. In (51) there is no Infi that
governs Aux; the only available Infi node is the one associated with the lexical
verb. Hence the structure in (52), where both adverbial and pronominal clitics attach
to the lower Infi (and therefore to the lexical verb, as a result of V-to-I movement):
the expected linear order would then be Aux-Comp-Pron-Adv-V-I,
which can
indeed be observed in "biclausal" Tenses (see Section 1.2.1.). However, in the
auxiliary constructions illustrated in (47) and (48), pronominal clitics cannot
intervene between the auxiliary and the verb (see (47)c-d). The absence of this
word order may be analyzed as being due to the fact that V-preposing is obligatory;
as a result, the pronominal clitic (generated in the medial position) necessarily ends
up in the postverbal position, which can be directly observed for 'her':
NP 0
a. a (mai)
vaz-ut-Op
tp
t(Adv)V-i Ion
tv
ep
tv
ep
John
tp
t(Adv)V-i Ion
(again) see-her
John
The proposed analysis explains why in the future and conditional paradigms the
infinitive lacks the complementizer particle a 'to 1 : 2223 since the preposed V+Infl
occupies the Comp position, the complementizer element a cannot be generated
(recall that Move I-to-C is an instance of substitution into an empty X position).
Compare (52)b to (53), where a is present and correlatively the clitic occupies its
base preverbal position.:
(53)
fr a jigni.
S e e Chapter 3, where it is shown that the analogy between a and to is not perfect.
A n alternative hypothesis can be envisaged: auxiliaries would be generated in the position
characteristic of a. But this assumption does not account for other important properties of
auxiliary structures, and in particular does not explain the distribution of clitics.
23
Note furthermore that the clitic climbing needed here is strictly local: the clitic
climbs over the complex X element formed by its host (Infi and the verb raised to
24
T o choose between the two analyses, one may also take into account two phenomena that used
to exist in older (see also dialectal) varieties of Romanian (the examples are from Bredemeier
1976: 38-39):
(i)
Tu ai aruncat-o.
you her have thrown-her
c eu cind am luat
that I when her have taken
(ii)
L-am vzutu-l/o voi lua-o/l-a splatu-l/m-oi bucura-m
him-have [I] seen-him/ her will [I] take-her/ him-has washed-him/ myself-would [I] enjoy
myself
Examples (i) show that could appear in front of the auxiliary; in (ii) clitics appear twice, both in
the pre-auxiliary and the postverbal positions, but notably not in between the auxiliary and the
verb.
25
T h i s supposes the existence of a Morphological Component between S-structure and
Phonological Structure (see Halle (1989)).
^Consider the following examples, which illustrate the future tense paradigms constructed with
the subjunctive (see Chapter 1):
(i)
Am s- due acas.
[I] have s-him take home
'I'll take him home'
(ii)
s-1 due acas.
[I] s-him take home
'I'll take him home.'
(iii) *L-am s due acas. (compare (i))
(iv) *L-o s due acas. (compare (ii))
(i)-(ii) do not involve V-preposing to Comp, as indicated by the presence of s, which occupies
the Comp position, and by the fact that the clitic occupies the pre-verbal position. Since Vpreposing does not apply, the clitic does not incorporate to the verb, which may explain the
ungrammaticality of (iii)-(iv) if we assume that the rule - be it syntactic or phonological responsible for the pre-auxiliary position of the clitics in (48)a-b can only apply to merged clitics.
Infi) and the elements that have incorporated to it via Restructuring Incorporation
(the auxiliary):
(55)
(28)
The important fact about this linear order is that the adverb and the pronominal
clitics are not adjacent to each other, as they are in non-auxiliary constructions
(Pron Adv V). The reordering in (28) is exactly what we expect under the
hypothesis that the lexical verb is proposed via Move I-to-C, a rule which takes
clitic adverbs along and strands pronominal clitics.
Consider now the concurrent hypothesis in (50), which seems to easily
accomodate the position of pre-auxiliaiy clitics. They are adjoined to Infi (or to IP);
the postverbal clitic 'her' occupies a syntactically irrelevant position. Under this
analysis the following questions are left unanswered: (a) Why is phonological
cliticization obligatory? (b) Why is the order Pron Adv Aux Vimpossible? (c) Why
are the orders Aux Adv V Pron and Pron Aux Adv Vcorrect? The first question
refers to the contrast shown in (49)b and (49)c, which indicates that phonological
cliticization is obligatory. Since phonological cliticization has been analyzed above
as an overt manifestation of clitic merging, this means that clitic merging necessarily
applies in auxiliary constructions, which in turn means that V-preposing necessarily
applies. But this generalization cannot be captured under the structural hypothesis
given in (50) because in this case no V-preposing applies.
Let me now comment briefly on the points in (b) and (c), which are related: if it
is correct to assume that pronominal clitics adjoin to (the IP constitutent headed by)
a "high" Infi node, then we would expect adverbial clitics to be able to adjoin to that
same Infi node. The predicted linear order would be *Pron Adv Aux V, but this
sequence is completely ungrammatical. One could of course stipulate that clitic
adverbs necessarily attach to lexical verbs (this would derive the correct Pron Aux
Adv V), but this is incorrect, as indicated by the impossibility of (57):
(57)
If adverbial clitics attached to lexical verbs, nothing could rule (57) out.
According to the analysis proposed here, the ungrammaticality of (57) is due to
the fact that the modal a putea 'can, may' does not govern a CP/IP complement, but
a VP; therefore there is no "low" Infi node to which the adverb could attach. Given
the idea that pronominal and adverbial clitics attach to the same I(P), the
ungrammaticality of (58) is also correctly expected:
(58)
*pot vedea-o
[I] can see-her
V-I-pron-Aux t A u x [ c [ c ty-i-pronltr W i
- -t v ]]]
pedepsi-l-a$
punish-him-would-lst
The dotted lines indicate the leftward merging of pronominal and auxiliary clitics, to
be distinguished from the standard head to head movement indicated by continuous
lines. What interests us here are the last two operations, namely the movement of
vzutu-I past the auxiliary am: the string ty-i-pron is properly governed because
the auxiliary has cliticized to the left, thus avoiding a potential ECP violation. Note
that the fact that on the last step the clitic is taken along (and not bypassed by the
verb, as on the second step) is to be expected given our analysis: as a result of Vpreposing, the clitic merges with the verb on its left and therefore can no longer be
left behind.
Conclusions
To summarize, the main distributional characteristics of Romanian clitic clusters
have been accounted for by assuming a syntactic theory of clitics, according to
which syntactic clitics are X elements that are subject to Functional Coindexation
and/or Restructuring Incorporation. Clitics such as conjunctions and the negative
adverb are generated in functional head positions, whereas pronominal and
adverbial clitics adjoin to (I)P. The hierarchical configurations underlying clitic
clusters are constrained by general principles such as C-command (between the
clitic and its trace, in case there is any), as well as by conditions depending on the
syntactic category (N, Adv or V) of each clitic. Since syntactic clitics occupy a
syntactic position which is distinct from the position of the host, clitic clusters allow
for certain internal reorderings, which have been analyzed as being due to a rule of
V-preposing (which may bypass not only pronominal clitics, but also clitic
auxiliaries). This rule triggers clitic merging, which gives rise to rigid constituents:
a merged clitic leaves its IP-adjoined position and merges with the verb, and
27
According to this analysis, V-preposing is local, on both the first and the second step of the
derivation. This analysis cannot cover examples such as ascultatu voi l-a(i vreodat 'listened you
him-have some time', which are sometimes quoted in the literature (see Rivero (1988a, to
appear)). I suspect that these (very marginal) examples are due to some kind of analogical
interaction with other constructions.
This late survival of inverted conjugations, which disappeared quite early in the
other Romance languages (pre-literary period for French, 6th-7th century), may be
due to the fact that until the 17th century Romanian displayed only the series of
clitic pronouns illustrated in (43), which had to cliticize phonologically onto another
word (compare the clitics in (41), which are syntactic clitics that need not cliticize
phonologically). But the conditions on rightward phonological cliticization are quite
restrictive: in particular it cannot apply if the word starts with a consonant. Compare
leftward cliticization, which can trigger the insertion of a final vowel (see 1-am vzut
vs vzutu-1-am), a phenomenon comparable to the insertion of t in the French
subject clitic inversion, see va-t-il partir?, etc. V-preposing may have been a means
of providing a host on the left of the clitic. It is only during the 17th century that the
"syntactic non phonological" clitics given in (41) appeared, by the development of
the prothetic vowel i. Henceforth, the rule of V-preposing is no longer required and
it gradually disappears. In present day Romanian the paradigm in (60) is completely
obsolete (almost unrecognizable); however, this type of construction survives (in
highly literary style) in subjunctives (see (37)b above) and in compound tenses:
2^Under its current formulation, Wackernagel's law says that the inverted conjugations were
bound to appear in case clitics were sentence initial; it does not say that they could not appear in
contexts in which sentence initial XPs (which presumably occupy the (Spec, CP) slot) were
present. Within the X' theory proposed in Chomsky (1986b), the C position is available even if
the (Spec, CP) slot is filled by wh-elements.
(62)
miratu-m-am
miratu-te-ai...
[compVHipNPcltv]
I
I
These examples have been analyzed as involving auxiliary verbs that take CP
complements 1 characterized by Move (V-)Infl to Comp; since at S-structure the
Comp position is occupied by the preposed inflected verb, the obligatory absence
ofa is expected if we assume that this particle normally occupies the Comp
position. In this section further evidence will be provided in favour of this
hypothesis.
1
The bare infinitive in (l)-(2) should be distinguished from the one that shows up after a putea 'to
can':
(i)
pot vedea.
[I] her can see
The distribution of clitics, as well as a number of other characteristics, indicate that the bare
infinitive in (i) is not a CP, but a VP complement (see Chapter 1).
3. 1. 1. 1. is an Infi element
The Infi status of a is supported by a morphological observation: Romanian
infinitives do not present any specific morphological inflection 2 that could be
analyzed as the realization of Infi: -a, -i or -e in a mine 'to eat1, a cid 'to read', a
spune 'to say', belong to the lexical root of these verbs (see voi mincap 'you eat',
voi citip 'you read', voi spunep 'you say'). It would then be reasonable to assume
that the absence of infinitival verbal inflections is made up for by a different
realization of Infi, the particle a; originally a preposition (going back to the Latin
ad), a is nowadays exclusively an infinitive marker. Romanian infinitives would
then appear to be comparable to English infinitives, which present a prepositional
Infi element, to, correlated with the absence of morphological inflections on the
verb.
The analogy between a and to infinitives is also favoured by certain aspects of
their distribution; they may appear after certain aspectuals and in subject clauses:
(3)
These examples indicate that a infinitives, just like to infinitives, are sentential
projections which occupy argument positions. Compare bare infinitives, which, in
both English and Romanian, do not count as arguments, but only as verbal
projections, and therefore appear as complements of auxiliary verbs. To say that a
infinitives count as sentential projections does not settle the status of a: it could be
generated either under Infi, as to in English, or under Comp, as de in French. The
former hypothesis is supported by the following observation: unlike de, and like to,
the Romanian particle necessarily accompanies the infinitive in the examples above.
If we assumed a to be under Comp, we would expect it to be absent (on a par with
de in French), because infinitival Comp positions are in general empty in
environments of this type. Infinitival Infi nodes on the other hand are in general not
empty.3
Another potential argument in favour of the Infi status of a is related to raising
configurations. Kayne (1984) has analyzed the impossibility of de in the French
example in (4) as being due to the fact that de is a Comp element whose CP
projection constitutes a governing category for the embedded subject position;
therefore an NP-trace that occupies this position cannot be bound by the raised
subject, as required by principle A of binding theory:
(4)
^Recall that the "long infinitives" in -are have become nominal forms; compare -r infinitives in
French.
3are infinitives are not IPs headed by an empty Infi, but either VPs (in English; cf also footnote
1 for Romanian) or V(+I)-headed CPs (in Romanian).
In the English gloss to is allowed, which is expected under the hypothesis that to is
under Infi: infinitival Infi does not count as a governor for the NP subject, whose
governing category is therefore the main clause. We could then be enclined to adopt
the same analysis for the Romanian example in (5), which contrasts with (4), and
seems parallel to its English gloss:
(5)
a. A nu mai ajuta ar fi
prostie.
*avreodatgindi
to sometimes think
c. a gindi
vreodat
d. peine parier l'italien apres cinq ans d'tudes dnote un manque de don
pour les langues
hardly to speak Italian after five years study denotes lack of gift for
languages
If we assume adverbs such as peine 'hardly' to be generated in a pre-VP position,
the word order in the French example in (6)d can be analyzed as being due to the
fact that in French infinitives the verb does not raise to Infi (see Pollock (1989)).
The Romanian example in (6)c could then be analyzed as involving V-raising to
Infi, past the adverb.
This analysis in fact undermines the idea that a is under Infi, because no position
is available in between Infi and VP, as a landing site for the raised verb. Recall
indeed that a restrictive theory of incorporation has been assumed here (see
Chapters 1 and 2), which disallows adjunction for Head Movement:4 this means
that the verb cannot raise and adjoin to a. According to the theory adopted here, an
incorporated sequence may be obtained in two ways: (a) affix incorporation is
triggered by morphological subcategorization (see Lieber (1980)); (b) cliticization
relies on Restructuring Incorporation (see Chapter 2), a reanalysis mechanism that
affects adjacent functional elements that occupy distinct syntactic positions. For a
number of reasons that we have already discussed with respect to Romanian
auxiliaries (see Chapter 1), the particle a cannot be analyzed as an affix; in other
words it presents no morphologically subcategorized position to which the verb
4
I agree with Rizzi and Roberts (1989) that the Clitic Placement of pronominals constitutes an
exception to this restrictive theory. This is not too problematic, since it is not clear whether Clitic
Placement should be viewed as a movement rule or as base adjunction.
85
might raise. We are thus left with the second type of incorporation, one for which
we need a syntactic landing site for the raised V.
One may of course assume two Infi positions, one for a and the other one for the
raised verb, but this would constitute a significant weakening of our X' theory. This
solution would also raise major difficulties for comparative syntax, because it
would require an explanation of why Romanian has at its disposal two Infi nodes,5
as opposed to English and French. These problems could be solved if we assumed
that a is not dominated by Infi but by Comp: Infi would then be available for the
verb to raise to.
Consider next the relative position of a/to and the negative adverbs nu/not
(7)
a. a nu vorbi
to not speak
b. *nu a vorbi
not to speak
In Romanian nu necessarily appears between a and the verb; it cannot come in front
of a. This distribution constitutes a further argument in favour of the Comp status
of a: the ungrammaticality in (7)b can be understood if we assume that the
projection headed by a constitutes a barrier for the relation between the negative
head and the variable it binds in IP. We know that IP does not constitute a barrier
for this relation: nu normally precedes the elements which are generated under Infi.
CP on the other hand does constitute a barrier for the relation underlying negative
quantification:
(8)
*tiu nu c a plecat
[I] know not that [he] has left
5
Note that for Romanian indicatives we need only one Infi node, which raises further questions,
concerning the special behavior of infinitives (and subjunctives, see 3.2.). Are we allowed to
postulate different configurations for IP, depending on the verbal mood? We are bound to say
"No" if we try to maintain a constrained theory of syntax.
6
Under the assumption that to is under Infi, in the same position as modals, we expect to not go
(parallel to will not go), which is ungrammatical.
which Case is assigned (by the preposition). The grammaticality of the Romanian
example can be understood if we assume a infinitives to be CP constituents, which
as such are allowed to take on a nominal status.
The hypothesis that a is under Comp is further supported by the data in (10); a
violation of the doubly filled Comp filter may be invoked, which prohibits the
coexistence of wh-elements (i.e., relative and interrogative pronouns) with [-WH]
Comp elements (i.e., subordinating conjunctions):
(10)
The English examples in (11) are grammatical, because tois an Infi element:
(11)
The doubly filled Comp filter operates in Romanian, as indicated by the following
impossibilities:
(12)
a/b.
e/f.
^In (14) the clitics not only can, but they must attach to the bare infinitive:
(i)
*se are cu cine distra
(same as (14)a)
(ii)
*nu }tiu cu cine trimite la mare
[I do] not her know with whom send to the seaside
This distribution is not surprising, because the (Spec, CP) node being occupied by a wh-element,
it determines a maximal projection of a sentential type, over whose boundary the clitic cannot
climb. The underlying structure of (14) would be an IP constituent governed by an empty Comp;
the wh-phrase occupies the (Spec, CP) position. The type of bare infinitive that shows up in (14)
thus appears to be different both from the one in auxiliary structures, and from the one that
appears after a putea 'can' (see footnote 1).
9ut see Rizzi's (1982) analysis of di.
lOwelsh presents elements introducing embedded tensed clauses (y 'that') and relative clauses (a
'who' and y 'that') which have been analyzed as preverbal particles (see Harlow (1983); Rouveret
(1987, 1990)) on the basis of constituency tests that indicate that they form a constituent with the
following verb.
(15) CP
XP
c
c
IP
ENFL
VP
V-INFL
NP&
V'
tv
In (15) the NP subject is generated as a sister of V', and stays in that position at Sstructure ((Spec, IP), which we have not represented, is an A'-position which in
indicatives is accessible to left dislocated elements; for reasons to be presented in
3.2., this position is not generated in infinitives and subjunctives). The verb moves
to INFL (or more precisely to Tense); 11 a is generated under Comp, and its Infi
characteristics are due to the fact that Comp is adjacent to Infi: 12 these positions
become indistinguishable, due to the rule of Functional Coindexation introduced in
the previous chapters.13 I delay a precise implementation of this idea till Section
3.2.2., when a unifying account of infinitivals and subjunctives will be proposed.
The IP structure characteristic of English and French differs from (15) in that at
S-structure the NP subject appears under (Spec, IP), thus separating Infi from
Comp. Since Comp and Infi are not adjacent,14 Functional Coindexation cannot
apply, so that functional elements such as to or de will unambiguously behave as
either Infi or Comp.
Let us now come back to the examples in (6), which illustrate an important
difference between Romanian on the one hand, and English and French on the
other. According to Pollock (1989), French infinitives do not raise out of VP (to
1
Romanian infinitives do not present any Tense contrast (a minca vs a fi mincat 'to eat/ to have
eaten' is an aspect contrast), and therefore we may be inclined to think that they lack Tense
features. This is probably a wrong conclusion, because it is reasonable to think that verbal
projections are necessarily governed by Tense. In other words,, we may assume that nominal and
verbal projections do not differ from each other by intrinsic features that would characterize the
head of the projection itself; NPs and VPs would be intrinsically indistinguishable, say XPs.
Their NP vs VP status would be due to the difference in the functional categories that govern
them: VPs would be XPs governed by Tense; NPs would be XPs governed by Det. Coming back
to Romanian infinitives, they count as verbal projections, and therefore must be assumed to be
governed by an abstract Tense element (notated Infi here), even if this element is phonetically
empty.
12
Note that independently of the exact status of a, its projection (i.e., a infinitives) counts as a
CP projection: on the assumption that a is under Infi, Comp is necessarily empty; under Haider's
(1988) theory, this is a case of "matching projection" of type CPAP.
^Functional Coindexation: Coindex adjacent functional X categories.
14
This analysis supposes that the adjacency condition on Functional Coindexation is sensitive to
empty categories such as PRO or NP-trace.
89
Infi). 15 If the structure in (15) is correct for Romanian infinitives, we must concede
that they necessarily raise out of VP, 1 6 on a par with finite verbs.
This peculiarity of Romanian infinitives could be derived as a consequence of
the structure in (15), if we assume that Affix-hopping is local, and in particular that
it cannot bypass an empty NP subject; therefore, given the position of the NP
subject in (15), the raising of the verb to Infi is the only way for Infi to attach to the
verb, and this attachment is a central morphosyntactic requirement (at S-structure
Infi cannot surface separated from V). Compare the situation of infinitival clauses
in a language like French, characterized (according to Sportiche (1988)) by the fact
that at S-structure Infi is separated from the verb only by the trace of the NP subject
that has raised to (Spec, IP); this trace may be assumed to be invisible for a number
of syntactic phenomena, 17 among which Affix-Hopping. This may explain why the
verb can stay inside VP: the affix is allowed to climb down to V over the trace of
the subject.
3. 1.2. 2. Lexical subjects in infinitives
The distribution of lexical subjects in infinitives provides evidence in favour of the
hypothesis that (15) is the correct underlying structure of Romanian IP constituents.
Romanian infinitives do not present any AGR feature (as opposed to the
Portuguese inflected infinitive), but they nevertheless take lexical subjects:
(16)
This behaviour contrasts on the one hand with finite lexical verbs, which in French necessarily
raise to Infi, and on the other with auxiliaries such as have and be, which are allowed to raise to
Infi only in infinitives.
16
For Italian see Belletti (1990).
17
T h e fact that this trace is invisible may be related to the status of the position that it occupies:
one may assume that it falls outside the A/A' distinction: it is not an -position, if A-positions
are defined as potential Case positions; but it is not an A'-position either.
l^This hypothesis is necessary in order to account for the fact that a clitic doubled direct object is
necessarily preceded by a preposition, which may be lacking if the direct object is not clitic
doubled: the role of the preposition would be to assign Case to the direct object in those contexts
in which the Case features are realized on the clitic, and hence are no longer available for the
object NP (see Jaeggli's (1982) restatement of Kayne's generalization).
but only towards a position which is assigned no Case under government; the Spec
position is precisely such a position.
Since in SVO languages nominative Case cannot be assigned under government,
the Spec-head agreement strategy is the only possible one: the AGR features on the
verb take inherent Case, which can be transferred to the subject NP under (Spec,
IP). Compare clitic doubling chains: Objective case is taken by the clitic, and it
cannot be transferred to the direct object position, because this position is not a
Spec position (it is normally assigned Case under government); hence the necessity
of a preposition which assigns Case to the doubled direct object.
One crucial element of this proposal is that AGR features on the verb are not
required for Nominative case assignment under government. An interesting piece of
corroborating evidence comes from Italian infinitives, which are a l l o w ^ to take
lexical subjects only if the Aux-to-Comp rule applies (see Rizzi (1982)), which
moves the auxiliary past the NP subject. Recall that in Italian, Nominative Case is
normally assigned in the preverbal position (see Rizzi (1982)). Since this
assignment depends on the existence of AGR features on the verb, it cannot apply
in Italian infinitives. The Aux-to-Comp rule has as a result a government
configuration, which allows Nominative Case assignment even if AGR features are
absent.
Now compare the distribution of lexical subjects in Romanian infinitives: no
Aux-to-Comp or AGR features are necessary, which provides evidence in favour
of the structure in (15), in which Nominative case is assigned under government.
3. 1. 3. Control structures
Another major problem of Romanian grammar will now be considered: infinitives
allow both lexical subjects and obligatory control:
(18) a. Am inceput a citi "Cei trei mupchetari".
[I] have started to read "The TTiree musketeers"
b. Am mincat inainte de a pleca.
[I] have eaten before of to leave
ate before I left.'
c. Ion a plecat inainte de a ajunge mama acas.
John has left before of to arrive mother home
'John left before mother arrived home.'
In (18)c a lexical subject is allowed. This possibility is difficult to reconcile with the
idea that in (18)a-b the subject is PRO. Compare French and English, which are
characterized by the complementary distribution between lexical subjects (which
can only show up in finite clauses) and PRO (which can only appear in non-finite
clauses). It is generally assumed that this complementary distribution is the only
theoretical possibility: lexical subjects must take Case and in order to do so they
have to occupy governed positions, whereas PRO occupies an ungovemed position
(this follows from its being specified as both an anaphor and a pronoun, see the
"PRO theorem" in Chomsky (1981)).
Let us then try to account for the Romanian data without introducing a theoretical
inconsistency: we should not admit, for instance, positions that are both governed
(which would allow for lexical subjects) and ungovemed (which would allow for
PRO). Two types of solution can be imagined: (a) one possibility would be to
assume that the structure underlying (18)c is different from that underlying (18)a-b;
(b) the second alternative is to admit that in Romanian infinitives the controlled
subject is not PRO.
We shall rapidly outline an implementation of the first solution, and show that it
is inadequate. The structure in (15), proposed above, can be assumed for Romanian
sentences that contain postverbal lexical subjects or pro, but PRO cannot be
supposed to appear under NPS, since this position is governed by (V-)INFL. It is
however possible to bring a different position into play, one not yet indicated,
namely (Spec, IP):
(15')
NP
VP
NPc
V-I
A
v;
I
tv
NPo
controlled subjects are PRO elements, despite the fact that lexical subjects are also
allowed.
We thus conclude that a structure like (15') cannot be the correct representation
of controlled infinitives in Romanian. Consequently, we must turn toward the
second possibility: controlled infinitival subjects in Romanian are not PRO
elements. This line of inquiry, which is independently needed to account for
controlled subjunctives, will lead us to develop a different theory, one in which
control effects do not depend on the presence of PRO. Following Manzini (1983),
Bouchard (1984) and Borer (1987), we shall define control as an anaphoric type of
relation (see Chapter 4). This approach will also allow us to account for subject
raising in Romanian infinitives (and subjunctives).
3. 2. 1. 1 S is an Infi element
The most obvious arguments in favour of the Infi status of s are related to
examples (20)a-b:
(20)
22
Subjunctive inflections are identical to indicative inflections, with the exception of the third
person singular (see elpleac 'he leaves'). Besides the present subjunctive, Romanian also presents
an invariable perfect subjunctive, constructed with the perfect auxiliary fi "be' discussed in 1.6.:
(i)
eu / tu /el ...s fi mincat
I / you / he ... s be eaten
'that I / you... have eaten'
23
Since s does not present any counterpart in English, we are going to preserve it in the glosses.
The English reader may translate it by that (in sentence initial position), or leave it out altogether
(when s ispreceded by a dislocated constituent).
No other element can separate s from the verb, not even the subject (see (20)b).
The strict adjacency requirement that holds between s and the verb cluster may be
taken to indicate that sif belongs to the verb cluster. In other words, s is some kind
of Infi element.
Constituency tests such as coordination lend support to this claim:
(22)
Aux position is occupied by the perfect auxiliary fi 'be', and under Adv we find a very
restricted class of clitic adverbs: mai 'again', prea 'too' (see Chapters 1 and 2). Note that
pronominal and adverbial clitics appear in the same order in the other verbal moods:
(i)
Pe Ion il }i vede primar.
pe John [she] him already sees mayor
John she already sees as mayor
(ii)
Pe Ion 1-au avansat.
pe John [they] him-have already promoted.
25
T h i s possibility is however not general across languages: *je sais [cpque [jptu viens] et
[jpMarie part]] know that you come and Mary leaves' (see Godard (1989))
3. 2. 1. 2. Si is a Comp element
A number of other properties distinguish s from Infi particles: (a)s is invariable;
(b) s can head an embedded clause; (c) its position is leftmost, necessarily
preceding clitics and negation:27
(24)
The fact that s i is clause-initial in (24)a suggests that s i occupies the Comp position
(if we assumed s to stay under Infl, we would have to postulate an empty category
under Comp; or else we would have to assume an obligatory string vacuous
movement of s (V-I) into Comp).
According to a clear empirical generalization, the Neg head subcategorizes IP
(see Zanuttini (1989)), but not CP complements. By virtue of this generalization,
and on the assumption that s is under Comp, we expect the ungrammaticality of
(24)b, on a par with (25):
26
The ungrammaticality of (22)c would be parallel to (i)-(ii), borrowed from Kayne (1975):
(i)
*Paul la [dteste et considre comme b6te]
(ii)
*Jcan vous [parlera et pardonnera]
Kayne quotes grammatical examples such as Paul les lit et relit sans cesse Paul them reads and
rereads again and again', and suggests that they may be due to the fact that lire et relire seems to
be analyzed as some kind of complex verb.
27\Ve should also add that s goes back to the Latin conjunction si.
(25)
a. S triasc Romania.
s live Romania
'long live Romania'
b. S se intimple ce s-o intimpla.
s se arrive what se-may arrive
'arrive what may'
(27)
a. Triasc Romania.
live Romania
b. Intimple-se ce s-o intimpla.
arrive-se what se-may arrive
c. * S intimple-se ce s-o intimpla.
d. * Intimple-se s ce s-o intimpla.
There are two correlated differences between (26) and (27): the presence vs the
absence of the particle s and a difference in word order: in (26)b the pronominal
clitic se occupies the canonical clitic position, whereas in (27)b se necessarily
follows the verb. This data can be captured by a rule that moves the verb to the
initial position, and correlatively prohibits the insertion of s (see the
ungrammaticality of (27)c-d, due to the presence of s). This analysis suggests that
s is under C.
This argument is similar to den Besten's (1977,1983) argument in favour of the
idea that the rule "Subject-Clitic V' Inversion"28 moves V' (in den Besten's notation
V' designates the inflected verb accompanied by negative and pronominal clitics) to
Comp (see also Kayne (1984); Rizzi and Roberts (1989)). The data in (28)a-c,
which den Besten borrows from Dubuisson and Goldsmith (1976), shows that
Subject-Clitic V' Inversion correlates with the deletion of the complementizer (this
argument is relevant for those French dialects that allow doubly-filled Comps):
(28)
28
For Emonds (1976) this rule consists in the rightward movement of the clitic.
97
a. Vive la France.
live France
b. Ainsi soit-il.
thus be it
The kind of verb movement that we assume for (27) and (30) cannot apply in
embedded clauses because V-headed CP constituents count as V projections, and in
general embedded clauses must count as (nominal) arguments of the main verb;
lexical complementizers would satisfy this requirement (see Kayne (1982)). 32 The
ungrammaticality of (32) as opposed to (31) is directly accounted for under these
hypotheses:
29A
slightly different account of the obligatory deletion of s in (27)c-d can be found in Rivero
(1988a, to appear), who assumes that s occupies an Infi position, in between the C position and
an AGR/Tense position. The absence of s in (27)c-d would be due to V movement (V goes
through Infi on its way to Comp). The difference between Rivero's hypothesis and mine is
partially terminological (I label Comp the position that she labels Infi), and partially substantial:
within her analysis V movement does not end up in Infl/Comp, while within my analysis it does.
3
Note that in root subjunctive clauses ca is ungrammatical.
31
This does not mean that the same type of V movement is assumed for (27) and (30). The
parallelism solely concerns the correlation between V movement and the absence of the (otherwise
obligatory) complementizer in subjunctives. Note also that in French Que vive la France (where
que and subject-verb inversion coexist) is grammatical.
32
I n Chapter 1 it has been proposed that V movement to an embedded C does apply in
Romanian auxiliary constructions. This is so because auxiliaries subcategorize for a V projection.
CPN
XP
CPN
IP N
NP
I'
I
Infi
VP
'
NP
NP S
V-Infl
ts
V'
NP
tv
It is obvious that the Romanian subjunctive marker s does not fit into this
structure, following either of the two possible hypotheses concerning its categorial
status. Let us assume first that s is generated under Infi; correspondingly ca would
be under C (see example (20)a). According to this hypothesis two important
peculiarities should be noticed, which distinguish Romanian subjunctive clauses
from their Romance counterparts: (a) the (Spec, IP) position is not necessarily
occupied by NP subjects, but by any kind of constituent: verbal complements, PPs,
adverbs, etc. (see examples (21), which were discussed before from another point
of view); in other words the (Spec, IP) position is an A'-position; (b) two distinct
33several dislocated XPs are allowed, which recalls Cinque's (1990) Clitic Left Dislocation
(compare the English-type Left Dislocation, which necessarily involves only one constituent):
(i)
A vrea ca pin miine Ion s termine cartea asta.
[I] would like ca until tomorrow John s finish book this.
Infi positions should be assumed: one for Tense (which also serves as a landing
site for V-to-I movement), and the other one for s. We may alternatively assume
that s is generated under C. This hypothesis also entails some unusual
assumptions: (a) Comp takes an IP complement which does not present a (Spec,
IP) position (this must be assumed, because no XP constituent can intervene
between s and the elements that cluster around V-Infl (the negative adverb nu,
pronominal and adverbial clitics, the auxiliary fi); (b) dislocated constituents,
including the subject, occupy the (Spec.C) position; (c) we need two distinct C
positions, one for s and the other one for ca.
The conclusion we can draw so far is that both hypotheses strongly suggest that
the (Spec,IP) position is not an -position in Romanian subjunctives. 34 This
amounts to saying that the representation of subjunctive IP constituents shown in
(33) is inadequate for Romanian. We may instead adopt (34), which is
independently supported by the characteristics of Romanian auxiliary structures and
by the structure of infinitival clauses (see Chapters 1-2, and Section 3.1):
(34)
IP
[jpV-Infl [ VP NP S [ v t v ]]]
A
The difference between (34) and (33) concerns the -position of the NP subject:
Spec of IP in (33), Spec of VP in (34)). V raising to Infi bypasses the NP subject,
which accounts directly for the postverbal position of the subject in s subjunctives;
the (Spec,IP) position, which would be an A'-position, has been left out in (33).
The distribution of s can be described correctly by assuming that s governs the
IP constituent shown in (35), where (Spec, IP) is necessarily empty. This is a
minimal characterization of s, which is independent of its exact (Comp or Infi)
status; in (35), the label XP is meant to point out the systematic ambiguity of s:
(35) XP
X
IP
V-I
NP
V'
I
NP
I
tv
34()ne might argue that the subject does occupy the (Spec, IP) position; the necessary adjacency
that holds between s and the verb cluster would be obtained by the movement of the verb
(accompanied by the other functional elements) towards a position adjacent to s. This hypothesis
cannot be accepted, because in a structure such as (33) there is no landing position (in between s
and (Spec, IP)) for the verb; see also 3.1.1.2 for a remark concerning a potential analysis in terms
of incorporation.
IP (T/AGR/VP)
NP
VP
V-T-AGR t v ...
The configuration given in (36) underlies the French-type of clause structure; the
complex X category V-T(ense)-AGR is the result of successive movement of V
towards Tense and AGR; the verb has raised to a position which is
morphologically subcategorized by Tense. At S-structure the distinction between
V, Tense and AGR is no longer relevant, and the whole sequence of elements bears
the categorial label Tense; the resulting sequence does not bear the label AGR,
35
A S already noted in Chapter 1, we take AGR to be a nominal element (comparable to
pronominal clitics) which does not project a maximal category, as opposed to Tense and Neg, but
for the sake of discussion this hypothesis (proposed by Pollock (1989)) is provisionally accepted.
36
W e shall leave aside a possible alternative, according to which the verb and the verbal
inflections are separate constituents at S-structure; they would only merge at some Morphological
Level.
37
These constituents are syntactic words insofar as they may feed certain syntactic rules such as
V-second (compare with words resulting from phonological cliticization (I'll leave, le livre du
gargon, which do not feed any syntactic rule), but they are postsyntactic
because they are
themselves the result of certain syntactic rules, such as V raising to Infi. Note however that V to
Infi is triggered by morphological subcategorization (see Lieber (1980)), and therefore this rule
probably applies before purely syntactic rules.
38
T h e notions of "merged", "matching" or "extended" projections are not central to our theoretical
framework, but they have been recently proposed in a number of studies (see Platzack (1983);
Vergnaud (1987); Haider (1988) and Grimshaw (1991)).
101
because I assume that AGR does not govern TenseP, but rather adjoins to Tense
(see Chapter 1). We might then preserve the label Tense in order to refer to all the
elements which at S-structure end up dominated by Tense.
The incorporation process just described characterizes affixes, which are
morphologically subcategorized for a V slot to which the verb raises. In Chapters 1
and 2 evidence has been provided in favour of another type of incorporation,
Restructuring Incorporation. As with affix incorporation, distinct X categories are
generated in distinct syntactic positions, but they end up incorporated to each other,
by virtue of the following mechanisms ((37)a corresponds to (38) of chapter 1, and
(37)b to (12) of Chapter 2):
(37) a. Functional Coindexation: Coindex adjacent functional X categories.
b. Restructuring Incorporation: Coindexed adjacent (functional) X
categories merge into one X category.
The main motivation of Restructuring Incorporation has been dealt with in Section
2. 1. 3.1 sketch here the basic idea. In examples such as (38)b the string Neg-clV+9 has been moved by Move I:
(38) a. tu {ne le vois} pas
b. (nelevois)-tutnele vois pas?
The fact that the bracketed sequence in (38) can be moved indicates that it forms a
constituent, and I assume Restructuring Incorporation to be the relevant mechanism
behind this process: by Restructuring Incorporation several distinct heads end up
forming one X constituent, which can then be moved as a whole.
Note that Restructuring Incorporation is not an obligatory rule. The well-known
type of examples in (39) show : (a) that clitics can be left behind by Move I-to-C
(see (39)a); this means that Restructuring Incorporation has not applied prior to
Move I; (b) that Neg blocks this kind of movement (see (39)b-d):
(39)
a.
Ecris-le.
b.
Nel'ecnspas.
c.
*n'ecrisj-leejpas
d . *6cnsj ne le ej pas
39xhe sign "-" indicates a relation between bound morphemes; "+" indicates Restructuring
Incorporation.
