Sei sulla pagina 1di 28

Remarks and Replies

Movement Out of Control Idan Landau


This article is a comprehensive critique of the reductionist view of control advocated in recent minimalist studies, most notably Hornstein 1999. The core of this view is the claim that obligatory control should be collapsed with raising, and nonobligatory control with pronominal coreference. I argue that Hornsteins theory (a) overgenerates nonexisting structures and interpretations, (b) fails to derive a wide range of well-known raising/control contrasts, and (c) involves unstated stipulations belying the appeal to Occams razor. Keywords: control, raising, NP-movement, A-chains, Occams razor

1 Introduction Recent treatments of control within minimalism share the conviction that obligatory control (OC) is to be understood in terms of syntactic movement (Martin 1996, ONeil 1997, Hornstein 1999). In its most radical form, this conviction amounts to the thesis that OC should be collapsed with raising, both instantiating NP-movement. This is the position advanced by Hornstein (1999). It is held that strong conceptual considerations, such as simplicity of the grammar and avoidance of redundancy, argue for eliminating both PRO and the control module from the grammar. This can be done, the thesis goes, without any loss of empirical coverage. In this article, I take issue with this reductionist view of control. The specific model that I examine is the one proposed by Hornstein (1999). The choice is motivated both by the explicitness of this model and by its radical ambitions. Though the issues that arise can be most sharply discerned and addressed by contrasting this model with alternative, more traditional views of control, most of my critique is applicable to the other proposals mentioned above, insofar as they reduce OC to movement. In arguing against Hornsteins proposal, I will not try to defend any particular theory of control; rather, my critique will rest upon minimal assumptions that are shared by the majority of approaches to control, but crucially, not by Hornsteins. This is feasible precisely because for all their differences, those approaches recognize some fundamental split between raising and control, a split that Hornsteins theory minimizes to triviality.

I would like to thank Danny Fox, Sabine Iatridou, and two anonymous reviewers for comments and suggestions. During work on this article, I was supported by the Kreitman postdoctoral fellowship at Ben Gurion University.
Linguistic Inquiry, Volume 34, Number 3, Summer 2003 471498 2003 by the Massachusetts Institute of Technology

471

472

REMARKS AND REPLIES

The arguments to follow fall into three categories. First, I will show that Hornsteins system overgenerates nonexisting structures and interpretations, most of which must be independently blocked. Second, I will show that the claimed empirical benefits of the theory are less than compelling, and in fact undermine some of its assumptions. And third, I will show that a wide empirical range of raising/control contrasts (unmentioned by Hornstein) makes an overwhelming case against the reductionist view. The concluding remarks address some general methodological issues that underlie the debate. 2 Hornsteins Theory Hornstein develops his theory of control in response to what he perceives to be eliminable inadequacies in previous accounts, specifically the Government-Binding (GB) and the early minimalist (null Case) theories. Those, he claims, are either unmotivated stipulations or redundant mechanisms. In the first category, Hornstein lists (a) the requirement that an argument chain bear exactly one -role; (b) the ban on movement to a -position; (c) the ban on movement to a nonc-commanding position; (d) the view that -roles are not checkable features; (e) the postulation of the dubious null Case. In the second category, he lists (a) the distinction between PRO and traceindeed, the very idea that PRO exists; (b) hence, the traditional wisdom that raising and control are substantially different; (c) the control module, that is, the part of grammar that is responsible for the choice of controller in OC and the interpretation of PRO in general. The theory proposed by Hornstein dispenses with all those assumptions. Instead, it paints the following picture. -roles are features transferred from predicates to DPs, and there is no limit on the number of -roles a DP can accrue. There is no null Case; the subject position of the infinitive is Caseless. Control is nothing more than raising to a -position. As such, it is subject to the Minimal Link Condition (MLC). This result is supposed to capture the Minimal Distance Principle (MDP), that is, the lack of subject control across an object, at least as a markedness constraint. Thus, both PRO and the control module are eliminated. A sample derivation of OC is given in (1) (expanded from Hornsteins (45)). (1) a. Mary hopes to win. b. [ IP Mary [ VP Mary v hopes [ IP Mary to [ VP Mary v win]]]] The array contains a lexical DP (Mary), embedded and matrix predicates (win and hope, respectively), and appropriate functional heads (light vs, infinitival to, finite I(nfl), etc.). Mary first merges with the embedded vP, checking the external -role of win. It then raises to the embedded Spec,IP position, checking the EPP feature of to. Next, it raises to the matrix vP, checking (i.e., receiving) another -role, this time that of hope. Finally, it raises to the matrix Spec,IP position, where it checks nominative Case and EPP features. All the intermediate steps leave full copies of the raised DP (Hornstein adopts the minimalist view of traces as copies). The resulting chain contains two -roles and one Case. Object control into complements works in a similar fashion. The second type of OC that Hornstein considers is subject control into an adjunct. A sample derivation is given in (2) (Hornsteins (41)).

REMARKS AND REPLIES

473

(2) a. John heard Mary without entering the room. b. [ IP John [ I past [ VP/VP[ VP John [heard Mary]] [ Adjunct without [ IP John [ I ing [ VP John [entering the room]]]]]]]] The adjunct is built up first, with John checking the external -role of enter and then raising to Spec,-ing. In the matrix clause, Mary merges with heard, checking the internal -role. This phrase then merges with the adjunct, forming an adjunction structure.1 At this point, John raises from the adjunct to the matrix Spec,VP, checking the external -role of heard, and finally to the matrix Spec,IP, checking nominative Case and EPP features. At LF, Mary raises to the outer Spec,VP (or AgrO ) to check accusative Case. Hornstein shows that all alternative derivations from the same array fail to converge, including object control. The raising of John from the adjunct targets a position that does not c-command (nor is c-commanded by) the adjunct. This is an instance of sideward movement, allowed by assumption. Mary does not c-command the adjunct either. If the MLC only compares distances between nodes that stand in a c-command relation to one another, then the raising of John does not violate the MLC (unlike the case of object control into complements, where the MLC is operative). As for the very possibility of escaping an island, Hornstein refers to Nunes 1995, where such movements are sanctioned, provided they are sideward movements. I return below to this issue. Hornsteins theory of nonobligatory control (NOC) is very simple. Essentially, NOC applies in all environments where raising is impossible (e.g., from subject gerunds). In such cases, a last resort operation saves the structure by inserting a small pro in the controlled position, much like the operation of do-support. The presence of pro explains the pronominal properties of NOC. This suffices as a general overview of Hornsteins theory. There are more details to be filled in, which I will discuss when considering their empirical consequences. In evaluating Hornsteins theory, I will contrast it with what I take to be the standard view of control. This view consists of the following assumptions: (3) The standard view of control a. PRO exists, and it is distinct from NP-trace. b. Hence, control involves two argument chains, while raising involves one. c. The control module exists. The assumptions in (3) are as minimal as one can get. Thus, they are not committed to any particular view of PRO (anaphor, pronoun, A variable, etc.)only to its nonidentity with NPtrace. Nor are they committed to any particular view of the control modulethe choice of controller and the interpretation of PRO may involve predication, lexical entailments, pragmatic factors,

1 The structure in (2b), Hornsteins (41), in fact reflects a different derivation from the one Hornstein describes in prose (and in his (40))namely, one in which the raising of John precedes adjunction. This vagueness is not crucial, as Hornstein indicates (footnote 31) that he takes no position on the issue of movement between unconnected subtrees.

474

REMARKS AND REPLIES

and so ononly to the existence of some component of grammar, distinct from the MLC, that determines these properties. Doubtless, there exist theories of control other than Hornsteins that deny some of the ingredients in (3) (although, as far as I know, no theory denies all of them). Yet it is safe to say that the vast majority of control theories, certainly those situated within the GB tradition, accept this much (see, e.g., Chomsky 1981, 1986b, Bresnan 1982, Manzini 1983, Bouchard 1984, Koster 1984, Lebeaux 1984, Borer 1989, Clark 1990, Sag and Pollard 1991, Williams 1992, Petter 1998, Landau 2000). As I will argue, nothing more is needed in order to counterpose to Hornsteins theory, a coherent alternative that is superior over a wide range of empirical phenomena. 3 Internal Problems Just as Hornsteins main arguments against the standard view are conceptual, so one may adduce conceptual arguments against Hornsteins own proposal. Most significantly, Hornsteins claim that -roles are checkable features, on a par with -features (see also Manzini and Roussou 2000), introduces massive redundancy into the grammar: all thematic information, previously thought to be accessed only at the semantic interface, is now mirrored by a system of formal features, accessed throughout the derivation. Moreover, unlike any other morphological feature, the purported -features are clearly relational. A DP is plural by virtue of bearing a certain feature, regardless of its environment; however, whether it is a theme or an experiencer is entirely context dependent. Questions of convergence immediately arise. Given that only animate DPs can function as experiencers, does the feature mismatch in Sincerity fears John result in crashing? If the answer is positive, why do we have a clear sense that this deviance is fundamentally different from that of *Children fears John? If the answer is negative, what is left of the empirical content of the analogy between -features and -features?2 Nonetheless, I will not pursue this line of conceptual argumentation any further. The reasons are both methodological and empirical. I discuss my methodological reservations about conceptual arguments in the concluding remarks. Quite independently, there are enough empirical problems that should be discussed before conceptual issues are addressed. I realize that this way of prioritizing conceptual and empirical issues shifts the focus of discussion away from Hornsteins original intention; yet the proper prioritization too, I submit, should be open to debate. 3.1 Control across Passive The basic claim of Hornsteins analysis is that lexical DPs can freely move (A-movement) from the subject position of control infinitives, as in standard raising constructions. Thus, both (4a)

2 Aware of these complex issues, Hornstein claims (footnote 17) that verbs are grouped by adicity[; in] other words, verbs are categorized by their thematic status, giving rise to feature-based paradigms. Notice first that thematic categorization is not other words for adicity: verbs of the same adicity often assign different -roles, and conversely, the same -role can be assigned by verbs of different adicity. Even so, it takes more than thematic similarity to justify verb groupings. What is needed is a demonstration that the obtained grouping is linguistically significant. Thus, one would need to show that, for example, all verbs that assign a theme role, and only those verbs, participate in some grammatical process. As far as I know, research in lexical semantics has yielded very few generalizations of this form.