(39')
CP
C
NegP
Neg
IP
CI
IP
V-I
The structure in (391) is quite general across languages (subject positions are left
out, because they are not relevant here). The relative position of Neg, Comp and
pronominal clitics is probably motivated by the relation between Neg and Tense
(see Zanuttini (1989) and Laka (1989)). This structure corresponds to (39)b in a
straightforward manner (C stays empty).
TTie example in (39)a can be analyzed as being produced by the movement of the
inflected verb to Comp (see den Besten (1983) and Chapter 2 above), past the
pronominal clitic. The example in (39)c is ruled out because the verb has moved to
a nonexistent position. By the structure preservation constraint, syntactic positions
cannot be created in the course of the derivation; a head element such as the verb
can only be moved to a head position provided by the base rules, and there is no
such position between the negative head ne and the clitic-verb sequence. (39)d is
especially interesting: it can be ruled out by the Head movement constraint if we
make the crucial assumption that ne is a head which blocks the movement of the
verb to the higher Comp position. We thus arrive at the conclusion that ne functions
as a head element distinct from (V-)Infl.
Summing up, examples (39) show that Restructuring Incorporation does not
necessarily apply. Examples like (38)a would then be structurally ambiguous,
allowing for two distinct representations, with and without Restructuring
Incorporation. What is then the categorial label of the highest projection, i.e. which
element counts as its head, for the representation without Restructuring
Incorporation? It is natural to assume that even in the absence of Restructuring
Incorporation a strong coherence exists between the various functional categories
clustering around the verb. One may invoke a general requirement on functional
categories: they must necessarily "connect" with lexical categories. The procedure
of Functional Coindexation stated in (37)a is one possible type of connection.
Note that Functional Coindexation is a necessary, but not sufficient condition for
Restructuring Incorporation. It can then be assumed that Functional Coindexation
applies before the Move I-to-C rule that underlies examples (39)a. The contrast
between (39)a and (39)d indicates that functionally coindexed elements, e.g., ne,
the clitic le and the inflected verb, do not become non-distinct from each other. In
particular, ne is a potential antecedent governor of the trace of a raised V+I. This
explains the ungrammaticality of (39)d, where ne blocks the movement of the verb
to Comp, by the Head to Head movement constraint. The blocking effect of the
intervening clitic in (39)a is circumvented via Merging (see Chapter 2).
It is thus clear that the elements, or the sub-strings, composing a functionally
coindexed sequence (or an "extended projection") are not necessarily subject to
restructuring incorporation, but may remain distinct from each other. Note that the
adjacency condition that governs both Functional Coindexation and Restructuring
Incorporation forces the assumption that there are not several distinct positions such
as (Spec, AGR), (Spec, Tense) or (Spec, Neg) available for the external argument
(compare Pollock (1989), for whom such positions are crucial). It is assumed that
only the highest functional projection has a Spec position.
Let us now come back to s subjunctives. The crucial empirical generalization
outlined previously was that s is adjacent to the sequence of elements clustering
around the inflected verb. The representation in (35), repeated here, is neutral with
respect to the exact status one may want to assign to s and its projection:
(35) XP
X
IP
I
V-I
VP
NP
V'
NP
tv
Farkas (1988) proposes a different analysis, which is close to the one adopted here: s would
lead a "double life", both as a Comp (in the syntax) and as Infi (in the morphology). This double
analysis is stipulated by two distinct subcategorization rules, within Sadock's (1985, 1991)
autolexical framework. In what follows we show that the double status of sneed not be
stipulated, but is the result of functional incorporation.
independently of the exact status that we may assume for s, its projection has
either the status of a matching CP/IP constituent, or that of an extended projection
of Infi, or that of IP.
Why then does (Spec,IP) not block Functional coindexation/ Restructuring
incorporation? (This question arises only on the hypothesis that s is under Comp.)
More precisely, what prevents the (Spec,IP) position from dominating lexical
material, in which case the adjacency condition would not obtain, and Functional
coindexation/incorporation would not apply. To answer these questions we shall
assume that A'-positions do not count for our adjacency condition: thus, by the base
rules s takes an IP complement which may present a (Spec,IP) position; but since
in Romanian (Spec,IP) is an A'-position, it does not prevent s from incorporating
onto (Neg-Cl-)V-Infl. 41 As a result of incorporation intervening A'-positions
delete, and therefore no lexical insertion can apply to them.42
3. 2. 3. On certain differences between subjunctives and indicatives
We will now turn to the reason why Romanian does not present an element
comparable to s in indicative clauses. If we assume s to be under Infi, no answer
is found to the question: why do subjunctive clauses present two Infi nodes,
whereas indicatives present only one? Alternatively, if we assume s to stand under
Comp, we may compare it to the complementizer characteristic of indicatives:
(41)
These examples show that c may appear either in front of the verb cluster (see
(41)a,b) or in front of a left dislocated constituent, which may be the subject or
some other constituent. Given our hypothesis concerning the clause structure of
Romanian, an example such as (41)c must be analyzed as involving the left
dislocation of the subject, on a par with the other dislocated constituents in (41)d-e.
This distribution indicates that unlike s, which is ambiguous between Comp
and Infi, c is distinct from Infi: a (Spec, IP) position may intervene between c and
the verb cluster. This difference between subjunctives and indicatives is not
4
' T h e technical details of this process are not relevant for our present purpose; one may assume a
successive application of Restructuring incorporation: pronominal clitics merge with V-Infl, then
Neg merges with Cl-V-Infl; finally, s i merges with Neg-Cl-V-Infl.
4 2 A n alternative technical implementation would be to assume that A'-positions are generated late
in the derivation, after the incorporation has applied; no A'-position can be inserted inside an
incorporated string.
accidental; under a slightly different form, we find it in classic Arabic, for instance:
the subjunctive Comp element necessarily precedes the verb, whereas the indicative
Comp element necessarily precedes an NP (see Ayoub (1982)).
This difference between subjunctives and indicatives may be understood by
taking into account other well-known differences, which concern the Tense features
characteristic of each of these two moods. It is currently assumed that Comp hosts
abstract Tense features (see den Besten (1983); Stowell (1981)). The question is
then, what is the difference, if any, between the Tense features under Comp and
those under Infi? It is reasonable to assume that the Tense features in Comp
represent the speech moment, while the Tense features that appear under Infi
specify the temporal coordinates of the event that is referred to in EP ( (1987)).
Let us notate these elements by T and Tense respectively, and examine the relation
between them. The subjunctive mood expresses an event which is dependent on the
main clause. Correlatively, the temporal reference of subjunctive clauses is
dependent on the Tense of the main clause: in order to assign temporal reference to
Romanian subjunctives one must take into account the Tense in the main clause.
Indicatives on the other hand may be assigned temporal reference without any
recourse to the main clause. 43 These two types of Tenses are currently referred to
as "anaphoric" and "deictic" (see Partee (1984)), and certain authors have tried to
treat them in terms of the [+anaphoric] and [-anaphoric] distinction (see Bennis and
Hoekstra (1988) and Fassi-Fehri (forthcoming), who use this distinction in quite
different ways). The present discussion is descriptive, and does not presuppose any
complete theory, [-an] Tenses are interpreted by reference to the speech moment,
which amounts to saying that [-an] Tenses are necessarily distinct from the T
features in Comp. This necessary distinction between [-an] Tenses and T can be
captured by assigning distinct Tense indices to Comp and Infi. It is natural to
assume that the existence of two distinct Tense indices blocks Functional
Coindexation, and this in turn blocks Restructuring Incorporation; hence the clear
Comp status of c in Romanian indicatives.
Consider now the referential properties of subjunctive Tenses: given our
distinction between Tense and T, the label [+an] is unsuitable, since it suggests
that in subjunctive clauses the Tense features generated under Infi bear the same
index as the T features under Comp. But this is not what we want to capture: the
peculiar property of subjunctive Tense is that its reference is not constructed with
reference to the speech moment, but rather with reference to the Tense features of
the Infi element of the main clause. This means that the Comp of subjunctive
clauses does not host T features: T would break the relation required by the
subjunctive, between the embedded and the main Tense features. We may then
assume that at LF subjunctive Tense raises from Infi to Comp (since Comp does
not dominate T features, it is an accessible landing site for the Tense features
generated under Infi), from where it can enter the required relation with the Tense
3This difference between subjunctives and indicatives is quite general across languages. Note
however that the contrast may be more or less marked: French and English for instance present the
phenomenon called "sequence of tenses", by which the tense of an embedded indicative clause is
dependent on the Tense of the main clause (note however that this dependence is formal, rather
than referential: a certain Tense in the main clause requires a certain Tense in the embedded
clause). In Romanian the sequence of Tenses is much less constrained. French subjunctives on the
other hand do present a degree of independence with respect to the Tense of the main clause.
44
Note that this relation itself is not strictly speaking "anaphoric", because the subjunctive Tense
features do not bear the same index as the Tense of the main clause, but only depend on them.
the doubly filled Comp filter. But this conclusion cannot be reconciled with the
evidence provided above in favour of the Comp status of s. This problem can be
solved by providing an account for the ungrammaticality of (43) which does not
rely on the doubly filled Comp.45
The basis of the analysis that follows is that a infinitives occur in typical NP
positions, to the exclusion of VP positions; in other words, a infinitives can only
function as arguments, but not as predicates. The ungrammaticality of (43) would
be due to the fact that wh-elements necessarily enter predicative constructions, but a
infinitives cannot function as predicates. S subjunctives on the other hand are
allowed to function as predicates,46 hence the possibility of their co-occurring with
wh-elements in Comp.
This hypothesis captures other contrasts between a infinitives and s
subjunctives (note however that (44)b used to be grammatical in older stages of
Romanian):
(44)
a. Pot cinta.
[I] can sing
b. *Pot a cinta
(45)
Pot s eint.
45
This is needed independently, given the possibility of "long" wh-movement with a infinitives.
In (i) we must assume that the intermediate wh-trace coexists with a:
(i)
Ce nu gtii dae te-au rugat a scrie?
what don't [you] know whether [they] asked you a write?
S subjunctives also function as nominal projections, as indicated by the fact that they can be
freely used instead of a infinitives, see in particular all the examples in 3.1., which stay
grammatical if we substitute the subjunctive for an a infinitive.
this case no other element is needed (in other words the Greek counterparts of (47)
are grammatical). (49) should then be viewed as a language-particular requirement:
(49) Lexical elements must be governed.
The principle in (49) is natural: it requires that any linguistic element should be
determined with respect to the other elements of a given structure; an ungoverned
lexical element is hors structure, and as such illicit. The condition in (49) may be
understood as a visibility condition comparable to the Case filter, we shall call it the
government filter. These two filters are independent of each other. There are indeed
contexts which pass the government filter, but not the Case-filter. In languages such
as French, clitic doubling is excluded by virtue of the Case filter (the clitic absorbs
Objective Case, which is no longer available for the lexical NP in the direct object
position); no violation of the government filter can be invoked, if we assume with
Borer (1981, 1984) that clitic doubled NPs occupy their canonic A-position.
The government filter is relevant for the converse situation, illustrated in (47)a-b,
examples which do not show any violation of the Case filter, their ungrammaticality
may be due to the fact that when they occupy the preverbal position miine and pe
Ion are not governed (the definition of government needed here is that based on the
strict definition of c-command, stated in terms of branching nodes, see Reinhart
(1976)).
Here are (47)a-b again with a rough indication of their underlying structure:
(47) a. * vreau [ e [ miinej [xps# vin Ion tj]]]
[I] want
tomorrow
s i come John
b. *doresc [ e [pe Ionj [^ps-lj examineze tj Popescu]]]
[I] wish
pe John s-him examine Popescu
According to the strict definition of C-command, miine and pe Ion C-command all
the constituents of IP, but are not themselves C-commanded by any of them, thus
violating the filter given in (49).47
Consider now the examples in (48). They are correct, because the constituents
miine and pe Ion are governed inside the subjunctive clause, by Infi and V,
respectively.
The definition of government adopted here leads to the following generalization:
the elements that occupy left dislocated A'-positions adjoined to a given XP
projection cannot be governed by elements belonging to XP. They may instead be
governed by elements outside XP. This is currently assumed for certain elements
such as heads and Specifiers; we simply extend this assumption to dislocated
^^Note however that the following examples are well-formed, despite the fact that miine and pe
Ion are not governed:
(i)
[jpMtine [j.va veni Ion]]
tomorrow will come John
(ii)
[jp pe Ion [j.il va examina Popescu]]
pe John him will examine Popescu
Our government filter does not apply in main clauses, but this is not really a problem, given
well-known independent differences between embedded and main clauses.
Examples (52), which differ minimally from (51) in the presence of ca, are
incorrect. 49 This ungrammaticality may be taken to indicate that ca necessarily
subcategorizes for a CP that presents a (Spec,CP).50
48
This possibility depends on lexical specifications : certain main verbs are able to govern the
subject of a subjunctive clause, in a way similar to English cases of Exceptional Case Marking,
cf. I made John cry (on this see Rivero (1988b)).
49
In colloquial Romanian, examples (52) are acceptable (which is why two question marks have
been used instead of *; but since we are speaking of two distinct socio-dialects, we could
maintain * in (52)), which indicates that in this dialect ca s functions as a variant of s.
^According to the hypothesis that s subjunctives are IP projections, the ungrammaticality of
(52) could be captured by an obligatory string-vacuous rule that would move s to Comp (cf.
Farkas 1982; Dobrovie-Sorin 1987), but such a rule has no clear status.
(2)
(3)
These examples illustrate three distinct semantic relations between the embedded
and main clauses, which depend on the semantic properties of the main verb, i.e.
pertaining to argument structure. The data in (l)a-b are instances of "obligatory
control": the subject of the embedded clause is necessarily coreferential with an
element (either the subject or an object) in the main clause. Following Williams
(1980); Mohanan (1985); and Kster (1984,1987), I believe that obligatory control
should be kept distinct from "optional" and "arbitrary" control (see (l)c and (i)-(iv)
1
Another fundamental question will be left open: what led the infinitive to disappear gradually?
(2') illustrates the "obviation" effect: certain verbs (in particular volition verbs)
govern a subjunctive clause if the embedded subject is disjoint in reference from the
main subject; but when the embedded subject is coreferent with the main subject,
the infinitive rather than the subjunctive is obligatory. Example (2) shows that
Romanian subjunctives are not subject to the obviation effect.
The examples in (3) are subject raising constructions, which differ from (1) by
the fact that the main verb cannot be said to select a lexical subject, but only a
dummy subject, as it does in (4):
(4)
The examples in (3) are derived from (4), via the raising of the embedded subject
towards the subject position of the main clause. The examples in (1) and (3) are
currently assumed to be characterized by distinct syntactic configurations: in
particular the embedded subjects are assigned distinct syntactic types (in technical
terms, PRO and NP-trace for (1) and (3) respectively).3 Different as they may be,
however, these two types of empty subjects have in common the fact that they can
show up only if the verb is either in the infinitive (in Romance languages including
Romanian) 4 or in the subjunctive (in Romanian). These facts raise a double
question: what is the common property of control and raising configurations and
why is this property compatible with the subjunctive in Romanian, but not in the
other Romance languages? Note that the examples in (l)-(3) involve s
subjunctives. Our analysis will have to explain why ca subjunctives (see 3.2.6 in
Chapter 3) are illicit in the constructions illustrated in (1) and (3); as to the
examples in (2), they allow ca infinitives, but in this case the interpretation is quite
different.
We shall answer the above questions by examining in turn the three supposedly
distinct constructions illustrated in (l)-(3). For each of them, our starting point will
be the standard analysis that is assumed for the infinitival constructions that
illustrate them, for instance in French and English. We shall then try to see in what
2
Compare the opposite properties of optional control illustrated in examples (i)-(iv), borrowed
from Kster (1987: 111):
(i)
It is impossible [e to help Bill]
(ii)
John proposed to Mary [e to go to the movies]
(iii)
It is difficult for Mary [e to help Bill]
(iv) John thinks [jpif is impossible [jpc to shave himself]]
3As to (2), several analyses have been proposed (cf. Section 4.5. below).
4
In Romanian the infinitive may show up instead of the subjunctive very marginally in (1) and
somewhat more easily in (3); in (2) the subjunctive alone is possible in modern Romanian.
way we must modify this analysis in order to extend it to Romanian, where the
subjunctive can be used (instead of the infinitive or in free variation with it).
Our main conclusions support the view that obligatory control is governed by
principles belonging to binding theory (see Koster (1978a, 1984, 1987); Bouchard
(1984); Manzini (1983); Borer (1987)). We shall also provide evidence in favour
of the idea that obligatory control and subject raising rely on the same
configurational relation (see Kster (1978a, 1984, 1987)).
accessible to lexical subjects. This problem also concerns Romanian infinitival null
subjects, which alternate with lexical subjects. Thus, given the possibility illustrated
in (5)a-b, it is difficult to assume that the null subject in (6)a-b is PRO:
(5)
(6)
We must then give up the idea that in Romanian, control effects are triggered by
PRO.
4. 3. On the contextual identification of anaphors
A different theory of control has been proposed by Koster (1978a, 1984, 1987);
Bouchard (1984); Manzini (1983); Borer (1987). Despite a number of diverging
ideas, these proposals have in common the attempt to reduce control theory to
binding theory, by assuming that PRO has the status of an anaphor.
Manzini is however apart in as far as she assumes that PRO is not a standard
anaphor, but an "anaphor without a governing category" which is subject to a
revised principle A: an anaphor without a governing category must be bound in the
governing category of its domain. Manzini's theory may well tum out to be correct
for optional control. As to obligatory control, I will assume that the other authors
above mentioned are correct in trying to reduce it to standard binding; in other
words PRO is a standard anaphor, and as such is subject to principle A of binding
theory:
(7)
7A number of alternative formulations of these definitions may be found in the literature, which
however do not concern us here.
(AGR, e) chain. Restated in this way, Borer's idea is vital for the Romanian data, in
which the controlled subject is identified by AGR features.
The present analysis however differs from Borer (or at least from a possible
interpretation of Borer's proposal), and approaches that of Bouchard and Kster,
insofar as it assumes that a controlled element (be it an empty category, AGR, or
(AGR, e)) is not intrinsically marked as an anaphor 8 but has the status of a
contextual anaphor.
The distinction between contextual and intrinsic anaphors is particularly clear if
we consider their relation to principle A. This principle has the status of a wellformedness condition on intrinsic anaphors; contextual anaphors on the other hand
are not constrained by this principle but rather defined by it:
(8)
a.
b.
c.
d.
In (9)a, but not in (9)b, me (this is true of all first and second person object clitics)
is bound in its governing category; this means that me is an anaphor in (9)a and a
pronoun in (9)b. To assume distinct intrinsic ("anaphoric" versus "pronominal")
features for me in (9)a and (9)b seems entirely stipulative; it seems more
satisfactory to say that the anaphoric versus pronominal status of me is contextually
determined: in (9)a principle A is obeyed, and this assigns an anaphoric status to
me, whereas in (9)b principle A is violated, and this assigns a pronominal status to
me. Compare the case of se, which is inherently marked as an anaphor, and as such
must obey principle A ((9)d is ruled out because it violates principle A).
To summarize our discussion, intrinsic anaphors are also necessarily
"contextual" anaphors (in as far as they are bound to participate in an anaphoric
relation), but conversely contextual anaphors are elements that participate in an
anaphoric relation, but are not intrinsically marked with the [+an] feature, and as
such are not necessarily subject to principle A of binding theory (and hence may
function as pronominal elements).
Let us now come back to the subjects of subjunctives in Romanian. Clearly, they
differ from se, and behave much like me/te:
(10)
8Borer explicitly mentions anaphoric inherent features for AGR, but her analysis may be
interpreted in terms of contextual identification.
I would like that Johni leave tomorrow, but Maryj wants very much that
(proj/j) leave the day after tomorrow
c. Eu ap vrea ca Ionj s piece miine, dar Manaj m-a convins c trebuie proj/j
s piece poimiine.
[I] would like that Johni leave tomorrow, but Maryj convinced me that
must proj/j leave after tomorrow
In (10)a-c the subject of the subjunctives does not function as an anaphor, because
it is not subject to principle A of binding theory; it behaves instead as a pronoun.
Compare the examples in (1), in which the subject of the subjunctives is allowed to
participate in a binding relation, which is imposed by obligatory control verbs. In
other words, Romanian controlled subjects are not intrinsic anaphors, but only
"contextual" anaphors: the inherent features of AGR are neutral with respect to the
[+an] and [-an] specification; AGR is compatible with both of them, the choice
depending on the context.
This distinction between intrinsic and contextual anaphors allows us to keep the
subject of Romanian subjunctives distinct from infinitival subjects, while
accounting for their common properties. This is a departure from Borer, who
assimilates PRO to "anaphoric AGR"; I take anaphoric AGR, in particular the
subject of Romanian subjunctives, to be a contextual anaphor, whereas PRO - or
rather the infinitival subject, since given the above analysis there is no PRO in our
grammar - is necessarily an anaphor. This does not imply the assumption of an
intrinsic [+an] feature that would characterize infinitival subjects: infinitival subjects
are empty categories that are identified by no features (compare pro, which is
identified by AGR), and therefore they must necessarily enter a relation with an
antecedent that provides them with features, which are necessary for the assignment
of reference. The relations that infinitival subjects have with their antecedents are of
different types: we assume obligatory control to be governed by binding theory;
optional (and arbitrary) control is probably captured by Manzini's theory, but we
shall leave this case aside.
To sum up, the theory of control that will be adopted essentially follows the
hypothesis according to which obligatory control reduces to binding theory. The
object of this theory is not a particular type of linguistic element (be it PRO, an
anaphoric empty category, or anaphoric AGR), but rather a particular type of
linguistic relation, namely anaphoric binding. For convenience, we shall speak of
the "controlled" or "anaphoric" element, but the reader should keep in mind that this
label does not designate intrinsic features.
The theory of control sketched above concerns only the syntactic aspect of the
phenomenon, but it also implicitly supposes a lexical aspect. It will be assumed
here that obligatory-control effects are imposed by the lexical properties of control
verbs (start, try, etc.). Principles pertaining to the interface between semantics and
syntax must explain why obligatory-control effects, i.e., the obligatory coreference
of the embedded subject with an element of the main clause, are necessarily
expressed by means of an anaphoric relation. At this point the syntactic aspect of
control comes in: anaphoric relations are the object of a syntactic module, binding
theory.
Coming back to the Romanian data presented in this section, it can be
summarized by the following generalizations: i) in Romanian no kind of subject
(neither in subjunctives nor in infinitives) is necessarily assigned anaphoric status.
This contrasts with French and English infinitival subjects, which are necessarily
anaphoric (i.e., either PROs or NP-traces; we leave aside Exceptional Case
Marking); ii) Romanian subjunctives (and infinitives, see 3.1.) may be controlled,
but need not be; (iii) Romanian indicatives cannot be controlled.
These generalizations indicate that for Romanian the "control problem" cannot
be the search for those properties that impose control. This is what is generally
done for English or French, where infinitives are necessarily controlled; the crucial
property is to find out that property of infinitives which accounts for their being
necessarily controlled. There is however a "control problem" in Romanian, namely
the understanding of the syntactic properties that allow control. To sum up, the
control question concerns the necessity of control in English and French, but only
the possibility of control in Romanian. For the latter language it is only possible to
state the syntactic conditions that allow a control configuration, but it is impossible
to state the conditions under which control is obligatory. These syntactic conditions
must be distinguished from lexical requirements: with both Romanian subjunctives
and French/English infinitives control properties are imposed by the lexical
properties of the main verb. Subjunctive clauses are either obligatorily or optionally
controlled, or not controlled at all, depending on the lexical specifications of the
main verb.
Our present task is then to answer the following two questions: a) why do the
subjects of subjunctives qualify as anaphors in Romanian, but not in English or
French; b) why is the definition of anaphors not satisfied by the subjects of
Romanian indicatives?
Before answering these questions, we will show that they are also posed by the
subject raising phenomenon and by the lack of obviation illustrated respectively in
(3) and (2).
4. 4. Subject raising
Let us now consider subject raising:
(11)
It is currently assumed that NP-traces, and in particular the traces of raised subjects,
have the status of anaphors, i.e., they must be bound in their governing category.
This explains why subject raising applies to infinitival subjects, but not to subjects
of tensed clauses: a tensed clause is a governing category for its subject, so that an
element which occupies the subject position cannot be bound by the subject of the
main clause. Infinitival clauses on the other hand do not count as governing
categories for their subject, and therefore an anaphoric relation between the
embedded and the main subject positions is legitimate. NP-traces are also subject to
the ECP; in other words they must be properly governed, and it is currently
assumed that this condition is met in those infinitives that do not project CP, but
only IP (an alternative equivalent formulation is to say that obligatory CP deletion
Control configurations are currently assumed (see Chomsky (1981)) to differ from
(12), in that CP is projected:
(13)
A structure such as (13) would ensure that e is not governed by the main verb, the
consequence being that e cannot be an NP-trace, but PRO. Kster (1978a, 1984,
1987) has argued convincingly that this hypothesis may be correct for optional
control, but not for obligatory control (and in particular not for try, start, etc.).
According to Kster, subject raising and obligatory control rely on the same
configurational structure, characterized by the absence of Comp. This hypothesis is
empirically supported9 by evidence concerning the distribution of infinitival
complements in Dutch.10
The Romanian data examined here provides further evidence in favour of the
same analysis. Before we turn to a precise discussion, it should be noted that
Koster's terminology may be confusing: the obligatorily controlled empty subject is
referred to as "governed PRO" or "anaphoric PRO", or "obligatory PRO". Given
the current theory of PRO these labels are contradictory. As far as I can see Kster
has preserved PRO in order to capture the control properties which distinguish
examples such as (13) from subject raising (see (12)). However, this distinction
is independently captured by th-theory; the underlying th-chains corresponding to
(12) and (13) are different:
(14) a. (NP, e)
b. (NP) (e)
The empty categories in (14)a-b both qualify as anaphors. That in (14)a is also
an NP-trace; this label is simply a descriptive statement of the fact that the empty
category and its NP antecedent belong to the same th-chain. To put it otherwise, the
label "NP-trace" captures information concerning the underlying th-chain, and so
does PRO (this element does not belong to the same th-chain as its antecedent). But
given the modular form of our present theory it is not necessary to establish classes
of linguistic elements on the basis of the type of chain in which they may
^There are of course a number of potential objections. It is a well-known fact (see Kayne (1984))
that in French the complementizer de is allowed to precede an obligatorily controlled subject, but
not the trace of a raised subject. If, following Kster, we want to give up the difference between
PRO and NP-trace, the differential distribution of de might perhaps be captured by taking into
account the difference in chain configuration: NP-traces, but not PRO, belong to the same chain
as their antecedent. For a discussion of other problems I refer the reader to Kster (1987: 1 19).
^Subject raising, obligatory control and exceptional Case marking behave alike (the embedded
infinitives show up on the left of the matrix verb, no Comp element is allowed, nor is the
infinitival particle te), and differ from optional control (the embedded infinitives are extraposed,
i.e., they appear on the right of the matrix verb).
participate. Indeed, NP-traces and reflexives are classed together as anaphors, i.e.,
as elements that are constrained by principle A, and no sub-classes are established
to capture the fact that NP-traces, but not reflexives, belong to the same th-chain as
their antecedent. Given this theory, I see no reason why two distinct types of empty
categories should be postulated for NP-traces (in particular those resulting from
raising) and obligatorily controlled subjects; they both qualify as anaphors, but
differ with respect to th-theory.
To summarize, Koster's approach correlates two hypotheses: (a) an analysis of
control in terms of anaphoric binding and (b) the hypothesis that control and
subject-raising rely on the same underlying configuration. Such a correlation seems
necessary, since NP-raising is standardly assumed to rely on anaphoric binding.
But see Borer (1987), who tries to avoid this consequence of her hypothesis, by
stipulating that raising verbs subcategorize an IP complement, whereas control
verbs subcategorize CP characterized by Move (V-)I to Comp. Such a stipulated
difference between raising and control is not supported by empirical evidence in
Romanian.
Let us now tum to the subject of Romanian subjunctives. The data illustrated in
(11) introduces a new problem into the picture we have just drawn: what counts as
an NP-trace is not a "simple" empty category (as in the standard case illustrated in
(12)), but a null subject that we have notated (AGR,e). The relation between the
raised NP and its trace is mediated by AGR, in that AGR is part of the chain which
links the NP-trace to its antecedent. Rather than an (NP, e) chain we have an (NP,
AGR, e) chain; the empty category appears to count both as a pronominal
(identified by AGR) and as an NP-trace.
This type of chain cannot be postulated under the standard assumption that nonhead members of chains are necessarily empty categories (see Chomsky (1981)).
The Romanian data indicates that this assumption must be modified by allowing
certain lexical elements, such as verbal inflections, as non-head members of chains.
This is not an isolated case: in Chapter 5 (see also Rizzi (1986b)) it will be assumed
that middle si constructions rely on chains of the type (NP, se, e) and in Chapter 6 a
type of wh-movement will be discussed, which relies on an A'-chain of the type
(wh,cl,e), where the wh-trace is "doubled" by a clitic. It thus appears that AGR and
clitics may qualify as legitimate intermediate chain-links; note that these elements
occupy A'-positions, and therefore the condition concerning chain-links may be
relaxed so as to allow for the presence of overt elements that occupy A'-positions
(but not A-positions).
4. 5. Obviation
Consider example (15)a and compare it to its French counterparts in (15)b-c:
(15) a. Ion vrea s piece devreme miine.
John wants that (he) leaveSubj early tomorrow
b. *Jeanj veut qu 'ilj parte tot demain matin.
c. Jeanj veut qu'iljparte tot demain matin.
(15)b illustrates the obviation effect: the subject of a subjunctive clause cannot corefer with the main subject.11 Two main types of explanation have been proposed.
We may assume that the governing category of subjunctives is the main clause (this
characteristic would be due to the particular type of Tense features that characterize
subjunctives); hence (15)b would violate principle B, because the embedded
subject, a pronominal, would be incorrectly bound by the matrix subject (for this
type of proposal see Jakubowicz (1985); Kempchinsky (1985); Picallo (1985)).
We may alternatively assume that subjunctive clauses define a governing category
for their subject. Under this hypothesis the binding conditions are not violated in
(15)b: the embedded subject is free in its governing category (the embedded
clause). The ungrammaticality of (15)b has then to be captured by additional
principles (Bouchard (1984); Suffer (1986); Everaert (1986)).
We will adopt the latter hypothesis, and propose that (16) is the principle
responsible for the obviation effect:
(16) Use an anaphor instead of a pronoun whenever possible.
(16) can be viewed as a particular case of the Avoid Pronoun Principle: 12 if in a
given configuration we have the choice between a pronominal subject and an
anaphoric subject, the second option will necessarily be selected. The
ungrammaticality of (15)b can then be accounted for in the following way: (15)b
does not rely on a legitimate anaphoric relation in French, and there exists such a
legitimate anaphoric configuration, that in which the embedded clause is an
infinitive.
Given (16), the lack of obviation effects that characterizes Romanian subjunctive
clauses reduces to the property discussed in Section 4.3. above, namely the fact that
they are accessible to anaphoric binding. In other words, if we are able to answer
the question stated at the end of Section 4.3., we solve not only the control
problem, but also the obviation problem raised by Romanian subjunctives.
Our analysis differs from the one proposed by Rivero (1987), who considers
that in those examples illustrating the lack of the obviation effect, the subject of
Romanian subjunctives is a pronoun whose governing category is the embedded
clause. This hypothesis raises an important difficulty: Romanian subjunctives also
allow for subject raising, and in this case Rivero is forced to assume that the
governing category of subjunctives is the main clause. Moreover, under Rivera's
analysis of the obviation effect the contrast between Romanian and the other
Romance languages is unclear (see also the discussion at the end of this chapter).
11rhis generalization should probably be relaxed: a number of interesting exceptions do exist (see
Ruwet (1984) and the discussion at the end of Section 4.7.).
12
Chomsky (1981) uses the Avoid Pronoun principle in order to account for the fact that in
English, Genitive pronouns may be dropped in certain NPs. Jaeggli (1982) uses the same
principle to capture the fact that empty subjects and object clitics are used instead of
phonologically realized pronominal subjects and non-clitic pronominal objects respectively.
Chomsky (1981: 142, fn. 45) also reports a suggestion by J. Gudron, who proposes to derive the
obviation effect from Avoid Pronoun.
125
splec
proj la
munte.
[I] would like that tomorrow s [I] leave for the mountains
This type of Romanian example is comparable to well-known marginal exceptions
concerning obviation effects studied by Ruwet (1984):
(23)
?/ei voudrais bien/j'aimerais bien que jej sois enfw autoris6 partir en Israel.
(A number of French native speakers insisted that a question mark [or two]
should precede Ruwet's examples.)
Given (16), the problem in (22) and (23) is that a pronoun is used instead of an
anaphor. This possibility seems to be related to the fact that the event referred to in
19 T his anaphoric relation is broken by ca; subject raising is excluded in ca subjunctives, as
pointed out by Grosu and Horvath (1987)):
(i)
* To(i elevii. s-au nimerit
pa exercifiul sta t. s-1 gregeasc].
(ii)
all the students happened [that this exercise s -it [they] fail]
*Bombele.pot
ca in orice moment t. s explodeze],
(iii)
your children seem [that the teacher s[they] are angry with]
This ungrammaticality is expected, for the reasons presented above (see the impossibility of
control in ca subjunctives).
the embedded clause is viewed as distinct from the event referred to in the main
clause: the "distance" between the two events is marked by the conditional mood,
by the passive, by negation, etc. We may suggest that anaphoric binding is not an
obligatory choice for this type of looser relation; coreference between two
pronouns belonging to two distinct governing categories is sufficient. Beyond this
similarity, the Romanian cases of the type in (22) differ from the French situation
by being completely grammatical, and extremely productive.
One may wonder whether the present account is not completely circular for
Romanian: we postulate that the [AGR,e] chain is pronominal in ca subjunctives
(see (23)) and anaphoric in s subjunctives (see (2)), but there does not seem to be
any empirical argument in favour of this analysis. This is indeed true if we take into
account only (22) and (2), i.e., if we consider the obviation problem in isolation.
But the obviation problem cannot be understood independently of control. And
obligatory control gives us an empirical argument in favour of the proposed
analysis of obviation: the cases of obligatory control show that an anaphoric
relation is precluded in ca subjunctives; therefore we cannot assume that (22) is a
case of binding.
20
Rivero discusses also a problem concerning the focalization of the subject of subjunctives (cf.
Picallo (1984)), which I have left aside because the grammaticality judgments are not clear to me.
between CP and IP. 21 Rivero admits that this coindexation is not sufficient to
account for the Romanian data, and therefore two further stipulations are
introduced: a) for certain dialects of Romanian functional categories such as Comp
do not have the status of governors for minimality; b) an indicative Comp projects a
category which counts as an inherent barrier.
According to the analysis proposed here the cases in which functional categories
do not count as minimal head governors are due to the mechanism of Restructuring
incorporation (or Functional Coindexation). The behaviour of s subjunctives is
due to the fact that s, unlike ca and c, is allowed to incorporate onto/be
functionally coindexed with V-Infl. This analysis directly accounts for the contrast
between s subjunctives on the one hand, and ca subjunctives and indicatives on
the other. The adjacency condition, which constrains both Functional Coindexation
and Restructuring incorporation, allows the Comp node not to count as a minimal
head governor for Infi in a language characterized by an IP-constituent in which VInfl occupies the initial position. Thus, the specific properties of Romanian
subjunctives may be derived from the specific structure of IP constituents in the
language.
An important argument in favour of the coindexation between Comp and Infi does exist,
namely the fact that the subjunctive Comp elements (ca and s) are phonetically distinct from c,
which heads indicative and conditional embedded clauses. We may however observe that the
relationship between Comp and Infi is not restricted to Romanian; even in languages that do not
present any overt manifestation of agreement between Comp and Infi, we must assume an abstract
relation between these two elements (cf. Safir and Pesetsky (1981); Safir (1981)), which could be
formalized in terms of coindexation.
de partir.
(3)
a. * Va procedat la anchet.
[it] will be proceeded to an inquiry
b. *Va R vorbii de dumneavoasir sptmina viitoare.
[it] will be spoken about you next week
c. *A fost ajuns/parvenit la un compromis acceptabil.
[it] was arrived at / reached an acceptable compromise
d. *A fosi tras in el.
[it] was shot at him
e. *I-a fost cerut/ordonai/sugerai de a pleca (spiece).
[it] himoat was asked/ ordered/ suggested to leave
f. *A fost apoi alergaipin la gar.
[it] was then run to the station
g. *A fost deja construit pe piloti (in acest loc).
[it] was already built on piles (in this place)
h. *A fosi apareiai la ora 4 dimineafa.
[it] was set sail at 4 h in the morning
Es wurde getanzt/gespielt.
it was danced/ played
The existence of passive intransitives seems problematic for the current hypothesis
(see Chomsky (1981)) stated in (5):
(5)
The passive phenomenon triggers (a) the non-assignment of the external throle and (b) the non-assignment of objective Case.2
Given the idea that intransitives are not assigners of objective Case, the nonassignment of objective Case required in (5) would apply vacuously in passive
intransitives, and this is not allowed (see Burzio (1981, 1986) and Marantz
(1984)). In a number of recent studies this problem is solved by assuming that (5)a
is the only general property of passives; the property stated in (5)b is
parametrizable, the various possible choices being responsible for the
crosslinguistic variation illustrated in (l)-(4):
(6)
The passive morpheme must receive (or trigger the non-assignment of)
abstract Case in certain languages (see English, Romance languages other
than French), but not in others (see Norwegian, Ukrainian, etc.)
Alternatively, it has been suggested that the properties of intransitives may vary:
'The behaviour of ergatives/unaccusatives will be discussed in subsequent sections, see in
particular 5.3.2.
2
(5) is supposed to characterize the common properties of various syntactic structures (see copula
passives, morphological passives, English middles, Romance middle/passive si, etc.); the precise
relation between a particular syntactic configuration and (5)a-b is not made explicit. In a number
of recent studies (see in particular Belletti (1982), Jaeggli (1986a); Marantz (1984)) it is assumed
that the properties stated in (5) derive from the fact that the "passive morpheme" absorbs the
relevant th-role and Case.
Both of these proposals 3 (see Marantz (1984) and Baker (1988) for (6), and
Jaeggli (1986a) for (7)), seem inadequate for a number of reasons. TTie first
objection is simply their ad-hoc character: given the standard definition of passives
in terms of Case, any counterexample is handled by a modification, or
parametrization, of the assumption itself, with no independent evidence in favour of
such a proposal. Moreover, (7) is also empirically inadequate. Thus, intransitives
do assign Case unproblematically whenever they take a cognate object, and this
possibility is extremely common crosslinguistically; in particular, those languages
that preclude passive intransitives freely allow cognate objects, see the Romanian
examples in (8)a-b:
(8)
As any overt NP, cognate objects are subject to the Case Filter. It thus appears that
objective Case can be assigned by an intransitive whenever it is needed, i.e.,
whenever a direct object appears. Since objective Case can be assigned in (8),
intransitives can be viewed as virtual assigners of objective Case,4 contrary to the
hypothesis concerning the parametrization of the Case-assigning properties of
intransitives. The assignment of objective Case thus appears to be irrelevant for the
problem at hand. In other words, the Case problem raised by passive intransitives
is only apparent; the real reason for the impossibility of (3) remains to be
discovered.
Another important observation concerns the passivization of transitives with null
prototypical objects:
(9)
These two alternative parameters have the same empirical consequences with respect to passive
intransitives, but not with respect to the (im)possibility of accusative NPs in impersonal passives.
4
I assume that even those intransitives that do not take cognate objects are virtual assigners of
objective Case.
The ungrammaticality of (13)c indicates that the null prototypical object of the verb
eat is not visible in the syntax. Compare cognate objects, which do behave as
visible direct objects with respect to both causatives and Benefactives.7 The data in
(12) and (13) very strongly suggest that null prototypical objects are not projected
in the syntax of French: the argument position corresponding to them is saturated in
the Lexicon.8
^Examples (ii) and (iv), which present cognate objects contrast with (i) and (iii):
(i)
*Marie fait vivre Jean.
Marie makes live Jean
(ii)
Marie fait vivre des moments p6nibles Jean.
Marie makes live difficult moments Jean
'Mary makes John live difficult moments.'
(iii)
*Je ne veux pas qu'il me vive comme un bohimc.
I do not want that he mejya[ lives as an artist.
(iv)
Je ne veux pas qu'il me vive une vie de bohdme.
I do not want that he mej) a i lives an artist's life.
similar proposal can be found in Rizzi (1986)a, who examines constructions of the type
illustrated in (i)-(iii):
(i)
Cela conduit pro [PRO conclure que...]
(ii)
* This leads pro to [PRO conclude that...]
(iii)
This leads
us to [PRO conclude that...]
According to Rizzi the fact that a control relation is allowed in (i), as opposed to (ii), indicates
that an arbitrary object pro (which appears in the same position as us in (iii)) can be projected in
the syntax in French (and in Italian) but not in English. In the latter language arbitrary objects of
this type would be saturated in the Lexicon. Somewhat paradoxically, Rizzi's proposal both
supports and contradicts my own claim. It supports it insofar as it is assumed that certain
arguments need not be projected in the syntax, but can be saturated in the Lexicon. The problem
is that according to Rizzi this process does not apply in French: in this language arbitrary object
pro is projected in the syntax, and this contradicts my claim concerning null prototypical objects.
One possible solution would be to say that arbitrary pro is not necessarily projected in the syntax
(on a par with prototypical objects); it would only be projected when the syntactic context requires
it (e.g., the necessity of a controller in (i)).
It seems difficult to find comparable evidence for Romanian: this language does
not present the French type of causatives (and in particular no " (or Dative)
insertion"); on the other hand the constraint on the distribution of Benefactives is
not as strict:
(14)
The absence of evidence in favour of the idea that transitives with null prototypical
objects behave syntactically on a par with intransitives does not mean that null
prototypical objects are projected in the syntax. Such a conclusion would be forced
if evidence could be produced in favour of the idea that transitives with null
prototypical objects behave like transitives with overt objects. In the absence of
such evidence, I will assume the null hypothesis, according to which null
prototypical objects are saturated in the Lexicon not only in French, but also in
Romanian (and in German).
This hypothesis is supported by the fact that the languages considered here
present no crosslinguistic difference concerning the distribution of null prototypical
objects in active constructions. This is true not only for examples such as (ll)a-c
(which show that French, Romanian and German all allow null prototypical
objects), but also for those cases in which null objects are not allowed:
(15)
a. * Ich habe bis spt in die Nacht getroffen/ beim A bschreiben berrascht.
I have till late in the night met/ caught in the act of cheating.
b. *J'ai rencontrS /surpris.
I have met/ caught in the act
c. *Am intilnit/ surprins.
The contrast between (ll)a-c and (15)a-c is easily explained: eat and drink allow
null objects, because verbs of this type assign a canonical, or prototypical
interpretation to their direct object ('something that can be eaten/ drunk'). This is not
the case of verbs like meet or surprise, catch in the act, which suppose an
individualized direct object (as noted in the grammatical tradition, the meaning of
these "absolute" transitives is incomplete in the absence of a direct object).9 The
crucial observation for the present discussion is that no crosslinguistic variation can
be observed for active transitives with null objects: depending on the semantics of
the verb, null objects are either allowed or they are forbidden in the three languages
examined here (see (11) and (15) respectively). It could be that a given French verb
does not have an exact counterpart in Romanian, e.g. the French verb accepts a null
object, whereas its closest Romanian translation does not. This would not be a
counterexample to the present claim that the distribution of null prototypical objects
in active constructions is not subject to syntactic crosslinguistic variation.
9
It comes as no surprise that the passivized forms of absolute transitives behave on a par with
their active counterparts:
(i)
*Es wurde getroffen/ berrascht.
(ii)
*I1 a 6te rencontr/surpris devant l'universit.
it was met/ caught in the act in front of the university
These examples are obviously grammatical if es/il are referential, but this interpretation is
irrelevant here.
This does not mean that null prototypical/arbitrary arguments are necessarily
saturated in the Lexicon; they may be projected in the syntax, provided that no
individuated reference is assigned to their syntactic realization. Starting with
Section 5.1.5., the main purpose of this chapter will be to examine cases of this
type.10
In sum, it has been shown that in active constructions null prototypical objects
are not realized in the syntax of French, and by the null hypothesis this analysis has
been extended to Romanian (and German). In other words, transitives with null
prototypical objects behave exactly as intransitives from the syntactic point of view.
Conversely, intransitives are frequently found with cognate objects, and in this case
they behave on a par with transitives. This means that subcategorization properties
do not allow us to distinguish between transitives and intransitives (both types of
verbs are allowed, but not obliged, to take a direct object).11 Only their selectional
requirements differ: transitives normally select individuated objects (certain
transitives are nonetheless compatible with nonindividuated or "prototypical"
objects), whereas intransitives select a cognate object, i.e., an NP whose lexical
features are identical to the lexical features of the verb itself.12 Because of their lack
of lexical specification, cognate objects normally remain implicit. In active
constructions they can be realized in the syntax only if they are modified by an
adjective: he used to live a peaceful life (cognate objects function semantically as
modifiers of the verb, but syntactically they behave as direct objects, as shown in
footnote 7).
The idea that intransitives and transitives cannot be distinguished on the basis of
their subcategorization properties seems to contradict a long-standing tradition in
both general linguistics and generative grammar. Despite its well-established
position, the notion of transitivity is far from clear. Thus, the term "transitive"
designates different empirical data from one author to the other: in an extremely rich
study of transitivity in French, Blinkenberg (1960) distinguishes intransitives from
"indirect transitives" (i.e., verbs that subcategorize indirect objects or PPs) and
"direct transitives" (i.e., verbs that subcategorize direct objects). In this taxonomy
the class of intransitives groups together "absolute" intransitives (i.e., verbs such as
telephone, which do not subcategorize) and unaccusatives. In more recent studies,
transitivity tends to cover both transitives and what we traditionally call
"intransitives", whereas the label "intransitivity" is attributed to the class of
unaccusatives (see Hopper and Thomson (1980), and Hoekstra (1984)). It thus
appears that, contrary to the current view, the dichotomy transitive vs. intransitive,
if at all relevant, does not distinguish between the verbs that subcategorize a direct
object and those that do not. It is important to note that the languages examined here
do not grammaticalize the transitivity distinction: no specific marker singles
intransitives out. Compare the case of unaccusatives (see Perlmutter (1978, 1989)
10
One case in point has been analyzed by Rizzi (1986a). In Section 5.1.5. I will discuss the
difference between the principle in (18) and Rizzi's rule of "Assign arb".
1
Unaccusatives should be kept apart: they impose a coindexation relation between the subject
and object positions.
12
This analysis of intransitives is similar to that of Hale and Keyser (1991), who propose that
intransitives rely on the incorporation of an NP into a null V. Under both proposals, a VP headed
by an intransitive presents an underlying direct object in its lexical representation; this object is
not necessarily projected in the syntax. This analysis is supported by the fact that in a number of
languages intransitives do not exist: the Japanese counterparts of English intransitives are formed
with a semantically-empty verb and a semantically full NP; in other languages intransitive verbs
are necessarily accompanied by overt expletive (dummy) objects.
se with (in)transitives
137
and Burzio (1981, 1986)), which in a number of languages are singled out by the
type of auxiliary (they take eire, and not avoir as a perfect auxiliary), as well as by a
specific distribution (occurrence in i/-impersonals in French, en/ne extraction in
French and Italian).
The parentheses I use in (in)transitives signal the idea that the supposedly
distinct classes of transitives and intransitives belong in fact together. The syntagm
passive (in)transitive designates on the one hand passive "intransitives", and on the
other hand passive transitives with null prototypical objects.
The conclusion of the foregoing discussion is that the notion of transitivity is
irrelevant when applied to lexical items (in other words no difference can be
established between run and eaion the basis of their subcategorization properties).
One may nonetheless use the term "transitive" when referring to syntactic
configurations: a sequence in which a direct object position is associated to the verb
may be said to be transitive, and one in which such an association does not hold
will be said to be intransitive. Thus, the sequence made up of an "intransitive" verb
and a cognate object is "transitive", whereas sequences with null prototypical
objects are "intransitive".
5. 1. 3. The passivization of (in)transitives and the relation between the lexical and
the syntactic components
Since lexical items cannot be said to be either transitive or intransitive, transitivity
cannot be stated as a condition on passivization.13 As will become clear below, this
does not mean that we abandon the idea that the object position is necessarily
visible in passive constructions. What must be given up is the idea that an object
position is visible in passives only if it is visible in the corresponding active
constructions. This line of inquiry can only be pursued under a particular
understanding of the Projection Principle (see Chomsky (1981: 29)):
(19) Representations at each syntactic level (i.e., LF, and D- and S-structure) are
projected from the Lexicon, in that they observe the subcategorization
properties of lexical items.
This principle can in fact be understood and used in various ways, as stated
informally in (20):
(20) a. The Projection Principle is a kind of structure preservation condition
(syntactic structures cannot be modified in the course of derivation; in
13lt is well-known that transitivity is not a sufficient condition for passivization, but conversely
it is currently assumed that it is a necessary condition. However, as there is no clear test for
transitivity, we are doomed to circularity. Since at least in certain languages intransitives do
passivize, one must assume that they are transitives; intransitives would be "pure" intransitives in
the languages in which they cannot passivize. The type of parameter proposed by Jaeggli (1986a)
follows this line of reasoning: to say that German intransitives assign structural Case amounts to
assigning them a certain degree of transitivity. The circularity is clear: German intransitives are
quasi-transitives because they passivize, but no other clear test can be provided in favour of their
quasi-transitivity.
139
like the null prototypical object of intransitives, the null cognate object of
intransitives is not necessarily projected in the syntax, although it may be so
whenever its presence is required, as in the case of passives.
To summarize, I assume: (a) that null prototypical/cognate objects are projected
in the syntax of passives, although they are not in the corresponding active
structures; (b) the presence of the empty category imposed by passives is
compatible with the lexical structure of intransitives. The hypothesis that the object
position is visible in passive intransitives has important advantages: (a) a unitary
analysis of passives is preserved (compare Marantz (1984) and Baker (1988) who
stipulate that the passive [morpheme] should be defined in two distinct ways for
transitives and intransitives); (b) it allows us to capture the parallelism between the
passivization of transitives with prototypical objects and the passivization of
intransitives. Finally, this hypothesis opens the way to a possible explanation of the
well-known but little-understood crosslinguistic differences. If we were to assume
that null prototypical/cognate objects were saturated in the Lexicon we would
expect no crosslinguistic variation. If instead they are projected in the syntax, they
are predicted to be legitimate or illegitimate, depending on the satisfaction of
licensing conditions, which in turn depend on the syntactic properties of a given
language. In line with current research (see Rizzi (1986)a), I assume that Case is
the formal licenser for pronominal empty categories; this hypothesis might pose a
problem given the traditional idea that intransitives do not assign Case, but it has
already been shown that this assumption is incorrect (as indicated by the existence
of overt cognate objects). Besides formal licensing, null objects (be they arbitrary,
prototypical or cognate) are also subject to content licensing. In the sections that
follow, it will be shown that the possibility of passive (in)transitives depends on
the satisfaction of both these two licensing requirements.
5. 1. 4. French impersonals and passive (in)transitives
The syntax of passive intransitives is related to the syntax of impersonal
constructions, which show important crosslinguistic differences; these may derive
from other differences, but for the purposes of the present discussion it will suffice
to begin with the (non-primitive) crosslinguistic differences between impersonals.
According to Pollock's (1981, 1986) analysis of French il impersonals, the
postverbal NP is not coindexed with il. In other words, examples (26) rely on chain
configurations of the type in (27), with two distinct chains. Note that (27)
corresponds to il impersonals in general (see (26)a), and to passive impersonals
(see (26)b) in particular:
(26)
(27)
141
(il) (NP) 15
It has been proposed by Reuland (1985) and Chomsky (1986)a that expletives
cannot survive at LF (because of the "Full Interpretation" principle); in ilimpersonals this forces the movement of the postverbal NP towards the position
occupied by il.
5. 1.4. 1. Formal licensing of object pro in French impersonal passives
By the null hypothesis, passive (in)transitives rely on essentially the same
representation as passive impersonals in general. The only peculiarity of (28)a-c is
the fact that the object position is occupied by an empty element (compare (26)b,
which presents an overt NP in the same position):
(28)
rcemment.
trich6pro
hier
sot.
acceptable.
(il) (pro)
As discussed above, the presence of an empty category in the object position is not
due to the lexical properties of the main verb, but rather to the defining properties of
passives (see (25)). Note that the presence of an empty category in the object
position is independently forced by Reuland's and Chomsky's above-mentioned
principle of Full Interpretation: for the expletive il to be deleted we need another
element susceptible to move to the position occupied by il.
In (28) and (29) I have indicated the empty category that occupies the object
position by pro. This is indeed the only possibility: variables and NP-traces are
necessarily bound (by elements in A- and A'-positions, respectively), and there is
no such binder in (28); we cannot assume PRO either, since the object position is
governed by the verb.
15
The chain configuration in (27) relies on the idea that il is not a syntactic clitic, but stays under
[,]; if we assumed that French subject pronouns are syntactic clitics, we would have to
assume a chain configuration of the type (e, il, AGR) (NP), where e is under (,) and il is a
clitic adjoined to Infi. The difference between the two possibilities is irrelevant here.
Let us now consider the licensing conditions to which pro is subject (see
Chomsky (1982) and Rizzi (1982, 1986a)): pro must be formally licensed by Case
features 16 and the content of pro must be licensed by its antecedent. The question
concerning the Case of pro in passive (in)transitives reduces to the more general
question concerning the assignment of Case in French impersonal passives. Since
Case is assigned unproblematically to the direct object position in French
impersonal passives (see Appendix 1), it is natural to assume that pro is formally
licensed in (28).
5. 1.4. 2. Content licensing of object pro
Let us now turn to the differences in acceptability that can be observed in (28).
Since pro is formally licensed, the marginal acceptability of (28)a-b can be due only
to the licensing conditions on the content of pro (no antecedent of pro is present in
(28)a-b). Note that this example becomes completely ungrammatical if we suppress
the adjunct, which indicates that adjuncts may function as content licensers of the
null object.17
What does it mean to say that adjunct modifiers are content licensers for the
object position? It is clear that what is at stake is not a relation of the type found in
standard binding, where two categories are assumed to bear the same index (see the
relation between a trace and its antecedent, or any other kind of anaphoric relation).
On the other hand, it is also clear that some kind of relation holds between verb
modifiers and the object position. Evidence in favour of this idea is related to
examples that are ungrammatical due to an incompatibility between objects and
adverbs. The ungrammaticality of (30)a might be due to a constraint imposed by
beaucoup 'much': this adverb seems to require that the th-position associated with
the predicate boire be non-individuated; this constraint is not met in (30)a, where
the object is a specific NP. Note that by itself, beaucoup is not incompatible with a
specific NP (see (30)b). It thus appears that a complex relation holds between
verbs, adverbs and direct objects:
(30)
16
The technical details of the mechanisms which contribute to the formal licensing of pro are not
specified in Chomsky (1982). In particular it is not clear whether Nominative Case is assigned to
pro or born by AGR: "In a pro-drop language, pro with Case can be left in subject position
governed by AGR, since its content will be determined by AGR with Case, i.e. the pro INFL"
(Chomsky 1982: 86). The Case requirement on object pro has been proposed for independent data
by Rizzi (1986a).
17
Within a different theoretical framework, N. Riviere (1979) considers that the presence of this
complement satisfies the transitivity requirement characteristic of passives. My own analysis,
which postulates an empty category in the object position, supposes a stronger version of the
transitivity requirement. This approach has the advantage to capture the crosslinguistic variation of
passives without parametrizing transitivity in various languages.
weak licensers of objects, insofar as they simply supply them with a spaciotemporal background.
Let us now ask why example (33) is correct, with no adverb:
(33) Jean a bu.
John has drunk
A straightforward answer can be given: it has been assumed above that in active
constructions prototypical objects are not instantiated in the syntax, and therefore
no th-identifier is needed. The necessity of th-identification is thus related to the
syntactic instantiation of the object position (which calls for content licensing),
which characterizes passives, as opposed to active constructions.
Note that the type of French examples illustrated in (28)c (see also (1)) are fully
grammatical for most speakers, whereas (28)a-b (see also (2)) are highly marginal.
The difference between the two types is obvious: the verb in (28)c subcategorizes
for a PP complement; a subcategorized PP appears to function as a better thidentifier for object pro than an adjunct, which is not surprising.
To conclude this section, let us consider certain interpretive differences between
passive intransitives and active structures with arbitrary subjects such as on 'one':
(34)
Passive intransitives allow an overt Agent PP, unlike on 'one' constructions. This is
expected under the proposed analysis of passive intransitives: the Agent th-role is
not projected onto the (,) position, and therefore it may be realized as a PP
complement (compare one constructions: one is assigned the Agent th-role, and
therefore no Agent PP can appear). 18 This clear structural difference between
passive intransitives and on constructions correlates with a subtle difference in
interpretation: in passive intransitives a particular state of things is described as
taking place, and some (specified or unspecified) individual is involved in that
activity; in on 'one' constructions, on the other hand, a particular activity or state of
things is viewed as being initiated by an indeterminate Agent/ Cause.19 If we want
to capture this difference, passive intransitives should be assigned glosses such as
18
Note that according to my analysis of middle/passive se (see Section 5.2.) the Agent th-role is
neither realized in the (,) position, nor "absorbed" by se. This should allow the presence of
an Agent PP, a possibility which is attested, but only marginally (compare the utter impossibility
of Agent PPs in on 'one' constructions). The marginality (indeed the impossibility) of Agent PPs
with middle/passive se is still in need of an explanation, but I believe that this phenomenon is
not due to the fact that the Agent th-role would be present in the verbal morphology or under
(,).
Kratzer (1991) gives interesting examples that show that German man 'one' constructions and
passive intransitives are not interchangeable in certain contexts.
"there is dancing/ eating/ talking... going on" or "there is being danced/ eaten/ talked
..", which are characterized by the presence of NPs that represent the
nominalization of the intransitive verb. This interpretation of passive intransitives is
related to the structure assumed here, and more precisely to the fact that the object
position is necessarily visible: the postulated empty category takes over the lexical
specifications of the verb, and leaves the latter minimally specified. The
passivization of intransitives thus appears to trigger an operation which is exactly
the opposite of the process of "thematic incorporation" that can be assumed for
intransitives (see Gruber (1976) for the incorporation that characterizes motion
verbs): a verb such as mentir can be viewed as having been obtained by the
incorporation of the noun mensonge into the verb dire (for other examples see Hale
and Keyser (1991)); the passive il m'a 6t6 menti pro sans vergogne brings
mensonge back to life at LF (the interpretation is "il m'a 6te dit des mensonges"),
due to the presence of pro.
5. 1. 5. An indexing condition on the chains underlying null prototypical arguments
So far a theory of passive (in)transitives has been proposed, which has been
illustrated by an analysis of the French data. In this section, an explanation will be
proposed for the ungrammaticality of passive (in)transitives in Romanian.
5. 1. 5. 1. Subject inversion and passive (in)transitives in Romanian
Like the other pro-drop Romance languages, Romanian does not present the type of
impersonal structures characteristic of French. Examples (35)-(36) are not
"impersonals" strictly speaking, but rather instances of an inversion phenomenon,
which allows NP subjects to occur in the postverbal position. Depending on the
type of verb, the postverbal subject may occupy either a postverbal subject position,
or the object position (with unaccusatives and passives):
(35)
The difference between the two chains in (35')a comes from a different notation of
null subjects: pro is in fact an abbreviation of (e, AGR), where AGR represents the
fi-features realized on the verb, which identify the empty category under [,]).
(35')a illustrates Rizzi's hypothesis that postverbal subjects coexist with a null
subject under the [NP, IP] position. According to the alternative given in (35')b, no
preverbal subject coexists with postverbal subjects; whether we assume or not that
AGR is part of the chain is irrelevant to our analysis.
The structure underlying the free inversion of the subject with unaccusatives and
passives (see (36)a-c) differs from the simple inversion cases in (35), insofar as the
postverbal subject occupies (or at least may occupy) the object position (see Burzio
(1981,1986)). 20 The subject position is occupied by an empty category, pro. Given
the existence of competing analyses concerning the position of null subjects, we
have at least two options for the underlying structure of (36)a-c:
(36') a. (pro, NP)
(e, AGR, NP)
IP
I (AGR) VPN
V
NP
b. (AGR, e,NP)
IP
I
I'
I (AGR) VPN
NP
V\
NP
I
NP
20This hypothesis explains why ne cliticization is possible from this position, as shown in (i), as
opposed to the postverbal position of intransitive verbs (see (ii)):
(i)
Nej furono arrestati molti ej.
of them were arrested many
(ii)
*Ne telefonano molti.
of them telephone many
This kind of evidence cannot be used for Romanian, which lacks a counterpart of ne. The minimal
hypothesis is that the Romanian example in (35) has the same underlying structure as Italian
(36)a, in which the direct object position is involved, just like in French impersonal passives.
(36')a is based on Rizzi's (1982) analysis, according to which pro/ e occupies the
[,] position. According to an alternative analysis (see Dobrovie-Sorin
(1987)), null subjects occupy the postverbal position; more precisely they occupy
the VP-internal subject position characteristic of overt postverbal subjects; hence
(36')b should be assumed, where the null subject occupies (Spec, VP). Note that
even if there is strong evidence in favour of the idea that in Italian, postverbal
lexical subjects do not occupy a VP-internal position, it is not clear that null
subjects behave on a par with overt subjects. In other words, (36')b may be
postulated even for Italian; one advantage is that in (36')b the relation between
AGR and the empty category that it identifies obeys strict C-command (i.e., Ccommand defined in terms of branching nodes).
Beyond their differences, (36')a and (36')b are alike insofar as they contain only
one th-chain, which bears the th-role assigned to the direct object position occupied
by NP; Nominative case is shared by the members of the chain (alternatively, we
may say that Nominative case is transferred from AGR [or pro] to NP, as proposed
by Rizzi (1982)). This is unlike French il impersonals, which rely on two distinct
chains.
Note now that chains such as (36')a or (36')b violate principle C of binding
theory: the NP in object position is illicitly bound by the null subject; this violation
is independent of the exact location of the null subject (either [Spec, IP] or [Spec,
VP]), because in any case the null subject C-commands the NP in the object
position.21
To solve this problem, let us assume that the chain underlying passives and
unaccusatives is neither (36')a nor (36')b but rather (37), where the overt NP
occupies either the [Spec, IP] or the [Spec, VP] position, and the empty category
stands in the direct object position:22
21
A possible solution to this problem is to assume that principle C does not apply to elements
belonging to the same chain: principle C is currently assumed to be relevant for "referential
expressions", and it is natural to assume that the notion of "reference" is relevant only at the level
of th-chains and that it cannot concern members of chains. This proposal is unacceptable, because
principle C can be shown to be relevant inside chains (see 5.1.6.4.). Principle C should therefore
be stated as a configurational principle (that governs the distribution of non-anaphoric lexical NPs)
rather than as a condition on referential expressions. But in this case (36')a-b would violate
principle C.
22
This suggestion itself raises two kinds of difficulties. The first concerns the idea that at Sstructure postverbal subjects occupy the [Spec, VP] position. While this hypothesis is quite
plausible for Romanian (see chapters 1-4 here, and Dobrovie-Sorin (1987), Motapanyane (1989))
and Spanish, it is much more problematic for Italian, which does not allow the order V-S-O. Note
however that the impossibility of this word order may be due to a locality condition on the
assignment of Objective Case, which would be blocked by an intervening subject (for this
suggestion see Dobrovie-Sorin (1987)). In those configurations in which Objective Case is not
assigned, in particular in passives (see also active constructions with intransitives and
unaccusatives), nothing would block a structure where the subject NP is under [Spec, VP] and the
NP-trace under [NP, VP]. The second problem bears on the well-known asymmetries between
postverbal subjects in general and postverbal subjects of passives and of inaccusatives: according
to Burzio (1981, 1986) ne extraction is allowed from the [NP,V'] position, but not from a
postverbal subject position. An alternative account seems possible, however: we could try to show
that ne-extraction is sensitive to the derivational history of the subject chain. In this case, ne
extraction would be allowed from an NP that has moved from the object position to a postverbal
subject position, but it could not apply out of an NP that has been base-generated in the
(37)
(NP, e)
(pro, e)
For the sake of exhaustiveness, let us consider the alternative hypothesis, according
to which the relevant chain underlying (38) would be of the form given in (36')a,
with pro in the object position; hence (pro, pro). In this chain each instance of pro
must be licensed by Case, and the only possibility is for them to share nominative
Case, probably assigned by AGR (note that Case sharing, or Case transmission,
applies in (36')a between pro and the postverbal subject NP). Rizzi (1986a) has
proposed that the Case provided through chain formation (which would suffice for
the Case filter) is "inconsequential for the licensing of pro", whose licensing is
strictly local and involves a direct relation with a governing (and Case-marking)
head. Under this proposal, chains of the type (pro, pro) are ruled out. It thus
appears that, due to the licensing conditions on empty categories, Romanian (see
also Italian) passives with null prototypical objects must be analyzed as relying on
chains of the type in (39).
postverbal subject position. This type of analysis has already been proposed for Genitive en in
French: Couquaux (1979) has accounted for the extraction of Genitive en from the preverbal
subject position in predicative constructions by assuming that the subject of the verb be is
generated as the subject of a small clause embedded under the verb be. I leave the technical
implementation of this suggestion open for further research. Note that the extraction of Genitive
en and that of quantitative en may well be similar insofar as they are both sensitive to whether the
chain underlying the subject contains the object position. However, the two rules differ insofar as
the extraction of quantitative en is more restricted: it can only apply if the S-structure position of
the subject is postverbal.
properties can only be relevant at the level of the chain. I shall therefore restrict my
attention to chains:
(42)
falls under (42)a; the procedure of free assignment of indices allows nonindexing.
But one may ask why the indexing constraint does not rule out the LF
representations of French passive intransitives. The question arises if we assume
that at LF the null prototypical object moves towards the (Spec, IP) position, thus
deleting il, as required by Chomsky's (1986a) Full Interpretation Principle. This
movement would give rise to a two-member chain, which would necessarily be
indexed, in violation of (43). To solve this problem, we must assume that it is not
the object, but some other element, e.g. a PP or an adverb, that replaces il at LF.
The same conclusion will be reached in Section 5.2.3.
In sum, the proposed explanation of the ungrammaticality of (in)transitive
passives in Romanian involves the following steps: (a) an empirical observation: a
chain of the type (pro, e) is necessarily interpreted in terms of individuated
reference; (b) an explanation of this observation, which relies on conventions
(42)b-c; (c) a condition (see (43)) on the syntactic representation of null
prototypical/cognate objects; more generally, this condition also applies to other
referentially non-individuated arguments.
In subsequent sections, further evidence will be provided in favour of (43). For
the moment, let us concentrate on the observation stated in (a) and the proposed
explanation for it. A similar observation has been made by Rizzi (1986)a: example
(45) cannot be interpreted as the passive counterpart of (44): the null derived
subject cannot be arbitrary; it can only be interpreted as referring to a specific set of
people:
(44)
Turning now to the explanation itself, let us first examine why A'-chains have been
left out of the reach of (42)b. A standard example of A'-chain is provided by clitic
chains (the A'-chains underlying wh-movement and QR are irrelevant here), to
which can be added the chain (AGR, e), underlying null subjects (on the
parallelism between clitics and AGR see Rizzi (1982) and Chomsky (1982)). It is
generally the case that clitics and null subjects are assigned individuated reference,
and therefore (42)b seems to apply not only to -chains, but also to A'-chains.
Note that the idea that (42)b could apply to A'-chains is a restatement (and an
extension) of Rizzi's (1986)a hypothesis, according to which the definite
interpretation characteristic of (45) would be due to the fact that the null subject is
licensed by Infi (or rather AGR). 23 Stated in this way, this proposal is too limited
in scope, or simply ad-hoc: it captures the observed phenomenon, but does not try
to establish a more general property that would cover the observed phenomenon,
but not only that phenomenon. My account in terms of the indexing properties of
-chains has the advantage to deal correctly not only with (in)transitives passives,
but also with the passivization of arbitrary pronouns, of idiomatic expressions, of
verbs that take sentential arguments, etc. (see Section 5.1.6.). Within this
23
Within my theory of reference, which assumes that the notion of reference is only relevant at the
level'of the chain, Rizzi's idea could be reformulated by saying that a chain that presents an AGR
element is necessarily assigned indices, and therefore counts as referentially individuated.
a. II se la coule douce
he himselfoat it flows sweet
'He takes it easy.'
b. Vous l'emporterez facilement
you it will take away easily
'You will get away with it easily.'
(i)
(ii)
(iii)
(i), (ii) and (iii) respectively, borrowed from Rizzi (see his (24)c, (24)d and (26)b):
It is difficult [PROarb' to hope [PROarb' to win the race]].
It is difficult [PROarb' to hope [that PROarb" winning the race] will be easy]
In questo dipartimento, [PROarb' costringere___arb" a [PRO lavorare]] e difficile.
in this department, [PRO to force
[PRO to work]] is difficult
to be assigned "the usual types of referential indices". Rizzi concludes that the "arb"
subscript that he uses in order to signal the arbitrary interpretation of object pro or
of PRO should not be viewed as an index, but rather as simply indicating a
collection of features (+human, +generic, +/-plural...) that enter into the
composition of the arbitrary interpretation. This is clearly an undesirable
assumption for a constrained theory of reference, because we end up with two
technical means to assign reference: referential indices on the one hand, and a rule
that assigns "arb", which appears to be disconnected from the indexing procedure,
on the other.
My own proposal seems to fare somewhat better, insofar as I assume a relation
between the absence of indices and the arbitrary interpretation. However, I also
need two types of indices, which can be descriptively labelled "chain-internal" and
"chain-external". The former would be associated with referential properties,
whereas the latter would not count for the assignment of reference, but only for the
computation of coreference or anaphoric relations that hold between chains,
independently of whether the relevant chains present internal indices. For a further
discussion, see Appendix 1.
Conclusions
To conclude, let me briefly compare the present approach to the current account of
passives in terms of Case absorption (see Chomsky (1981); Jaeggli (1986a)),
which to a certain extent is also a condition on the visibility of the object position.
The difference between the two approaches concerns the way in which this
visibility requirement (or Case-assigning properties) is stated in the grammar.
Inherent to the current approach is the idea that transitivity is a feature of a given
class of verbs. The clear counterexample of passive intransitives is handled by the
assumption that in certain languages intransitives would present peculiar Case
properties (they would, in fact, function as transitives), compared to standard
intransitives.
Compare the approach proposed here, according to which intransitives are
defined in the same way in all languages: they select cognate objects which
normally remain implicit; whenever a cognate object is syntactically realized, the
intransitive verb assigns objective Case to it. In the preceding discussion, the
conclusion has been reached that there exist no clear tests that allow us to
distinguish between two classes of verbs, transitives and intransitives. Therefore,
transitivity cannot be viewed as a property of a lexical class, and we cannot impose
a constraint to the effect that only transitives can passivize. In other words,
intransitives are compatible with passivization in all languages.
I have also assumed a unique, non-parametrized analysis of copula
passivization, which crucially relies on the visibility of the direct object position,
which holds in particular for passive (in)transitives (compare Marantz's (1984) and
Baker's (1988) parametrization of passives). Besides allowing us to preserve a
unique analysis for passives, the hypothesis that the object position is visible in
passive (in)transitives also allows us to account for the crosslinguistic variation of
passive (in)transitives. The possibility or impossibility of (in)transitive passives
across languages depends on whether the chain that comprises the object position
obeys the indexing constraint stated in (43). This, in turn, depends on the particular
structure (and the indexing properties attached to it) that underlies impersonals
The proposed analysis of passive (in)transitives predicts that the contrast between
Romanian passives with postverbal subjects and French /7-passives can be
observed for all these elements. In what follows it will be shown that this
prediction is correct.
5. 1.6. 1. French on and Italian si
Before going into the analysis of a number of constructions that allow us to
compare Romanian and French, I will deal with a phenomenon concerning on 'one'
in French (the same holds for subject si in Italian). This phenomenon cannot be
discussed for Romanian, because this language does not present any overt pronoun
comparable to on. This element is a clitic pronoun which presents two distinct
interpretations: one which is synonymous of the first person plural (translatable by
we) and an arbitrary one (translatable by they''or one).
Cinque (1988) has observed the following interesting fact: in non-generic
contexts, the arbitrary interpretation cannot appear in passives or with
unaccusatives; in (49) for instance (borrowed from Cinque) the only possible
interpretation of on is 'we':25
(49) a. On est partis tout de suite.
one has left immediately
b. On a 6t6 punis tort
one has been punished unjustly
25
Grevisse quotes examples like on est venu voler la pharmacie 'one came to steal in the
pharmacy', where on is arbitrary, despite the unaccusative verb venir 'come'. I do not have a clear
proposal to make, but it seems to be the case that in examples of this type the subject of venir is
necessarily assigned an Agent interpretation, which turns the verb into an intransitive (for further
discussion see Section 5.3.2.).
155
26
Note that on cannot show up in impersonals in French, because this element is a subject clitic,
which cannot appear in the object position. Therefore the only passives that we can construct with
on are of the type given in (49)b, i.e., passives associated with NP-movemenL
27
T h e discussion in this paragraph is based on examples borrowed (or adapted) from Ruwet
(1983) and Simatos (1987).
2
^This kind of example is relatively rare, for reasons which remain unclear. It seems that the rarity
of (50) is not due to an incompatibility between idiomaticity and passives, but rather to one
between idiomatic expressions and the impersonal construction. It is thus possible to find less
rigid idiomatic expressions, which accept the movement of the direct object in passives, but only
very marginally accept impersonal passives:
(i)
Le pli a vite pris.
the fold has been quickly taken
(ii)
Un role important a jou6par...
an important part has been played by ...
(iii)
VII a vite pris le pli.
it has been quickly taken the fold
(iv)
??/i a jou6 un rdle importanipar...
it has been played an important role by ...
a. Le tournant estpris.
the turning is taken
b. Le pli est pris.
the fold is taken
'The habit has been acquired.'