REMARKS AND REPLIES

475

and (4b) involve NP-movement (for convenience, I use the trace notation, although copies are intended). (4) a. John1 seems [ IP t1 to have [ VP t1 won the game]]. b. John1 [ VP t1 hopes [ IP t1 to [ VP t1 win the game]]]. According to Hornstein, the single difference between (4a) and (4b) is the extra step in the matrix VP of (4b), serving to pick the matrix agent -role. The grammar itself places no restrictions on the number of -roles a chain can bear. This system seems to generate, incorrectly, sentences like (5a), with the interpretation (5b) and derivation (5c). (5) a. *John was hoped to win the game. b. It was hoped that John would win the game. c. John1 was hoped [ IP t1 to [ VP t1 win the game]]. (5c) is structurally indistinguishable from (4a), namely, simple raising that creates a chain with one -role. The combination of raising and passivization of the embedded subject is of course attested in exceptional-Case-marking (ECM) constructions. (6) a. John was expected to win the game. b. John1 was expected [ IP t1 to [ VP t1 win the game]]. The question then is this: how can Hornsteins account rule out (5c) while still allowing (4b) and (6b)? In fact, the problem has been noted by Brody (1999), and Hornstein (2000) offers a solution. The idea is that unlike ECM complements, control complements are CPs, and CPs are phases, blocking extraction. Hornstein writes, Assume that some operation, say incorporation, can void the CP phase derivationally and say that this is prevented from occurring in passive verbs (p. 137). According to Hornstein, there is independent evidence that passives cannot support incorporated C, as they do not license that-deletion (7b); in that regard, they resemble adjectives (7c). (7) a. John fervently believes (that) theres a man here. b. Its fervently believed ??(that) theres a man here. c. Its unlikely ??(that) theres a man here. There are several problems with this argument. First, the facts are dubious. There is no general prohibition against that-deletion under passives/adjectives. (8) I was told/certain (that) Mary would come. Second, the appeal to phase voiding is obscure. Why should this be prevented from occurring under passive verbs? But suppose not only that the facts were solid, but also that a principled reason were given for the ban on C-incorporation onto passive verbs. These assumptions nullify Hornsteins analysis, given that control is compatible with passive in other environments. (9) a. Mary was persuaded to leave. b. Mary1 was persuaded t1 [t1 to leave].

476

REMARKS AND REPLIES

(10) a. It was decided to leave. b. It was pro1 decided [t1 to leave]. (9b) and (10b) are the structures that Hornsteins account would assign to the personal and impersonal passives in (9a) and (10a), respectively.3 The alleged ban on C-incorporation into passives cannot distinguish (5a) from (9a)/(10a). No doubt some modification can fill the gap, but at this point, I think, one should return to the standard account, until a more promising alternative is offered. To recall, according to that account, movement out of CP must pass through Spec,CP, which is an A-position, and movement from an A- to an A-position (the target of passive) is impossible (the ban on improper movement). If control involves no movement, then the issue of improper movement does not arise for (9a)/(10a). Only (5a), which has no control source, involves such movement, and is correctly ruled out. In short, if control is raising, and if DPs can raise across a passive verb, there is no way to exclude control across a passive verb. Hornsteins (1999) analysis overgenerates such structures, while his (2000) proposal solves the problem at the cost of undergenerating other structures. An anonymous reviewer proposes an alternative implementation of Hornsteins analysis to avoid these problems. The proposal extends Manzini and Roussous (2000) analysis of adjunct control as a parasitic gap configuration to complement control. The infinitive is generated as a VP specifier, the controller is generated as a complement, and verb raising yields the surface word order. (11) [ TP John1 tried2 [ VP[ CP e1 to leave] [ V t2 t1 ]]]. In this derivation, e1 and t1 satisfy connectedness and do not c-command each otherthe licensing conditions on parasitic gap constructions. (5a) is blocked since it violates connectedness: there is no matrix extraction to license the gap inside the left branch. This alternative overcomes one problem at the cost of raising at least five others. First, one wonders why classical A-movement does not license parasitic gaps, whereas the movement to subject in (11) does. In fact, standard A-movement can license two parasitic gaps simultaneously (12a), yet the A-movement in (11) fails to license a second parasitic gap (12b), suggesting that the first gap is also not parasitic. (12) a. Which report1 did John file [without reading e1 ] [in order for Mary to copy e1 ]? b. *John1 tried [[e1 to leave] tV t1 ] [in order for Mary to look for e1 ]. Second, the idea that the controller is generated as a deep object is tantamount to the claim that control verbs are unaccusativesurely incorrect for most control verbs. Furthermore, one would have to posit two distinct thematic mappings for identical verbs: the surface subject would be generated as a VP complement in John preferred to have dinner first, but as a VP specifier in John preferred dinner first. Third, the sequence tried to leave is not a constituent in structure

Notice that (10a) is not a case of NOC, but one of OC by an implicit agent.

REMARKS AND REPLIES

477

(11), leaving unexplained familiar VP-hood tests (e.g., Try to leave though John did, he could not find his way to the exit). Fourth, as recognized by the reviewer, connectedness is violated not only in (5a) but also in standard grammatical raising constructions (e.g., (6a)), where there is no thematic source for the raised DP in the matrix clause. A structural distinction between raising and control is then needed, which is independent of Hornsteins analysis. But this is exactly the point I made above; hence, the alternative version offers no advantage over the original one. Fifth, the same logic that blocks (5a) should block any extraction from the infinitivea left-branch islandthat is not accompanied by a matrix extraction, clearly a false prediction (13a). Conversely, matrix extractions should be able to license parasitic gaps in the embedded object position as well; after all, this is the canonical gap position in other parasitic gap constructions. This prediction, too, is disconfirmed: (13b) cannot be interpreted as (13c). (13) a. Who2 did [ TP John1 try [ VP[ CP e1 to meet t2 ] [ V tV t1 ]]]? b. *Who2 did [ TP John1 promise [ VP[ CP e1 to consult e2 ] [ V tV [ VP t1 tV t2 ]]]]? c. Who is the person X such that John promised X to consult X? Whether the infinitive is in a complement or a specifier position, the raising analysis of control seems to generate more problems than it solves.4 3.2 Sideward Movement from Complements A novelty in Hornsteins theory is the mechanism of sideward movement. The article itself does not elaborate on the properties of this mechanism. Long footnotes (32 and 33) hint at the theory of Nunes (1995), with possible extensions. Obviously, such a powerful mechanism needs to be considered with caution; however, the brief comments in Hornsteins article do not permit such consideration. So let us simply grant that grammars employ sideward movements and restrict our attention to the implications for control theory. Recall that Hornstein appeals to sideward movement in order to account for OC into adjuncts. Yet nothing in the mechanism of sideward movement itself dictates that it must take place from adjuncts. An obvious question to ask, then, is whether it takes place from complements. In other words, what rules out the following derivations, where the controller is embedded inside a matrix DP? (14) a. *Johns1 friends prefer [t1 to behave himself]. b. *We urged Johns1 friends [t1 to talk about himself]. c. *People biased against John1 constantly attempt [t1 to incriminate himself]. Hornstein does not address this question. As in the previous section, the problem here is one of overgeneration. The system allows mechanisms A and B to apply in different environments, but

4 In what follows, I continue to focus on Hornsteins original proposal, with occasional references to points where it (dis)agrees with the alternative implementation (11) (see footnotes 5, 7, 12, 15, 20).