Within the analysis proposed here, the moved idiom chunks are necessarily
assigned some kind of individuated reference, and thereby differ from their active
counterparts. For arguments going in this direction I refer the reader to Simatos
(1987, 1990).
Turning to Romanian, it is to be expected that in this language the passive
construction is ungrammatical with idiomatic expressions, even if the idiomatic NP
occupies the postverbal position:
(53)
157
29
Recall that according to Rizzi's analysis of postverbal subjects examples (53) rely on chains of
the type (pro, NP). As proposed in Section 5.1.5.1. above, this hypothesis has to be abandoned,
because (pro, NP) chains violate principle C of binding theory.
The grammaticality judgments shown in (55) and (56) are parallel to those
discussed above for passive intransitives (and passive transitives with null
prototypical objects), and for passivized idioms. It is desirable to extend the
analysis proposed for those constructions. Our main task is to show that the
configuration underlying passives with sentential complements is comparable to
that proposed for passive (in)transitives.
Following a well-established hypothesis (that goes back to Emonds (1976)), let
us assume that sentential complements d o not occupy N P positions, and in
particular, that they do not stand in the direct object position, 3 1 but in some
extraposition slot. I will assume that the object position associated with an
extraposed sentential complement is occupied by a pronominal empty category,
pro. 3 2 The exact S-structure position of the sentential complement is irrelevant here;
for concreteness, let us assume VP-adjunction:
30
One can find examples such as (i)-(iv), with a limited number of verbs, and with quite marginal
acceptability:
(i)
V.A fost acceptat ca mama s piece dup noi.
[it] has been accepted that mother leave after us
(ii)
VA fost constatat c mul[i elevi erau absenfi.
[it] was observed that many students were absent
(iii)
VA fost descoperit c niciunul din noi nu era in stare s...
[it] was discovered that none of us was able to ...
(iv)
VA fost sugerat c/subliniat c...
[it] was suggested that/ underlined that...
This possibility will be discussed below.
31
Within a derivational framework, Emonds has proposed a rule of necessary extraposition
(induced by the necessary deletion of the empty associated to S', by the rewriting rule NP ->
[[] S']); within a representational framework one may directly base-generate sentential
complements in "extraposed" positions and let the object position be occupied by pro. Any other
possible representation could be filtered by conditions on representations: as suggested by
Emonds, an empty head noun could not survive as a sister of S' (possibly because of ECP); on
the other hand, NP should not be allowed to exhaustively dominate S' because phrase structure
rules are necessarily of the type A -> C, and not of the type A -> B.
^According to Stowell (1981), this empty category is an NP-trace, resulting from the rightward
movement of the sentential complement. But NP-traces are currently assumed not to be marked
for Case: they result from the movement of an NP from a Caseless position towards a position to
which Case is assigned. Stowell's hypothesis is fundamentally different from this standard
assumption. The sentential complement does not move in order to be assigned Case, but in order
to avoid Case assignment: according to Stowell the movement of sentential complements out of
NP positions is forced by the Case Resistence Principle, which postulates that the projection of a
syntactic category which assigns Case (CP is taken to count as a V projection) cannot occupy a
position to which Case is assigned. But within this hypothesis the trace resulting from movement
bears Case, and this is not allowed by the definition of NP-traces.
(57)
159
VP
VP
CP
l \
V pro
The configuration underlying the Romanian examples in (56) would then be of the
type given in (58), where pro and e occupy the [Spec, VP] and [NP, V'] positions,
respectively:
(58)
IP
V-I
Vpass
tV
prOj
e;
CP
To account for the ungrammaticality of (56) one must show that (58) is ill-formed.
This ill-formedness cannot be due to the formal licensing of the chain (pro, e)
because such a chain is allowed in Romanian passives. The ungrammaticality must
therefore be imputed to the presence of a sentential complement. We may assume a
constraint on the (pro, e) chain associated with sentential complements: it must be
referentially non-individuated, and, given the theory proposed in 5.1.5., this in turn
requires that no index be assigned to it. This condition is natural if we want to keep
the referential properties of (certain) sentential arguments distinct from the kind of
reference characteristic of NPs.
The ungrammaticality of (56) is due to the fact that the underlying chain, (pro,
e), is necessarily indexed, by virtue of (42)b (because this chain contains more than
one element), and this violates the condition that characterizes the pro associated
with sentential complements.
Consider now the French impersonal passives in (55), which rely on the
configuration given in (59)b, comparable to the structure in (59)a, characteristic of
impersonal passives with NP complements:
(59)
a.
NP
IP
VP
b.
IP
NP
I'
VP
VP
tv
CP
NP
pass
tv
pro
CP
Vpass
In (59)b pro is legitimate, because it is formally licensed by the Case assigned by
the verb. 33 Since in (59)b pro is the only member of its chain (the associated
sentential complement is not a member of the chain), it is allowed to stay
unindexed, and thus no violation of the condition imposed by sentential
complements arises. As to the condition of content licensing, it is the sentential
complement that serves as a content licenser for pro.
Note that the requirement imposed by sentential complements is trivially
satisfied in active sentences in both French and Romanian, because in active
sentences a pro in the direct object position is the only member of its chain;
therefore, by the free indexing procedure, pro will be assigned no index, as
required by sentential complements.
Let us now examine a number of examples in which sentential complements are
marginally allowed with Romanian passives:
(60)
33Sentential complements (or more precisely their pronominalized form) provide an interesting
confirmation of Pollock's (1983) hypothesis according to which French impersonal passives are
able to assign (Objective) Case to the postverbal position; in (i) the passive is compatible with a
clitic pronoun marked for accusative Case:
(i)
comme il (me) l'a rappeld / ddcidd / jug6 / suggerd,...
as il (meDat) it has been reminded/ decided/ estimated/ suggested ..
Note that Case-assignment cannot be tested for object NPs: because of the Defxniteness Effect, a
definite NP is ungrammatical in *il a 6t6lu le livre dont tu m'asparl 'il has been read the book
about which you talked to me'); hence the ungrammaticality of *il l'a 6t6 lu 'il it has been read'.
The grammaticality of (i) indicates that when le stands for a sentential complement it does not
count as "definite" (the notion of "referentially individuated" is distinct from the notion of
"definite").
In (60) the sentential complement is associated with a chain of the type (pro.e),
which is necessarily indexed. Within the analysis proposed here, this means that
examples (60) are interpreted in terms of individuated reference. One may use two
tests to establish whether sentential complements are interpreted in terms of
individuated reference: (a) glosses using the fact and (b) pronominalization using
the indeterminate plural pronoun acestea 'these, these facts' (Romanian does not
present any closer counterpart of it):
(61) a. Ion a amintit faptul c...
John reminded (mentioned) the fact that ...
b. Ion a aflat faptul c...
John learned the fact that..
(62) a. Ion a spus acestea fr nici-o urm de regret
John said that (lit. 'those (words)') without any regret (sorrow)
b. Acestea au fost spuse/amintite fr nici-o urm de regret
that was said without any regret (sorrow)
Note that the two tests do not completely correlate with each other: faptul is
possible only with factives (i.e., with the verbs that presuppose the truth of their
complement), whereas pronominalization applies to a larger class.
With certain verbs, glosses like those in (61) and (62) are illicit, which indicates
that such verbs do not allow a referentially individuated interpretation for their
sentential complements:
(63)
Examples (64)a mean that John said that..., or that John thought that.... Examples
(64)b, on the other hand, are understood as saying that John examined/ believed
something that somebody else had said.
The verbs in (63)-(64) are precisely those that do not allow the passive:
(65)
The ungrammatically shown in (65) is due to the fact that in Romanian passive
constructions, the sentential complement necessarily takes an interpretation in terms
of individuated reference, but verbs like a considera, a crede or a pretinde do not
allow this interpretation.
The French examples in (66) are comparable to the Romanian examples in (60),
insofar as they are also interpreted in terms of individuated reference:
(66)
a. Que les deux tours soient bien distincts est confirmspar le fait qu'ils sont
combinables entre eux.34
that the two constructions be distinct is confirmed by the fact that they can
be combined with each other
b. ?Que les gens arrivent en retard a finalement 6t6 accepte par tout le
monde.
that people arrive late has finally been accepted by everybody
The sentential complements in (66) stand in an S-initial A'-position (like the other
sentential complements), and are associated to a chain of the type (pro, e) which is
necessarily indexed.
Note that the verbs that allow (66) also allow glosses using the fact or a
pronominal complement:
(67)
a. Jean l'a
dit/rappelS/suggSrS/confirms
John it has said/ reminded/ suggested confirmed
b. Ca a 6t6 dit / confirms / rappelS / suggSrS plusieurs reprises.
it was said/ confirmed/reminded/ suggested several times
c. Jean a dit/rappelS/confirms le fait que...
John said/ reminded/ confirmed the fact that...
In (67)a object NPs such as cela, ga or le are allowed, and these pronominal NPs
can be moved to the subject position, as shown in (67)b. The reference of this kind
of pronoun is weakly individuated: a prepositional content which can be
pronominalized counts among the "individuals" that constitute the background of
our linguistic productions.
Those verbs that do not allow individuated sentential complements (see (68)) are
precisely those that do not allow the movement of the sentential complement (see
(69)):
(68)
163
(60)). Relevant for the present discussion are those verbs that do not allow this
interpretation: such verbs are allowed to appear in French impersonal passives
because in this language the object pro in impersonal passives is the only member
of its chain. In Romanian, on the other hand, the same verbs cannot be used in
passive constructions,35 because the empty category in the direct object position of
passives is necessarily part of the chain that includes the subject position; such a
chain is necessarily indexed, and therefore it is assigned individuated reference,
which is not allowed.
A short note is in order concerning the English counterparts of the constructions
discussed above:
(70)
Examples (70) cannot be assigned the analysis proposed here for French, since
English does not present impersonal constructions relying on two distinct chains.
We must therefore analyze (70) as involving a chain of the form (it, e), where the
empty category e occupies the direct object position. According to the present
theory, this chain should be necessarily indexed, and as such it should not allow the
interpretation in terms of non-individuated reference. I leave this problem open for
further research. 36
Williams (1979) has observed that some of the verbs that can appear in (70) do
not allow the sentential complement to appear in the pre-verbal position (cf. *thatS
was held (reasoned)). The foregoing discussion of comparable examples in French
extends to English.
5. 1. 6. 4. The French stylistic inversion
Let us now consider the French stylistic inversion, which clearly contrasts with the
corresponding il impersonals. In particular, the stylistic inversion is incompatible
with both (in)transitive passives and passive verbs with sentential complements,
see (71) and (72) respectively:
(71)
35
The same generalization holds in Spanish, with certain exceptions which are mainly
counterparts of the exceptions discussed above for Romanian.
^Following Zwicky (1971), Emonds (1976, Section IV.2) notes that verbs such as mumble,
whine, etc. cannot be passivized:
*It was whined/mumbled/shrieked/growled/quipped by Morris that night was falling.
This impossibility is expected under the present account, because the chain (it, e) to which the
sentential complement is associated is necessarily indexed.
a. *le jour quand a rappete que les enfants devaient aniver 1'heure
the day when was reminded that the children had to arrive on time
The French stylistic inversion differs from impersonal passives by the fact that the
postverbal subject agrees with the verb, which indicates that the postverbal subject
is coindexed with the empty category in the [NP, IP] position 3 7 (see Kayne and
Pollock (1978); Kayne (1983); Pollock (1983)):
(73)
37
The contrast between (i) and (ii) constitutes further evidence in favour of the idea that passives
in the stylistic inversion rely on the coindexation of the subject and object positions. According
to Pollock (1986), examples (i) are ungrammatical because a variable (the empty category that
occupies the direct object position) violates binding condition C: it is bound by the empty
category in the subject position. II impersonals on the other hand are grammatical (see (ii)),
because the variable is not coindexed with it
(i)
a. *Combien de coupables. aurais-tu aim6 [que e. soient condamn6s e.]?
how many culprits, would you have liked [that e. be condemned e.]
b. *Combien de criminels. Pierre a-t-il [que. aient relax6s e.]?
how many criminals. Pierre has he denied [that e. have been released e.]
(ii)
how many criminals Pierre has he denied [that iL have been released e.]
There is in fact another possible analysis for examples (i) in a representational model (i.e. a
system in which one does not worry about the input of wh-movement): the second empty category
would count as an -trace, and the variable would occupy the preverbal subject position. The
observed ungrammaticality would still be explained: it would be due to the fftai-trace effect.
Note that principle C incorrectly rules out examples such as (iii)-(iv), in which the postverbal
subject is bound by the empty category in the subject position:
(iii)
J'aurais aim6 que ej soient condamrts [cent coupablcs]j.
I would have liked that e. be condemned [one hundred culprits].
(iv)
Since (iii) and (iv) are correct, we must assume that the postverbal NP does not occupy a position
which is C-commanded by the empty category; in other words, in the stylistic inversion the
postverbal subject does not occupy the [NP, V'] position but rather some A' position. This means
that the chain configuration underlying (iii) and (iv) is not just (e, NP) but rather one which
involves three elements: the lexical postverbal NP, and two empty categories, which occupy
respectively the [NP, IP] and [NP, V'] positions. This analysis of the stylistic inversion also
predicts that examples (i) are ungrammatical: by definition, variables cannot occupy A'-positions.
Compare examples (ii), which are correct because the variable occupies an -position and does not
violate principle C.
165
Assuming this sketchy analysis of the stylistic inversion, let us return to the
ungrammaticality shown in (71), which indicates that (in)transitive passives are not
compatible with the stylistic inversion. The reason might be that in the absence of a
lexical subject, chains of the type (e,e) are illicit in a non-pro-drop language.38
Examples (72) are also ungrammatical, but the reasons are somewhat obscure. It
is also not very clear why idiomatic expressions are ungrammatical in (75):
(75)
38
The case of German impersonals, to be discussed below, offers an interesting apparent counterexample to this claim.
The ungrammaticality of (76)b can probably be derived from the type of impersonal
structures characteristic of German, and a violation of the Case filter may be
invoked: the passive verb cannot assign objective Case. This indicates that the
configuration underlying German impersonal passives cannot be of the type in
(77'), built on the model of French il impersonals: (77') is illicit for the same
reasons that rule (76)b out (object pro is not formally licensed, because it cannot be
assigned Case):
(77)
We may assume that the abstract representations underlying (77) and (78)a-b are
those in (80)a-c, built on the model of examples (79)a-c (the only difference is that
the subject position is occupied by an empty category). Note also that in examples
(79) the verb does not agree with es but with the postverbal subject. This
constitutes further evidence in favour of the idea that in (76), (77) and (79) es does
not occupy the [,] position; hence the chain in (80)d, in which es is absent (it
is currently assumed that es is an impersonal Topic):40
(80)
39
Cardinaletti (1990) argues that es is a displaced subject (which forces her to assume that the
Topic position is some kind of -position). What she means is that es must somehow relate to
the subject position, but it is not clear that a standard kind of movement can be invoked.
40
A S different from Cardinaletti (1990), I do not assume es to be part of the same chain as the
empty category under (,). The main motivation against this idea is (a) the lack of agreement
between es and the verb, shown in (79) and (b) the fact that I assume Topic to be an A'-position.
167
'Each of the elements of this chain is nonetheless formally licensed: e is governed by the lexical
verb and pro is governed by wurde (on the idea that the elements that occupy C are licensing
heads in Germanic languages see Platzack (1987) for Icelandic, and Tomaselli (1987) for German).
Passive se supposes an implicit Agent (see ce livre se lit facilement 'this book se reads easily'),
whereas middle se does not suppose any (see cette brauche s'est cass6e ce matin 'this branch se
broke this morning'). I believe that these two "types" of se rely on the same underlying
configuration; the difference between them is essentially due to the lexical properties of verbs (an
that this language does not allow copula passive with (in)transitives), whereas
French does not allow them (while permitting copula passive (in)transitives). This
state of facts is quite surprising, given the current analysis, according to which the
passive se construction presents ail underlying structure comparable to that of
copula passives (see Belletti (1982), taken up in Chomsky (1981, pp. 270-271)).
Within the analysis proposed here for the passivization of (in)transitives, the first
question to answer is why the indexing condition on the chains underlying null
prototypical objects is satisfied in Romanian passive se (in)transitives (as opposed
to Romanian copula passives). The ungrammaticality of French passive se
(in)transitives will be traced back to a violation of principle A of binding theory.
The analysis extends to idiomatic NPs and clausal complements.
5. 2. 1. Copula passives and passive se
According to Belletti (1982), taken up by Chomsky (1981), the se that appears in
(83) has the characteristic properties of "passive" morphemes: it absorbs the
external th-role and objective Case.
(83)
The conceptual disadvantage of this proposal is that the "passive" past participle
morpheme (in copula passives) and "passive" se are treated as synonyms; 43
correlatively, "passive" se and "reflexive" se are homonyms. This type of analysis
is, I think, a disguised instance of generative semantics: by postulating a passive
morpheme which "absorbs (or causes the absorption of) the external th-role", we
introduce semantic information into the syntactic representation. Further, the
semantic information that is thus introduced in the syntax is precisely what we are
trying to understand.
Let me then suggest an analysis within which semantics is interpretive, that is, it
can be read off structures which have been generated by a truly autonomous
syntax, one which makes no reference to the passive meaning of a given structure.
The syntactic representation should take into account the distinct formal properties
of the constructions under discussion: the verb BE (+past participle) for copula
passives, and se for se passives. None of these elements should be stipulated as
being a "passivizer" (i.e., as absorbing objective case and the external th-role).
TTie passivizing properties of se can indeed be derived from its anaphoric status.
Being an anaphor, se is by definition bound by the NP subject; being an object
clitic, it binds an empty category in the object position. These two relations are
standardly assumed to be notated by coindexation, so that we end up with the
indexing configuration given in (84), where NPj and ej occupy respectively the
implicit Agent is absent or present in the lexical representation of verbs). In other words, I believe
that "passive" se and "middle" se refer in fact to the same linguistic element.
43
Note that according to J.C. Milner (class lectures) complete synonymy is probably nonexistent
in general; morphological synonymy is even more questionable than lexical synonymy.
169
subject and the object position, and se occupies an A'-position, attached to the
inflected verb. Depending on the exact position of the verb relative to the subject
NP, se may stand in front of the lexical NP, but this does not bear on the present
analysis:
(84)
NPisejei
The chain in (85)a presents a number of variants; see in particular the French impersonal
constructions discussed below.
45
Rizzi (1976a) has tried to extend the analysis in terms of subject si to middle-passivizing si.
with those authors (see Hyams (1986)) who have explicitely acknowleged that the
constructions relying on this type of si are active constructions46 (comparable to
on- constructions in French), to be distinguished from those relying on
middle/passive si: like any nominative clitic, subject clitic si identifies an empty
category in the subject position, but bears no relation with the object position;
correlatively, subject si cannot be said to be an anaphor (no antecedent can be
defined for subject si). Insofar as subject si is not an anaphor, it stands outside
any possible unification with the other si's, which are anaphoric elements (see also
Section 5.3.): in particular, the configuration in (84) does not hold for subject si.
The term "impersonal si" is imprecise, because it could be used to refer to
examples such as il s'est vendu beaucoup de livres, which are impersonal (insofar
as they present an expletive subject), but do not involve subject/Nominative si. I
shall therefore adopt the label "subject" si instead of "impersonal" si; "passive" si
will refer to both impersonal passive si (see il se boit beaucoup de caf6 dans ce
pays 'il se drinks much coffee in this country') and the passive si associated with a
preverbal subject (see (83)).
Leaving aside for the moment the discussion of subject si, let us concentrate on
examples (87), and examine whether, contrary to the above-mentioned current
assumption, they can be analyzed in terms of passive se. One problem raised by
this hypothesis is the supposed incompatibility between (in)transitives and
passivizing strategies, but this has already been settled in previous sections.
Let us then assume that the underlying structure of examples (87) is comparable
to that in (83), the only difference being that in (87) the object position is occupied
by a null subject. (88)a-b correspond respectively to (83) and (87):
(88) a. (NP, se, e)
b. (pro, se, e)
This type of chain is similar to the one proposed above for passive (in)transitives,
insofar as the direct object position is visible. This analysis is forced here by the
anaphoric properties of se.
The perfect grammaticality shown by the Romanian examples in (87) is
unexpected: (88)b is a chain that contains more than one element, and as such, it
should necessarily be indexed (by the indexing convention stated in (42)b), and
consequently it should be assigned individuated reference (by virtue of (42)c); this
is incompatible with null prototypical/cognate objects, which cannot be assigned
individuated reference. This contradiction should either lead to ungrammaticality or
force the referentially individuated reading (e.g., "a song was sung"). But instead,
examples (87) are grammatical and interpreted as involving cognate/prototypical
objects.
Within the analysis proposed for copula passives, the interpretation characteristic
of (87) can only be accounted for if we can provide a principled reason for why
chains of the type in (88)b are not indexed. I will build on an obvious peculiarity of
these chains, namely the fact that they contain se, which is a lexical anaphor. Let us
assume the following:
46
Belletti (1982) distinguishes nominative and accusative si but does not explicitely assume an
active configuration for nominative si.
The use of anaphoric indices allows us to preserve the fundamental intuition behind
(42)b, namely the idea that a multi-member chain is marked by indexing (which is a
way of indicating that the members of the chain belong together). The undesirable
consequence of (42)b is avoided by the idea that anaphoric indexing d o e s not
necessarily correlate with individuated reference. A technical implementation and
motivation for (89)a-b will be provided in Appendix 1.
The idea that chains such as (88) are not assigned individuated reference also
explains the grammaticality of passive se with clausal complements:
(90)
47
A fost vindut.
has been sold
(pro, e)
(pro, AGR, e)
(AGR, pro, e)
The question is why in (i)-(iv) the empty category does not function as both the trace of AGR and
of pro, which would force us to assume a complex chain and as a consequence the deletion of
indices. Such an analysis would destroy most of the results obtained in Section 5.1. Fortunately,
however, chains of the type in (iii)-(iv) are clearly not complex, because AGR is exclusively
related to the subject position. In other words, in (iii)-(iv), pro is the only antecedent of e, and
AGR is the only antecedent of pro. Compare si (passive si in particular) which is an object clitic,
and as such necessarily binds an element in the object position; the "complexity" that arises in
passive si configurations is that the same element in the object position is also bound, via se, by
the subject NP.
173
une
enquete.
un compromis
acceptable.
manifestants.
sugg6rera
de
partir.
surpilotis
(k cet
endroit).
4h du
matin.
assiette.
a. II se raconte
beaucoup
de betises ces
temps-ci.
id.
^principle (95) leaves open the possibility to move clitic-chains as a whole. But note that if (se,
e) is moved as a whole at LF, we do not get rid of the violation of principle A: since se itself is
moved, it will still lack an antecedent at LF.
The descriptive generalization seems to be that idiomatic NPs cannot move at LF,
which is not surprising, since idioms are frozen expressions, which should be
interpreted as a whole, with the idiom chunk next to the verb. I would simply like
r6cemment
in Italian
(99)
We may imagine two other possible underlying representations for (99)b, depending on various
analyses of passives:
(i)
*Adesea proj se este [ej trdat ej de pneteni fal$i].
often pro. se is [e^ betrayed e. by unsincere friends]
(ii)
*Adesea proj/j sej este [ej/j tse trdat ej/j de prieteni fal$i].
According to (i)-(ii) passives share the small clause analysis of other copula constructions. They
differ from other copula constructions by the fact that with passives the object position is visible,
and coindexed with the subject position of the small clause. If (i) is the correct underlying
representaion of (99)b, this example is to be excluded on a par with (99')a: se cannot satisfy its
definition (since there is no object position in the main clause); on the other hand proj and the
object ey are too far from each other, being separated by subject ej. (ii) differs from (i) by the idea
that at D-structure se is generated attached to the lexical verb itself, rather than to be. In this case,
the defining property of se can be satisfied in the small clause embedded under be, hence the i
indices. The j indices on the other hand are due to the past participle (or passive) morpheme itself.
This indexing configuration is illicit on a par with (99')b, because the 1-Advancement
Eclusiveness Law is violated inside the small clause.
52
Sandfeld et Olsen (1936) register examples of this type, but they note that Iordan considers that
this construction is "fautive", i.e., incorrect in standard Romanian:
(i)
Se nume$te pe Popescu profesor.
se appoints pe Popescu professor
Sandlfeld and Olsen do not register any example corresponding to (104), which would present an
Accusative clitic with passive se. They do give a number of examples where the verb does not
agree with the postverbal subject, but it is clear that the Romanian examples are not counterparts
of (103); they display the lack of subject-verb agreement which characterizes substandard
Romanian:
(ii)
A scris copiii.
has written the children
In this example copiii is the subject of a trimis, even if it does not agree with the verb. These data
show a simplification of the verbal paradigm, in which the third person singular and plural forms
are identical. An example such as (103) in Italian is completely different, because the lack of
subject-verb agreement indicates that the postverbal NP does not function as a subject, but rather
as an object.
179
181
The possibility of these constructions is probably not due to se itself, but rather to
the lexical properties of certain unaccusatives. Thus, in a number of languages,
unaccusatives may passivize: 57 the French parvenir 'arrive at' and descendre
'descend, go down' must be classified as unaccusatives, because they take etre (note
in particular je suis (*ai) descendu jusqu' 30 mttres de profondeur. am (*have)
descended up to 30 meters under ground'), but they can nevertheless be found in
impersonal passives:
(110) 111 a parvenu un accord satisfaisant.
it has been arrived at a satisfactory agreement
II a 6t6 descendu jusqu' 30 mitres de profondeur.
it has been descended up to 30 meters under the ground
The problem raised by (109) is exactly the same as the one raised by (110): the
indexed representation in (110') is an LF representation which is illicit by virtue of
the 1-Advancement Eclusiveness Law:
(109') a. Ien proj/j sj-a ajuns ei/j tirziu acas
(110') a. J/j/j a dte parvenu ej/j a un accord satisfaisant.
In (109') the i indices are imposed by se, and the j indices are due to the lexical
properties of unaccusatives; similarly in the LF representation given in (110'), the
movement imposed by the passive coexists with the one imposed by unaccusatives.
The only way out is to assume that in impersonal copula passives and passive se
constructions, unaccusatives lose their lexical specification and behave on a par
with intransitives.
The shift from unaccusative to intransitive behaviour may be subject to
crosslinguistic variations (factors such as volition, human activities/ states, etc. may
be relevant in certain languages, but not in others),58 but basically the same process
is at work. Note also that even with respect to the other current tests for
unaccusatives (see Perlmutter (1978, 1989)), certain verbs behave in a paradoxical
way:
(111) a. Je voudrais savoirpourquoi Jean a manqu6hier.
I would like to know why John missed (the class) yesterday
57
See German impersonal passives with sterben 'to die' and bluten 'to bleed' in contexts in which
a volitional Adv is present. Note that a volitional interpretation is a necessary, but not a sufficient
condition for unaccusatives to allow passivization: gehen 'to go', nahen 'to come closer' and
kommen 'to come' are clearly volitional, but they do not passivize. In some languages the
passivization of unaccusatives is not even a marked, or marginal possibility, but a current option.
Oskaragoz (1980) gives examples that show that in Turkish, impersonal passives can be
constructed with unaccusatives such as dii$ 'fall' and biiyii 'grow up' which are clearly not
volitional (but are restricted to human beings). In Lithuanian even the restriction to humans does
not hold.
^According to Givon (1981) there exists at least one language, Ute, which allows all verbs to
passivize. This means that in this language no verb is lexically marked as belonging to the
unaccusative class; or else the shift from unaccusatives to intransitives is completely free in Ute.
183
(iii)
*cintatacas, Ioana..
sung at home, Ioana ...
(iv)
*Muncit, Ioana a incereat s se odihneasc.
worked, Ioana has tried to take a rest
Note, however, that this test is not entirely reliable: certain unaccusatives are ungrammatical with
"absolute" past participles:
(v)
*Lipsit, Ion n-a aflat vestea cea bun.
missed, John has not heard the good news.
This paradigm indicates that intransitives (see (113)b) compare with transitives (see
(113)a) and contrast with unaccusatives, psych verbs, copula constructions and
passives (see (113)c-g).
The generalization underlying the contrast between (113)a and (113)c-g seems
clear: in [-finite] contexts passive si is allowed (see (113)a), whereas subject si is
not (see (113)c-g). Note that examples of the type in (113)a are grammatical only if
the postverbal NP agrees with the verb. In the absence of agreement the si is
necessarily subject si, which gives rise to ungrammaticality, on a par with (113) cg. ((132 is Cinque's (79)a)
(114) *Sembra essersi venduto poche automobili.
[it] seems si have been sold (sing) few cars
Let us now try to understand the data: in particular, why subject si is not allowed in
(113)c-g and (114). As noted by Cinque, it is clear that the possibility of assigning
nominative Case cannot help us in distinguishing between the structures that rely
on passive si and those that rely on subject si: even if passive si presumably takes
accusative Case, the assignment of nominative Case is nonetheless necessary in
these constructions (see (113)a), in order to identify the pro element under [,],
What Cinque does not notice is that one may assume that it is the assignment of
nominative Case to si itself which is the relevant fact (the presence of nominative
pro is irrelevant). My proposal is that the ungrammaticality in (113)c-g is due to the
descriptive generalization stated in (115). It is important to note (this remark is due
to R. Kayne, p.c.) that (115) holds for the subject clitics that appear in the Italian
dialects (these elements are clearly syntactic clitics,61 cf. Rizzi (1986c), on a par
with subject si).
(115) Nominative clitics (be they phonological or syntactic) are necessarily
identified by AGR.
The condition stated in (115) allows us to account for the contrasts in (113) in
terms of the distinction between subject/Nominative si and passive/Accusative si.
Returning now to examples (113), the c-g examples are ungrammatical because
in these cases si can only be nominative (unaccusatives, adjectives and passives are
incompatible with passive si), and as such must be identified by AGR (as stated in
(115)), but there is no AGR. Example (113)a is grammatical because it allows the
presence of passive si (which takes objective Case) correlated with a nominative
pro, neither of which need be identified by AGR features (on the presence of pro in
[-finite] clauses see Rizzi (1982); Chomsky (1982)).
61(115) also holds for French subject clitics (whose clitic status is clear at Phonological Form,
but debatable in the syntax, cf. Kayne (1984)), and in particular for on:
(i)
Jean '6tant pas la maison, Marie est allSe au cin6ma.
John not being at home, Mary went to the cinema
(ii)
Lui 'tant pas la maison, Marie est all6e au cin6ma.
him not being at home, Mary went to the cinema
(iii)
*Onn '6tant pas la maison, Marie est all6e au cin6ma.
on ['we'] not being at home, Mary went to the cinema
(iv)
*II '6tant pas la maison, Marie est a116e au cin6ma.
he not being at home, Mary went to the cinema
The crucial observation for our present concerns relates to (113)b, whose main
verb is intransitive. Its grammaticality indicates that this example must be analyzed
as relying on passive si and not on subject si. We thus reach the conclusion that at
least in [-finite] clauses, si intransitives rely on passive si in Italian. In finite clauses
si intransitives are structurally ambiguous between a configuration relying on
passive si and one involving subject si.
5. 3. 4. An alternative analysis
I owe here an explanation for not adopting Cinque's (1988) distinction between
[+arg] and [-arg] si. The first reason is theoretical: [arg] features do not count
among the primitive notions of the variant of GB theory adopted here (see the
Introduction). Moreover, I believe that the distinction in terms of [arg] features is in
fact an artifact, which makes no clear prediction for independent data. This
distinction must be set against the background assumed by Cinque: si is analyzed
as a clitic (base-generated in an A'-position under Infi, as shown in (116))
coindexed with [,] (and AGR). TTiis definition raises a problem for the thcriterion in a number of cases, and more precisely with passives, unaccusatives,
etc., which assign no th-role to the [,] position at D-structure. Consider (116),
which corresponds to Cinque's (24)b; I have deleted the features that Cinque
assigns to the various elements:
(116)
IP
AGR
si
NP
parte
"If si is an argument, by the th-Criterion under the Projection Principle, it must be
associated with a th-role at every level of representation, including D-structure"
(Cinque (1988: 527)). But this requirement is not satisfied in (116), where the verb
belongs to the unaccusative class, and as such assigns a th-role to the [NP,V']
position, and leaves [Spec, IP], or rather [Spec, VP] with no th-role (the same
problem has been noted by Burzio (1986); Belletti (1982) and Manzini (1986)).
Cinque's solution is to assume that in (116) si is not an argument (hence the feature
[-arg] that would characterize it), as opposed to the si that shows up with transitives
and intransitives, which assign a th-role to [NP, IP], and thus allow the presence of
a [+arg] si. Thus, si dorme is compatible with two distinct representations,
depending on the [-arg] or [+arg] features that can both be assigned to si with
intransitives; correspondingly, we would get either an "argument" pro element, or a
A non-trivial assumption made here is the idea that at D-structure th-roles are allowed to be
assigned to an A'-position (si occupies an A'-position, and it is directly assigned a th-role, leaving
the [NP, IP] position with no th-role).
6
3 N o t e that like si, the on that shows up in (117)c-d is necessarily interpreted as "we" in nongeneric contexts. Compare (117)a-b, where on is ambiguous between "we" and the arbitrary
interpretation.
leaves us with two problems. The first concerns various theoretical refinements that
should make our theoretical framework compatible with data of the type in (116): a
reformulation of the th-Criterion, a more precise definition of the notion of
argument, or a modification of our view of D-structure are in order. The second
problem is empirical: since the [arg] distinction between the two types of si cannot
be adopted, we must look for another explanation of the contrasts in (113). And
this problem has already been solved above. Si doime can be analyzed as a passive
si construction: in this case si is an anaphoric object clitic, which, as such, (a)
identifies an empty category in the object position and (b) is bound by the subject
NP. Examples (113)c-g are ungrammatical because they present verbs which can
only be constructed with subject si (see the discussion in 5.3.1.), but subject si
cannot show up in [-finite] clauses. According to this analysis, si dorme is
systematically ambiguous between a configuration involving passive si (analogous
to the Romanian se doarme) and one which involves subject si.
Appendix 1: Indexed elements and indexed chains
An important part of the proposed analysis relies on a number of indexing
conventions which may look quite stipulative. Let me then try to justify (42)a-b and
(81), repeated here:
(42) a. Indices are assigned freely to one-member chains.
b. Indices are assigned necessarily to -chains that contain more than one
member.
(81) Indices can only be assigned to chains that contain (a) an overt element in an
-position or (b) elements marked with pronominal (or "strong") fi-features.
Conventions (42)a-b and (81) are conceptually different: (42)a-b deal with the
cardinality of a chain, whereas (81) deals with the nature of the elements
composing a chain. Note, on the other hand, that (81) is a necessary condition on
(42)a-b. Let us then try to make the relation between (42) and (81) explicit; what
we need is a relation between the assignment of indices to elements and the
assignment of indices to chains. A possible technical implementation is given in
(118)a-b, which apply on S-structure configurations:
(118) a. Assign an index to overt [+N] categories.
Overt [+N] categories:
(a) overt NPs in -positions (pronouns and full NPs);
(b) elements that occupy A'-positions and bear pronominal or
strong fi-features: pronominal clitics, pronominal AGR.
b. Coindex empty categories: empty categories inherit the index of their
antecedent.
The index assignment stated in (118)a applies "freely" in the sense that any [+N]
overt category is assigned a random index (i, j, ...t) or no index at all (this option
will be notated by 0). In other words, non-indexing is a subcase of free indexing.
However, we do not want to allow coindexation as a subcase of free indexing: the
indices assigned through (118)a are random, but necessarily distinct from each
other.
Before showing how (118)a-b work, let me mention that (118)a-b are close in
spirit, but technically different from Chomsky's (1980) indexing algorithms known
as "contraindexing of nonanaphoric NPs" and "coindexation by movement". The
coindexation of anaphoric NPs will be introduced into the discussion shortly.
Within the theory proposed here, (118)a is the only procedure of index
assignment, (118)b can be viewed as a procedure of index-inheritance, or indextransmission, that applies between an indexed element and an empty category,
which cannot be directly assigned an index. If there is no indexed element, the
procedure of index-inheritance itself cannot apply, and the chain will be left
unindexed. Thus, by (118)a-b we obtain the results that were previously captured
by means of (81), which is no longer necessary. A case in point is provided by
German impersonal passives (see the discussion in Section 5.1.6.5.), which rely on
chains of the type (e, e) that contain no element that could be assigned an index; the
first empty category is a null subject, which is not assigned an index, because
German does not present "strong" AGR (German is not a null subject language);
the second empty category is a null object, which occupies the preverbal position,
because German is an SOV language:
(119) Es wurde e e getanzt.
it was e e danced
Note furthermore that none of the conventions in (118)a-b corresponds to the
generalization stated in (42)b, which was crucial for my account of the
impossibility of Romanian passive intransitives. The crucial phenomenon that (42)b
was intended to capture was the necessary indexation of chains of the type (NP, e).
This necessary indexation can be shown to follow from the principles that
govern the identification of empty categories. By applying our indexing
conventions to copula passives, we obtain configurations as those shown in (120):
(120) a.
b.
c.
d.
the licensing requirement of empty categories (more precisely empty anaphors); the
principle stated in (42)b can be dispensed with.
It is interesting to compare -chains to A'-chains such as clitic chains or (AGR,
e). I am going to argue in favour of representations like those in (121)b,d and
(122)b,d, which differ crucially from (120)c by allowing non-indexing. The
indicated indexing is necessary in order to capture the interpretation correctly: in
(121)c le functions as a predicate, and as such it should not bear an index; in (122)c
the null subject is an expletive which should not be assigned any reference.
(121) a. Mane l'a rencontr e hier.
Mary him has met e yesterday
b. (li,ei)
c. Intelligente, Mane Vest sans le savoir.
intelligent, Mary it is without it knowing
d. G0,e0).
(122) a. Pleac e miine.
[he, she] leaves tomorrow
b. (AGRj.ei)
c. Plou e.
[it] rains
d. (AGR0, e0) or simply (AGR, e)
What we have to explain is the difference between NP-traces and clitic-traces
(AGR-trace is a particular type of clitic-trace). It has been shown above that NPtraces are necessarily coindexed (non-indexing is disallowed), and we now want to
show that clitic-traces allow non-indexing (together with coindexing). If we assume
as above that co-indexing is necessary for the binding of empty anaphors, we reach
the conclusion that clitic-traces are not bound by their clitic-antecedent (see (121)d
and (122)d, which are not indexed). In other words, the proposed theory of
indexation leads us to the conclusion that the relation between the clitic and its trace
is not a binding relation. And this is consistent with the fact that clitic-traces do not
count as anaphors (they are not viewed as being subject to principle A), and
principle A is stated in terms of -positions exclusively. The question is why this
should be so: why do we not assume a "super-principle" A which would be
relevant not only for -positions, but also for A'-positions? Also, why do we not
assume a definition of empty anaphors that would subsume NP-traces and clitictraces? A principled reason in favour of the current theory, and against the
suggested generalization of binding theory can be found if we distinguish between
two relations that appear to be coalesced by the chain-formalism, namely a relation
between positions, and a relation between elements. Consider again the A-chain
characteristic of passives: it basically involves a relation between two positions, the
subject and object positions, as shown in (123). The relation between two elements
(an antecedent NP and an NP-trace) is a consequence of the relation between the
two positions, which is established (by virtue of the properties of the verb BE,
combined with the properties of past participles) independently of the nature of the
elements that instantiate these positions:
(123) ([NP, IP], [NP, V'])
Consider now clitic-chains. In this case no relation can be established between two
distinct positions independently of the elements that instantiate them: the A'position occupied by the clitic is not projected independendy of the presence of the
clitic. It thus appears that the relation between a clitic and its trace is a relation
between an element (the clitic) and the position occupied by its trace, rather than a
relation between positions. In other words, a clitic-chain does not function as a
relation between two positions, but rather as a discontinous element. The fact that a
clitic-chain is freely (as opposed to necessarily) assigned an index may be
interpreted as being due to the fact that a clitic-trace is licensed by the clitic in the
absence of any index, by virtue of the close relation that holds between the clitictrace and the clitic. Everything works as if clitics occupied the positions of their
traces. It is in fact reasonable to assume that insofar as they are A' chains, clitic
chains allow reconstruction at LF: clitics are reconstructed into the position of their
traces. The notation pro instead of (AGR, e) or (cl, e) captures the idea that one
(discontinuous) element is involved. Note also that the notion of "identification" is
currently used to refer to the relation between the clitic and its trace. The case of Achains is crucially different: the NP-antecedent and the NP-trace cannot be viewed
as a truly discontinous element, because the relation between them is first and
foremost a relation between two positions. Correlatively, there is no possibility to
refer to an -chain as a whole by using just one cover-term comparable to pro.
Correlatively again, the relation between the two elements of an -chain is not one
of identification, but one of binding.
To summarize, it has been shown that indexing convention (42)b (i.e., the
necessary assignment of indices inside -chains, and correlatively the prohibition
of non-indexing), need not be stated, but is due to the licensing of empty anaphors;
it does not apply to clitic-traces, which do not qualify as anaphors and may be
assumed to be licensed at LF, via reconstruction.
Let us finally consider the chains underlying passive se/si in Romance
languages. According to my analysis, the relevant chain is of the type (NP, se, e) or
(se, NP, e), with two overt [+N] categories, each of which is assigned a distinct
index by (118)a. For illustration, consider example (124):
Cette question s'est discut6e e hier.
this question se has discussed e yesterday
(125) a. Cette questionj/0 sj/0'est discutde e hier.
b. Cette questionj/0 sj/0'est discute ej/0 hier.
(124)
The indexed configuration in (125)a is due to convention (118)a: cette question and
se are assigned distinct indices, because these two elements are distinct [+N]
categories. The move from (125)a to (125)b is the result of an application of
(118)b: the empty category in object position necessarily inherits the index of the
object clitic that binds it. The question is now whether it is possible to pass from
(125)b to (126):
(126) (NPj, sej, ej)
The coindexation shown in (126) can be viewed as necessary: se is an overt
anaphor, which by definition is subject to principle A. But note that necessary as it
may be, any mechanism that allows us to go from (125)b to (126) does not obey
the indexing procedures assumed so far: in other words, the anaphoric status of se
runs against free indexing, and in fact leads us to delete some of the indices
assigned by applying free indexing. I am going to assume that representation (126)
cannot be correct, because index deletion and reindexing are powerful devices,
which should be avoided.
Procedures of index deletion and reindexing have already been proposed in the
literature, see in particular Higginbotham (1980), who deals with bound pronouns.
This quite drastic procedure can be avoided by assuming, following Chomsky
(1980), that lexical anaphors are assigned not referential but rather anaphoric
indices. The distinction between two types of indices allows us to distinguish
between two conceptually distinct relations: the relation between an empty category
and its antecedent (or, in derivational terms, the coindexation relation motivated by
movement) and the relation between a lexical anaphor and its antecedent.
Correlatively, empty categories (which lack any kind of intrinsic features) 64 are
kept distinct from lexical anaphors, which bear anaphoric intrinsic features which
do not associate with referential properties, but only indicate a relation to an
antecedent.
To the conventions in (118)a-b we should then add a specific procedure for
overt anaphors. Anaphoric indices,65 just like referential indices, are transmitted to
C-commanded empty categories. The indexing conventions in (118)a-b thus
become (127), to which I add the well-formedness condition on chains stated in
(128):
66 A .
-chains do allow this possibility: cases in point are clitic doubling configurations, where the
doubled NP occupies either an -position or a left dislocation position (see Chapter 6 below).
These chains are currently interpreted in terms of individuated reference, which given the
assumptions adopted here may be taken to indicate that referential indices are assigned, and this
violates (128), since by virtue of (127)a, the clitic and the doubled NP must be contraindexed. We
may assume that index deletion and reindexing are allowed to apply in clitic chains: the clitic
necessarily imposes its index to the -position to which it is related; in case that position is filled
by an overt NP, the latter will have to take up the index of the clitic. This imposed indexation
probably relates to the fact that clitic-doubling chains not only allow the referential interpretation,
but also seem to be necessarily associated to iL
67
N o t e that given the proposed analysis, lexical anaphors give rise to different indexing
configurations, depending on chain-configuration: for the reasons discussed above, referential
indices are not assigned to the chains underlying middle/passive se configurations. Compare
reflexives:
(a)
Jean se lave e.
John se washes
(b)
Jeanj se; lave e
(c)
Jeanj/ sei/ lave e
(d)
Jeanj/ sei/ lave ej/
(e)
(Jeanj/) (sei/ ei/)
From (a) we obtain (b) by free contraindexing; (c) and (d) are the result of anaphoric coindexation
and coindexation of empty categories, respectively; (e) is obtained by a procedure that freely
assigns chain-structure (see Rizzi (1986b)): several -positions may belong to the same chain
provided only one of them is assigned a th-role. The indexing configuration in (e) does not violate
(128), because it involves two distinct th-chains, each of which contains identical indices.
(D)
The hypothesis stated in (C), which can be found in Marantz (1984), Baker (1988)
and Afarli (1989), is in fact empirically indistinguishable from the one adopted by
Jaeggli (1986a). It is indeed difficult to see in what way one could verify whether
in French, Case is assigned both to the passive morpheme and to the postverbal NP
(as assumed by Jaeggli), or just to the latter (as proposed by Baker and Marantz).
The suggestion in (C) is also meant to capture the fact that in certain languages
passives can be constructed with (in)transitives. Note, however, that passive
(in)transitives and impersonal passives of the French type (where the postverbal
NP appears to bear a Case other than nominative) do not necessarily go together:
German allows the former, but not the latter construction. Therefore, it seems that
the two phenomena cannot be captured by one single parameter. Under Jaeggli's
hypothesis we must indeed assume a parameter specific of intransitives: in certain
languages (see German, as opposed to Romance languages other than French)
intransitives would be able to assign structural Case.
The hypothesis stated in (D) seems more interesting. It is indeed quite clear that
the difference between French and Romanian impersonal passives is a particular
case of the more general difference between impersonal constructions. French ilimpersonals can be used not only with passives, but also with unaccusatives and
passive se:
( 1 3 2 ) a.
II anivera
trois
personnes.
de
livres.
This question cannot be settled here. However, it seems plausible to think that it is the chain
property that is crucial: the existence of the type of chain configuration characteristic of French
impersonals may be due to the fact that in this language nominative Case is necessarily assigned
to the (NP, IP) position. If we assume that nominative Case cannot be transmitted to the (NP,V)
position, we derive the necessity of direct Case assignment to the (NP, V') position (see Appendix
2). In pro-drop Romance languages, nominative Case can be directly assigned to positions other
than (NP, IP), and therefore no expletive chain is needed. Thus, the chain property characteristic of
French impersonals can be derived from a condition on nominative Case assignment. To put it in
more intuitive terms, the existence of French impersonals is related to the fact that this language
lacks free subject inversion, which, to varying degrees, is available in the pro-drop Romance
languages. Note that the lack of free inversion is a necessary but not sufficient condition on the
implemented correctly, the ability of impersonal passives to assign Case to the (NP,
V') position would not be stated as a property of passives themselves, but rather as
a more comprehensive characteristic, which concerns all the structures characterized
by the non-assignment of the external th-role (see examples (132)a-b). It thus
appears that the Case properties of impersonal passives need not, and most
probably should not, be parametrized (they are either consequences of the Case
properties of impersonals or derive from the chain configuration characteristic of
impersonals).
It is not clear what the exact Case assigned to the (NP, V') position is in French
impersonal passives: we may consider that passives are allowed to assign objective
Case 69 (under this option Case absorption should not be stated as a characteristic of
passives); we may alternatively assume that passives absorb objective Case,
another Case being assigned to the (NP, V') position (see Belletti (1988)). I tend to
prefer the former hypothesis (in a number of languages morphological accusatives
are indeed allowed with passives), but the choice between the two possibilities is
not relevant here. What is relevant is the fact that the NP which occupies the direct
object position is the only member of its chain, and that Case (be it objective or
partitive) is assigned in that position .
existence of impersonals of the French type (see English, which lacks free inversion, but does not
present the relevant kind of impersonals).
^Recall that examples such comme il me l'a sugg6r6 'as there to me was suggested it' are
correct in French, with an Accusative clitic.
Part of the material discussed in this chapter was presented at the 1985 GLOW Conference in
Brussels, and published in Dobrovie-Sorin (1985). The present version, considerably revised and
enlarged, benefited from the comments of A. Belletti, G. Cinque, A.Grosu, J. Guiron, O.
Jaeggli, J.C. Milner, L. Rizzi, I. Simatos and two anonymous reviewers for Linguistic Inquiry.
Special thanks go to R. Kayne; some important results presented here are answers to his
insightful remarks and questions. Much recent work on Logical Form (more particularly,
Hornstein (1984) and Williams (1986, 1988)) converges with the theory presented here. Space
constraints prevent me from comparing these approaches systematically.
^The Government-Binding framework dissociates these two relations, insofar as each of them
obeys distinct constraints: quantifier-variable configurations are subject to constraints on
quantifiers, on variables, and on the relation between the two; movement configurations on the
other hand are subject to locality conditions.
objects
a. (i)
(ii)
b. (i)
(ii)
(2)
a. (i)
(ii)
b. (i)
The contrast between care 'which' and cine, ce 'who, what' is not an isolated fact but
is paradigmatic of Romanian wA-structures in general. With wA-phrases such as cip
'how many', a doubling accusative clitic is optional, but its presence or absence
correlates with an important semantic contrast. With pied-piped wA-phrases the
distribution of doubling clitics depends on the definiteness of the wA-moved
constituent. These observations will be illustrated in later sections.
Note that the obligatory clitic that shows up in (1) and (3) is not a resumptive
pronoun, as indicated by the fact that these wh-structures are sensitive to islands (cf
Steriade (1980)), just like cine structures:8
5
(accus 'whom').
^This pronoun is inflected for gender, number and Case:
whom',
pe cine'pe whom'
(i)
(ii)
(ii)
(4)
(5)
It thus appears that the two types of Romanian w/j-structures both qualify as
movement configurations. Any account of Romanian -structures has to answer
two questions: (a) What is the property that distinguishes the wft-elements that take
obligatory accusative clitics from those that preclude them? and (b) What parameter
is responsible for the presence of this particular type of w/i-structure (the one that
takes obligatory clitics) in Romanian as opposed to its absence in the other
Romance languages and in English?
6. 1. 1.2. Wh-structures and the clitic doubling parameter
Let us briefly review the already existing proposals (Steriade (1980), Aoun (1981),
Borer (1981, 1984)), which deal almost exclusively with question (b) and try to
relate the obligatory presence of accusative clitics in certain w/j-constructions to the
clitic doubling of direct and indirect objects, which is allowed in Romanian9 and
Spanish,10 as opposed to French and Italian:
Chomsky (1986)), and subjacency violations do not result in complete ungrammatically. The
well-formedness of (i-ii) thus cannot be advanced as an argument against a movement analysis of
Romanian wh-structures. Cinque (1984b, 1990) and Obenauer (1984-1985) have shown that
unlike standard wh-structures, the ones that violate islands present not a "standard" variable, but a
"pronominal variable" (i.e., an empty category that qualifies as pro and is bound by a quantifier).
Rizzi (1990) derives a similar conclusion independently, as a result of his "relativized
minimality" constraint (the distinction is stated differently, in terms of antecedent government vs
binding. The former relation connects elements which are not referentially indexed (in particular
variables and quantifiers), whereas the latter involves elements that bear referential indices).
Besides these general properties of Wh-island Constraint violations, I should note that in
Romanian they might not even count as subjacency violations because this language allows two
wh-elements to appear in S-initial position (see Comorovski (1986), who presents other
phenomena that depend on the possibility of (iii-iv)):
(iii) Cine ce a vzut?
who what has seen
(iv) Care pe care 1-a vzut?
which one pe which onei has seen himj
Given (iii) and (iv) Wh-island Constraint violations can be explained by assuming that wh-traces
are allowed to coexist with lexical wh-elements in Comp.
9A doubling clitic is optional or obligatory, depending on the type of complement (direct/indirect
object; +/-specific), and on certain semantic features of its head noun (+/-human). It is not my
(6)
(7)
a. L-am
vzutpelon.
him-(we) have seen pe John
'We saw John.'
b. /-am
dat crfi biatului.
him-(we) have given books boyj) at
'We gave books to the boy.1
a. *Nous l'avons vu Jean.
we him have seen John
b. *Nous lui avons des livres Jean.
we him have given some books to John
According to "Kayne's Generalization", the contrast between (6) and (7) does not
affect a unitary theory of cliticization. It depends only on the existence or absence of
a preposition able to assign Case to the doubled NP: a lexical NP object may be
doubled by a clitic only if the NP is preceded by a preposition. 11 In the GB
framework, this generalization can be derived from the Case Filter,12 as noted by
Jaeggli (1982). Like all nominal expressions, clitics need Case; they absorb the
Case assigned to the argument position. Hence, this Case is no longer available, and
a lexical NP in the direct object position is ruled out by the Case Filter.13 The
examples in (6) are grammatical, since the doubled NP is assigned Case by an
independent means: the preposition pe (compare a in Spanish) or morphological
Case (dative Case in Romanian).
There are two main possible types of S-Structure representation that we can
ascribe to clitic-doubling sequences:
aim to account for the entire paradigm of clitic doubling in Romanian (for a detailed presentation
of the data see Farkas (1978), Steriade (1980) and Dobrovie-Sorin (1987)).
l^For the clitic-doubling phenomenon in Spanish, see Jaeggli (1982), Hurtado (1984), Sufler
(1988). The clitic doubling of indirect objects is correct in both standard Spanish and River Plate
(R.P.) Spanish; the clitic doubling of direct objects is possible only in R.P. Spanish.
11
Note that this generalization should be reformulated in order to account for (6)b, where the
doubled NP is not assigned Case by a preposition, but presents morphological inflections for
dative Case. What seems to count is the fact that there exist explicit Case markers: these may be
prepositions or morphological Cases (see also accusatives in Greek, which are morphologically
marked and can be clitic-doubled). Note that Kayne's generalization does not explain why French
and Italian indirect objects (marked by or a) cannot be doubled (in formal speech). We can
assume that Romanian and Spanish, unlike French and Italian, allow for the double assignment
of dative Case (to the clitic and to the doubled NP). Stated in these terms, this suggestion merely
describes the data. A real explanation of this phenomenon should relate this difference to other
contrasts between the two groups of languages.
12
Following Vergnaud (1982), Chomsky (1981, 49) assumes that the Case filter rules out any
NP that has phonetic content and no Case.
^Following Aoun, I note that clitic-doubling chains show that the Case filter should be kept
distinct from the th-criterion, contrary to Chomsky's hypothesis that the Case filter is just a
visibility condition for th-role assignment. One Case position is sufficient for a th-chain to be
visible for th-assignment, but one Case position is not sufficient in (clj, NPj) chains, which
contain two nominal elements ; each of them must be marked for Case. French impersonal
passives constitute another case in point: their subject needs Case, although it is not part of a thchain (cf Pollock (1981)).
Borer (1981; 1984) assumes the schematic configuration given in (8)a, where
(Prep) NP occupies the -position of the direct object. Jaeggli (1986) provides
further evidence in favour of (8)a, which I will also adopt here. The S-Structure
representation given in (8)b-adopted by Aoun (1981) for River Plate Spanish. 14 is
based on the idea that in certain languages a clitic-doubled direct object occupies an
A'-position, the -position being taken by the clitic trace.
The question now is, What is the relation, if any, between clitic doubling and the
presence of clitics in certain w/i-structures? Both Borer and Aoun assume that the
representation of wA-structures is related to that of nonmovement structures; hence
(9)a and (9)b, corresponding to (8)a and (8)b, respectively: 15
(9)
a. w h j d j e j
b. whjcljejej
a.
*Pecinej
li-ai
vzutej?
Pecinei
ai
vzutei?
In (1 l)b the empty category in the direct object position is a variable (it occupies an
-position and is marked for Case, since no clitic is there to absorb Case), and it is
correctly bound by cine.
The foregoing analysis of the distribution of clitics in ce/cine structures can be
summed up by the following empirical generalization, which derives without
stipulation from the general principles of the grammar: wh-variables cannot be
doubled by accusative clitics.16 Crucially, languages that admit clitic doubling are
nonetheless subject to this generalization.17
6. 1. 2. Romanian
wh-structures
quantification
Given this analysis of cine structures, the obligatory presence of clitics in care
structures forces us to the conclusion that care structures (and the other Romanian
w/j-structures that take obligatory accusative clitics) do not rely on quantification.
Independent evidence supporting this hypothesis comes from the facts concerning
weak crossover and parasitic gaps.
Consider the following weak crossover constructions:
(12) a.
ey?18
16
This generalization can be maintained even if one assumes Aoun's representation given in (9)b,
which we put aside on theoretical grounds: no element in (9)b satisfies the well-formedness
conditions on variables (the clitic trace has no Case and the second empty category occupies an
A'-position).
17
This is essentially Jaeggli's (1982) generalization, based on data from River Plate Spanish:
(i)
el hombre que/a quien vi
the man that / a whom (I) saw
(ii) *el hombre a quienj loj vi
the man a whom, (I) saw himi
18
An LI reviewer observes that (12)a improves when the clitic i-'him' is supplied (note however
that this doubling, marginally possible in colloquial speech, is not acceptable in standard
Romanian), but then the interpretation becomes identical to that of care questions: 'Which one did
you meet?', rather than "Whom did you meet?'. This remark does not invalidate our hypothesis,
(13)
(15)
Unlike the examples in (14), care structures (just like the pied-piping examples in
(15)c-d, which take obligatory accusative clitics) do not license parasitic gaps. This
according to which clitics are excluded from quantifier-variable configurations. When doubled by
a clitic, cine is necessarily reanalyzed as care (that is, as a non-quantifier), which accounts both for
the change in interpretation and for the lack of weak crossover effects.
19
An LI reviewer points out that strong crossover violations appear with both cine and care
structures. This does not constitute a problem for the proposed analysis, because "weak crossover"
violations are crucially different from "strong crossover" violations: the first are due to wellformedness constraints on variables, whereas the second can be analyzed as Principle C violations
on reconstructed representations (see Dobrovie-Sorin (1987)).
indicates that these examples contain no variable (since parasitic gaps are generally
licensed by a variable).20
To sum up, the following cluster of properties distinguish between two types of
Romanian wA-structures: (a) obligatory absence versus obligatory presence of
"doubling" accusative clitics; (b) weak crossover effects versus lack of weak
crossover effects; (c) licensing versus preclusion of parasitic gaps. Weak crossover
and parasitic gaps being diagnostic properties of quantification structures, the
inescapable conclusion is that some Romanian wA-structures do not rely on
quantification, in contrast to cine structures. This property converges with the
distribution of accusative clitics.
6. 1. 3. The quantifier status of wh-phrases
Our next task will be to understand this contrast. We will see that it derives from a
difference in the quantifier nature of the wA-phrase: some Romanian wA-phrases,
but not all of them, function as syntactic quantifiers.
Each of these notions can indeed be defined separately:
(16)
205
inherit them from a lexical category, the wA-quantifier,22 which is generated in one
of two possible positions: under (Spec,N') or under NP. We must therefore assume
a mechanism of transmission of qu-features between lexical quantifiers and their
maximal projections. In what follows we will see that two factors are responsible
for this transmission: (a) the position in which the quantifier is generated and (b)
the lexical properties of the quantifier.
The contrast between the two types of Romanian wA-structures is in certain
cases due to the different positions in which wA-elements are generated. This can be
illustrated by the difference between care and cine.
Cine (just like who in English and qui in French) is a pro-NP element; that is, it
is exhaustively dominated by the NP node. This structural characteristic of cine has
important bearings on the quantifier status of the NP node that dominates cine: the
quantifier features that characterize the lexical element cine are attached to the NP
node that exhaustively dominates it; if in addition this NP occupies a "scope"
position (that is, the position of XPi in (16)c), it counts as a syntactic quantifier, and
as such it will have to bind a variable (by the condition against vacuous
quantification):
(17)
Pecinejaivzutei?
a. Pecare
biatl-ai
vzutf
1-ai
vzutl
22
Given the structural position of care, its quantifier features are attached to the
(Spec.N') node. The NP node that dominates care could acquire quantifier features
only if these features percolated from (Spec,N') up to NP. The obligatory presence
of an accusative clitic indicates that care N' does not function as a syntactic
quantifier, which in turn indicates that care does not transfer its features to its
maximal projection. In other words, care is a "restricted quantifier" (the restriction
is defined by N'); its domain of quantification is limited by the NP to which it
belongs. Care ranges not over the class of elements that can fill the A-position
characteristic of direct objects, but over the class of elements that satisfy the
referential properties defined by the lexical properties of N' (or of the antecedent of
[N1 e]): the class of boys in (18)a. According to this hypothesis, the LF
representations of (18)a-b are not of the type shown in (17'), but rather of the type
shown in (18'); in (18')b e is identified by its antecedent (the class of boys,
students, and so on):
(18') a [NPi for which , is a boy] yousawhimj
b [ for which , is e] you saw himi
The difference between cine and care is comparable to Cinque's (1986) distinction
between dislocated bare quantifiers, which necessarily bind a variable, and
"nonbare" quantifiers (or "quantified NPs"), which do not enter a quantifiervariable configuration. Note however that the dichotomy between bare and nonbare
quantifiers does not necessarily correlate with a contrast in quantifier status. It is
true that bare quantifiers are necessarily syntactic quantifiers, because their inherent
qu-features attach to the NP node itself; they present no restriction (that is, no N'
sister node), and therefore their domain of quantification cannot be restricted to the
NP to which they belong. It is however not the case that a nonbare quantifier
always functions as a syntactic "nonquantifier": the presence of a lexical item under
the N' node makes it possible for the quantifier under (Spec,N') to restrict its
domain to the NP, but it does not impose this restriction; nonbare quantifiers can be
restricted, but they are not necessarily restricted. In other words, NPs of the form
[NPtSpec.N' Q] N'] may or may not function as syntactic quantifiers, depending on
the percolation of qu-features: the wA-quantifiers generated under (Spec,N') are
lexically marked for a particular choice among the following three possibilities: (a)
the wA-quantifier necessarily transfers its wA-features to the dominating NP node,
which will thereby function as a quantifier, and the empty category it binds in S will
be a variable; (b) the wA-quantifier does not transfer its wA-features, with the result
that the dominating NP node does not function as a quantifier and cannot bind a
variable;24 (c) the wA-quantifier optionally transfers its wA-features, with the result
that the dominating NP node optionally functions as a quantifier and optionally
binds a variable.
These three possibilities all exist in Romanian: qu-feature transmission from
(Spec,N') to NP is obligatory, impossible, or optional, depending on the lexical
properties of the quantifiers themselves (feature transmission is also subject to
quantifiers referred to in points (a) and (b) are known as "weak" and "strong", respectively
(cf Milsark (1977) and Barwise and Cooper (1981)). According to the terminology used by these
authors, the quantifiers referred to in (c) are weak; I will show that in fact they may take either a
weak or a strong reading (see also Section 3 below).
207
locality conditions; see Section 6.1.3.2). The choice formulated in point (b)
characterizes care structures, as opposed to ce-structures:
(19) a. Ce elev ai putea tu suporta?
what student could you stand
b. *Ce elev 1-ai putea tu suporti
(19)a is characterized by the presence of a syntactic variable (the empty category in
the direct object position is marked for Case), which must be bound by a quantifier,
since (19)a is grammatical, we must assume that ce elev 'what student1 counts as a
syntactic quantifier. The ungrammaticality of (19)b can be attributed to vacuous
quantification: ce elev has no variable to bind (in the presence of the clitic no licit
variable is available in S). The ungrammaticality of (19)b thus indicates that ce
'what', a nonbare quantifier, is characterized by the option (a): ce necessarily
transfers its qu-features to the dominating NP. This contrast between two nonbare
quantifiers, care (N') and ce (N'), shows that the contrast between syntactic and
nonsyntactic quantifiers is not only structural (bare versus nonbare) but also lexical.
The contrast in the distribution of clitics correlates with a systematic semantic
difference between ce TV'and care N': care structures can be used only if a certain
set of students has already been mentioned or is implicit in a given dialogue; cestructures suppose no such shared knowledge between the two speakers. This
semantic difference can be derived from the difference between the LF
representations given in (18') and (19'):
(19') for what x , x a student, you could stand
According to (18'), care does not bind a variable in S but has a quantifying domain
restricted to the class of elements defined by the nominal head. In other words, the
quantifying domain is independent of the sentence in which the quantified NP is
used. On the other hand, the representation given in (19') indicates that ce does bind
a variable in S: the quantifying domain of ce (Nf) is defined both by the lexical
properties of N' (or its antecedent) and by the properties of the rest of the
sentence.2^
difference in LF representation is associated with a well-known difference in pragmatic
meaning: in (18), but not in (19), the set of elements defined by the nominal head is
pragmatically interpreted as constituting "shared knowledge" among the dialogue protagonists
(see Pesetsky's (1987) distinction between "discourse-linked" and "discourse-free" wh-elements,
which is comparable to Kripke's (1977) distinction between presence and absence of "pragmatic
reference"). Like Pesetsky, and unlike Kripke, I believe that this distinction is semantic, and not
pragmatic. The difference in pragmatic meaning (discourse "linking" or "freedom") can indeed be
derived as a consequence of the difference between the two types of LF configurations given in
(18') and (19'), which are themselves due to the different semantic properties of care and ce
respectively. The LF representation given in (19') does not allow for any set of elements to be
defined independently of the sentence in which the quantifier is used. In (18'), on the other hand,
the quantifier ranges over a set of boys, independently of the rest of the sentence; since it does not
depend on the rest of the sentence, this domain of quantification can be restricted -must be, as a
matter of fact, because natural languages tend to restrict domains of quantification as much as
possible -only by the pragmatic context (the information that constitutes "shared knowledge"
between speaker and hearer). To sum up, certain "restricted" (or "strong") wh-elements (the which
Let us now turn to the third type of w/i-quantifiers mentioned in point (c), which
is also instantiated in Romanian:
(20)
The clitic is optionally present in (20), which indicates that cip' 'how many'
optionally transfers its qu-features to the dominating NP: a quantification
configuration underlies (20)a, but not (20)b. The difference in quantification
properties that we assume between (20)a and (20)b correlates again with a semantic
contrast: (20)a asks a question about the number of individuals who are students
(no particular set of students is presupposed) and who have been examined by the
addressee; on the other hand, (20)b is a question concerning a subset of students,
included in a larger set, which constitutes shared knowledge between speaker and
hearer.
6. 1. 3. 2. The locality of feature percolation
Consider next the following contrasts:
(21)
(22)
The examples in (21)a and (22)a are grammatical: [NP ce elev] and [NP ci{i elevi]
are quantifiers (by inheritance of qu-features) that correctly bind variables in the
direct object position. The ungrammaticality of (21)b and (22)b indicates that pe ce
elev and pe cip elevi do not correctly identify the variable; in other words, they do
not count as quantifiers. Their nonquantifier status is obviously not due to their
lexical properties, since (21)a and (22)a are correct, but may be triggered by the
presence of the preposition pe. It is then tempting to treat (21)b and (22)b as
violations of a locality constraint on feature percolation: [pp pe [NP ce elev]] is not
a licit quantifier, because pe blocks the percolation of qu-features from ce up to the
dominating PP. More precisely, the dummy preposition pe does not L-mark NP,
which therefore functions as a barrier (see Chomsky (1986)), which blocks the
percolation of qu-features.
Note that pe does not block the transmission of qu-features in (23):
type) are pragmatically interpreted as being "discourse-linked", whereas "unrestricted" (or "weak")
wh-elements (the who type) are pragmatically interpreted as being "discourse-free".
. (i)
head them.27 In Sections 6.2 and 6.3 I will show that the principles proposed here
also account for certain constraints to which the clitic doubling of quantified NPs in
situ is subject.
6. 1. 4. Wh-structures and quantification: lexical wh-phrases versus structural
quantifiers
6. 1.4. 1. Equivalents of care in other languages
Let us assume that the principles of the quantification theory proposed above do not
constitute a parametrized option characteristic of Romanian but instead belong to the
general principles of language. 28 We would then expect systematic syntactic
27
According to Steriade's (1980) account, Romanian wh-traces are replaced by pronominal copies
(similar to Perlmutter's (1972) "shadow pronouns") that inherit the [+/-specific] features of the
moved constituent. Granting this assumption, the distribution of clitics in wh-structures is
predictable entirely from the conditions under which pronominal objects may or must reduplicate;
more precisely, the contrast between the two types of wh-structures comes down to the contrast
between specific vs nonspecific pronouns, which respectively require and exclude clitic doubling.
Within Steriade's approach, the correlation between specificity and clitic doubling is stated as a
primitive, and left unexplained. It is my intent to explain it, or rather the more general correlation
between the distribution of clitics and quantification structures:
(i)
a. Nu (*l)-am vSzutpe nimeni.
not (*him) [I] have seen pe nobody
b. Ion (*l)-a indlnit pe cineva.
John (*him)-has met pe somebody
The exclusion of the clitic in (i) is parallel to the exclusion of clitics in cine structures and the
other wh-structures that behave in the same way. This parallelism between wh-structures and the
in situ cases of quantification does not oblige us to derive one paradigm from the other. It is
instead possible to show that both configurations are governed by the same principles of
quantification theory (see Section 3).
Within Steriade's account the parallelism between the clitics that "double" wh-traces and the
clitics that double pronominal NPs in situ is stipulated in order to account for the fact that in whstructures the clitic is obligatory (as it is for pronominal NPs in -positions (see the examples in
(ii)-(iii)), whereas the doubling of specific NPs other than pronominal is optional (see (iv-v)):
(ii) Ion 1-a examinat pe el.
John him-has examined pe him"
(iii) *Ion a examinat pe el.
John has examined pe him
(iv) Ion 1-a examinat pe vecin.
John him-has examined pe neighbour"
(v) Ion a examinat vecinul.
John has examined the neighbour".
The pronominal versus nonpronominal contrast illustrated in (iii) and (v) is irrelevant within our
approach. We can nonetheless explain the difference between the obligatory presence of clitics in
care structures (cf (1)) versus their optionality with specific NPs in situ (cf (iv)-(v)), by bringing
into play the different positions (A'-versus -position) occupied by the doubled NP (see Section
6.3.3).
28
According to a quite natural, generally accepted assumption, semantic properties are universal.
That is why, on conceptual grounds, we do not want to stipulate semantic differences between the
(27) a.
b.
c.
d.
Romanian care (and the other wh-elements that take obligatory clitics) and its equivalents in other
languages (at least in those languages where pairs such as who, what (N1) versus which (N') exist).
29i assume the internal structure of the Comp node proposed in Chomsky (1986); S' is the
maximal projection of Comp (hence labeled CP (Complementizer Phrase)), which is rewritten as
in (i); S, taken to be the maximal projection of Infi, is symbolized by IP:
(0
CP = [Spec.C'lC'C IP]]
The head position C of CP hosts complementizers like that and que; the (Spec,C) slot is filled
by wh-phrases.
3 0 A s A. Belletti and L. Rizzi note (personal communication), any principle like (28) should be
adequately qualified in order not to extend to verb-second languages, where it is clearly not the
case that any NP in (Spec,C') acquires a quantifier status. We might therefore be forced to restrict
This definition is structural, since it is the position of an element, and not its
intrinsic features, that defines its quantifier status. The lexical difference between
which and who can be detected when these elements stay in situ, because in this
case their lexical features are alone responsible for their quantifier status; when they
are in (Spec,C'), the difference between which and who is neutralized by virtue of
(28), which attributes a quantifier status to whatever element appears in (Spec,C').
(28) accounts for the well-known discrepancy between S-Structure and LF
representations that characterizes pied-piped structures:
(29)
(28) to wh-phrases and empty categories (see also footnote 33, where a reformualtion of (28) is
suggested).
31
In (i) the pied-piped constituent is indefinite and therefore may function as a quantifier (see
Sections 2 and 3 below). This accounts for the absence of the clitic:
(i)
Ion, strmofi ai cruia incercasem in zadar s gsesc
John, ancestors of whom (I) had tried in vain to find
3
^More interestingly, one might try to render the behaviour of Romanian compatible with (28)
by assuming that in this language wh-phrases do not occupy the (Spec.C') position (the slot
characteristic of left dislocations might be invoked instead). This hypothesis is supported by
evidence concerning the constituent structure of S and S' in Romanian (see Dobrovie-Sorin
(1987)).
33
An LI reviewer suggests an alternative formulation of the parameter (28)/(31), stated directly in
terms of "null operators": Romanian would lack them, as opposed to the other Romance
languages. On this hypothesis, the apparent quantifier behaviour of which, quel, cul, quale
would follow from their ability to coexist with a null operator in Comp, an ability that their
Romanian counterpart does not share. As far as the data analyzed here are concerned, the two
alternative parameters seem to be equivalent.
^Comparatives are not relevant in the present connection, because they do not involve the
movement of direct objects, which alone concerns us here. Besides, Romanian comparatives
involve not a null operator, but a lexical one (which as such normally binds variables), decit
'than' (lit. 'of-how much').
35
On the distribution of ca and s, see Dobrovie-Sorin (1987).
36
Note that the Romanian counterparts of that relatives do not become grammatical even if
supplied with an accusative clitic. This impossibility is not completely clear to me. A way is
needed to rule out two distinct possibilities: (a) a resumptive pronoun strategy and (b) a
"movement" strategy comparable to care configurations. A promising line of inquiry may work
out the hypothesis that the Comp node is not involved in Romanian relatives, which could be
derived from the particular sentence structure of Romanian.
^According to the derivational analysis, that relatives are obtained by the deletion of the whelement in Comp; the empty category resulting from deletion is assigned a quantifier status at
LF, which ensures the correct binding of the variable in S. This analysis can be restated in
representational terms. According to the X' theory proposed in Chomsky (1986), two slots are
available in Comp, so that an empty category can be generated under (Spec.C), even if the head
for the old man whati (you) look at ei is not and ordinary man
'For the old man you are looking at is not a ordinary person.'