478

REMARKS AND REPLIES

oddly we never encounter them in the same environment. Still, nothing in the way they are stated excludes this possibility. What is needed, then, is an independent principle C, which will constrain the operation of A and B to just the grammatical cases. Short of that principle, the postulation of either A or B can hardly be called explanatory.5 3.3 Implicit Controllers It is well known that controllers may remain implicit (Kimball 1971, Bresnan 1982, Epstein 1984, Williams 1985, 1987, Manzini 1986, Rizzi 1986a, Brody and Manzini 1987, Roeper 1987, Lasnik 1988, Chierchia 1989, Clark 1990, Higginbotham 1999, Landau 2000). The question is not whether implicit control exists, but how to deal with it. The familiar cases involve implicit passive agents or implicit goals. For the purposes of the present discussion, I restrict attention to the latter. Implicit dative control seems to be a universal option, particular languages differing in how pervasive the phenomenon is. Typical examples are the following: (15) English a. John said/shouted (to the visitors) to return later. Italian (Rizzi 1986a) b. Il generale ha ordinato (ai soldati) di partire. the general has ordered (to-the soldiers) DI to-leave The general ordered (to the soldiers) to leave. Hebrew c. Gil himlic (le-Rina) lirot rofe. Gil recommended (to Rina) to-see doctor Gil recommended (to-Rina) to see a doctor. Although no agreement has been reached concerning the proper analysis of these facts, some important insights have been established. Rizzi (1986a) has argued that implicit dative controllers should be represented in the lexicon (as thematic slots) but not in the syntax; specifically, they are not small pros. The reason for this distinction is that although implicit datives can control, they cannot bind anaphors (in nongeneric contexts). Hornstein does not discuss implicit control, but it should be clear that Rizzis distinction cannot be maintained in his system. If control is raising, then controllers must be syntactically represented, because only syntactic entities can be moved. This forces Hornstein to reduce all cases of implicit control to control by pro. Of course, this may well be the case; perhaps Rizzis

5 In the adjunct cases, the topmost link in the chain does c-command all the other links. Suppose one were to rule out (14ac) by reference to such a condition. This would fail to rule out other examples. (i) *John1 persuaded friends of t1 [t1 to nominate himself]. (ii) *Who1 did Mary persuade friends of t1 [t1 to nominate himself]?

The alternative analysis in (11) correctly rules out those cases (but not others, to be discussed below), where no matrix movement licenses the parasitic gap in the infinitive.

REMARKS AND REPLIES

479

binding/control asymmetry can be explained in terms other than the pro/thematic slot distinction. The burden of proof, however, resides with Hornstein. At any rate, implicit control raises even harder problems for Hornsteins theory. This is because nothing forces the presence of pro in the relevant constructions. Recall that for Hornstein, pro needs no Case (hence its occurrence in NOC infinitives), by virtue of lacking phonetic content. This means that verbs that license dative pro have optional dative Case features. In particular, unlike the accusative feature of hit, which must be checked (explaining *John hit), the dative Case feature of such verbs can be dropped, so nothing would go wrong if no DP checked it.6 Consider again implicit dative control. Following Hornsteins analysis, suppose we select an array without pro, containing John, said, to, return, later, assorted functional categories . We then construct the derivation in (16a), with the interpretation in (16b). (16) a. [ IP John1 [ VP t1 said t1 [ IP t1 to [ VP t1 return later]]]]. b. John said to himself to return later. In (16a), the chain headed by John bears three -rolesthe agent of return, the goal and agent of saidand one Case, the matrix nominative. Notice that dative Case being an optional feature of said, John need not check that feature (if it had, then presumably it could not check nominative Case on the matrix I). No principle assumed by Hornstein seems to be violated in this derivation. Yet clearly (16a) does not have the reflexive reading of (16b), so somehow it must be blocked. But how? In fact, the problem is quite general. Whenever a control verb allows an implicit controller, Hornsteins theory predicts that reflexive interpretations will be possible. But the opposite is true: a disjoint reference reading is always forced. Thus, the understood matrix goal argument in cases of type (15) is always disjoint from the matrix subject. (17) a. Hornsteins prediction: Implicit dative controllers may be bound by the matrix subject. b. In reality: Implicit dative controllers must be disjoint from the matrix subject. That (17b) is descriptively correct has been observed by several researchers (Williams 1985, Chomsky 1986b, Brody and Manzini 1987). It seems that with respect to Condition B (and C) of the binding theory, implicit arguments behave like pronouns. More generally, the syntax of natural language places heavy restrictions on reflexivizing operations. When lexical, they are restricted to only a few verbs (mostly describing bodily actions; see Reinhart 1997); when syntactic, they require some morphological marking (Reinhart and Reuland 1993). Hornsteins mechanism of movement to -positions is extremely powerful, generating nonexisting reflexive readings in a variety of cases (in section 4.1, I return to another case of this sort). Specific restrictions,

6 Indeed, this difference in the obligatory/optional status of the Case feature on the verb is invoked by Hornstein to explain why certain verbs can be reflexive while others cannot (wash vs. see). In section 4.1, I return to this proposal.

480

REMARKS AND REPLIES

then, must be invoked in order to filter out ungrammatical cases. On this issue, as on the preceding one, Hornsteins article is silent.7 3.4 The Status of the Minimal Distance Principle Hornstein presents it as a major advantage of his theory that it derives a true generalization about the choice of controller in OC constructions. This generalization, known as the Minimal Distance Principle (MDP) since Rosenbaum 1967, states that the DP closest to PRO is the controller. The MDP correctly predicts subject control in (18a) and object control in (18b), but incorrectly predicts object control in (18c). (18) a. John wanted to leave. b. John persuaded Mary to leave. c. John promised Mary to leave. Essentially, the MDP tradition (Rosenbaum 1967, Bach 1979, Bach and Partee 1980, Larson 1991, Manzini and Roussou 2000) offers two ways to handle the promise-type exceptions: (a) label them as exceptions and downplay their relevance to the core grammar, or (b) construct sophisticated analyses that make those cases regular at some hidden level (e.g., Larsons double object proposal). Hornsteins proposal belongs to the first category. The regular cases like (18b) follow from the MLC, which substitutes for the MDP. The latter is treated as a markedness condition, making cases like (18c) highly marked, as evidenced by their late acquisition (Chomsky 1969). The MDP tradition has always been a minority view in generative grammar. Most control studies, of all theoretical stripes, firmly deny that OC controllers are picked by distance (Chomsky 1981, Manzini 1983, Comrie 1984, Koster 1984, Melvold 1985, Farkas 1988, Sag and Pollard 1991, Petter 1998, Wurmbrand 1998, Landau 2000). This is not due to some pressure for orthodoxy; rather, a careful look at the facts shows that no semantics/pragmatics-free theory of controller choice is tenable. Although cases of type (18c) are not as common as those of type (18b), they are far too systematic to be dismissed as highly marked exceptions. Basically, they consist of two subgroups: verbs of commitment and verbs of request for permission. The first subgroup contains such verbs as promise, commit, vow, and threaten; the second subgroup contains such verbs as ask, beg, plead, and petition. There is considerable ambiguity with the second group (control shift), which is sensitive to pragmatic factors (e.g., authority relations). Different languages exhibit different degrees of freedom in this respect. Finally, verbs like propose also allow subject control across an object.

7 Notice that the alternative implementation in (11) fares no better on this point. If the infinitive is a VP specifier, both matrix arguments are generated below it. Thus, the gap at the dative position and the gap at the embedded subject position satisfy anti-c-command and connectedness. The fact that the higher trace at the matrix agent position c-commands the trace at the dative position is of no consequence.

REMARKS AND REPLIES

481

(19) a. We1 vowed to our leader [PRO1 to be loyal]. b. The prisoner1 asked the guard [PRO1 to smoke one more cigarette]. c. John1 proposed to Mary [PRO1 to help her with the arrangements]. All in all, there is a rich body of research concerning the various factorssemantic, pragmatic, and parametricthat affect the choice of controller (see the literature cited above).8 The issues are hard, but not impenetrable, and some important insights have emerged. None of those insights are retained, or even addressed, by analyses such as Hornsteins, which assign the choice of controller to formal locality conditions. Whatever one makes of the issues involved, ignoring them is not a viable option. At the very least, it is clear that a theory that does not derive the MDP is, ceteris paribus, better off than one that does.9 3.5 Delimiting the Category of Nonobligatory Control The class of NOC constructions is defined by Hornstein as the elsewhere case: whenever OC fails, NOC obtains. OC fails in all and only the environments from which the subject of the infinitive cannot raise. Hornstein discusses three such environments: sentential subjects, extraposed clauses, and interrogative complements. In all these cases, pro is inserted as a last resort operation. (20) a. It was believed that [pro shaving] was important. b. It is impossible [pro to win at roulette]. c. John told Sam [how pro to hold oneself erect at a royal ball]. The first question to ask is whether Hornstein correctly delimits the class of NOC constructions. My answer will be doubly negative: some NOC constructions are left out, and some OC constructions are erroneously included. An NOC construction that is left out involves initial adjuncts. What makes these instances of NOC is that the controller is not uniquely determined by sentence grammar. Rather, factors such as logophoricity and topicality determine which of the available discourse referents are possible controllers. Sometimes, there is no syntactic controller (21a) (from Kawasaki 1993); when there is one, it need not be the local subject (21b); it can also be split (21c) (from Bresnan 1982). At all times, though, it must be a possible logophoric antecedent (21de) (from Williams 1992).

8 A summary of the various approaches can be found in Landau 2000:chap. 5. In a recent response to Hornstein 1999, Culicover and Jackendoff (2001) document a wealth of semantic nuances in the choice of controller, all of which are beyond narrow syntactic approaches. 9 In Landau 2000, I attempt to reconcile promise-type cases with locality by appealing to the Principle of Minimal Compliance (Richards 1997). The PMC sanctions a violation of the MLC just in case the offending operation follows a licit operation, and both are triggered by the same attractor. Should that idea prove untenable, the fact that my analysis derives the MDP would be a cause for worry, not pride.

482

REMARKS AND REPLIES

(21) a. [After PROarb pitching the tents], darkness fell quickly. b. Mary1 was baffled. [Even after PRO1 revealing her innermost feelings], John remained untouched. c. Mary1 lost track of John2 because, [PRO1,2 having been angry at each other], he had gone one way and she the other. d. [PRO1 having just arrived in town], the main hotel seemed to Bill1 to be the best place to stay. e. *[PRO1 having just arrived in town], the main hotel collapsed on Bill1 . This case is particularly damaging to Hornsteins typology. Recall that Hornstein classifies rightedge adjuncts under OC (e.g., (2)). Movement out of these adjuncts is licensed because it is an instance of sideward movement. However, movement out of initial adjuncts, as in (21), to some matrix position, would also be sideward movement. Yet no matrix argument, if any exist, is an obligatory controller in (21). Intuitively, it is clear where the difference lies: the matrix subject c-commands a right-edge adjunct but not an initial adjunct. The OC/NOC contrast should somehow be keyed to this contrast. But this contrast is of no consequence to Hornsteins theory. Movement out of right-edge adjuncts does not target the surface subject position Spec,IP, but rather the thematic subject position Spec,vP. The latter, crucially, does not c-command the right-edge adjunct; if it did, the movement would not be sideward and consequently would not be allowed. In other words, given that the target for movement c-commands neither right-edge nor initial adjuncts, Hornsteins theory has no way to distinguish them and consequently predicts the latter to fall under OC. Consider next the inverse case, where OC constructions are misanalyzed as NOC. This is the case of interrogative complements (20c). Here, Hornstein follows the common wisdom in the control literature, taking the binding of oneself to be diagnostic of NOC.10 In Landau (2000), I show that the common wisdom is false and that interrogative complements instantiate partial control, which is a variety of OC. In fact, such complements show all characteristics of OC. For example, they do not allow long-distance control, a fact already noted in Mohanan 1985 and Chomsky 1986b, from which (22ab) are taken, respectively. (22) a. John said that Mary asked [how PRO to feed herself/*himself]. b. I thought they wondered [how PRO to feed themselves/*myself]. The controller in these cases must always be included in the reference of PRO. This property of partial control is found with the majority of control verbs (e.g., The chair decided to gather during the strike). Evidence for this is the disjoint reference effect in (23)strikingly, even in the presence of oneself.11

The two exceptions I am aware of, where interrogative complements are grouped with OC, are Chomsky and Lasnik 1977 and Lebeaux 1984. 11 For an extensive discussion, see Landau 2000:chap. 2. I return in section 5.5 to further implications of partial control for Hornsteins theory.

10

REMARKS AND REPLIES

483

(23) a. John1 wondered [who PRO to introduce his1 fiancee/*him1 to]. b. John1 asked [how PRO to talk to Mary/*him1 about oneself]. Descriptively, Hornsteins theory is in no worse position than most control theories, which misanalyze interrogative complements as cases of NOC. Theoretically, however, it is particularly unsuited to deal with the facts above. In this theory, islandhood entails NOC. To accommodate interrogative complements in the OC class, Hornstein must either deny the islandhood of interrogative complements or abandon the link between islandhood and NOC. The former is not a viable option; the latter undermines his account of (20ab).12 Either way, the distribution of NOC constructions does not follow from Hornsteins theory without additional, as yet unknown machinery.13 3.6 The Interpretation of PRO in Nonobligatory Control Following earlier proposals (Bouchard 1984), Hornstein claims that the subject of the infinitive in NOC has pronominal interpretation. In fact, it is a pronoun, simply lacking phonetic content. Evidence for this is the flexibility in controller choice, possibility of split control, strict readings under ellipsis, and de re interpretation. These properties are uncontroversial; what is controversial is the conclusion. Thus, what the facts tell us is that PRO in NOC does not behave like an anaphor. Does it necessarily follow that it is a pronoun? Hornsteins conclusion would follow only if there were no other options. In fact, there is a third possibility, one that has been proposed and defended in the past. The interpretive properties of NOC have been investigated in detail in studies of the socalled super-equi construction (Grinder 1970, Kuno 1975, Lebeaux 1984, Williams 1992, Landau 2001). An insight emerging from these studies is the striking parallelism between the distribution of PRO in NOC and picture-NP anaphora. More generally, it has been noted that PRO in NOC behaves like a logophor, displaying sensitivity to perspective, center of consciousness, and so
12 An anonymous reviewer suggests that the alternative derivation in (11) is compatible with the view that interrogative complements fall under OC, on the assumption that movement of the control type is not sensitive to wh-islands. The movement in (11), however, generates a parasitic gap, by the reviewers own proposal. That parasitic gaps are subject to all island constraints is well documented (Chomsky 1986a), and wh-islands are no exception. (i) ??What did John file without asking how Mary had written? (ii) What did John file without knowing that Mary had written? If the reduction of control to parasitic gaps is to be meaningful at all, they should be subject to the same constraints. The fact that OC traverses wh-islands (with perfect acceptability) therefore casts serious doubt on the analysis in (11). 13 Brody (1999) argues that locality is not exclusive to OC, but constrains NOC as well. (i) John told Mary [how PRO to teach herself/*himself/(?)oneself]. (ii) John thinks that Mary dislikes [PRO teaching herself/*himself/(?)oneself]. However, as (22)(23) clearly show, interrogative complements fall under (the partial control class of) OC constructions. The same is true of gerundive complements (iii), as opposed to gerundive subjects (iv). (iii) *Mary1 dislikes [PRO teaching her1 ].

(iv) Mary1 thinks that [PRO teaching her1 ] would be a mistake. The claim that NOC observes locality was first made by Grinder (1970), who proposed the Intervention Constraint for super-equi constructions. However, subsequent research has shown that restrictions on the choice of NOC controllers cannot be stated syntactically and are better understood in terms of logophoricity; see discussion in the next section.

484

REMARKS AND REPLIES

on. Given that logophors are subject to more stringent antecedence conditions than pronouns, they are expected to be more restricted in distribution than the latter. This is indeed the case. In the following examples, the intended antecedent does not satisfy the semantic/pragmatic conditions on logophoric antecedents. Consequently, PRO is excluded but an overt pronoun is allowed. (24) a. John sued Mary1 for divorce because it was no longer possible [for her1 /*PRO1 to support him]. b. Johns1 wife thought that [for him1 /*PRO1 to indulge himself in drinking] is inappropriate. c. Johns1 friends think it is illegal [for him1 /*PRO1 to feed himself]. (Chomsky 1986b:125) d. [His1 /*PRO1 having shaved already] shows that Mary arrived more than five minutes after John1 did. (Lebeaux 1984) It is hard to tell how to account for these facts in Hornsteins theory. One possibility might be to posit a null logophor instead of pro. This would weaken the logic of the last resort argument, which is supposed to rest on independent evidence for last resort resumptive pro in weak islands (see Hornsteins footnote 42). At any rate, no account that disregards these well-established facts can be considered complete.14 4 Claimed Benefits A fair evaluation of a theory should consider not just the cases it neglects, but also those it claims to be among its positive results. In this section, I will discuss three results that Hornstein presents as independent benefits of his theory; results that come for free once the control-is-raising analysis is adopted, but pose nontrivial difficulties to its rivals. I will argue that the first result is empirically inadequate because of overgeneration, the second is theoretically unsound, and the third, although valid, is too idiosyncratic and peripheral to warrant any general conclusions. 4.1 Reflexive Verbs Once PRO is viewed as an NP-trace, and NP-movement is allowed to target -positions, the possibility is opened up that control applies in other environments, not necessarily infinitival ones. Following suggestions by Howard Lasnik and Alan Munn, Hornstein argues that inherently reflexive verbs should be treated this way.

14 In this connection, it is worth pointing out that split antecedents are possible not only in NOC but also with a few OC verbs, contrary to Hornsteins claim (repeated in Hornstein 2000).

(i) John proposed to Mary [PRO to help each other]. (Koster and May 1982) (ii) Ich habe ihm angeboten [PRO einander zu helfen]. (Wurmbrand 1998:184) I have him offered PRO each-other to help I offered him to help each other.