This type of relative, headed by ce 'what', is a marked construction, which does not
belong to the core grammar of contemporary Romanian: it is little productive (the
example under (32) is a literary quotation; ce relatives are somewhat more frequent
with non-human antecedents), and presents a number of idiosyncratic properties.38
The absence of an accusative clitic 39 indicates that the ce-relative in (32) is a
quantifier-variable structure. Two different hypotheses are compatible with this
idea: (a) ce is a lexical quantifier, as suggest! by its lexical identity; (b) ce is a
complementizer of the that-type (cf. Horvath and Grosu (1987) and Horvath and
Grosu (1987)). I will not try to choose between these two possibilities; in terms of
either of them ce-relatives are analysed as quantifier-variable configurations, and
this explains why they do not take obligatory accusative clitics. Ce relatives would
thus appear to fit well into the proposed analysis of the distribution of clitics in whstructures. However, if they were to be analysed as thai-relatives, we would have to
admit that null operators exist in peripheral areas of Romanian.
The lack of clefts 4 0 is another outstanding characteristic of Romanian, which
constitutes independent evidence in favour of the proposed analysis of care
position C is filled with lexical complementizers, as in i/iai-relatives. By (28) the empty category
generated in (Spec,') acquires quantifier status.
38
Ce-relatives take optional clitics in "short wh-movement" configurations (see (i)), and
obligatory clitics in "long wh-movement" (see (ii)-(iii)) and "parasitic gap" configurations (see
(iv)). (examples (i)-(iii) are from Sandfeld and Olsen (1936,112).)
(i)
cartea ce ai citit-(o) ...
the book what [you] have read (it)
(ii) ui inel ce zicea c *(il) are dela mo$i
a ring what (he) said that (he) has *(it) from elders
(iii) nifte nofe ce am Infeles dela d-1 Ionescu c *(le) posedaji
some notes what 0) understood from Mr Ionescu that (you) possess *(them)
(iv) mogneagul ce privegti fr a-*(l) necunoafle
the old man what (you) look at without to recognizing *(him)
The ungrammaticality of (ii-iv) cannot be due to any constraint on wh-movement in general,
because the corresponding interrogative sentences arc correct, and the clitic is necessarily absent:
(v) Ce zicea c are delamo$i?
what (did he) say that (he) had from (his) elders?
We must then admit that the ungrammaticality of (ii-iv) is idiosyncratic; that is, it is not
derivable from the general principles of language (or from any parametrized option characteristic
of Romanian).
39
Quite clearly, the accusative clitics that may show up in ce-relatives (see footnote 38) should be
analyzed not as the obligatory clitics characteristic of care structures but rather as resumptives.
Resumptives are indeed found in two different types of contexts: (a) they may optionally appear
instead of a variable; (b) they are used in contexts in which variables are illicit (for instance, by
virtue of the island constraints). The first case is illustrated by (i) of footnote 38, and the second
by (ii-iv), in which a constraint other than islandhood must be invoked.
4^The meaning of clefts can be expressed in Romanian by "pseudoclefts":
*EcuIoncamplecat
43
The contrast between (36) and (l)a is a true minimal pair, unlike the contrast between (l)a and
(i)-(,ii), which is frequently, but misleadingly quoted:
(i)
el hombre que/a quien vi
the man that / a whom (I) saw'
(ii) * el hombre a quien lo vi
the man a whom (I) saw him'
Examples (i)-(ii) are that-lype relatives (or iv/io-type relatives; compare a quien, which is
marginally acceptable in (i), instead of que), relying on a quantifier-variable configuration. As
discussed in the preceding sections, the paradigm (i)-(ii) is as a matter of fact parallel to
Romanian interrogatives in cine/ ce (the ungrammaticality of (ii) and (iii) is due to vacuous
quantification):
(iii) Pe cine (*l)-ai indlnit?
pe whom did you meet (*him)
^Sufier (1988) gives the following grammatical examples, in which interrogative partitives are
clitic-doubled:
(i)
cul de las dos candidatasj laj entrevistaron?
which of the two candidatesj (did they) interview herj?'
(ii) cules de ellosj losj interrogaron?
which of them; (did they) question themj?'
(iii) A cuantas de las actricesj lasj reconocieron?
how many of the actresses, (did they) recognize themj?'
Sufler considers that the contrast between the ungrammaticality of (36) and the acceptability of (iiii) is due to the fact that partitives are +specific, whereas NPs of the wh N' form are -specific. It
is however difficult to assume intrinsic nonspecificity for cul '. According to our analysis, the
acceptability of (i)-(iii) is due to the fact that the "structural quantifier" parameter does not
necessarily operate for partitives, whereas it does for wh N's independently of their intrinsic
quantification features. One might suggest that, unlike other wh-phrases, Spanish partitives are
allowed to stay out of the (Spec, C') position (the examples under (i)-(iii) would then be
assimilated to left dislocations,; see Section 2).
According to our account, the contrast between Romanian and Spanish derives from the
mechanism that underlies wh-structures. On this point our analysis is close to Steriade's (1980)
pre-GB proposal, stated as a stipulation: Unlike Romanian, River Plate Spanish wh-structures
involve, not pronominal copies of the extracted constituent, but just empty variables. If our
analysis is correct, this stipulated difference between the Romanian and Spanish types of whmovement depends on two different values of the structural quantifier parameter, which is
independently motivated (see the discussion of null operator constructions).
Steriade's correct intuition got lost in both Aoun's (1981; 1985) and Borer's (1981; 1984)
analyses of the contrast between Romanian and Spanish. Borer considers that the presence of
clitics in wh-structures is the unmarked situation for a clitic-doubling language (see Section 1.1.2
above): wh-variables would be properly governed by doubling clitics. Romanian is taken to
illustrate this generalization, but the River Plate Spanish data constitute a problem. Borer solves
it by assuming that clitics function as proper governers only if they bear the same Case as the
empty category they are supposed to govern. This requirement would be satisfied in Romanian,
but not in Spanish, the relevant parameter being a difference in the Cases assigned to
"prepositional accusatives": Romanian pe would assign the accusative, whereas Spanish a could
only assign the dative (besides being an accusative marker, a is indeed the preposition
characteristic of indirect objects). Spanish clitic-doubling sequences would then be characterized
by a Case discrepancy between the accusative clitic and the doubled NP, which would be dative.
Because of the Case harmony requirement imposed on clitic government, an accusative clitic
would not be able to properly govern the (supposedly dative) variable left behind by the whmovement of the a NP, hence the exclusion of accusative clitics in Spanish wh-structures. Jaeggli
(1986) shows that direct objects preceded by a are not dative, which invalidates most of Borer's
explanation.
According to Aoun the contrast between Romanian and River Plate Spanish would depend
on two different choices from among four possible types of clitic pronouns. Romanian would
have at its disposal accusative non-R clitics (clitics that may double an NP which occupies the Aposition of the direct object); River Plate Spanish, on the other hand, would have at its disposal
only R-clitics (clitics that necessarily bind an empty category in the direct object position, the
doubled NP occupying an A'-position). This difference would account for the contrast in whstructures: in Romanian wh-movement is allowed (because the resulting variable occupies an Aposition), but it is excluded in Spanish, because the variable would illictily occupy an A'position. Aoun's hypothesis is stipulative: the difference between accusative clitics in Romanian
versus Spanish cannot be shown to be relevant in other parts of the grammar. Moreover, Aoun's
hypothesis predicts that, because of their nonreferential status, accusative clitics should be
possible in cine structures, contrary to fact.
4
^Move NP is just a particular case of Move XP, where XP is a maximal projection. I prefer to
restrict attention to Move NP because I am dealing here with the wh-movement of direct object
NPs.
219
trecut
pe John not-(I) have met the girl which saw him last year
b. *Pe Ion am plecatinainte
s- examineze
Popescu.
47
Cinque discusses other differences, among them the following: (a) the left-dislocated element
of CLLDs can be of any maximal category (in the sense of X' theory), whereas ELDs essentially
allow for left-dislocated NPs only; (b) there is no theoretical limit to the number of left-dislocated
constituents in CLLDs, whereas ELDs do not allow more than one left-dislocated constituent; (c)
in CLLDs the S-intemal element can only be a clitic, whereas in ELDs an emphatic pronoun can
also appear; (d) there is obligatory "connectivity" (Cinque's connectivity resembles
"reconstruction": the dislocated element behaves as if it occupied the -position with which it is
coindexed) between the left-dislocated constituent and the S-internal position in CLLDs, but not
in ELDs. Let us just illustrate the generalization given in (c):
(i)
*Pe Maria nu vreau s-o mai vd peeatit triesc.
pe Mary (I) not want to-her any more see pe her how long (I) live
(ii) 1(Cit despre) Maria, nu vreau s-o maitvdpe ea tit triesc.
(As for) Mary, (I) not want to-her any more see pe her how long (I) live
'As for Mary, I don't want to see her any more in my whole life.'
Note that care structures behave like CLLDs:
(iii) *fata pe care nu vreau s-o mai vdpe ea tit triesc
the girl pe which (I) not want to-her any more see pe her how long as (I) live
Whereas the presence of the emphatic pronoun pe ea results in ungrammaticaility in both CLLDs
(see (i)) and care structures (see (iii)), the ELD illustrated in (ii) is marginal, but nevertheless
possible. The impossibility of emphatic pronouns in CLLDs may be considered an argument in
favour of the movement hypothesis: the input structure of examples like (i) and (iii) is
ungrammatical; Romanian does not admit doubling sequences that include a clitic, an emphatic
pronoun, and a lexical NP. The (relative) acceptability of emphatic pronouns in examples like (ii)
can be accounted for by adopting the standard analysis of ELDs: in contrast to CLLDs, they do
not rely on movement, or on any other syntactic relation (at S-Structure the left-dislocated
element is not coindexed with any element in S); the clitic is in this case a resumptive pronoun..
^ C o m p a r e ( 3 9 ) with (i)-(iii), which show very mild (if any) island violations:
(i)
(Cit despre) Ion, -am intilnit fata care 1-a vzut ultima dat.
(as to) John, not-(I) have met the girl which him-has seen the last time
(ii) (Cit despre) Ion, am plecat inainte s-1 examineze Popescu.
(as to) John, (I) have left before that-him examine Popescu
(iii) (Cit despre) Ion, s-1 ajufi e plcere.
(as to) John, to-him help is a pleasure.
49
Given our hypothesis concerning the absence of the null operator strategy in Romanian, we
correctly expect that Romanian lacks Topicalization (more concretely, the example in (37) is
ungrammatical without a clitic). We should note that the English-type topicalization is
object are obligatory (see (37)); (b) parasitic gaps are not licensed; (c) no weak
crossover effects can be observed (for illustration, see Cinque (1990) and
Dobrovie-Sorin (1987)). The CLLD of referential expressions thus appears to rely
on the same syntactic mechanism (A'-movement without quantification) as care
structures.
6. 2. 2. The CLLD ofquantiedNPs (QNPs)
Recall that the quantificational status of Romanian
structures depends on the
percolation of qu-features from wA-elements to the wft-phrases that dominate them.
An interesting generalization can be reached if we are able to show that the
proposed typology extends to quantifiers other than w/j-elements.
The examples in (40), which illustrate the CLLD of "bare" quantifiers such as
nimeni 'nobody', cineva 'somebody' and ceva 'something', are parallel to the whstructures that involve "bare" wft-quantifiers such as cine 'who', ce 'what' (see the
examples in (2)):
(40) a. Pe nimeni -am suprat
pe nobody not-(I) have annoyed
b. Ceva ai s descopen i tu.
something (you) will discover you too
c. Pe cineva trebuie s superi tu in ecare zi.
pe somebody (you) must upset every day
Doubling clitics are excluded:
(41) a. *Pe nimeni nu 1-am suprat
b. * Ceva ai s-1 descopen i tu.
c. *Pe cineva trebuie s-1 supen tu in ecare zi.
unacceptable in other Romance languages. In Italian, for instance, only focused NPs are subject to
topicalization (see Cinque (1984a; 1990)):
(i)
Gianni inviterd (non Piero).
John, I'll invite [not Peter]
If Gianni is not stressed, an accusative clitic is obligatory in Italian:
(ii) Gianni, lo inviterd domani [non oggi],
John, I'll invite him tomorrow [not today].
Two questions arise, which will not be answered here: how can we explain the restriction
concerning focus in Italian, and what is the parameter that distinguishes Italian from English?
Crucially, Romanian lacks even the weaker variety of topicalization that characterizes Italian.
Focused direct objects are allowed in S-initial positions, but an accusative clitic is still
obligatory:
(iii) Eu pe Popescu 1-am vSzut (nu pe Ionescu).
me pe Popescu (I) him-saw (not pe Ionescu)
(iv) Eu romanul sta 1-am citit (nu pe cellalt).
me, this novel it (I) read (not pe the other one)'
I will not propose an analysis for this structure. What is relevant here is simply the fact that (iii) and
(iv) do not involve quantification. This is also true of focused NPs in situ (they are not subject to
QR).
221
223
52xhese parallelisms do not mean that w/i-stnictures and CLLDs are completely identical; a
number of differences are expected, related to the quantification properties attached to the
interrogative illocutionary force. Although the data are not completely clear to me, it seems
difficult to find non- wh QNPs corresponding to -phrases such as ce elev 'what student', which
were shown to function necessarily as quantifiers (see Section 1.3); a thorough examination of
compound forms of ce 'what' such as orice (elev) 'whatever (student)', fiece (elev) 'any student'
would be necessary.
a. Caut secretar.
(I) look for a secretary
b. caut pe secretar.
her (I) look (for) pe a secretary
(49) a. Am vzut secretar.
(I) have seen a secretary
b. Am vzut-o pe secretar.
(I) have seen-her pe a secretary
Example (48)a displays a well-known ambiguity, which is characteristic of
indefinite NPs not only in Romanian but quite generally, in French, English, and
the like: under the specific reading of the indefinite NP the speaker is looking for a
certain secretary (in this case secretar could be replaced by a definite expression,
or a proper name); under the nonspecific reading the speaker is looking for any
person who is qualified as a secretary. Example (48)b, where the indefinite NP is
clitic-doubled, selects the specific interpretation.
I believe that the specific versus nonspecific ambiguity is characteristic of
indefinites (and of weak NPs; see below) in general, regardless of the context in
which they appear. In particular, it can be observed in "nonopaque" (or
"nonmodal", in Jackendoffs (1972) terms) contexts, as in (49)). The specific
reading of (49)a-b is essentially the same as the specific reading of (48)a-b. Under
the nonspecific interpretation,55 (49)a can be understood to mean 'the number of
53
This generalization also holds for River Plate Spanish. This constitutes an important argument
against Aoun's (1981) parametrized theory of cliticization, since it shows that Romanian
accusative clitics are R-clitics, just like River Plate Spanish clitics, in contradiction to the non-R
nature that Aoun assumes for Romanian accusative clitics in vWi-structures (see footnote 45
above).
54
S o m e Romanian speakers reject (48)b as belonging to the "popular style"; they feel it is
incorrect from the point of view of normative grammar, although Romanian grammars do not
explicitly mention this type of example. The same remark holds for the other (clitic-doubled)
QNPs preceded by pe. For these Romanian speakers QNPs behave like bare quantifiers: the
exclusion of pe, correlated with the exclusion of doubling clitics, is directly accounted for by
assuming obligatory QR. For other speakers, including myself, such examples are grammatical,
and no significant difference in style is perceived. This second category of speakers all agree that
the only possible reading is the specific one.
55
The nonspecific reading of (49)a is nonetheless clearly distinct from that of (48)a: the existence
of a secretary is presupposed in (49), but not in (48). This difference is due to the different type of
verb: in (48) the variable resulting from QR is in the scope of a verb that creates "opacity".
secretaries that I have seen is equal to one 1 , 56 no implicit reference being made to
the identity of the secretary.
The grammaticality of (48)b and (49)b is unexpected on May's (1977) analysis
of the ambiguity characteristic of (48)a. By QR we obtain the LF representation
(48')a, which according to May corresponds to the specific (or "transparent", in
May's own terms) interpretation:57
(48')
a. [$[NPO secretarjlt^cautej]]
a. (i)
(ii)
(iii)
b. (i)
(ii)
(iii)
reference to their identity (or the identity of a group they might belong to). These
interpretive properties, which again raise the above-mentioned problems for May's
analysis, can be correctly accounted for under our assumptions. The nonspecific
reading of (50)a for example derives from the LF representation in (50')a, obtained
by QR, on a par with (48')a. The semantic interpretation associated with this LF
representation defines a clausal set for the quantifier: mulfi 'many' quantifies over
the set of children that have been fooled. In (50)b on the other hand, QR cannot
apply (it would yield the illicit representation (50')b); hence, multf can operate only
on a set of children that is defined independently of the rest of the sentence.61 We
thus have an explanation for why the nonspecific reading is impossible in cliticdoubling configurations.
On the proposed analysis QR must be assumed to apply in contexts where the
NP selects the nonspecific interpretation; examples like (51) are therefore
represented at LF as shown in (51'):
(51) a. Caut un elev care s tie engleze$te.
(I) look for a student which speak su f,j unct j ve English
b. Caut doi elevi care s vorbeasc englezete.
(I) look for two students which speakSubjunctive English
(51') a. [[ u n e ^ e v c a r e ?tie englezejte] [<j caut ej]] 62
We correctly expect that in these contexts clitic doubling is excluded (because (52')
is illicit):
(52) a. *I1 cautpe un elev care s $tie englezete.
b. *Ii caut pe doi elevi care s vorbeasc engleze$te.
(52') a. Q[stNPiPe u n e ' e v c a r e
englezete][ilj caut e j ]
To sum up, the proposed analysis explains how the semantic interpretation can be
constrained by the syntax. We assume the null hypothesis concerning the rules of
grammar: QR applies optionally, resulting in distinct LF representations, which
account for certain semantic ambiguities; if QR is blocked by syntactic constraints
(in particular by clitic doubling), only one LF configuration is available, which
corresponds to the specific reading. The application of QR also depends on the
lexical specifications of the different quantifiers and more precisely on their ability
to transfer qu-features to the NP that dominates them. The application of QR is
indeed subject to the condition stated in (53):
(53) QR applies to an NP that bears gu-features.63
mentioned earlier, this semantic difference triggers a difference in pragmatic meaning.
62It
should be possible to show that QR (and consequently the nonspecific reading) is triggered
by subjunctive relatives, but I will leave this question aside.
63A
different definition of QR may be assumed: QR raises determiners (out of their original
position, be it (Spec,N') or NP), and not NPs (see Dobrovie-Sorin (1988), (1989)). This
restatement is more interesting from the theoretical point of view, and it may turn out to make
different empirical predictions (see Chapter 7). However, the data discussed here do not
discriminate between the two formulations.
(i)
(ii)
65A
complete analysis of the paradigm (54), which cannot be made here, should provide a
theoretical motivation for the requirement on QNPs with empty heads (the necessary
identification of the empty category is probably the relevant principle) and for the fact that clitic
doubling can override it. It would also be interesting to compare Romanian with French and
Italian (see Cinque (1986)), two non-clitic-doubling languages that present paradigms comparable
to (54)a-c.
over the set defined by the head noun). Note that clitic doubling is acceptable in this
case for all Romanian speakers, whereas some of them reject the clitic doubling of
weak NPs (see footnote 54).
6. 3. 3. Scope ambiguities
Let us now examine the following examples:
(55) a. Fiecare profesor va examina zece elevi.
each professor will examine ten students
b. Topjudectorii cunosc doi avocap'.
all the judges know two lawyers
c. Top' copiii
au vzut un actor celebru.
all the children have seen, a well-known actor
These sentences are characterized by a well-known ambiguity, which is generally
assumed to derive from different scope assignments. According to one possible
interpretation, the ten students examined may differ from one teacher to another, in
other words, (55)a says something about a number of students that is greater than
ten; in (55)b each judge does not necessarily know the same two lawyers; in (55)c
each child may have seen a different actor. The representations generally assumed
(since May (1977)) for this type of reading are the ones given in (55'), where the
NPs set in bold characters are assigned narrow scope:
(55') a. [fiecare profesorjj [zece elevi]: [ej va examina ej]
[each teacher]} [ten students]j [ej will examine ej]
b. [top judectorii] j [doi avocap]; [ej cunosc e;]
[all the judges] j [two lawyers]j [ej know ej]
c. [top copiii] j [un actor]j [ej au vzut ej]
[all the children]} [an actor]j [ej saw ej]
The examples in (55) are compatible with a second series of possible readings:
'there are ten students A, B,..., J, such that each teacher will examine A,B,.. J ' (55)a;
'there are two lawyers A and B, such that all the judges know A and B' (55)b; 'there
is a (certain) actor such that all the children saw him' (55)c. According to May,
these interpretations can be represented by the LF configurations given in (55"),
where the NPs set in bold characters have wide scope:
(55") a. [zece elevi]j [fiecare profesor]} [ej va examina ej]
[ten students]j [each teacher]} [ej will examine ej]
b. [doi avocap]j [top judectorii]j [ej cunosc ej]
[two lawyers]j [all judges]} [ej know ej]
c. [un actor]j [top copiii]j [ej au vzut ej]
[an actor]; [all children]} [ej saw e;]
a. Fiecareprofesoriivaexamina
pezeceelevi.
b. Topjudectorii ii cunosc pe doi avocap.
c. Top copiii 1-au vzutpe un actor celebru.
In (57)b the quantifier mulp 'many' is not affected by the negation; the only possible
reading is 'there are many students that I didn't examine', which means that (57)b is
not the opposite of have examined many students'. In (57)a on the other hand, the
quantifier many is affected by the negation and a possible gloss is have examined
few students'.
If we try to account for these readings in terms of different scope assignments at
LF, we are again confronted with an illicit representation, (57")b, that corresponds
to die correct sentence (57)b:
(57") b. 0[pe mulp elevi]j [nu]j [ej ij-am examinat ej]
[pe many students]j [not]j [ej themj (I) have examined ej]
We are then forced to abandon the widely accepted hypothesis according to which
the so-called wide scope interpretation relies on QR at LF. It is instead possible to
show that this construal reduces to the specific reading, which according to our
6 6 A similar paradox emerges if we adopt May's (1985) Scope Principle. Since for present
purposes the two formalisms are equivalent, we will stick to the older theory.
"is in the scope o f ' (= the referential properties of are affected by the
quantificational properties of ) if (a) is a variable and (b) is ccommanded by . 68
In (58) "relative scope" is defined not as a relation between two quantifiers but as a
relation between a variable and an element that c-commands it. The idea that the
"narrow-scope" interpretation is necessarily associated with the presence of a
variable is natural: variables are referentially open expressions,69 and as such they
are accessible to external quantification.
Let us now come back to our examples. According to (58), a narrow-scoped NP
is an NP that is subject to QR (which within the framework adopted here is the
only grammatical mechanism that generates variables). The following type of LF
representation should then be adopted for the narrow-scope interpretation of the
NPs set in bold characters in (55):
(55^) a. [zece elevi]j [fiecare profesoq va examina ej]
[ten students]j [each teacherj will examine ej]
In (55^) the variable ej is c-commanded by the subject NP; therefore, according to
(58), the narrow-scope reading is allowed for the object NP. On this analysis, the
unavailability of the narrow-scope interpretation for clitic-doubled QNPs (see (56))
is straightforward, since LF representations of the type in (56^) are illicit (on a par
with (56")):
(563) a . 0[pe zece elevi]: [fiecare profesorj ii; va examina e:]
67
In the recent literature similar principles have been independently formulated by Aoun and Li
(1989) and Williams (1988), whose proposal was inspired by Hoji (1985) and Huang (1982).
68A
stronger requirement might be necessary (possibly Condition A of the binding theory, see
Sloan (1988)). I leave this aside because the relevant empirical data are outside the scope of this
chapter. I simply suggest that such a stronger condition might be derived from the following
proposition: a variable can be affected by a c-commanding NP if and only if the c-commanding
NP is not excluded by the range of the variable, where the "range" of the variable is the c-domain
of the quantifier raised by QR. This proposition need not be stipulated; it derives directly from
the basic idea that underlies my approach: variables can be affected by external quantifiers just
because they count as open expressions; but they count as open expressions only in a certain
domain, their "range". Outside this domain, they count as closed expressions.
69
This idea becomes even clearer if we assume the formalism suggested in footnote 63 and
developed in chapter 7: if QR applies to (Spec,N'), we obtain open formulae (see also footnote
59), which as such are accessible to external quantification.
0ft>r the relative position of postverbal subjects and direct objects in Romanian, see DobrovieSorin (1987).
7
' N o t e that the meaning of "wide scope" used here is equivalent to "scope over", to be kept
distinct from the meaning "not narrow-scoped" which we discussed with respect to (56).
72 I t
is clear, however, that the quantification properties of wide-scoped NPs should be specified;
plurals and referential QNPs can take wide scope, whereas singular nonquantificational
expressions cannot.
If we assume that clitic-doubled NPs cannot be raised by QR, the variable resulting
from the QR of the subject NPs will not be c-commanded by the object NPs at LF;
therefore the narrow-scope interpretation of the subject NPs in these examples
cannot be accounted for by (58). 73 Two solutions may be considered. The first
would be to say that the narrow-scope interpretation of the subject NPs in (60)a-b
does not pertain to LF. The second would be to assume that besides QR (which
according to (53) applies exclusively to NPs that bear gu-features) we need a
second rule, which raises NPs, even if they do not bear gu-features. This rule may
be viewed as a case of pied-piping; we may also recall Heim's (1982) rule of NP
Raising, which raises all NPs, regardless of their quantification properties. The
resulting LF representations will be distinct: QR gives rise to a quantification
relation, which involves a variable bound to a quantifier, NP Raising does not leave
any variable and the raised NP does not count as a quantifier (this rule is therefore
allowed to raise clitic-doubled NPs, unlike QR), but can be construed as having
scope over a variable that it c-commands.74
Let us now compare (55) and (56) to (43) and (44), whose (a)-sentences are
repeated here:
(43)
73
dislocated QNP. That is why, by virtue of (58), the dislocated object QNP is
construed as taking narrow scope. As for (55), the relevant relation is established at
LF, between the variable in object position and the subject NP (see the LF structure
given in (55^)).
Note however that the parallelism between QNPs in situ and QNPs in A'positions is not perfect: (55) allows the referential or wide-scope interpretation of
the italicized NPs, in contrast to (43). This asymmetry between QNPs in situ and
dislocated QNPs is due to their different S-Structure positions. Nondoubled QNPs
in situ are ambiguous, because in the absence of a doubling clitic QR may, but need
not, apply. Dislocated QNPs on the other hand cannot assume an ambiguous
interpretation: in the absence of a clitic, an empty category in the object position
qualifies as a variable, and this necessarily correlates with the
nonspecific/nonreferential reading and the possibility of narrow scope; conversely,
if a clitic is present, no variable is available, which correlates with the
specific/referential and wide-scope readings.75
6. 3. 4. Prepositional accusatives and specificity
Consider the following examples:
(61)
75
a.
Cautpeo secretar.
(I) look for pe a secretary
b.
*Cautpe un elev care s tie engleze$te.
(I) look for pe a student which know s u ^j u n c t j v e English
c. (i) Am pclit pe multf copii, dar pe tine -am reuiL
(I) have fooled pe many children, but pe you not-(I) have succeeded
(ii) Am dus pe dou franpizoaice la gar.
(I) have taken pe two French women to the station
(iii) Am plimbatpe dou pnetene pnn Gmigiu.
(I) have taken for a walk pe two friends around Cijmigiu
d.
* Cautpe doi elevi care s vorbeasc engleze$te.
(I) look for pe two students which speak s u j,j u n c t j v e ' English
The same asymmetry between the in situ and the adjunction schema of quantification can be
observed in (i)-(ii) versus (iii)-(iv), where the clitic is respectively optional and obligatory; note
that fiecare 'each' and oricare 'any' are compound forms of care:
(i)
Invit-) pe oricare elev.
invite (him) pe anywhich student
'Invite any of the students.'
(ii) Trebuie s-(i) examinm pe fiecare in parte.
[we] must that-(them) [we]examineSubj pe eachwhich individually
'We must examine each of them individually.'
(iii) biatul pe care 1-ai vzut
the boy pe which [you] him-have seen
'the boy you saw'
(iv) * biatul pe care ai vzut
the boy pe which (you) have seen
235
comparable generalization holds for Spanish (see Jaeggli (1982)). More precisely, certain
speakers reject prepositional accusatives (marked by a) with QNPs; other speakers tend to accept
them, but only on the specific interpretation.
77
This account is stated more elegantly if we assume the definition of QR suggested in footnote
63: assuming that QR raises not NPs but quantifiers, the locality conditions would block the
movement of the quantifier itself. The suggested QR rule does have an S-Structure counterpart:
French examples such as j'ai beaucoupj lu ej de livres have a lotj read [ei of books]' would be a
case in point. The proposed analysis of prepositional accusatives then makes the following
interesting prediction for a language that would exhibit both prepositional accusatives and an SStructure QR of the French type: the latter should be allowed with nonprepositional accusatives,
and excluded with prepositional accusatives. This is exactly the case of Bengali, as I have been
informed by P. Dasgupta (personal communication; see Dasgupta (1988)).
78
Our results come close to Fodor and Sag's (1982) conclusions, according to which referential
expressions "do not have any scope" (in other words, they do not enter quantifier-variable
configurations). One of Fodor and Sag's arguments is based on the following examples:
(i)
John overheard the rumor that each of my students had been called before the dean.
(ii) John overheard the rumor that a student of mine had been called before the dean.
Example (i) shows that QR obeys the Complex NP Constraint: although each is normally
interpreted in terms of wide scope, (i) does not have the interpretation according to which for each
student x, John overheard the rumor that that student had been called before the dean. On the
assumption that indefinites suppose QR of the existential operator, we would expect the scope of
this quantifier to be restricted to the embedded clause, contrary to fact. Example (ii) is indeed
compatible with the wide-scope interpretation: Mary is one of my students and John overheard
the rumor that she had been called before the dean. This interpretation does not square with the
hypothesis according to which indefinite NPs are necessarily submitted to QR, if QR is sensitive
to islands. Fodor and Sag thus come to the conclusion that the "referential" interpretation of
indefinite NPs should not be represented in terms of quantification.
237
The grammaticality of (63) indicates that the LF representation given in (63') is licit,
on a par with the S-Structure representation of (62):
(63')
We can also note that the different semantic restrictions associated with clitic
doubling analyzed in Sections 6.3.2. and 6.3.3 do not show up for quantified
indirect objects:
(64) a. Saie-(i) unui elev care s $tie englezete.
write-(him) aoat student which that (he) knowsut,j English
b. Ion (ii) va cere unei secretary s-1 ajute.
John (her) will ask aoat secretary (that) (she) help him
'John will ask a secretary to help him.'
c. Nu (le)-am sens multor elevi.
(I) have not (them) written manyoat students
The nonspecific interpretation is possible in all these examples, in spite of the
presence of the clitic. According to the analysis proposed in Section 6.3, this
reading relies on QR at LF; the LF configurations resulting from QR must then be
ruled in, on a par with (63').
These data can be subsumed under the following generalization, which holds not
only for Romanian but also for Spanish79 (standard, as well as River Plate):
variables can be doubled by dative clitics (as opposed to accusative clitics).
According to Jaeggli (1982), the contrast in clitic doubling between direct and
indirect objects is due to the difference in categorial status between indirect and
direct objects (PP versus NP). 80 It is obvious that this hypothesis cannot be
extended to Romanian, where indirect objects are NPs morphologically marked for
dative Case; if the categorial label were crucial, Romanian datives should behave
like direct objects and contrast with Spanish indirect objects, but this is not the case.
We may instead treat the difference between direct and indirect objects in terms of
In both Romanian and Spanish direct objects are assigned objective Case under
government by the verb, whereas indirect objects are "inherently Case-marked".82
These two types of Cases are crucially different. Objective Case is a structural
Case, assigned at S-Structure by verbal categories to a given syntactic position,
independently of the particular th-role assigned to it, and independently of the type
of element (empty or lexical) that occupies it. Objective Case can be viewed as the
identification of a syntactic position by the verbal head: objective Case renders the
direct object position visible, even if no lexical element occupies it at S-Structure.
Inherent Cases, on the other hand, are presumably related to particular th-roles
and assigned at D-Structure. Given these standard assumptions concerning inherent
Case, the following question arises: Can empty categories resulting from movement
be inherently Case-marked? The answer is most certainly no: inherent Case is not
assigned at S-Structure, the only level at which traces show up. In other words,
inherent Case is relevant for lexical elements, but not for empty categories.83 This
means that inherent Case cannot help in distinguishing between the different types
of empty categories. 84 More precisely, the Case requirement that enters the
definition of variables (see Chomsky (1981)) makes sense for structural Case in
81
A proposal by Jaeggli (1986) takes the same direction for Spanish itself. According to Jaeggli,
dative clitics, unlike accusative clitics, are optional absorbers of Case: they may be assigned
dative Case, but need not be. In a clitic-doubling construction the clitic does not absorb dative
Case, which is therefore available for the doubled lexical NP or for the clitic trace. Consequently,
a dative clitic trace can qualify as a variable. What seems crucial in Jaeggli's hypothesis is the
definition of clitic traces: (a) a clitic that has absorbed Case necessarily determines a pro; (b) a
clitic that is not marked for Case is unable to determine a pro in the argument position, which
will therefore be filled by a variable. But these two definitions cannot be empirically tested; nor
can the idea that dative clitics are optional absorbers of Case. As a matter of fact, underlying
Jaeggli's (1985) analysis is the implicit idea that dative Case is not relevant for the identification
of indirect object empty categories. This more radical hypothesis will be defended here.
82
The inherent Case markings corresponding to indirect objects are phonetically realized in both
these languages: by means of morphological Case in Romanian and by means of a preposition in
Spanish.Vergnaud (1982) has shown that in French functions as a Case marker when used in
front of indirect objects; this hypothesis could be extended to Spanish, but the matter is
controversial (see Jaeggli (1982) and Demonte (1987)). For ease of reference, I will speak of
dative Case when referring to indirect objects in Spanish, be they nominal or clitic.
83
This conclusion may also derive from the idea that inherent Cases are not assigned, but just
"marked"; and "marking" can only be relevant for lexical categories.
8
^Cinque (1990) assumes that in Italian, indirect object empty categories do not fall under the
typology of empty categories because of their PP status. The Romanian data indicate that this
characteristic of indirect objects is due not to their PP status but probably to the fact that they are
not assigned structural Case.
and quantification
239
general, and for objective Case in particular, but not for inherent Case; as a
consequence, the presence of a dative clitic (on which inherent Case is realized) will
be irrelevant for the variable status of the empty category it binds. This explains
why dative clitics are optional in quantification structures.
6. 4. 2. Romanian relatives and interrogatives containing cruia
Let us now tum to the contrast between (66) and (67):
(66) a. Cruiajif-ai
trimis bani ej?
whichp at himj) at -have (you) sent money e
'To which one did you send money?'
b. [Crui biat]jij-ai
trimis bani ej?
[whichoat boy] himDat have (you) sent money e
'To which boy did you send money?'
c. biatul cruiajij-am
trimis bani ej
the boy whichDat himD a f(I) have sent money
'the boy to whom I sent money'
(67) a. *Cruiaj ai trimis bani ej?
b. * Crui biatj ai trimis bani ej?
c. * biatul cruiaj am trimis bani ej
The analysis of care structures proposed in Section 6.1 can be extended to crui(a):
the maximal projection that dominates this wA-element does not count as a syntactic
quantifier; consequently, cruia cannot license an indirect object variable. The
ungrammaticality of the examples in (67) is due to the fact that ej cannot be
identified contextually: it cannot be a clitic trace, because no clitic is there to identify
it; it cannot be an anaphor, because it is not -bound; it cannot be a variable,
because cruia does not qualify as a quantifier. 85 The examples in (66) are
85
I n the early 1900s the following sentences were possible in Romanian (examples from Nilson
(1969)):
(i)
acel Cuza cruia datorim toate liberttile...
that Cuza whichjj a t (we) owe all the liberties
(D. Zamfirescu 1938)
(ii)
ideea c Turcia nu va putea s dureze, ci pindesc
primejdii crora nu va putea s
reziste...
the idea that Turkey will not be able to last, that dangers watch her whichjj a t (she) will
not be able to resist
(Iorga 1927)
For Steriade (1980) the possibility of (iii), in which the dative pronoun is not cliticized, explains
the possibility of (i)-(ii), in which the clitic is absent:
(iii) Cinti, c dau eu un bac$i lie.
sing for give I a tip y o u p a t
'Sing, and I'll give a tip to you'
(Sadoveanu 19,5, quoted by Sandfeld and Olsen (1936,1,126).
Popescu, unor adoratori ai cruia ar fost greu s (le) mrturisim situajia delicat in care ne
aflam,..
Popescu, s o m e ^ j admirers of whom [it] would have been difficult (themd a t) tell the
delicate situation in which (we) were,...
87m the Italian counterparts of these examples the dative clitics are optional (see Cinque (1984a;
1990)). This difference between Romanian and Italian is somewhat unexpected, given the
otherwise neat parallelisms between the two languages (see the Appendix). According to Cinque,
the Italian data indicate that the indirect object may function, at least in certain contexts, as a PP,
which is never the case in contemporary Romanian.