REMARKS AND REPLIES

485

(25) a. Mary washed. b. Mary1 washed PRO1 . c. [ IP Mary1 [past [ VP t1 [wash t1 ]]]]. In this derivation, Mary checks both the internal and external -roles of wash. Thus, the same mechanism, doubly -marked chains, accounts for OC and reflexive verbs.15 The question immediately arises how to block this mechanism from applying to any transitive verb, overgenerating forms like (26a) with the interpretation (26b). (26) a. Mary saw. b. Mary saw herself. Hornsteins solution (footnote 21) appeals to Case theory. Verbs like see must assign accusative Case. A reflexive derivation of (26a) would leave either the accusative Case of saw or the nominative Case of IP unchecked. By contrast, accusative Case is optional on wash; when it is suppressed, a reflexive derivation becomes possible. But why is it obligatory? Notice that (25a) must be interpreted reflexively, and not as a case of indefinite object drop. If the Case feature is optional, and pro needs no Case, then Hornsteins account permits, incorrectly, a derivation that yields Mary washed pro. Suppose we somehow overcome this problem. Still, for Hornsteins account to be nonvacuous, there must be some independent way to determine whether an accusative Case feature on a verb is obligatory or notother than the (non)existence of a reflexive entry for that verb. Hornstein spells out none, but the only conceivable way is to check whether the verb allows indefinite object drop. In the absence of a (phonetically realized) object, accusative Case cannot be checked. If the sentence is still grammatical, we may conclude that the Case feature on the verb is optional. Hornsteins analysis thus makes the following prediction: (27) If a transitive verb allows object drop, it should license a reflexive reading. However, (27) is strongly disconfirmed. In appropriate contexts, all the verbs in (28)a small samplemay appear without overt objects. Yet none takes on an inherently reflexive meaning like wash and shave. (28) John ate/watched/cursed/taught/preached/drew/cleaned. As noted in section 3.3, the problem runs deeper. It is just not up to syntax to reflexivize verbs in the manner Hornstein envisions. As Reinhart (1997) shows, inherently reflexive verbs like wash are lexically derived. The process is not productive, explaining both the scarcity of such forms and the crosslinguistic variation in the class of reflexive verbs. Moreover, it is marked

15 The alternative analysis in (11) does not extend to reflexive verbs, as the two gaps stand in a c-command relation, hence neither can be parasitic.

486

REMARKS AND REPLIES

in the morphology (zero in English, but overt in most languages). Syntactic reflexivization, a totally different operation, involves the merging of lexical anaphors. This operation is indeed productive, as one expects syntax to be. Hornstein conflates the two options, ending up with an unreasonably powerful mechanism, which must be severely constrained by ad hoc stipulations. The approach not only yields wrong results, as in (28), but seems fundamentally misguided. 4.2 De Se Interpretation The de se/de re contrast emerges in situations where a subject of an attitude verb is misinformed about his or her own identity. A typical example involves a war hero who suffers from amnesia and remembers nothing of his wartime experiences. Suppose this person (hereafter, the unfortunate) sees a TV program describing his own exploits and is impressed with the courage exhibited by that person, who he does not know is himself. The unfortunate comes to believe that this hero will get a medal. Under this scenario, (29a) is true but (29b) is false. (29) a. The unfortunate expects that he will get a medal. b. The unfortunate expects to get a medal. This situation arises because (29a) can be satisfied by de re beliefs about a certain individual (denoted by the unfortunate), but (29b) can only be satisfied by de se beliefs about the self. Hornstein claims that the movement analysis also accounts for the required de se interpretation of OC PRO (p. 80). This is so because the control chain in (30a) creates a compound monadic predicate of the form (30b). (30) a. John hopes to leave. b. x.x hopes x leave c. Everyone1 hopes that he1 will leave. In fact, as Chierchia (1990) and Higginbotham (1992) observe, the mechanism of variable binding does not distinguish de se from de re interpretations. Thus, both (30a), under Hornsteins account, and (30c), under anyones account, contain at LF the -abstract in (30b) (modulo tense distinctions); yet (30c) supports a de re reading, whereas (30a) does not. How to account for the de se interpretation is an extremely difficult question for semanticists and philosophers. At the moment, Hornsteins proposal does not seem to contribute to resolving this question. 4.3 Wanna-Contraction The well-known phenomenon of wanna-contraction distinguishes PRO from wh-traces. PRO, like NP-trace, does not block contraction, whereas wh-trace does. (31) a. Who do you want PRO to/wanna see twh? b. *Who do you want twh to/*wanna vanish? c. Johns going tNP to/gonna leave. If PRO is a Case-marked expression, Hornstein points out, as the theory of null Case proposes,

REMARKS AND REPLIES

487

it should pattern like wh-trace and block contraction. That it behaves like NP-trace argues that it is more like non-Case-marked NP-trace than like Case-bearing wh-trace (p. 76). This supports the reduction of control to raising. The argument is valid, but of limited interest. First, note that Hornstein is arguing against a highly theory-internal hypothesis, namely, the null Case proposal.16 For earlier GB analyses, the idea that PRO is Caseless was the norm. But even granting that, the very significance of wanna-contraction is questionable. After all, despite the extensive literature on this phenomenon (a fact of historical but not necessarily theoretical significance), this is a particular fact about the behavior of one or two verbs, in one languagehardly a solid basis for general conclusions. 5 Raising versus Control Rosenbaums (1967) discovery that raising and control constructions are fundamentally distinct, although superficially similar, has been a cornerstone of generative grammar for more than three decades. It is, I believe, a major achievement of the field, on a par with the discovery of the unaccusative/unergative distinction and the NP/DP distinction. Like other true discoveries, it has the hallmark of unexpected fecundity: once it was recognized, more and more facts were discovered that followed from it. The raising/control contrast has provided insight into many crosslinguistic phenomena that otherwise would appear unrelated. Hornsteins theory does not obliterate the raising/control contrast. Rather, it trivializes its import. According to this theory, the single difference between the two constructions is the fact that the raised DP is -marked once in raising but twice in control. As Hornstein observes, this is indeed a radical departure from the standard view. Therefore, it should be the focus of empirical testing, designed to decide between the two alternatives. However, Hornsteins article sidesteps the whole issue. Except for one fact, there is no mention of the rich literature on the raising/control distinction, no consideration of empirical consequences, and no attempt to cope with problematic data. The purpose of this section is to fill in some empirical content (mostly familiar) that is highly pertinent to the debate. The one contrast between raising and control that Hornstein does cite is the well-known observation (due to Rosenbaum) that nonthematic NPs can raise but not control. (32) a. The shit seems to have hit the fan. b. *The shit expects to hit the fan. c. There seems/*expects to be a man in the garden. These facts are unproblematic for Hornsteins theory. In control, the matrix verb has an external -feature to check. But idiom chunks and expletives cannot check that feature because they cannot

16

Even a proponent of this proposal could maintain that null Case is too impoverished to block contraction.

488

REMARKS AND REPLIES

bear the relevant -role. So the features remain unchecked at LF. Terminology aside, this is the standard account.17 These examples indicate what kinds of raising/control contrasts are within the reach of Hornsteins theory, and what kinds are not. Basically, the theory has only one contrastive feature to appeal to: the thematic distinction. This is pretty slim grounds. However, as noted above, generative research has uncovered many systematic differences between the two types of constructions; these go far beyond the thematic distinction. We will see that Hornsteins theory is incapable of dealing with all these facts without additional stipulations. 5.1 Complementizers That raising and control complements have different sizes is an old claim. For various reasons, theoretical as well as empirical, it has been maintained that raising complements are IPs and control complements are CPs. I will focus here on one of the reasons for this claim, which is stated in (33), presumably a universal generalization. (33) Control complements may be introduced by complementizers; raising complements are never introduced by complementizers. Infinitival complementizers have been recorded in many languages: de in French, di in Italian, om in Dutch, att in Swedish, aL in Icelandic, me in Hebrew, for in Belfast English, i in Welsh, and so on. Strikingly, they are found only in control contexts, not raising contexts. Even when such a complementizer appears to introduce a raising complement, closer examination reveals a control alternant. Consider the following pair in Hebrew: (34) a. Rina xadla (me-)leacben et Gil. Rina stopped (from-)to-irritate acc Gil Rina stopped irritating Gil.

17 Hornsteins theory can also account for Rosenbaums second observation, namely, that flipping the arguments in passive preserves truth conditions in raising but not in control. (i) a. The doctor seemed to have examined John. b. John seemed to have been examined by the doctor. (ii) a. The doctor wanted to examine John. b. John wanted to be examined by the doctor. The contrast stems from the fact that the distribution of -roles among arguments is identical in (iab) but not in (iiab). Similarly, Hornstein can explain Mays (1985) observation that raising subjects can be interpreted with narrow scope with respect to the main predicate, but control subjects cannot. (iii) a. Someone from New York is likely to win the lottery. (ambiguous) b. Someone from New York is eager to win the lottery. (unambiguous) For May, the matrix subject in (iiia) may lower at LF to its base position, leaving an uninterpreted (hence deletable) trace behind. Since the matrix subject in (iiib) does receive a matrix -role, the trace of lowering may not delete, and it remains unbound at LF. Hornstein could either translate this story to the copy terminology or require that all -roles be expressed at LF (with the result that the higher copy in control must be the one feeding interpretation).

REMARKS AND REPLIES

489

b. Ha-muzika ha-roeset xadla (*me-)leacben et Gil. the-music the-noisy stopped (*from-)to-irritate acc Gil The loud music stopped irritating Gil. In Landau (2002), I show that me- is an infinitival complementizer associated with negative complements. The verb xadal stop, cease belongs to a class of aspectual verbs that are systematically ambiguous between raising and control (Perlmutter 1970). Typically, only animate DPs can control. Notice that the complementizer may appear only when the matrix subject is animate (34a). The inanimacy of the matrix subject in (34b) forces a raising analysis, which in turn rules out the presence of a complementizer, in accordance with (33).18 Hornsteins analysis would be hard pressed to make sense of these facts. Why should the presence of an external -feature on the matrix verb license a CP projection in the complement? The two facts seem completely unrelated. By contrast, under the standard analysis raising is sensibly linked to IP and control to CP: NP-movement cannot traverse CP boundaries (see section 3.1). If one denies that control and raising are distinguished in terms of movement, those links are indeed lost. 5.2 Unaccusative Properties As the Unaccusative Hypothesis generated new insights into the nature of NP-movement, a point of intersection with the raising/control distinction became obvious. Raising predicates, which lack an external argument, are unaccusative, whereas control predicates are not.19 Thus, only the former ` are expected to pattern with unaccusative verbs vis-a-vis the standard tests for this property. In this section, I review two such cases, from French and Italian. The crucial point to notice is what the tests do and do not diagnose. These tests are not sensitive to properties of the matrix position where the raisee/controller is found. This is highly problematic for Hornsteins theory, since the nature of that position (thematic or not) is the single feature in this theory distinguishing control from raising. Rather, the tests diagnose a lower syntactic position, inside the infinitive, linked by a chain relation to the matrix subject. As we will see, no such position is diagnosed in the control cases, which fail the relevant tests. Conclusion: no chain is formed in control. 5.2.1 En-Cliticization Ruwet (1972) observed that raising verbs permit en-cliticization of the partitive complement of their surface subject on the embedded verb, whereas control verbs do

18 A similar correlation exists in Italian, where sembrare seem and parere appear are ambiguous between raising and (dative) control. As Kayne (1981) notes, only the latter variants allow the complementizer di. (i) a. Gianni sembra/pare (*di) essere partito. Gianni seems/appears DI to-be left Gianni seems/appears to have left. b. Mi sembra/pare di aver capito. to-me seems/appears DI to-have understood It seems/appears to me that I have understood. 19 This is an oversimplification. A few nonagentive control verbs (e.g., manage) are unaccusative. The point of interest is that no control verb exhibits the NP-movement property of unaccusatives.

490

REMARKS AND REPLIES

not. As is well known, en-cliticization is a diagnostic of NP-movement in French, found with passive and unaccusative subjects as well. (35) a. Le directeur du departement semble etre accepte. the head of-the department seems to-be accepted The head of the department seems to be accepted. b. Le directeur semble en etre accepte. ` (36) a. Le directeur du departement espere etre accepte. the head of-the department hopes to-be accepted The head of the department hopes to be accepted. ` b. *Le directeur espere en etre accepte. If the embedded subject position in control complements is an NP-trace, it is a mystery why this trace cannot be linked to a clausemate en, unlike all other NP-traces. By assumption, traces are copies with internal structure, so the subject copy can contain the en-copy as a subconstituent. By contrast, PRO being a simplex morpheme, it cannot host an internal copy of the clitic.20 5.2.2 Si-Reflexivization Rizzi (1986b) shows that NP-movement cannot skip a position coindexed with the moved NP, a consequence he derives from the Chain Condition and the -Criterion. Specifically, configurations of the sort [NP1 . . . si1 . . . t1 ] in Italian, where t1 is the trace of NP1 and si1 is a reflexive clitic, are excluded. This serves as an anti-unaccusative test in this language: all contexts that involve NP-movementpassive, impersonal, and unaccusative constructionsfail this test. Not surprisingly, Rizzi shows that raising and control contrast in the expected way. (37) a. *I due candidati1 si1 risultavano [t1 poter vicere]. the two candidates to-each-other appeared to-be-able to-win (The two candidates appeared to each other to be able to win.) b. I due concorrenti1 si1 sono promessi [di PRO1 essere leali]. the two competitors to-each-other were promised DI to-be loyal The two competitors promised to each other to be loyal. For Hornstein, if (37a) violates the MLC, so should (37b). At times, it has been suggested that the reflexive clitic si is really an expression of the external argument (rather than the dative), and it is the lack of such argument that rules out (37a). This analysis would indeed do some work

20 Ruwets (1972) conclusion is correct, even though his control examples are problematic, involving nonderived subjects (which do not allow en-extraction to begin with); the data and judgments in (36) are due to Marie Claude Boivin (personal communication). Some French speakers reject en-extraction out of animate DPs, finding (35b) unacceptable as well. However, speakers without this restriction report a clear contrast. Notice that the alternative analysis in (11) cannot explain this contrast either, even with the stipulation that raising infinitivals occupy a different position than control infinitivals do. En-extraction can apply to the base matrix position of the controller, prior to (remnant) movement to the subject position. There is no reason why the parasitic gap in the embedded subject position should not be a complex copy, identical to the remnant DP (recall that under the alternative analysis, the parasitic gap replacing PRO has all the properties of a normal NP-trace).

REMARKS AND REPLIES

491

for Hornsteins theorybut only half the work. Under such an analysis, it would still be the case that (37b) violates the MLC, this time with (the trace of) si intervening in Spec,VP rather than in the dative position. Notice that the problem is independent of the assumption that chains contain a single -rolean assumption that Rizzi adopts but Hornstein rejects. Whether si is part of the raising chain or not, it should count as an intervener. 5.3 Each-Association Burzio (1981) noted that control breaks the association of each with a lower NP, but raising does not ((38ac) are cited by Chomsky (1981:61)). (38) a. b. c. d. One interpreter1 each was assigned t1 to the visiting diplomats. One interpreter1 each seemed [t1 to have been assigned t1 to the visiting diplomats]. *One interpreter1 each tried [PRO1 to be assigned t1 to the visiting diplomats]. *One interpreter1 each said that [he1 had been assigned t1 to the visiting diplomats].

Notice that an NP-trace does not interfere with each-association (38ab), whereas PRO, just like an overt pronoun, does (38cd). A raising analysis of control loses this contrast. It is not obvious that this phenomenon can be handled by the simple constraint on reconstruction (see footnote 17), which Hornsteins theory might appeal to. This is so because the intended interpretation in (38c) is unavailable even when one interpreter has wide scope with respect to tried. To see this, consider the slightly different examples in (39). (39) a. One interpreter1 tried [PRO1 to be assigned t1 to every visiting diplomat]. b. *One interpreter1 each tried [PRO1 to be assigned t1 to every visiting diplomat]. c. One interpreter1 each seems [t1 to have been assigned t1 to every visiting diplomat]. (39a) has a reading (perhaps not easily accessible) with the scope relations every one tried. This means that no reconstruction is needed in order to get wide scope for the embedded universal; Quantifier Raising is sufficient. Therefore, the fact that (39b) is ungrammatical under any reading (including the one observed in (39a)) cannot be due to failure of reconstruction. Thus, the ban on reconstruction of control subjects, even if derivable in Hornsteins system, would not account for (39b). By contrast, a simple locality condition on each-association, together with the standard view on control, accounts for the facts. Suppose that each-association is clause bound, so that the DP denoting the set over which each distributes, or its copy, must be clausemate to each. This is satisfied in the raising cases (38b)/(39c), but not in the control cases (38c)/(39b)only under the assumption that the embedded subject in the latter is not a (copy) trace. Without that assumption, the patterning of control clauses with finite rather than raising clauses remains mysterious. 5.4 Case Concord SigurLsson (1991) argues convincingly that PRO in Icelandic bears Case. One particularly interesting argument he cites is based on two observations: (a) certain predicates determine quirky

492

REMARKS AND REPLIES

case on their subjects; (b) floating quantifiers agree in case with their subjects. Thus, in control constructions, it is possible to detect the Case of PRO by the Case of the floating quantifier. skolann]. (40) a. Strakarnir vonast til [aL PRO vanta ekki alla the-boys.NOM hope for to PRO.ACC lack not all.ACC in the-school The boys hope not to be all absent from school. lei Last ekki ollum skola]. b. Strakarnir vonast til [aL PRO the-boys.NOM hope for to PRO.DAT to-be bored not all.DAT in school The boys hope not to be all bored in school. r+Lnnie]. c. Strakarnir vonast til [aL PRO verLa allra getiL all.GEN mentioned in the-speech the-boys.NOM hope for to PRO.GEN be The boys hope to be all mentioned in the speech. The assumption that PRO bears Case is necessary in order to explain the case mismatch between the nominative matrix subject and the embedded floating quantifier. It also conforms with the generalization that case concord is clause bound. This alone undermines Hornsteins claim that the controlled position is Caseless (being an NP-trace).21 What is more striking, however, is that no case mismatch is observed in raising constructions (data from ONeil 1997:109, attributed to Hoskuldur Thrainsson). skolann]. (41) a. Strakarna virLast [t vanta ekki alla to-lack not all.ACC in the-school the-boys.ACC seem The boys seem not to be all absent from school. b. Strakunum virLast [t lei Last ekki ollum skola]. the-boys.DAT seem to-be bored not all.DAT in school The boys seem not to be all bored in school. r+Lnnie]. c. Strakanna virLast [t verLa allra getiL the boys.GEN seem to-be all.GEN mentioned in the-speech The boys seem to be all mentioned in the speech. Here, the quirky case determined by the embedded predicate shows up on the matrix subject. This raising/control contrast makes perfect sense under the standard view. In (40), the matrix and the embedded subject belong to two distinct A-chains; hence, they bear distinct cases. In (41), the matrix and the embedded subject belong to the same A-chain, which is Case-marked only once. Given that quirky case always overrides structural (nominative) case in morphological spellout, the former surfaces. Hornsteins theory is unable to deal with this contrast. The Icelandic facts furnish such a compelling argument against the reduction of control to raising because they so clearly reflect the chain structure of the construction. It would hardly make sense to argue that (a) A-chains can

21 That PRO is (dative) Case marked has also been argued for Russian (Comrie 1974, Neidle 1988, Baylin 1995, Babby 1998, Moore and Perlmutter 2000).