The systematic parallelism between Romanian CLLDs and Italian CLLDs suggests
that these constructions rely on the same syntactic rule. However, unlike Romanian,
Italian does not allow for the clitic doubling of direct objects. In other words, Italian
does not present any grammatical input for a movement transformation that would
have CLLDs as output. But this does not mean that a movement analysis cannot be
assumed:88 a movement analysis is accepted for passives, for instance, even though
the input structures are ungrammatical. As a matter of fact, no necessary relation
between clitic doubling and CLLDs can be established on theoretical grounds.
CLLDs are licit with respect to the various principles and conditions of the
grammar, independently of the existence of clitic doubling, which depends on the
existence of a preposition that assigns Case to the doubled direct object (see Section
6.1.1.). Italian lacks such a preposition, which accounts for the impossibility of
clitic doubling. How do we then account for the fact that left-dislocated direct
objects can be doubled by a clitic, with no violation of the Case Filter? Either of the
following hypotheses could answer this question: (a) NPs in dislocated positions
are not subject to the Case Filter, (b) a default Case is assigned in A'-positions. As
for the impossibility of clitics in Italian wA-structures, it can be explained by the
structural quantifier parameter: an NP in Comp has the status of an operator, which
must bind a variable, hence the exclusion of accusative clitics (recall that this
explanation is also needed for River Plate Spanish, a clitic-doubling language).
88
According to Cinque (1984a; 1990) the Italian CLLD does not rely on movement. Note
however that Cinque's main arguments (parasitic gaps, weak crossover) show that CLLDs do not
rely on W/i-Movement, and more precisely that CLLDs are not quantification structures. This
does not necessarily imply a nonmovement analysis. The difference between the movement and
nonmovement analyses of the CLLD may tum out to be purely terminological. If we choose to
assume that the only syntactic rule is Move (see Chomsky (1976)), the Italian CLLD is a
movement configuration (since it shows island violations, the CLLD counts as a rule of
grammar), even if there is no grammatical input for movement (see Cinque (1977)). A different
analysis of the Italian CLLD is necessary (a) if movement can be defined independently of island
sensitivity; (b) if we are able to show that the Italiai. CLLD does not satisfy this definition of
movement. In this case it would still be possible to account for the island sensitivity
characteristic of Italian CLLDs by adopting Cinque's (1990) Well-Formedness Condition on
Chains, which is stated independently of movement.
By virtue of (1), QR applies only if gu-features have percolated from the determiner
to the dominating NP. This constraint does not apply in (2): any NP whose (Spec,
NP) dominates a quantifier qualifies as a "quantified NP", and as such is affected
by QR.
In what follows, the feature percolation procedure will be replaced by a
movement rule:
(3)
DR raises1 determiners.2
(By "determiner" I mean any kind of functional element3 that is part of the DP 4
domain: cardinals and other "weak" determiners, as well as the definite article and
"strong" determiners (in the sense of Milsark (1977)) are considered
"determiners".)
* The central idea of this chapter was presented at the Tilburg LF Conference in 1988 and
published in Dobrovie-Sorin (1991). The present version has benefited from comments by Dan
Finer, Jim Higginbotham, Utpal Lahiri, Howard Lasnik, Richard Larson, Peter Ludlow, Hans
Obenauer, Elisabeth Villalta and Akira Watanabe. Special thanks go to Irene Heim and Anna
Szabolcsi for their careful reading of the various versions of this work and for the patience they
showed towards my clumsy use of formal semantics.
^The properties that I am concerned with here do not distinguish between May's derivational
model and a representational one, such as the Scope Assignment formalism (cf. Riemsdijk and
Williams (1981) and Williams (1986)): the new principles proposed here modify them both. For
ease of convenience the QR formalism will be used.
^Rule (3) brings to mind Heim's (1982, p. 133) rule of Quantifier Construal, by which
quantifiers (i.e., nominal determiners) are attached as leftmost immediate constituents of S.
According to Heim, Quantifier Construal applies to "strong" determiners such as every, but not to
indefinites (although Heim does not discuss the other "weak" determiners, she would probably
have treated them on a par with the indefinite article). As will be shown below, DR raises both
weak and strong determiners.
3
In line with Cardinaletti and Giusti (1990), and unlike Abney (1987), I assume that many is an
X element generated under Det rather than a Quantified Phrase.
4
For the purposes of the present discussion the DP hypothesis is not crucial; it is likely that DR
could have been proposed in the absence of the DP hypothesis. This is why I will use the label
NP for the constituent that is currently referred to as DP.
The rule of DR defined in (3) is the LF counterpart of the rule that operates at Sstructure in French (see Obenauer (1984-1985)):5
(4)
The evidence that I will provide in favour of the rule stated in (3) does not count as
evidence against (2). Rather, DR and QR should be viewed as different
instantiations of the rule Move that applies in the mapping between S-structure
and LF. Following Heim (1982), I will assume that QR represents in fact NPPrefixing (NPR(aising) henceforth), a rule which moves all-and not only
quantified-NPs and adjoins them to S.6
The present proposal raises an obvious question: Why do English, Romanian,
etc. not present the S-structure rule illustrated in (4) for French? Although no
explanatory answer will be suggested here (see Azoulay (1985)), let me simply
mention two constraints that seem to govern DR at S-structure: (a) DR can apply
only if the preposition de is present; compare (4)a-b, where livres is preceded by
de, with the ungrammatical *j'ai deux lu livres have two read books'; (b) DR
cannot affect "heavy" elements: *j'ai deux kilos mang6 de pommes 'Ihave two kilos
eaten of apples'. If the present proposal is correct these constraints do not hold at
LF.
Part of the central empirical data to be analyzed below has already been
accounted for in Chapter 6: prepositional accusatives and clitic doubling force the
specific reading of cardinals. The framework adopted here will allow us to provide
a finer analysis of the various possible interpretations of NPs in general, and of
cardinals in particular. It will also be possible to account for the restrictions that
"weak" islands (wh-islands, negation, factives) impose on the interpretation of whphrases. According to Cinque's (1990) extension of Comorovski's generalization,
only referential wh-phrases can escape out of weak islands. This generalization will
be revised here: it will be shown that certain non-referential wh-phrases do escape
out of weak islands; the only interpretation that is blocked is the "amount" one, to
be defined below.
7. 2. LF representations and NP denotations
I will assume the null hypothesis, according to which LF rules affect both weak and
strong determiners. The resulting LF representations will be filtered by two kinds
^Note, however, that according to Obenauer, beaucoup is base-generated in its S-structure
position, attached to V, rather than moved from within the NP.
^The motivation for letting QR function as NPR is simply that "quantificational" features are not
marked on NP, but on (Spec, NP), and therefore the quantificational status of an NP can be
"seen" by a rule only if transmission of features has taken place, from (Spec, NP) to NP. Insofar
as May does not assume any kind of feature percolation, his QR cannot be sensitive to
quantificational properties, and therefore it will affect any kind of NP.
(NPR)
(DR)
It is currently assumed that X elements can only adjoin to X elements, a condition that rules
out the representation in (5"). Thus, we are led to assume that this constraint, which has been
questioned even for overt S-structure movement (see Riemsdijk (1989)), is not active at LF.
In (6) the action of using is not viewed as applying to every ounce of butter, but to
an overall quantity (namely seven ounces) of butter. Similarly, in (5^) the action of
student-examination is not viewed as applying distributively to each individual
student, but globally to the overall quantity of students (this kind of nondistributivity, or rather non-individualization, should not be confused with the
collective reading, on which the NP can be construed as a plural individual).
Depending on the type of noun, this kind of interpretation will be more or less
easy to grasp (mass nouns can be interpreted only in this way; at the other end of
the scale, humans are more difficult to treat as overall quantities), but countables
clearly allow for it:
(7)
a.
b.
c.
d.
(DR)
(NPR)
Note that the derivation in which the landing site of NPR is higher than the raised
determiner is to be ruled out as a case of improper movement (the trace of the
determiner inside the open NP would C-command the determiner). We are then left
with the possibility illustrated in (53)b, where the NP is raised in between the raised
determiner and the S constituent. This derivation itself relies on the assumption that
the relative hierarchical order of S-adjoined constituents is free (see May (1977)).
One may wonder whether the LF shown in (53) could not be derived by NPR
followed by DR (see Heim's (1982) proposal for strong quantifiers such as every):
(5"') a. 0[tenj [NP e, students]j [ s John examined
^DRI *
NPR
[ffp]
(NPR+DR)
I will assume that this derivation is illegitimate (as indicated by the diamond): in its
adjunction position, the NP is not L-marked, and therefore it constitutes a barrier
for the raising of the determiner.8 Note that this is a case in which a derivational
model rather than a representational model seems to be needed: indeed, the
representation in (5'") does not differ from that in (53)b; it is only their derivations
that are distinct.
Coming back to (53)b, it has the format of restricted quantification. I assume
that the second part of the formula, i.e., the open NP, translates as the restriction,9
and the third part as the scope of the existential quantifier, which is necessarily
associated with D-Raised cardinal determiners. The corresponding reading involves
existential quantification over a set variable, with the raised determiner interpreted
as a cardinality marker of the set variable: "There is a set X of individuals (where
is a student) such that John examined x, and the cardinality of X is ten".10
According to the LF representation in (53)b, the sentence in (5) is true if (a)
there is an indeterminate number of students and (b) (at least) ten of them were
examined by John. The sentence is false if condition (b) fails; if condition (a) fails,
the sentence is either false or indeterminate, (i.e., lacks a truth value).11 This type of
reading can be referred to as "partitive" or "presuppositional", due to the condition
stated in (a): a presuppositional NP carries the presupposition that the set denoted
by N' is not empty, see de Jong and Verkuyl (1985) and de Jong (1987); thus, on
the presuppositional reading, the example in (5) presupposes the existence of a
8
This goes against Fiengo, Huang, Lasnik and Reinhart (1989), who assume that phrases
occupying A'-positions do not function as barriers (for discussion I refer the reader to the quoted
source).
9ln Dobrovie-Sorin (1992) I tried to argue that DR + NPR necessarily gives rise to unrestricted
quantification: as a result of DR the determiner is no longer a sister of N', and therefore N' would
no longer count as the restriction of the determiner. I now think that this is not correct: the format
of restricted quantification does rely on DR (see Heim (1982)). It is only the LF corresponding to
the referential reading that does not involve DR.
10
The analysis of cardinals (and indefinites) as denoting sets with the determiner functioning as a
cardinality marker for the set is currently assumed in formal semantics (Bartsch (1973), Scha
(1981), Keenan (1987) and Reinhart (1987)). My own proposal is the hypothesis that this
currently assumed semantic representation relies on a particular LF representation.
11
The choice between the two possibilities varies from one semantic model to another, with no
consequence for our present purposes.
group of students; it is asserted that the predicate "having been examined by John"
holds of part of them. 12
The proposed analysis of (53)b relies on a correlation between the format of
restricted quantification and the presuppositional reading, which I will not try to
motivate. Let me simply quote Hausser (1974), who assumes it on independent
grounds: "All quantifiers which have been determined to be existential
presupposition inducers will from now on be represented by means of restricted
quantification".
Note now that (5^)b can be rewritten in the format of unrestricted quantification,
as shown in (5 4 ):
(5 4 )
The move from (5 4 )a to (54)b is legitimate due to the semantic properties of the
raised determiner, namely the fact that it is symmetric. In (5)b student is not the
restriction of the determiner ten: the existential quantifier associated to ten binds a
set variable defined by the intersection of the set defined at the NP-level (the set of
students) with the set defined by the predicate (x: John examined x). 1 3 The
corresponding reading, let us call it "existential", would be "There are (is a set of)
ten individuals that are students and have been examined by John". This reading is
truth-conditionally different from the presuppositional one. Thus, it does not
presuppose the existence of a set of students, and therefore, on the existential
reading, example (5) may be true even if no set of students can be picked up from
context. Compare the presuppositional reading (see (53)b), on which example (5) is
false (or has no truth value) if no set of students can be identified in the context.
A short note is in order here, concerning the proposed "readings" of (53)b and
(5 4 ). It is probably not correct to assume that an example such as (5) "means" what
was stated as its readings. What we really want to say is that the presuppositional
12
An adequate context that would force the presuppositional reading of (5) is easy to provide;
e.g., Yesterday the professors of the English department examined 30 journalists, 20 students
and 40 retired people. John examined 10 students. The presuppositional reading is also
necessarily triggered by partitive NPs such as ten of the students, two of your children, etc. Given
the proposed framework, we would have to say that o/-phrases necessarily map into the restriction
of the determiner, but I leave the technical details for further inquiry. Note that partitives
constitute a problem for semantic frameworks such as those in Keenan and Stavi (1986) and
Keenan (1987). In order to account for the non-existential reading of partitives these authors have
to assume that two of the functions as a determiner (on a par with the, every,two, etc.), which is
at odds with the current syntactic analyses.
The rule of DR thus allows us to formalize an intuitive statement that can be found in
Reinhart (1987), for whom weak determiners operate on a clausal set (strong determiners would
instead operate on sets fixed at the NP-level). The analysis proposed here differs from Reinhart's
in that it allows for an "NP-level" interpretation of weak determiners: this reading corresponds to
those LF representations that rely on NPR alone, and therefore assume the format of "restricted
quantification". To summarize, a determiner operates on a clausal set if the relevant LF has the
format of unrestricted quantification.
(NPR)
This means that closed NPs are necessarily associated with the referential reading
defined as in (9):
(9)
According to this proposal, the referential reading is assigned to NPs that may have
been affected by NPR, but not by DR, nor by DR + NPR (both DR and DR +
NPR give rise to open NPs). Let me stress that the referential reading does not
depend on the application of NPR: referential NPs may be assumed to stay in situ,
unless scope properties force us to assume NPR. Thus, given the evidence
examined in this chapter, there is no need to assume NPR for the referential
interpretation of (5). The only reason I assume it here, is to stress the fact that LF
movement does not necessarily give rise to a quantificational, non-referential
interpretation. However, if for independent reasons, NPR must be blocked, the
referential reading would still be available (unless a derivation relying on DR
obtains). I take the Weak Cross Over effect to be a test for "(non-)referentiality".
This test should correlate with other well-known observations concerning the
consequences non-referentiality has on entailments (failure of the law of the
excluded middle, failure of the law of contradiction, etc.).
Note that the notion of referentiality used here should be kept distinct from the
pragmatic notions of specificity or D-linking, although certain correlations can be
established, as discussed in the following sections.
A possible paraphrase of the referential reading associated with (5l) is "having
been examined by John is a property of ten individuals who are students". This
means that the sentence is true iff (a) there are ten individuals, J, M,... (b) who are
students and (c) who have the property of having been examined by John. The
sentence is false if any of the conditions stated in (a)-(c) is not met. It is interesting
to note that the truth values of (5l) can be established by simply calculating the
denotation of the NP ten students, and by checking whether the set denoted by the
predicate was examined by John belongs to the set of sets denoted by ten students.
This is possible because this NP is closed, and therefore interpreted as denoting an
(NPR)
It will be shown that although the LF representation associated with bare quantifiers
does not involve an open NP (bare quantifiers are not subject to DR), it does
involve unrestricted quantification. This indicates that bare quantifiers function as
operators rather than referential NPs.
These results are summarized in (11):
(11)
LFs
NPR
DR
NO
YES
YES
YES
NO
YES
unrestr. Q restr./ unrestr. Q
16
DR + NPR
NPR of bare Q
NO
YES
unrestr. Q
(12)
NPR
DR+NPR (restricted quantification)
DR+NPR (unrestricted quantification)
DR
referential
presuppositional
existential
amount
It was proposed that weak NPs are compatible with four LF configurations, and
correlatively present four distinct interpretations. This differs from current semantic
analyses, which assign only one reading, namely the existential one, to weak NPs.
In addition to the differences in truth-conditions already pointed out, there are wellknown syntactic tests that distinguish between the various readings. Thus, both
presuppositional and existential NPs give rise to Weak Cross Over violations,
whereas referential NPs do not. There-contexts, on the other hand, provide us with
a test that distinguishes between presuppositional and existential NPs: only the
latter are allowed in there contexts. At this point in the discussion, the difference
between the existential and the amount interpretations seems speculative, but it will
become clearer when weak islands will be discussed (see Section 7. 7.): it will be
shown that weak islands do not block the non-referential reading in general (they
block neither the existential nor the presuppositional readings), but only the amount
reading.
7. 3. Locality conditions on DR
So far, it has been established that the DR hypothesis provides us with a
framework which is able to assign four distinct LF representations to weak NPs 17 ,
underlying a four-wise difference in interpretation (referential, presuppositional,
existential and amount). Evidence in favour of DR can be given by showing that it
obeys the general constraints of syntax. If the DR of a weak determiner is blocked,
NPR alone will be possible. The prediction made under the present account is that
the presuppositional, existential and amount readings of weak NPs are blocked in
l^The dash indicates that NPR does not rely on either restricted or unrestricted quantification (see
footnote 15).
17
Obenauer (1992a,b) gives independent syntactic evidence - based on the French "complex
inversion" and the past participle agreement - in favour of the idea that weak NPs give rise to at
least two distinct LF representations, comparable to my open and closed NPs.
(ambiguous)
(referential)
These examples all involve quantified NPs (QNPs) in the direct object position; in
(13) and (15) the QNP is preceded by the "dummy" preposition pe (compare a in
Spanish, cf. Jaeggli (1982)). Examples (13) show that the preposition is allowed
(and even obligatory for humans) with bare quantifiers. The pair (14)-(15) indicates
that with indefinites and cardinals the use of prepositional accusatives forces the
referential reading; the prepositional accusatives in (15)a-c are interpreted as
denoting individuals: "there is a friend and I met him; there is a secretary and I am
looking for her; there are three individuals (John, Mary and Pat) who are students
and each professor will examine them".
The correlation between prepositional accusatives and referentiality (currently
refeired to as "specificity") is a well-known, though poorly understood, empirical
18
The grammaticality judgments indicated in (15) are not shared by all speakers (this is true for
both Romanian and Spanish). For some speakers, examples (15) are ungrammatical (for a
suggestion as to why this holds, see footnote 23). The insertion of a doubling clitic does not
improve acceptability: those speakers who reject (15)a-b also reject their counterparts with
doubling clitics. The description of the data is complicated by the fact that in contemporary
Romanian the clitic doubling of prepositional accusatives has become almost obligatory (it was
optional at the beginning of the century). Therefore, the examples in (15) sound somewhat
archaic; I discuss them here because I try to separate the interpretive constraints imposed by
prepositional accusatives from those due to clitic doubling.
NPR
DR
^The fact that the NP is not L-marked is enough to rule (15 2 ) out. But note that extraction out
of NPs obeys an even stronger requirement, namely canonical government by the verb (see
Appendix and Pollock (1989)).
^Recall that the derivation NPR+DR was also ruled out previously (see the discussion of (5'")).
22
A s mentioned in footnote 19, certain speakers reject examples like (15) altogether. We must
assume that for these speakers weak quantifiers are obligatorily subject to either DR alone or DR
+ NPR; for those speakers who accept (15), weak quantifiers may also be affected by NPR alone.
23
Note also that in Romanian the NPs governed by pe (and in fact by any kind of preposition)
constitute barriers for the relation between the Genitive clitic attached to the verb and its trace,
which occupies the (Spec, NP) position characteristic of Genitives [the clitic in (i) is marked with
the Dative-Genitive morphological Case; the hypothesis that this clitic is Genitive rather than
Dative can be found in Steriade (1980) and Dobrovie-Sorin (1987):
(i)
a. Ij-am vzutfj^pprofesorulei].
(I) Genhimi-have seen [NPthe teacher ei]
saw his teacher.'
b. *Ij (l)-am vzut [pppe [NP profesortj]]
(I) Genhimj (himAccus)-have seen [pppe [NP teacher ej]]
saw his teacher.'
A similar construction exists in French, which is also blocked by the presence of dummy (and
lexical) prepositions:
(ii) a. J'enj ai rencontr6 [l'auteur ej]
I its have met the author
met its author.'
b. * J'enj ai 6crit [pp [Q^l'auteur ei]]
I its have written to the author
The analogy between the movement of determiners at LF and the S-Structure movement
illustrated in (i)-(ii) may be questioned, since genitives and determiners are not constituents of the
same type (they are XPs and X, respectively). However, the movement of genitives and that of
By applying DR to (17) we obtain (18), which is illicit for the reasons already
discussed above (when pe is present, NP constitutes a barrier for the raising of any
of its subparts):
(18) 0[cine]i Ion a examinat [pppe [ ei va]]
Osome [John examined [pppe [x person]]
If (18) were the correct representation for (17) we would expect this example to be
ungrammatical (because (18) is illicit) but this is not the case. Since (17) is fully
grammatical we must assume that DR does not apply to subconstituents of bare
quantifiers.25
Since DR does not apply, the only legitimate LF representation of examples
such as (17) is (19), obtained through NPR:
determiners are comparable insofar as both elements are subconstituents of an NP; moreover, even
if the movement of genitives is assumed to leave an NP-trace, the moved constituent is a clitic,
i.e., an X category.
24
This generalization seems to be subject to variation across languages: as pointed out to me by
Anna Szabolcsi, Hungarian presents morphologically marked accusatives which seem to be able
to function nonreferentially (they may incorporate to the verb).
25
I do not think that the morphological unanalyzability of nimeni 'nobody' and nimic 'nothing'
(ni may be analyzed as a negative prefix, but -meni is neither a free nor a bound morpheme; mic
may be related to the adjective mic 'small') can be invoked as an explanation for the fact that DR
does not apply. Thus, analyzable pronouns such as the English notxxly and nothing do not seem
to allow the DR of no.
(19)
( = (17) + clitic-doubling)
I met pe a friend
two distinct LF representations (QR could affect them or not, depending on qufeatures percolation), corresponding to DR + NPR and NPR, within the DR
framework developed here. In addition, however, it is now possible to postulate a
third type of representation, namely that relying on DR + N'-incorporation, which
underlies the "amount" reading. The distinction between DR + NPR and DR + N'incorporation is a welcome result, because it captures empirical differences
concerning both the relevant interpretations and the locality conditions to which
these two types of structures are subject (see Section 7. 7., where weak islands are
discussed).
According to the current semantic analyses, cardinal NPs are treated as
involving quantification over set variables with the determiner indicating the
cardinality of the set (see (Bartsch (1973), Scha (1981), Keenan (1987) and
Reinhart (1987)). This kind of higher order quantification can be correlated with the
LF resulting from DR + NPR (see Dobrovie-Sorin (1993)). However, the
implementation relying on -features percolation followed by QR was interpreted
as first order quantification, i.e., as quantification over individual variables. It
remains to be seen whether higher order quantification could be used within this
framework.
Note finally that if we correlate quantification over set variables to DR + NPR,
we obtain the correct result that bare quantifiers such as somebody and nobody
cannot give rise to quantification over sets, but only to quantification over
individuals. This difference between bare quantifiers and cardinals allows us to
account for empirical differences between these types of NPs, regarding their
behaviour in weak island contexts. Such differences are more difficult to capture
within the framework developed in Chapter 6: we should postulate distinct
interpretation rules to bare quantifiers and to NPs that bear percolated qu-features.
Other differences between the two approaches are theory-internal. Thus,
according to the proposal developed in Chapter 6, the impossibility of doubling
clitics in certain quantificational configurations derives from the hypothesis that
variables need Case. This requirement is stipulative because it is difficult to
understand why variables need Case.26 According to the proposal developed here,
the possibility of clitic-doubling is not related to Case theory, but depends on
whether the doubled NP can be assigned the referential interpretation. It is natural to
assume (22)a-b, from which the generalization in (23) follows:
(22) a. Non-referential NPs cannot be assigned pronominal features,
b. A clitic imposes its pronominal features on the doubled NP.
(23)
26It
is not clear why an empty category could not inherit the Case "absorbed" by the clitic that
identifies it, or the Case assigned to the wh-phrase itself (see *pe cinej lj-ai vzut ej'pe whom
him-have [you] seen e', where the wh-phrase is Case-marked by pe). In addition, the Case
requirement does not seem to apply to all the elements that one might want to define as variables.
Thus, one may assign the label "pronominal variable" to the empty categories resulting from
NPR (recall however, that bare quantifiers present a special behaviour; they in fact bind nonpronominal variables). It should then be stipulated that the Case requirement concerns only nonpronominal variables (the term "variable" used in Chapter 6 should be understood as meaning
"non-pronominal variable"; Borer (1984) is correct in saying that variables do not need Case if by
"variable" we mean "pronominal variable").
This indicates that it is incorrect to say that direct object pronouns marked [+human] are
obligatorily preceded by pe (see Farkas (1978)). We may then wonder what the real generalization
is. The remarks that follow build on the idea that pe is necessary with [+human] pronouns
whenever pe is possible: pe can be absent only if it must be so (the non-presuppositional/
existential interpretation is imposed in certain contexts, and this would conflict with the presence
of pe).
The question as to whether quantified pronouns are accessible to DR alone is left open for
further research.
raised, followed by the NPR of the resulting open NP). Quantified pronouns differ
from quantified NPs in that they appear to obligatorily (and not only optionally) be
subject to DR + NPR. The paradigm in (24)-(27) shows that the interpretations
associated with the LFs that our analysis postulates match with the interpretive
restrictions imposed by a given linguistic context. Bare quantifiers differ crucially
from quantified pronouns: they cannot function as closed/referential or
presuppositional, even if they are not affected by DR.
7. 4. DR and wh-movement
In Chapter 6 it was shown that Romanian exhibits two types of wh-structures. It is
now possible to capture the difference between them by assuming that in one case,
but not in the other, the DR rule defined in (3) applies in the mapping between Sstructure and LF.
Consider first the weak wh-quantifier ce 'what':
(28) a. Ce roman ai citit?
what novel have (you) read
b. *Ce roman 1-ai citit?
what novel it-have (you) read
This data can be accounted for if we assume that ce 'what' is obligatorily raised by
DR. More precisely, (29)a-b are obtained by applying reconstruction of the whphrase into the -position, followed by DR, itself followed by NPR:29
(29) a. [S'[whati] [S'tNPei novel]i [s
[NPei]]30
?x [x: novel; you read x]
'what object is such that it is a novel and you read it?'
b. 0[S'[whati] [S'tNPei novel]i [S-.iti...[NPei]]
29 D R alone is incompatible with the semantic properties of what N' (this quantifier must bind an
individual variable) but may apply to how many (see 7. 6. and 7. 7.3.2.).
30i t is not clear to which position the quantifier raises: since it is an X category, it cannot adjoin
to maximal projections such as S' (but this condition might not apply at LF).
I assume that because of its semantic properties, which cannot raise out of the whphrase.31 This means that representations of the type given in (29) are not available
for examples (30); only (30l), which relies on NPR, is legitimate:
(30l) a. [which boyj] [s' you saw[NPei]]
'Which (of these) boy(s) is such that you saw him?'
Since the wh-quantifier cannot be raised by DR, the wh-phrase is a closed NP, and
therefore counts as referential, 32 which as such is allowed to be clitic-doubled.
However, it is not clear why the variable must be clitic-doubled (see the
ungrammaticality of (30)b). The answer to this question must take into account
crosslinguistic data: in languages such as French, Spanish, etc. a doubling clitic is
not necessary (it is in fact excluded) with strong wh-phrases:
(31)
1 More precisely, DR applies, but the resulting LF is filtered out due to the semantics of which.
T o speak of referential wh-phrases seems paradoxical. For further discussion see Sections 7.7.3.
and 7.7.4.
32
Given this structure, the wh-phrase under [Spec, C'] is only an indirect antecedent
of the wh-trace, the relation being mediated by C, marked with the [+WH] feature.
One may then assume that it is C[+WH]. rather than the wh-phrase itself, that is the
true licenser of the wh-trace. It is natural to assume that any C marked [+WH]
functions as an operator, and that operators license A'-bound empty categories.
It is well known that in English, French, Spanish, etc., the configuration in (32)
is legitimate regardless of the type of element that stands under [Spec, C']. There are
two relevant cases here: (a) null operators are allowed in [Spec, C'] and (b) no overt
agreement holds between wh-phrases and wh-traces (on "wh-agreement" see
Chung (1982). Romanian is characterized in the opposite way: (a') null operators
are not allowed, and (b') certain object wh-phrases trigger "wh-agreement", which
manifests itself as an accusative clitic doubling the wh-trace.
There are two ways in which the crosslinguistic pattern of variation could be
captured. We could assume that Romanian is characterized by the same structural
configuration as French or English. The difference would be due to a parameter
concerning the licensing of C:
(33) a. In English, French, etc. C is licensed structurally by [Spec, C'],
independently of the content of [Spec, C].
b. In Romanian, C is licensed only if it is "identified".
Possibly because of their lack of lexical content, null operators would not function
as identifiers of C. As to lexical wh-phrases, they would license C by agreement,
which would percolate down to the wh-phrase itself. This line of inquiry will be
abandoned here because (33) seems highly stipulative to me, and the
implementation itself does not seem obvious.
Let us then try to derive the above-mentioned crosslinguistic differences as a
consequence of a structural difference between wh-constructions. The suggestion
that I would like to make is that Romanian wh-constructions do not rely on
movement to [Spec, C] but rather on adjunction to IP:
(34) IP
Since adjunction positions are not projected by the base rules, it is reasonable to
assume that null elements, in particular null operators, cannot show up in adjunction
positions. This accounts for the lack of null operators in Romanian.
7. 5. 2. A'-licensers
An account of "wh-agreement", i.e., of the distribution of doubling clitics in whstructures, can be proposed if we restrict the class of A'-licensers as stated below:
(35) An empty category is not licensed by a closed NP that A'-binds it.
[xPjtDet.J
IP
The examples in (39) are grammatical because the variable in the direct object
position is licensed by the raised quantifier: ce 'what' is lexically characterized as
being accessible to DR and nothing blocks this rule; as a result of DR, ce Ccommands the wh-trace, hence binds and legitimates it. The only difference
between (39) and (40) being the presence of the preposition pe, it is possible to treat
(40) as a violation of the locality constraint on DR which we have already examined
when dealing with quantified NPs in situ (see the discussion of (15)):
(40') 0[s [cei] [sai putea tu suporta [pppe [Npej elev]]]
[S [whati] [scould you put up with [pppe [Npej student]]]
In (40') the DR of ce 'what' violates the ECP (antecedent government is blocked
because what crosses at least one barrier (the NP, which is not L-marked by pe)).
In other words, pe blocks the DR of ce out of the dominating PP, and consequently
the wh-phrase counts as a closed, i.e., referential, NP, which cannot license the
empty category in the direct object position (see (35)).
Consider now (41):
(41) Pe cine ai invitat?
pe who have (you) invited
(41') [s [cinei] [sai invitat [pppe [Npei]]
[S [whoi] [sdid (you) invite [pppe [Npei]]]
In (41') cine is a bare quantifier, hence there is no barrier that could block the NPR
of cine out of [pppe [Npcine]]. Since bare wh-quantifiers function as operators
(see (36)), they are able to license their wh-traces.
It is interesting to note that a clear parallelism exists between the behaviour of
bare wh-quantifiers and that of bare quantifiers in situ. Thus, in Section 7. 3. 2. the
conclusion was reached that DR does not apply to any subpart of bare quantifiers,
which are raised as a whole by NPR (hence the possibility of pe). However, bare
quantifiers in situ do not function as referential NPs: they give rise to Weak Cross
Over violations and do not allow clitic-doubling. These characteristics also hold for
bare wh-quantifiers (see Chapter 6). Correlatively, bare quantifiers function as
licensers of A'-bound empty categories.
7. 5. 4. Wh-strategies and the constituent structure ofS and S'
One may now raise a further question: Why does Romanian not use the Move wh
to [Spec, C'] strategy? A possible line of inquiry may be suggested, which takes
into account independent results, concerning the constituent structure of Romanian
sentential constituents. In Chapters 1 through 3 it was shown that the -position of
the NP subject is not the [Spec, I'] node, but a position internal to VP. [Spec, I'] is
an A'-position, which is available for left dislocated and topicalized constituents,
and possibly for wh-phrases as well. Hence, Move wh to [Spec, I'] could
conceivably be assumed instead of the adjunction analysis shown in (34) and (38).
Note that according to both of these hypotheses, C and [Spec, C'] are not activated
in Romanian wA-structures.
In sum, it has been proposed here that the ECP may be satisfied as late as LF.
The licensing of A'-bound empty categories appears to be more constrained than is
usually assumed. The standard assumption is that an A'-bound empty category is
licensed, no matter what the nature of the antecedent is, provided it is local.
According to the revision proposed above, A'-bound empty categories cannot be
licensed by closed, referential NPs. They must be bound either by an operator or by
a clitic. The present discussion is mainly concerned with direct objects; it may be
that for other empty categories the licensing conditions are more permissive (see
Appendix).
(3-way ambiguous)
(reconstruction + NPR)
(reconstruction + DR) 33
(reconstr + DR + NPR)
Given the above discussion of Romanian wh-structures, we predict that LFs of the
type in (421) require the presence of doubling clitics. As to (42^) and (42^), they
are not overtly distinguished: in both cases a doubling clitic and the prepositional
marker will be disallowed, which indicates that the wh-phrase counts as nonreferential. In what follows it will be shown that the distinction is nevertheless
relevant for weak islands. The various possible interpretations of the LFs of (42)
will be made precise below; they will be shown to be parallel to the referential,
existential and amount readings already distinguished for weak NPs in situ.
Summary
The framework proposed here allows us to establish the following typology of LF
representations for wh-structures:
(43)
a.
b.
c.
d.
33
LFS of the type in (422) have already been assumed (Chomsky (1977), Heim (1987)). They
were supposed to be derived through the reconstruction of part of the wh-phrase. The rule of DR
provides a more principled derivation.
265
A'-binding holds under C-command between two elements that bear the same
referential index; since movement does not create referential indices, they can only
be carried over. This means that we need a mechanism by which referential indices
are assigned to/licensed on NPs, which will take them along under A'-movement.
Rizzi assumes that referential indices are licensed by referential th-roles.36 Given
these assumptions, the only elements that can be A'-bound are those that are
assigned referential th-roles; if an empty category is not assigned a referential throle, it must be licensed by antecedent-government, which is blocked in weak island
contexts (by virtue of Rizzi's relativized minimality).
7. 7. 2. Reconstruction, DR and the locality of wh-movement
As observed by Cinque (1989, 1990), Dobrovie-Sorin (1988, 1991), Frampton
(1990), Kroch (1990) and others, certain interpretations of wh-phrases are sensitive
to weak islands. Rizzi (1990) cannot deal with this problem because in the relevant
cases a referential th-role is available. According to Cinque's (1990) refinement of
Rizzi's theory, referential wh-phrases are allowed to move long distance, as
opposed to non-referential wh-phrases, which must be antecedent-governed (this
generalization is illustrated by Longobardi's observations concerning the
reconstruction of scope, by Cinque's own observations concerning the Italian clitic
left dislocation and by Pesetsky's (1987) and Comorovski's (1989a, b) D-linking
phenomena).
One problem raised by Cinque's generalization is that we do not know exactly
how to define referential expressions: we may wonder what the various cases listed
above have in common37; we may even ask whether they can really be subsumed
by the notion of "referentiality" and whether they are all sensitive to exactly the
same type of locality conditions. Since we do not know what "referentiality" means,
it is better to avoid using referential properties as a primitive notion, i.e., as some
kind of abstract feature that would characterize linguistic elements.
According to my own account, referentiality is read off a particular LF
representation (that relying on closed NPs), and not attached to individual linguistic
36
The label "referential th-role" is probably not very felicitous, but the notion that it is intended
to denote seems clear to me, even if Rizzi (1990) does not provide a clear definition. I would like
to propose the following one: a referential th-role is one which allows the construction of a set of
individuals (the label "referential" th-role -which I use for convenience -is completely irrelevant,
and could be replaced by anything else, say "th-role 1"). Independently of their linguistic
environment, verbs such as see, eat, write, etc. suppose sets of individuals of the type "objects
that are seen/ eaten/ written". An empty category that instantiates the object position of these
verbs can then be viewed as an individual variable. Similarly, sets of individuals that are
measured/weighed can be defined; NPs such as a garden, a room, a child etc. can be viewed as
corresponding constants. However, "1 kilo, 2 kilos, 100 grams etc." or "1 meter, 2 meters, 3
centimeters, etc." cannot be viewed as constants obtained by an assignment of value to individual
variables of the kind just mentioned. It is also clear that idiom-chunks are not constants
corresponding to individual variables. Finally, it does not seem possible to construct sets of
individuals corresponding to subcategorized adverbials (ways of behaviour, places to go to or
places to stay at).
37
Cinque proposes certain tests, in particular pronominal anaphora, but these tests are not applied
to all of the types of non-referential NPs considered.
38
This initial idea will be somewhat refined below. The sensitivity to weak islands will be
checked on the finer typology of wh-structures established in 7. 5. and 7. 6.