REMARKS AND REPLIES

493

be doubly Case-marked; and (b) that option is licensed only when the head of the chain is marked. Yet nothing short of these evidently ad hoc assumptions would seem to reconcile the contrast in (40)(41) with Hornsteins theory.22 5.5 Partial Control The final argument comes from my own work on partial control. In Landau 2000, I discuss a neglected property of control complements, namely, that the reference of PRO need not be exhausted by the reference of the controller. Such cases are perfectly natural when the speaker has some salient group in mind. For example (the notation DP1 . . . PRO1 indicates the partial control reading): (42) We thought that . . . a. The chair1 preferred [PRO1 to gather at 6]. b. Bill1 regretted [PRO1 meeting without a concrete agenda]. c. Mary1 wondered [whether PRO1 to apply together for the grant]. When one scans all the types of control infinitives, it turns out that most allow partial control and a few (a much smaller number) do not.23 The reason why partial control has been largely overlooked is not that a few control verbs allow it (most do), but rather that this possibility becomes apparent only when a singular controller cooccurs with an embedded collective predicate, not a common combination. In Landau 2000, I provide extensive documentation for the crosslinguistic solidity of this phenomenon, as well as a theoretical analysis. Without further detail, one can already see how damaging the very existence of partial control is to the thesis control is raising. Simply put: there is no partial raising. It is not even clear how to formulate a rule of NP-movement that would yield a chain with nonidentical copies.24 Indeed, partial readings are not found in raising contexts (43a). This is so even though the semantic type of raising complementsa propositional infinitiveis capable of supporting partial control, as the Italian example (43b) demonstrates. (43) We thought that . . . a. *The chair appeared to be gathering once a week. b. Il presidente crede di essersi riuniti inutilmente la notte scorsa. the night last the chair believes DI to-be-SI gathered in-vain The chair believes to have gathered in vain last night.

22 ONeil (1997), advancing another theory of the control is raising variety, invokes assumptions no less exotic to solve this puzzle: for example, the idea that inherent Case is checked by AgrS (supposedly absent in raising complements). The basic issue of a doubly Case-marked chain, however, is left unexplained. 23 Partial control is possible with desiderative, factive, propositional, and interrogative infinitives; exhaustive control is forced with implicative, modal, and aspectual infinitives. The split is anchored to the infinitival tense (see Landau 2000:chap. 2 for discussion). 24 In the same vein, Hornsteins logical formula (30b) is not suitable to capture the semantics of (42).

494

REMARKS AND REPLIES

Let me stress again that partial control can be neither dismissed as an exotic phenomenon nor relegated to marked environments. At the very least, Hornsteins theory must make room for cases of OC that are not reducible to raising. But if that much is conceded, then the whole project of eliminating PRO and the control module is voided.25 6 Conclusion Hornstein terms his article an exercise in grammatical downsizing (p. 69). It is worth evaluating how successful this exercise is. The number of independent empirical problems discussed in my critique is, on a conservative count, around 13. In the worst case, Hornsteins theory would have to invoke a specific principle to handle each of these problems, enriching the grammar with 13 new principles. In the best case, it would come up with one magic principle to handle all the problems at once. By ignoring all those empirical issues, Hornsteins article suggests that his theory approximates the worst case more closely than it does the best. Thus, the claim for downsizing should be regarded at present as a hope, not an achievement. I should stress that I do not take Hornsteins failure to deliver theoretical downsizing as a serious argument against his theory, perhaps on grounds of Occams razor. As I understand Occams razor, it is a conditional, which says, Other things being equal, avoid unnecessary assumptions (entities, constructs, mechanisms, etc.). Explaining the facts is the principal ingredient in the other things; that is what was once called descriptive adequacy.26 Reading Hornsteins paper, one is struck by the overwhelming weight bestowed on conceptual arguments: the author repeatedly stresses that certain theoretical moves are motivated (justified, vindicated, etc.) by the fact that they simplify the grammar, eliminate redundancies, avoid unnecessary stipulations, and the like. Virtually nowhere does he claim that they are motivated (justified, vindicated, etc.) because they successfully explain a problematic set of data that previous theories have failed to accommodate. This lack of concern for empirical adequacy is disturbing. Thus, I am not suggesting that we reject Hornsteins analysis of control because it introduces unknown complications and redundancies. As indicated above, the analysis is too impoverished to warrant such a conclusionperhaps there is one simple principle that elevates this analysis to a reasonable level of empirical adequacy. My point is that until such a principle is spelled out, and as long as so many empirical issues remain either damaging to or unresolved by the theory, considerations of simplicity and economy are simply premature.

25 The literature contains more raising/control contrasts. Rizzi (1980) observes that impersonal si-passives in Italian can be embedded under raising but not under control predicates. Hornstein (footnote 35) claims to have developed an account of this elsewhere, so I put it aside. Rizzi also observes that control infinitivals can be preposed, but raising ones cannot (e.g., To be a winner, John certainly wants/*seems); the same contrast is noted by Chomsky (1981:62) and Jacobson (1992). However, Jacobson observes that many object control verbs strongly resist complement preposing (e.g., persuade, order), and I might add a few subject control verbs as well (e.g., manage, condescend); see also Martin 1996. The whole issue, I believe, is related to the Case properties of the matrix verb and is orthogonal to the raising/control distinction. Other less than perfect contrasts that Jacobson cites (ellipsis/pronominalization of the infinitive) are also amenable to this treatment (see Landau 2000:chap. 2). 26 Note that explanatory adequacy is not a conceptual constraint but an empirical one, operating on a different level (i.e., the question of language acquisition).

REMARKS AND REPLIES

495

In a revealing passage, Hornstein writes, In my view, the theoretical adjustments needed to eliminate the PRO module are methodologically preferable to the theoretical stipulations they have replaced. Even if the gains were paltry, the burden of proof would be on those who favor maintaining the restrictions on -roles, chains, and merger that have been dispensed with (pp. 9394). This statement has the facts wrong and the logic backward. The empirical gains of Hornsteins theory are neither paltry nor nil; they are in fact negative, given the number of true generalizations about control that are lost in the theory. As long as this situation persists, the burden of proof resides with the new theory and not with the old ones. In this respect, the present critique is quite different from Brodys (1999) reply to Hornstein 1998. Brody is mainly concerned with eliminating movement as a mechanism for chain construction, and he deals with control only tangentially. The thrust of his argument is not the empirical deficiencies in Hornsteins analysis (although he does mention a few) but its redundancy, given certain assumptions about the interpretive component. In that respect, Brody and Hornstein share an underlying emphasis on considerations of simplicity and redundancy.27 By contrast, my critique rejects this methodology, suspending or at least diminishing conceptual considerations, in a search for empirical adjudication. It is important to keep in mind what this debate is about and what it is not about. One thing it is not about is the issue of innovative versus conservative approaches to scientific progress. Likewise, it is not about the legitimacy of conceptual arguments in science (rather, their proper weight). It is also not about the minimalist framework in general, or even the possibility of a minimalist analysis of control; I myself have proposed such an analysis in Landau 2000, albeit a nonreductionist one. More narrowly, the debate is not even about specific assumptions, like the idea of movement into -positions. For all I know, that might be a viable option in the grammar. The debate is about which of two competing sets of assumptions does maximal justice to the facts of control. The choice may be hard, but never intractable. For all its limitations, the criterion of empirical adequacy is still the best index we have for scientific progress. References
Babby, Leonard. 1998. Subject control as direct predication: Evidence from Russian. In Proceedings of FASL 6, ed. by Zeljko Boskovic, Steven Franks, and William Snyder, 1737. Ann Arbor: University of Michigan, Michigan Slavic Publications. Bach, Emmon. 1979. Control in Montague Grammar. Linguistic Inquiry 10:515531. Bach, Emmon, and Barbara Partee. 1980. Anaphora and semantic structure. In CLS 16: Papers from the Parasession on Pronouns and Anaphora, ed. by Jody Kreiman and Almerindo Ojeda, 128. Chicago: University of Chicago, Chicago Linguistic Society. Baylin, John F. 1995. A configurational approach to Russian free word order. Doctoral dissertation, Cornell University, Ithaca, N.Y.

27 Another typical example is the elaborate discussion in Hornstein 2000:fn. 28 of the question, Do lowering operations simplify or complicate the grammar? I see no objective way to decide on such matters.

496

REMARKS AND REPLIES

Borer, Hagit. 1989. Anaphoric AGR. In The null subject parameter, ed. by Osvaldo Jaeggli and Kenneth Safir, 69109. Dordrecht: Kluwer. Bouchard, Denis. 1984. On the content of empty categories. Dordrecht: Foris. Bresnan, Joan. 1982. Control and complementation. Linguistic Inquiry 13:343434. Brody, Michael. 1999. Relating syntactic elements: Remarks on Hornsteins Movement and chains. Syntax 2:210226. Brody, Michael, and M. Rita Manzini. 1987. Implicit arguments. In Mental representations: The interface between language and reality, ed. by Ruth Kempson, 105130. Cambridge: Cambridge University Press. Burzio, Luigi. 1981. Intransitive verbs and Italian auxiliaries. Doctoral dissertation, MIT, Cambridge, Mass. Chierchia, Gennaro. 1989. Structured meanings, thematic roles and control. In Properties, types and meanings II, ed. by Gennaro Chierchia, Barbara Partee, and Raymond Turner, 131166. Dordrecht: Kluwer. Chierchia, Gennaro. 1990. Anaphora and attitudes de se. In Semantics and contextual expression, ed. by Renate Bartsch, Johan van Benthem, and Peter van Emde Boas, 132. Dordrecht: Foris. Chomsky, Carol. 1969. The acquisition of syntax in children from 5 to 10. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1981. Lectures on government and binding. Dordrecht: Foris. Chomsky, Noam. 1986a. Barriers. Cambridge, Mass.: MIT Press. Chomsky, Noam. 1986b. Knowledge of language: Its nature, origin, and use. New York: Praeger. Chomsky, Noam, and Howard Lasnik. 1977. Filters and control. Linguistic Inquiry 8:425504. Clark, Robin. 1990. Thematic theory in syntax and interpretation. London: Routledge. Comrie, Bernard. 1974. The second dative: A transformational approach. In Slavic transformational syntax, ed. by Richard Brecht and Catherine Chvany, 123150. Ann Arbor: University of Michigan, Michigan Slavic Publications. Comrie, Bernard. 1984. Subject and object control: Syntax, semantics, and pragmatics. In Proceedings of the 10th Annual Meeting of the Berkeley Linguistics Society, ed. by Claudia Brugman, Monica MacCaulay, Amy Dahlstrom, Michele Emanatian, Birch Moonwoman, and Catherine OConnor, 450464. Berkeley: University of California, Berkeley Linguistics Society. Culicover, Peter W., and Ray Jackendoff. 2001. Control is not movement. Linguistic Inquiry 32:493512. Epstein, Samuel David. 1984. Quantifier-pro and the LF representation of PROarb . Linguistic Inquiry 15: 499505. Farkas, Donca F. 1988. On obligatory control. Linguistics and Philosophy 11:2758. Grinder, John T. 1970. Super equi-NP deletion. In Papers from the Sixth Regional Meeting of the Chicago Linguistic Society, 297317. Chicago: University of Chicago, Chicago Linguistic Society. Higginbotham, James. 1992. Reference and control. In Control and grammar, ed. by Richard Larson, Sabine Iatridou, Utpal Lahiri, and James Higginbotham, 79108. Dordrecht: Kluwer. Higginbotham, James. 1999. A plea for implicit anaphora: In Atomism and binding, ed. by Hans Bennis, Pierre Pica, and Johan Rooryck, 183203. Dordrecht: Foris. Hornstein, Norbert. 1998. Movement and chains. Syntax 1:99127. Hornstein, Norbert. 1999. Movement and control. Linguistic Inquiry 30:6996. Hornstein, Norbert. 2000. On A-chains: A reply to Brody. Syntax 3:129143. Jacobson, Pauline. 1992. Raising without movement. In Control and grammar, ed. by Richard Larson, Sabine Iatridou, Utpal Lahiri, and James Higginbotham, 149194. Dordrecht: Kluwer. Kawasaki, Noriko. 1993. Control and arbitrary interpretation in English. Doctoral dissertation, University of Massachusetts, Amherst. Kayne, Richard. 1981. On certain differences between French and English. Linguistic Inquiry 12:647686. Kimball, John P. 1971. Super equi-NP deletion as dative deletion. In Papers from the Seventh Regional Meeting of the Chicago Linguistic Society, 142148. Chicago: University of Chicago, Chicago Linguistic Society.

REMARKS AND REPLIES

497

Koster, Jan. 1984. On binding and control. Linguistic Inquiry 15:417459. Koster, Jan, and Robert May. 1982. On the constituency of infinitives. Language 58:117143. Kuno, Susumu. 1975. Super equi-NP deletion is a pseudo-transformation. In Proceedings of NELS 5, ed. by Ellen Kaisse and Jorge Hankamer, 2944. Amherst: University of Massachusetts, GLSA. Landau, Idan. 2000. Elements of control: Structure and meaning in infinitival constructions. Dordrecht: Kluwer. Landau, Idan. 2001. Control and extraposition: The case of super-equi. Natural Language & Linguistic Theory 19:109152. Landau, Idan. 2002. (Un)interpretable Neg in Comp. Linguistic Inquiry 33:465492. Larson, Richard. 1991. Promise and the theory of control. Linguistic Inquiry 22:103139. Larson, Richard, Sabine Iatridou, Utpal Lahiri, and James Higginbotham, eds. 1992. Control and grammar. Dordrecht: Kluwer. Lasnik, Howard. 1988. Subjects and the -Criterion. Natural Language & Linguistic Theory 6:117. Lebeaux, David. 1984. Anaphoric binding and the definition of PRO. In Proceedings of NELS 14, ed. by Charles Jones and Peter Sells, 253274. Amherst: University of Massachusetts, GLSA. Manzini, M. Rita. 1983. On control and control theory. Linguistic Inquiry 14:421446. Manzini, M. Rita. 1986. On control and binding theory. In Proceedings of NELS 16, ed. by Stephen Berman, Jae-Woong Choe, and Joyce McDonough, 322337. Amherst: University of Massachusetts, GLSA. Manzini, M. Rita, and Anna Roussou. 2000. A minimalist theory of A-movement and control. Lingua 110: 409447. Martin, Roger A. 1996. A minimalist theory of PRO and control. Doctoral dissertation, University of Connecticut, Storrs. May, Robert. 1985. Logical Form: Its structure and derivation. Cambridge, Mass.: MIT Press. Melvold, Janis. 1985. Getting PRO under control. Ms., MIT, Cambridge, Mass. Mohanan, K. P. 1985. Remarks on control and control theory. Linguistic Inquiry 16:637648. Moore, John, and David M. Perlmutter. 2000. What does it take to be a dative subject? Natural Language & Linguistic Theory 18:373416. Neidle, Carol Jan. 1988. The role of Case in Russian syntax. Dordrecht: Kluwer. Nunes, Jairo. 1995. The copy theory of movement and linearization of chains in the Minimalist Program. Doctoral dissertation, University of Maryland, College Park. ONeil, John H. 1997. Means of control: Deriving the properties of PRO in the Minimalist Program. Doctoral dissertation, Harvard University, Cambridge, Mass. Perlmutter, David M. 1970. The two verbs begin. In Readings in English transformational grammar, ed. by Peter Rosenbaum and Roderick Jacobs, 107119. Waltham, Mass.: Ginn. Petter, Marga. 1998. Getting PRO under control. The Hague: Holland Academic Graphics. Reinhart, Tanya. 1997. Syntactic effects of lexical operations: Reflexives and unaccusatives. OTS Working Papers in Linguistics. Utrecht: Utrecht University. Reinhart, Tanya, and Eric Reuland. 1993. Reflexivity. Linguistic Inquiry 24:657720. Richards, Norvin W. 1997. What moves where when in which language? Doctoral dissertation, MIT, Cambridge, Mass. Rizzi, Luigi. 1980. Comments on Chomsky, On the representation of form and function. In Proceedings of the June 1980 CNRS Conference on the Cognitive Sciences, ed. by Jacques Mehler. Paris: CNRS. Rizzi, Luigi. 1986a. Null objects in Italian and the theory of pro. Linguistic Inquiry 17:501557. Rizzi, Luigi. 1986b. On chain formation. In The syntax of pronominal clitics, ed. by Hagit Borer, 6595. New York: Academic Press. Roeper, Thomas. 1987. Implicit arguments and the head-complement relation. Linguistic Inquiry 18: 267310. Rosenbaum, Peter. 1967. The grammar of English predicate complement constructions. Cambridge, Mass.: MIT Press.

498

REMARKS AND REPLIES

Ruwet, Nicolas. 1972. Theorie syntaxique et syntaxe du francais. Paris: Editions du Seuil. Sag, Ivan, and Carl Pollard. 1991. An integrated theory of complement control. Language 67:63113. SigurLsson, Halldor A. 1991. Icelandic Case-marked PRO and the licensing of lexical arguments. Natural Language & Linguistic Theory 9:327363. Williams, Edwin. 1985. PRO and subject of NP. Natural Language & Linguistic Theory 3:297315. Williams, Edwin. 1987. Implicit arguments, the binding theory and control. Natural Language & Linguistic Theory 5:151180. Williams, Edwin. 1992. Adjunct control. In Control and grammar, ed. by Richard Larson, Sabine Iatridou, Utpal Lahiri, and James Higginbotham, 297322. Dordrecht: Kluwer. Wurmbrand, Susi. 1998. Infinitives. Doctoral dissertation, MIT, Cambridge, Mass. Department of Foreign Literatures and Linguistics Ben Gurion University P. O. Box 653 Beer Sheva 84105 Israel idanl@bgumail.bgu.ac.il

Potrebbero piacerti anche