The ungrammatically of (47)b has been analyzed as an ECP violation (where the
ECP is defined in terms of relativized minimality) by Rizzi (1990): Tie is a minimal
antecedent which blocks the antecedent government relation between combien and
its deepest embedded trace (A'-binding cannot apply because the trace of combien is
not assigned a referential th-role, see (44)). For our present purposes it does not
matter whether this explanation is correct (for a semantic account of weak islands
see Szabolcsi and Zwarts (1990) and Szabolcsi (1992)). The crucial point is that
whatever explanation is chosen for (47)b-c will account for the illicit status of
(46l)b-c in exactly the same way. The illicit character of (46l)b-c accounts in tum
for the fact that (46)b-c lack the amount interpretation.
7. 7. 3. The long wh-movement of non-D-linked wh-phrases
Let us recall now that within the framework proposed here we have at our disposal
a typology of wh-structures which distinguishes among five types of wh-structures
(see (43)a-e). It has been established that bare quantifiers, as well as structures
relying on DR + NPR, cannot be interpreted as referential. However, they will be
shown not to be sensitive to weak islands. The case of clefts will appear to be
similar. All these cases go against Comorovski's (1989a,b) and Cinque's (1990)
generalization that non-referential/ non-D-linked wh-phrases cannot be extracted
out of weak islands.39
7. 7. 3. 1. Bare quantifiers
According to the analysis proposed in 7. 3. 2., bare quantifiers do not undergo DR,
but they function nevertheless as non-referential. Note also that the interpretation of
bare quantifiers is clearly non-D-linked. The speaker does presuppose that the
hearer may answer by mentioning a particular individual, but this individual is not
part of the background of the dialogue:
(48)
39
Cinque's generalization is an extension of Comorovski (1989a, b), who argues that only Dlinked wh-phrases get out of wh-islands. Most of my examples are constructed with the negation
or factives; wh-islands seem to impose particular constraints, which I will leave aside for further
research.
7. 7. 3. 2. DR + NPR
7. 7. 3. 2. 1. The DR of what
In Section 7. 4. above it was shown that Romanian presents a minimal pair between
what N' and which N' expressions: the former cannot be clitic doubled, nor marked
with a prepositional Accusative, whereas the latter are obligatorily clitic doubled
and marked with a prepositional Accusative. This overt distinction correlates with
the well-known interpretive difference in terms of D-linking. But this difference in
D-linking does not appear to make a difference for the extraction out of weak
islands. Both (49) and (50) are well-formed:
(49) a. Ce carte regrep c a citit Maria ?
'What book do you regret that Mary has read?'
b. Ce carte n-a cititMaria?
'What book hasn't Mary read?'
(50) a. Care carte regrep c a citit-o Maria?
which book do you regret that has read-it Mary
b. Care carte n-a citit-o Maria?
which book hasn't read-it Mary
The data in (49) show that weak islands are freely violated by the non-D-linked
what ', and this is not expected under the D-linking constraint on extraction out
of weak islands.
Under the present proposal we must explain why the LF representations of (49),
relying on DR + NPR, are legitimate:
(491) a. Whati [[ ei book]i do you regret that Mary has read ej] ?
b. Whati [[ ei book]i hasn't Mary read ei] ?
The empty category in the direct object position is assigned a referential th-role, and
therefore, given Rizzi's theory, A'-binding can obtain; hence weak islands can be
freely violated (see (44)b).
7. 7. 3. 2. 2. The D-linked and non-D-linked interpretations of how many
Consider next the following examples:
(51) a. Cite femeij regrep c ai iubit ei ?
how many womenj do you regret to have loved ei
b. Pe cite femeij regrep c lej-ai iubit ei?
This data shows that cite/cip 'how many' expressions may optionally be clitic
doubled (correlatively a prepositional Accusative occurs). The overt difference
between these two possibilities corresponds to a difference in interpretation: (51)b,
but not (51)a, is D-linked. But here again we observe that the lack of D-linking that
characterizes (51)a does not block extraction out of weak islands. This data can be
understood if we assume the following LF representations:
(511) a. [cphow manyj [CPfej womenjj [do youregretto have loved ej]]
b. [cphow many womeni [do youregretto themi-have loved ei]]
(51 l)b, which presents an empty category bound by a closed NP, corresponds to a
"predicational" reading: the questionee is assumed to have at his disposal a Dlinked set of women (i.e., a set of women which constitutes common knowledge
between speaker and hearer), say {Mary, Joan, Pat}; for each woman the
questionee checks whether a certain predicate (that of having caused regret ...)
holds, and then he counts the positive answers.40 Note that this procedure of
answering a wh-question is equivalent to answering a set of yes/no questions : Do
you regret to have loved Mary/Joan/Pat?. This procedure depends on D-linking, or
more precisely on the referential interpretation of the wh-phrase. It is only if the
wh-phrase denotes an individual that truth conditions can be assigned by checking
whether a given predicate holds of a given individual (in more formal terms one
checks whether a property (a set of individuals) belongs to the set of properties (the
set of sets of individuals) denoted by the individual). Since yes/no questions
depend on the possibility of assigning truth-conditions in this way, a wh-question
may be equivalent to a yes/no question only if the wh-phrase denotes an individual.
Compare (51)a, which asks a question concerning the number of individuals that
are women and that the questionee regrets to have loved. This interpretation does
not involve D-linking: there is no discourse-salient set of women, but only a set
defined as the intersection of two predicates, [[x: you regret to have loved x] and
[x: woman]]. The questionee is asked to tell what the cardinality of that set is. The
corresponding reading relies on existential quantification: "there is a set of
individuals such that....; who/ how many individuals are there in the set?". The
difference between (51)a and (51)b could thus be characterized as relying on
existential quantification and on predication, respectively.
The difference between (51)a and (51)b can also be characterized in terms of
different presuppositions: (51)a presupposes you regret to have loved (the
question bears on the identification of the variable) whereas (51)b presupposes
there is (Mary, Joan, Pat} and the question bears on the truth of you regret to have
loved Mary, Joan, Pat.
Languages like French and English do not overtly distinguish between (51)a and
(51)b, and it is somewhat difficult to characterize in intuitive terms the difference
between the two readings. English speakers may think of the difference between
(52)a and b:
(52) a. How many women do you regret to have loved?
b. Which of these women do you regret to have loved? How many are they?
40
The relation of "predication" described here is probably what Szabolcsi (1992) refers to as the
"look-up" procedure (one looks at every individual in a given universe and checks whether it
exhibits the property expressed by a given predicate). Szabolcsi observes that "look-up is viable
because we assume that each individual is a 'peg', from which all its properties are hanging, cf.
Landman (1986)". A reasonable interpretation of this observation is that the look-up procedure is
viable only in D-linking contexts: the universe is then limited to individuals with "hanging
properties".
In sum, the Romanian examples in (51)a-b differ with respect to the presence of
closed or open wh-phrases (which underlies the (non)D-linking reading), but both
of them are grammatical in weak-island configurations. This indicates that weak
islands do not trigger a D-linking interpretation, which is expected under the
present proposal because nothing blocks a derivation that relies on both DR and
NPR. Recall that weak islands do block a representation relying on DR alone
because in this case the variable bound by the raised determiner does not occupy an
-position, and therefore cannot be assigned a referential th-role. This accounts for
the fact that the amount reading of how many is not available in weak island
contexts (see Section 7.7.2.1. above).
7.7. 3. 3. Clefts
Let us finally consider cleft sentences, which are currently analyzed as relying on
the presence of a null operator that licenses a variable in an A-position:
(53) It is John [Oi [Imetx i last year]].
Cleft sentences freely violate weak islands:
(54) a. It is John that I regret to have loved.
b. It is John that I did not love.
c. It is John that I wondered where I met
According to Cinque's (1989, 1990) generalization, weak-island violations correlate
with the referentiality of the wh-phrase; since null operators can by no means be
assumed to be referential, Cinque considers that the relevant referential element is
the antecedent of the null operator itself. I find this suggestion stipulative. Empirical
evidence against the idea that clefts rely on D-linking will be presented in the
following section.
7. 7. 4. The long wh-movement ofD-linked adjuncts
In sum, it has been shown above that certain non-D-linked wh-phrases may be
extracted out of weak islands. The relevant cases are bare quantifiers, clefts, and
wh-structures that rely on DR + NPR. This data shows that (55), which is Cinque's
(1990) extension of Comorovski's (1989) generalization, is too strong:
(55) Only referential /D-linked wh-phrases can be extracted from weak islands.
The only type of non-referential reading that is not available in weak island
contexts, is the amount reading.41 Rizzi's (1990) account of weak islands,
combined with the DR hypothesis proposed here, captures this data correctly; no
proviso concerning the referentiality of the wh-phrase itself need be added.
41
The same generalization can be found in Szabolcsi and Zwarts (1990), Szabolcsi (1992) and
Frampton (1990).
What is not expected under Rizzi's theory, however, is that D-linking allows
even Adjuncts to get out of weak islands:
(56)
a. ?For which of these reasonsj do you wonder if they can re you ei?
(example by Szabolcsi, p.c.)
b. ?On which of these daysj doesn't Mary remember whether she was
wearing a white dress ei?
c. Tin which bookshop jdid you forget when you last bought a book ei?
This data can be captured by saying that D-linking licenses referential indices on
wh-phrases (recall that according to Rizzi, referential indices are needed for A'binding, and referential indices can only be licensed by referential th-roles). A better
understanding of these examples is needed, but for our present purposes the
empirical generalization is sufficient:
(57) D-linked Adjuncts can get out of weak islands.
It is now possible to give further evidence in favour of the point I made above, that
bare quantifiers and clefts do not rely on D-linking. If they did, we would expect
(a) bare adjunct wh-phrases and (b) clefted adjuncts to be allowed to escape weak
islands:
(58)
(59)
Since these examples are ungrammatical, we are forced to conclude that neither bare
quantifiers nor clefts rely on D-linking. In the absence of D-linking, the
extractability out of weak islands depends on the presence of a referential th-role;
since adjuncts do not present such a th-role, they cannot be extracted in (58)-(59).
Conclusions
(1) The referential th-role condition (individual variables) captures correctly the
partition of the data that can be observed in non-D-linking contexts (adjuncts,
complements without a referential th-role, and the amount interpretation versus
arguments).
(2) In D-linking contexts Rizzi's condition may be violated: even adjuncts escape
from weak islands.
(3) For wh-phrases bearing a referential th-role, D-linking does not make a
difference as to extractability out of weak islands.
(Cinque (1990))
(Cinque (1990))
Examples such as (61) constitute a strong argument in favour of the idea that nonreferential elements cannot get out of weak islands. It is however clear that this
generalization is not correct for bare wh-quantifiers, which extract out of weak
islands quite freely:
(62)
a.
b.
c.
d.
This type of derivation has already been examined (see the discussion of (5"') and
discarded. The theoretical reason is that a raised NP occupies a position which is
not canonically governed by the verb. The empirical reason is that by allowing (71')
we would be let to allow a similar derivation for prepositional accusatives, which
would have the undesirable result of predicting that the existential reading would be
possible in that case too. 42
42
This outcome might be avoided by saying that a derivation comparable to that in (71') is illicit
for prepositional Accusatives because in this case the preposition is "dummy", and as such it does
not L-mark the NP; therefore, in a representation that would differ from that in (7) by the
presence of dummy pe (instead of the lexical devant 'in front o f ) the raising of the quantifier
would still be blocked. Note, however, that this cannot be the correct solution. To account for the
existential reading of (70) we would have to assume (7Cf):
We are now faced with the conclusion that nonreferentiality cannot be tied to the
LF rule of DR, and this seems devastating for the present proposal. Fortunately,
this conclusion, which is probably correct, does not, in fact, destroy our account. It
seems true that nonreferentiality must be tied to DR only for direct objects; DR
does not seem to be relevant for indirect objects and prepositional phrases. This is
not a very serious problem because it is known that indirect objects (and
prepositional phrases) behave differently from direct objects with respect to
quantificational structures (e.g., variables bound by open NPs can be clitic doubled;
for further discussion see Chapter 6).
Appendix 2: Strong determiners
The main purpose of this chapter was to analyze empirical data in which the
semantic interpretation of NPs appears to be partially determined by the syntactic
configuration. For reasons to be specified below, the interpretation of strong NPs is
less sensitive to syntactic configurations.
Following Heim (1982, p. 133), let us assume that LF configurations presenting
every have the tripartite structure shown in (72)b:43
(72)
e; student
I examined ej
(70')
But here the preposition is dummy (see Vergnaud (1985)), and as such should block DR, on a par
with the Romanian pe.
"^Recall that this representation is derived by NPR + DR within Heim's framework, and by DR
+ NPR within the framework proposed here. This difference is not relevant for the present
discussion.
cannot be rewritten as unrestricted quantification.44 For the same reason, DR + reincorporation cannot affect strong determiners.
Thus, the only possible representation for (72) appears to be the format of
restricted quantification shown in (72)b-c, which gives rise to the presuppositional
reading. The fact that strong NPs are necessarily interpreted as presuppositional is
clearly indicated by the well-known test of there contexts.45 More precisely, therecontexts indicate that strong NPs can be assigned neither the existential nor the
amount interpretations. The two possibilities left open, then, are the
presuppositional and the referential readings. The Weak Cross Over test indicates
that certain strong NPs, e.g., most ', or every N\ cannot be assigned the
referential reading,46 but only the presuppositional one:
(73) a. *Hisj mother loves every childj.
b. * Theirj mothers love most childrenj.
"Strong" NPs such as every and most thus clearly indicate that referentiality and
presuppositionality should be kept distinct. This does not mean that all strong NPs
are necessarily non-referential. Certain strong NPs can indeed be clitic-doubled,
which forces the referential reading (Weak Cross Over violations disappear).
To summarize, the rule of DR is sensitive to the semantic properties of
determiners. DR alone may affect weak determiners, but not strong determiners
(because they are not symmetric, strong determiners are not compatible with the
representation resulting from DR alone). As any rule of grammar, DR is optional,
and as such it does not apply in those cases in which its application would lead to
illicit representations. Also, identical LF representations (see the one obtained via
DR + NPR) may give rise to distinct semantic representations (restricted or
unrestricted quantification), depending on the presence of strong or weak
determiners. It thus appears that the interpretation of strong NPs is largely
determined by the semantic properties of their determiners. This is not the case of
weak NPs, which allow four distinct representations, depending on the rules that
map S-structures onto LFs.
^Every seems to allow the format of unrestricted quantification: Vx [x is a student = > John met
x]. Note however that the logical connective "implies" must be inserted, and one may assume that
the rules that map syntax to LF cannot insert logical connectives (the conjunctive connective is an
exception: it can be viewed as a natural translation of juxtaposition or adjunction configurations,
and as such it may be introduced at LF). Clearer cases in point do exist: most, for instance, can
only be represented in terms of restricted quantification.
45
The relevant type of example is well-known:
(i)
*There came most students.
(ii) There came two students.
46
T h e idea that strong and weak NPs do not differ from each other with respect to
"referential/quantificational" status contradicts the taxonomy proposed in Chapter 6, where strong
NPs are taken to function specifically/referentially. Since my main purpose was (and still is) an
analysis of weak quantifiers, this mistaken view did not affect the main results.
Bibliography
Abney, S. 1987, The English Noun Phrase In Its Sentential Aspect, PH.D., MIT,
Cambridge, Mass.
Afarli, T. A. 1989, "Passive in Norwegian and in English", Linguistic Inquiry
20:1,101-108.
Akmajian, S. and T. Wasow 1975, "The Constituent Structure of VP and Aux and
the Position of the Verb BE", Linguistic Analysis 1, 205-245.
Akmajian, S., S. Steele and T. Wasow 1979, "The Category AUX in Universal
Grammar", Linguistic Inquiry 10.1,1-64.
Aoun, Y. 1981, The Formal Nature of Anaphoric Relations, M.I.T. Thesis.
Aoun, J. 1985, A Grammar of Anaphora, MIT Press, Cambridge, Mass.
Aoun, Y. and D. Sportiche 1983, "On the Formal Theory of Government", The
Linguistic Review 2,3,211-236.
Aoun, Y. and Li, T.A. 1989, "Scope and Constituency", Linguistic Inquiry 20,2,
141-173.
Avram, L. 1988, '"Verb auxiliar' in limba romn: categorie lexical sau morfem
gramatical?", Studii cercetri lingvistice XXXIX, 5,439-445.
Ayoub, G. 1982, Structure de la phrase verbale en arabe standard, Third Cycle
Thesis, Paris VHI-Vincennes
Azoulay, A. 1985, Les tours comportant l'expression DE+ adjectif, Librairie Droz,
Geneve-Paris.
Baker, M. 1988, Incorporation: A Theory of Grammatical Function Changing, The
University of Chicago Press, Chicago.
Baker, M., Johnson, K. and Roberts, I. 1989, "Passive Arguments Raised",
Linguistic Inquiry 20, 2,219-253.
Bartsch, R. 1973, "The Semantics and Syntax of Number and Numbers", in J.P.
Kimball, ed., Syntax and semantics 2. New York and London: Seminar Press.
Barwise, J. and R. Cooper 1981, "Generalized Quantifiers and Natural Language",
Linguistics and Philosophy 4, 159-219.
Belletti, A. 1982, "'Morphological' Passive and Pro-drop: the Impersonal
Construction in Italian", Journal of Linguistic Research 2.1.
Belletti, A. 1988, "The Case of Unaccusatives", Linguistic Inquiry 19, 1-34.
Belletti, A. 1990, Generalized Verb Movement, Rosenberg e Sellier, Torino.
Bennis H. and T. Hoekstra 1988, 'The Tense Connection', GLOW Newsletter,
Foris, Dordrecht.
den Besten, . 1977, "On the Interaction of Root Transformations and Lexical
Deletive Rules", unpublished paper, University of Amsterdam [publ. version:
den Besten (1983)].
den Besten, . 1983, "Root Transformations and Lexical Deletive Rules", in
Abraham, W. (ed.), On the Formal Syntax of the Westgermania, Benjamins,
Amsterdam/ Philadelphia.
Blinkenberg, A. 1960, Le probleme de la transitivit6 en frangais moderne,
Munsksgaard.
Bonet i Alsina, . E. 1991, Morphology after syntax: pronominal clitics in
Romance, MIT dissertation.
Borer, H. 1981, Parametric Variation in Clitic Constructions, M.I.T. dissertation.
Borer, H. 1984, Parametric Syntax, Dordrecht, Foris (publication of Borer (1981)).
Bibliography
279
Cinque, G. 1986, "Bare Quantifiers, Quantified NPs and the Notion of Operator at
S-structure", Rivista di grammatica generativa, vol. 11.
Cinque 1988, "On Si Constructions and the Theory of ARB", Linguistic Inquiry,
19, pp 521-581.
Cinque, G. 1989, "'Long' Wh-movement and Referentiality", paper presented at the
Second Princeton Workshop on Comparative Grammar, Princeton University,
april 1989.
Cinque, G. 1990, Types of A'-Dependencies, Linguistic Inquiry Monograph 17.
Comorovski, I. 1986, "Multiple Wh Movement in Romanian", Linguistic Inquiry,
171-177.
Comorovski, I. 1989a, Discourse and the Syntax of Multiple Constituent
Questions, Ph.D. dissertation, Cornell Unversity.
Comorovski, I. 1989b, "Discourse-Linking and the Wh-Island Constraint", in
Proceedings of the 19th Annual Meeting, NELS.
Contreras, H. 1987, "Small clauses in Spanish and English", Natural Language
and Linguistic Theory 5, 225-244.
Couquaux, D. 1979, "Sur la syntaxe des phrases pridicatives en fran^ais",
Linguisticae Investigation's III.2, 245-284.
Couquaux, D. 1981, "French Predication and Linguistic Theory", in R. May and J.
Kster, eds., Levels of Syntactic Representation, 33-64, Foris, Dordrecht.
Culicover, P. 1976, Syntax, Academic Press, New York.
Dasgupta, P. 1988, "Bangla Quantifier Extraction, Unaccusative in Situ, and the
ECP", Linguistic Inquiry 19, 4.
Davidson, D. 1966, 'The Logical Form of Action Sentences', in D. Davidson 1980.
Davidson, D. 1980, Essays on Actions and Events, Clarendon Press, Oxford.
Demonte, V. 1987, "C-command, Prepositions and Predication",
Linguistic
Inquiry, 18,1, 147-157.
Diesing, M. 1990, The Syntactic Roots of Semantic Partition, Ph.D. dissertation,
Amherst.
Dobrovie-Sorin, C. 1985, "Redoublement clitique, relatives et interrogatives en
roumain et espagnol", Linguisticae Investigationes IX:2, 269-307.
Dobrovie-Sorin, C. 1986, "A propos du contraste entre le passif morphologique et
se moyen dans les tours impersonnels", Linguisticae Investigationes X:2, 289330.
Dobrovie-Sorin, C. 1987, Syntaxe du roumain. These de Doctorat d'Etat,
Universite de Paris 7.
Dobrovie-Sorin, C. 1988, "What Does QR Raise", Logical Form (Program and
Abstracts), Tilburg University.
Dobrovie-Sorin, C. 1990a, "Clitic Doubling, WH Movement and Quantification in
Romanian", Linguistic Inquiry 21, 3.
Dobrovie-Sorin, C. 1990b, "Auxiliaries and Sentence Structure in Rumanian", in
Mascaro and Nespor, eds. Grammar in Progress. GLOW Essays for H. van
Riemsdijk, Foris Publications.
Dobrovie-Sorin, C. 1991, "Mouvement wh et relations de quantification", in J.
Gueron et J.-Y. Pollock, eds. Grammaire generative et syntaxe comparee, Ed. du
CNRS, 15, quai A. France, Paris.
Dobrovie-Sorin, C. 1992, "LF Representations, Weak Islands and the ECP",
GLOW abstract.
Dubuisson, C. and J. Goldsmith 1976, "A propos de l'inversion du clitique sujet en
frangais", in A. Ford, J. Reighard and R. Singh eds., Actes du Sixieme Congres
Bibliography
281
Bibliography
283
Huang, J. 1984, "On the distribution and reference of empty categories", Linguistic
Inquiry, 15.
Huang, J. 1989, "PRO-drop in Chinese", in O. Jaeggli and K. Safir eds., The Null
Subject Parameter.
Huot, H. 1977, Recherches sur la subordination en frangais, These de Doctorat
d'Etat, Universite Paris 7.
Huot, H. 1981, Constructions infinitives du frangais: le subordonnant 'de', Droz,
Geneva.
Hyams, N. 1986, "The Interpretation and Acquisition of Italian Impersonal s f , in
H. Borer, (ed.),Syntax and Semantics 19, 219-240.
Hurtado, A. 1984, "On the Properties of LF", Cornell Working
Papers in
Linguistics, 5.
Jackendoff, R. 1972, Semantic Representation in Generative Grammar, MIT
Press.
Jackendoff, R.S. 1977, X' Syntax, MIT Press, Cambridge, Mass.
Jaeggli, O. 1980 "Remarks on To Contraction", Linguistic Inquiry 11.1, 239-245.
Jaeggli, O. 1982, Topics in Romance Syntax, Foris Publications, Dordrecht.
Jaeggli, O. 1986a, "Passive", Linguistic Inquiry 17, 587-622.
Jaeggli, O. 1986b, "Three Issues in the Theory of Clitics", in H. Borer (ed.),
Syntax and Semantics 19, Academic Press, New York.
Jaeggli, O. and K. Safir 1989, eds. The Null Subject Parameter, Kluwer.
Jakubowicz, C. 1985a, "Do Binding Principles Apply to Infi?", Proceedings of the
Northeast Linguistic Society 15, 183-206.
Jong, F.M.G. de, and H. J. Verkuyl 1985, "Generalized Quantifiers: The
Properness of Their Strength", in J. van Benthem and A. ter Meulen, eds.,
Generalized Quantifiers in Natural Language, Dordrecht, Foris.
Jong, F.M.G. de 1987, "The Compositional Nature of (In)definiteness", in E. J.
Reuland and A. G. B. ter Meulen, eds.
Kaisse, E. 1985b, Connected Speech: the interaction of syntax and phonology,
Academic Press, New York.
Kaisse, E. 1985, "Locating Turkish Devoicing", Proceedings of the WCCFL 5,
119-128.
Kayne, R.S. 1969, The Transformational Cycle in French Syntax, Ph. D M.I.T.
Kayne, R.S. 1975, Syntaxe du francais. Le cycle transformationnel, Paris, Le
Seuil.
Kayne, R. S. 1975 French Syntax, MIT Press, Cambridge, Mass.
Kayne, R. S. 1982, "Predicates and Arguments, Verbs and Nouns", GLOW
abstract, Paris.
Kayne, R. 1983, "Chains, Categories External to S, and French Complex
Inversion", Natural Language and Linguistic Theory 1, 109-37.
Kayne, R. 1984, Connectedness and Binary Branching, Foris Publications,
Dordrecht.
Kayne, R. 1987 "Facets of Romance Past Participle Agreement", ms., MIT.
Kayne, R. 1989 "Null Subjects and Clitic Climbing", in O. Jaeggli and K. Safir,
eds. The Null Subject Parameter, Dordrecht, Reidel.
Kayne, R. 1991, "Romance Clitics, Verb Movement and PRO", Linguistic Inquiry
22,4, 647-687.
Kayne, R. and J.-Y. Pollock 1978, "Stylistic Inversion, Successive Cyclicity, and
Move NP in French", Linguistic Inquiry 9, 595-621.
Bibliography
285
Manzini, . R. 1983, "On Control and Control Theory", Linguistic Inquiry 14,3,
421-467.
Manzini, M.R. 1986, "On Italian Si", in H. Borer (ed.), Syntax and Semantics 19:
The Syntax of Pronominal Clitics, Academic Press, New York.
Marantz, A. 1984, On the Nature of Grammatical Relations, Press.
Marantz, A. 1988 "Clitics, Morphological Merger and the Mapping to
Morphological Structure", in M. Hammon and M. Noonan eds., Theoretical
Morphology, Academic Press Inc., pp 253-270.
May, R. 1977, The Grammar of Quantification, M.I.T. Thesis.
May, R. 1985, Logical Form. Its Structure and Derivation, Press.
Milner 1982, Ordres et raisons de langue, Le Seuil, Paris.
Milner, J.-C. 1986, Introduction un traitement du passif, coll. ERA 642,
Universiti Paris 7.
Milner 1989, Introduction une science du langage, Le Seuil, Paris.
Milsark, G. 1977, "Toward an Explanation of Certain Peculiarities in the Existential
Construction in English", Linguistic Analysis 3, 1-30.
Mohanan, . T. 1985 "Remarks on Control and Control Theory", Linguistic
Inquiry 16,4,637-649.
Motapanyane, V. 1988, "La position du sujet dans une langue l'ordre
SVO/VSO", Rivista di grammatica generativa 14, 75-103.
Napoli, D. J. 1973, The two si's of Italian, Ph.D. dissertation, MIT.
Naro, A. 1976, "The Genesis of the Reflexive Impersonal in Portuguese: a Study in
Syntactic Change as a Surface Phenomenon, Language 52.4
Nilson, E. 1969, Les termes relatifs et les propositions relatives en roumain
modeme, Lund.
Obenauer, . 1984-1985, "On the Identification of Empty Categories", The
Linguistic Review 4,153-202
Obenauer, H.-G. 1992 a, Aspects de la syntaxe A-barre- Effets d'intervention et
mouvement des quntieurs, these d'Etat, Universite de Paris 8.
Obenauer, H.-G. 1992 b, "L'interprtation des structures wh et l'accord du participe
pass", in H.-G. Obenauer et Anne Zribi-Hertz eds., Structure de la phrase et
thorie du liage, Actes du colloque international du GDR 120 (CNRS), 9-11
mars 1990, Presses Universitaires de Vincennes.
Ouhalla, J. 1988,The Syntax of Head Movement, Study of Berber, PH.D.,
University College, London.
Ozkaragoz, I. 1980, "Evidence from Turkish for the unaccusative hypothesis", BLS
6,411-422.
Pan-Dindelegan, G. 1974, Sintaxa transformafional a grupului verbal in limba
romn, Bucuresti, Ed. Acad. RSR.
Partee, B. 1972, "Opacity, coreference, and pronouns", in D. Davidson and G.
Harman eds. Semantics of Natural Language, Reidel, Dordrecht, 415-441.
Partee, . H. 1973, 'Some Structural Analogies between Tenses and Pronouns in
English', The Journal of Philosophy 70, 601-609.
Partee, . H. 1984,' Nominal and temporal anaphora", Linguistics and Philosophy
7:2, 243-286
Partee, . H. 1988, "Many Quantifiers", Proceedings ofESCOL.
Perlmutter, D. 1972 "Evidence for Shadow Pronouns in French Relativization", in
The Chicago which Hunt, C.L.S.
Bibliography
287
Bibliography
289
Index of Names
Abney, S 56,242
Afarli, T 193, 194
Akmajian, S., 1, 45
Alsina, . E., 61
Aoun, ., IX, 50, 198, 200, 217, 224,
231
Avram, L, 46
Ayoub, G 105
Azoulay, ., 243
Baker, ., , XIII, 9, 13, 29, 38, 51,
52, 61, 128, 131, 139, 153, 194
Bartsch, R 246, 255
Barwise, J. 206, 226
Belletti, ., , XII, XIII, 8, 15, 51, 89,
130, 139, 168, 169, 170, 175, 178,
185, 195
Bennis H.,1
den Besten, H 12, 20, 22, 50, 52, 54,
96, 102, 105
Blinkenberg, ., 136
Bonet i Alsina . E., 61
Borer, ., VIII, 50, 93, 110, 114, 120,
198, 200, 217, 250
Bouchard, D 93, 114, 115, 120
Bredemeier, J., 23, 70
Bresnan, J., 106
Burzio, L 130, 136, 147, 169, 185,
Cardinaletti, ., 166, 242
Chomsky, N 1, 4, 8, 16, 24, 29, 30, 31,
37, 39, 45, 53, 58, 61, 80, 90, 91,
111, 114, 118, 120, 121, 130, 137,
141, 142, 151, 153, 168, 169, 180,
184, 188, 191, 196, 199, 200, 201,
203, 208, 211, 218, 237, 238, 241,
244, 253, 264
Chung, S 261,
Cinque, G IX, , XIV, 98, 128, 154,
183, 184, 185, 199, 206, 211, 218,
219, 220, 228, 233, 238, 240, 241,
243, 264, 265, 266, 268, 271, 273,
Comorovski, I., 199, 266, 268, 271,
Contreras, H., 37
Cooper, 206, 226
Couquaux, D 148, 175,
Culicover, P., 45
Dasgupta, P. 235, 253
Davidson, D., 143
Demonte, V., 238,
Diesing, M 249, 252
103,
219,
253,
163
103,
264,
121,
193,
Pesetsky, D XIV,
266
Picallo, C., 120, 124, 126
Pinker, S VIII,
Platzack, C., 4, 28, 100, 167
Pollock, J.-Y., XII, 4, 6, 7, 8, 12, 50, 82,
83, 84, 88, 92, 100, 103, 140, 160,
163, 164, 193, 200, 274, 253
Postal, P., 176
G.K. Pullum, 49, 51
Raposo, E., 15, 135
Reinhart, T 54, 90, 110, 203, 225, 226,
246, 247, 249, 255
Reuland, E., 141, 180
Riemsdijk, H 233, 242, 244
Rivero, M.-L., XII, 1, 11, 14, 21, 27, 28,
29, 30, 32, 39, 52, 64, 68, 69, 74, 79,
97, 103, 120, 126
Rivifere, N 130, 142
Rizzi, L 7, 12, 25, 29, 32, 50, 54, 61,
77, 82, 87, 91, 95, 96, 114, 120, 132,
133, 135, 136, 137, 140, 142, 145,
147, 148, 151, 152, 167, 169, 171,
184, 192, 199, 265, 266, 267, 268,
271
Roberts, I., , XIII, 7, 9, 16, 25, 29, 38,
47, 50, 52, 53, 54, 56, 61, 67, 82, 96
Rochette, ., 51, 123
Ross, J. R., 1, 45
Rouveret, ., 1, 52, 87, 107, 135
Ruwet, N 120, 125, 155
Sadock, J.M., 49, 103
Safir, K 106, 126
Sag, I., 236
Sandfeld, K 178, 197, 214, 239
Scha, R 246, 255
Shibatani, M.
Simatos, I., 155, 156
Sloan, K 231
Sobin, N 193
Speas, M., 37
Sportiche, D XI, 9, 37, 88, 92, 203
Stavi J., 247
Steele, 1, 45
Steriade, D 198, 200, 210, 217, 239,
253
Stowell, T 84, 105, 158
Sufier, M 120, 200, 216
Swart, H 274
Szabolcsi, ., 268, 270, 271, 274,
F. Zwarts , 268, 271
Thomson S.A., 136
Tiktin, ., 80
Tomaselli, ., 167
Subject Index
inherent, 238
nominative, 90
objective, 238
C-command, 54, 147
and scope, 231
Chains
and indexing, 150, 154, 171, 187
and raising, 119
clitic, 190
links of, 120, 125
passive se, 190
Clefts, 214, 271
Clitic, XI, 49, 65, 73, 95
adverbs, 11, 26, 42, 63, 66, 77
auxiliaries, 15, 46
chains, 190, 192
chains and reconstruction, 190
climbing, 76
clusters, 62-64, 78
traces, 174, 189
doubling, 90f, 192, 197, 199, 228
endoclitic pronouns, 78
host of, 51
Clitic Left Dislocation, 240
licensing of, 56,190
nominative, 184
phonological, 49, 70
placement, 50, 53
pronouns, 21, 60
R-clitics, 201f
splitting, 23
syntactic, 49
and V-to-I movement, 47
verb-clitic order, 52, 57, 60, 67, 96
COMP, 84, 95, 127
COMP/Infl adjacency, 87, 98
COMP/Infl ambiguity, 87, 93, 98-99,
106
doubly filled, 86, 107
Complementizer
indicative (c), 104, 124
infinitive (a), 82, 109
subjunctive (ca, s), 109, 123f
deletion, 97
Complex NP Constraint, 236
Control, 36, 91-93, 114-115, 125
obligatory, 117
of infinitival complements, 91
of subjunctive complements, 112
Crossover
strong, 203f
weak, 202, 248, 250, 252, 263, 277
Detransivization, 245
Determiner, 242
as operator, 261
cardinal, 246
partitive, 247
Phrase, 242
Raising, 242, 250
strong, 242,276
symmetric, 247
weak, 242, 247
Discourse Linking, 207, 248, 252, 258,
268ff, 272
D-linked Adjuncts, 271
D-linked WH-phrases, 271
Dutch, 119
ECP, see Empty Category Principle
Empty Categories
A'-bound, 260
identification of, 188
Empty Category Principle, 29, 237, 253,
263, 265, 268
and Long Head Movement, 28, 32, 57
Enclisis, 70, 75, see also Clitic
English, 1-6, 10, 12ff, 21, 36ff, 47, 81ff,
88ff, 112ff, 163
Excorporation, 53, 67
Extended Projection, 7f, 34, 103
Feature percolation, 204, 228, 242
locality constraints on, 208
Floating quantifiers, 14
Focus, 220
Movement, 244
French, XIII, 1, 6, 9-13, 26, 32, 36-38,
46-49, 52, 63, 80, 82, 84, 87, 95, 99,
106, 113, 116, 125, 128-130, 133136, 140-150, 154-157, 162-175,
179, 181, 193, 199, 205, 211, 218,
224, 243, 253, 260ff, 265, 270
Full Interpretation Principle, 141, 175
Functional
adjunction, 17
categories,
coindexation, 17, 33, 40, 56, 65, 88f,
101, 104, 123
projections, 7, 18
German, 165
Governing Category, 121
Government, 30
antecedent, 29, 265
canonical, 275
head, 265
theta , 30
direction of, 92f
proper government, 29
Governor
minimal, 123, 127
Subject Index
L-marking, 252
and th-government, 30
and Tense-marking, 30
Left dislocation, 109
and scope phenomena, 222
Clitic, 218
standard,218
Lexicalist Hypothesis, 50
Logical Form (LF), 196, 223, 242
and semantic interpretation, 225ff, 243
and scope assignment, 230
Merging, 18, 33, 57-59, 80
Modals, 2, 24
Move a, 217, 241,243
Minimality
relativized, 33, 265
Negation, 6, 23f, 55, 58f, 68, 84, 95,
100-102
Noun Phrases
amount reading of, 244, 277
closed, 248
existential reading of, 247, 275
interpretation of, 222, 243-247
NP Prefixing, 243
NP Raising, 233, 243
open, 245
partitive, 252
presuppositional reading of, 246, 277
referential reading of, 248
quantified, 223, 242
strong, 206f, 228,
weak, 206f, 224, 228
Nuclear Scope, 276
Null Operators, 213, 260
Obviation, 113, 120-121
lack of, 124, 125
Operators, 241, 249, 262
Parameters, Vllf, XIII
and levels of representation, VIII
and the Lexicon, VIII
null subject parameter, 135, 146ff, 167
structural quantifier parameter, 213
Parasitic Gaps, 203
Passives, X
and case absorption, 153, 194
and Case Assignment, 193
and the Case Filter, 131
and Case Theory, 129ff, 153
and external theta-roles, 129
and null prototypical objects, 131 f
copula, 128, 153, 168
impersonal, (in German), 165
with intransitives, 129ff, 137, 153
with transitives, 129ff, 137
295
296
The syntax of
Romanian