61.
(50) a. *Whoi does hei think ei is intelligent b. *Whoi does hei think that I like ei c.
Whoi does hej think ei is intelligent d.
Whoi ei thinks hei is intelligent (50a) and (50b) are examples of ""strong crossover"" in which a WP has crossed over the coindexed p.r.o.noun he while moving to the main-clause front. In (50a), the wh-trace is bound in its governing category while in (50b) it is A-bound by the p.r.o.noun he. The resulting expression in either case is ungrammatical. In (50c) and (50d), there is no crossover since either the p.r.o.noun is not coindexed as in (50c) or the p.r.o.noun has not been crossed as in (50d). In each case, the trace is bound by a WP from an A-bar position, not an A-position, showing that a wh-trace is A-free.
Hence a wh-trace is covered under Principle C.
The a.s.similation of an empty category in the cla.s.s of such obviously referring expressions as John and the men has important theoretical consequences, as we will see. The current point is that we are beginning to have a rather natural account of the chain (wh-, e). The original intuition that an empty category must be a dependent category is still upheld by the idea that wh-traces, I may now say, must be A-bar bound in order to be licensed; this, of course, cannot be said of other r-expressions discussed so far since they may not be A-bar bound. In fact, this property of a wh-trace lends further support to the idea that a WP must move to the Comp position; otherwise the trace cannot be licensed given Principle C. But this licensing requirement is not part of Binding theory and, as we will see in a moment, it is not even independently needed since it follows from a more restrictive condition called ""Empty Category Principle."" Each of the elements of the chain (wh-, e) is thus interesting: a WP is an NP that is not a genuine argument though it functions as a Subject or a direct Object; a wh-trace is also an NP that is an empty, and thus a dependent, element, though it is cla.s.sed as an r-expression. These features will have a natural explanation when we get to LF-interpretations.
So far the theory has not imposed any explicit criterion of dependency on wh-traces; we only know that these are A-free. Even if we uphold the implicit idea that wh-traces are A-bar bound, it still does not fully square with our original intuition that empty elements must be somewhat ""narrowly"" dependent since the only element a wh-trace is linked with is a rather ""distant"" element, namely, a WP in the A-bar position, especially if the WP has ""hopped.""
The general idea of ""narrow"" dependence of any category is largely captured in the concept of government since it imposes a local relationship 62
Chapter 2.
between two elements, a governor and a governee, where the governor is typically a lexical head and the governee a complement. However, a Subject is either ungoverned or is governed by Infl, not really a ""proper""
governor. This is particularly problematic for traces, since they do, on occasion, occupy the Subject position. Thus it is necessary to impose (51), the Empty Category Principle.
(51) A trace must be properly governed.
where ""proper government"" is defined as in (52).
(52) a properly governs b just in case a. a governs b and a A fN, V, A, Pg, or b. a locally A-bar binds b So for wh-traces, three possibilities of well-formedness follow: (i) when a WP moves from an Object position, the trace will be governed by a lexical head that is typically a verb, (ii) when the phrase moves from the Subject position, it adjoins to the next Comp position, from where it locally binds its trace, and (iii) if the phrase has moved/hopped further, then the trace of the second movement locally binds the first trace. Consider (53).
(53) a.
Whoi do you think Bill saw ei b.
Whoi ei thinks that Bill is intelligent c.
Whoi do you think ei (ei left) d. *Whoi do you think ei that (ei left) The first three examples ill.u.s.trate the listed options respectively; the last example shows the failure of option (iii). In (53a), the trace is properly governed by the verb see. In (53b), the trace is locally A-bar bound to the WP. In (53c), the trace of the second movement locally A-bar binds the first trace; note that the WP properly governs the second trace. In (53d), however, the element that intervenes between the two traces to block proper government, as seen clearly in the tree diagram in figure 2.5 (adapted from Sells 1985).
2.3.3.
LF.
After witnessing the elaborate licensing conditions on s-structures, we face the general problem of interpreting s-structures. I will ignore the issue of phonetic interpretation and turn directly to issues regarding semantic interpretation.
Over the last three decades, there has been a growing consensus that s-structures are not directly semantically interpreted. Rather, s-structures Linguistic Theory I
63.
Figure 2.5 That-trace feed into another level of representation to which, in turn, semantic interpretation is attached. The displacement problem and the consequent problem of indirect sound-meaning connections make the requirement for an additional level of representation after s-structures quite natural.
As noted, it was found that d-structure representations are unsuitable for feeding into the semantic component-for example, wh-elements need to move in order to be interpreted. The general idea is that, in order to uphold indirectedness, computation must branch. As usual, this issue has become moot with the advent of MP. Since MP does not postulate ""inner"" levels of representation, it proposes that phonetic computation and semantic computation proceed in parallel, although computation does branch at a point called ""Spell-Out,"" where the phonetic features are stripped away to form a separate phonological representation PF. However, Spell-Out is not a level of representation. Since LF and PF are the only levels of representation in MP, it changes the mechanisms of computation fairly drastically. For example, we can no longer treat multiple-wh constructions at dierent levels.
Keeping to G-B, another need for a separate level for semantic representation comes from a global constraint on phonetic and semantic representations. It is natural to expect that each of these levels will consist only of those elements that are relevant for the level concerned, such that if a level contains elements that cannot be interpreted at that level, then the representation will be rejected. Thus a principle of Full Interpretation 64
Chapter 2.
(FI)-a representation may not contain any vacuous element-may be invoked for the interface levels.
Consider pleonastic elements such as it and there. Since English requires a lexical Subject, these elements appear at the d- and s-structures, as noted. They must also appear at the phonetic interface since English sentences contain them. However, they cannot appear at the semantic interface since they cannot be given semantic interpretation. Thus, in order to obey FI, they must be deleted before a representation is ready for semantic interpretation, although there is much controversy on how the deletion of pleonastic elements is executed. The same applies to the vacuous preposition of in destruction of the city, which was needed solely for licensing the city, as noted. From the other direction, if semantic representation requires introducing covert elements, such as wh-trace, as a necessary condition for semantic interpretation only, then such elements cannot be allowed to either remain or be introduced at the phonetic interface. It is natural, then, that FI is satisfied only if computations branch at the s-structure.
Another argument for the existence of a distinct level of semantic representation is that meanings are largely ""wired-in"": we do not expect languages to vary significantly in the semantic component; languages vary in how they sound. What notion of ""semantics"" is at issue here? In the literature, ""semantics"" is standardly defined in terms of a list of phenomena: relative quantifier scope, scope of negation, modality, opacity, p.r.o.noun binding, variable binding, focus and presupposition structure, adverbial modification, and so forth (Hornstein 1995, 1). It is not immediately obvious that ""semantics,"" so characterized, ought to be ""wired-in."" Why should, say, the properties of variable binding or adverbial modification be invariant across languages?
It is frequently suggested by Chomsky and others that semantics, even in the restricted sense, is likely to be ""wired-in"" because ""the most fundamental principles of semantics are . . . remote from the data available to the child (situations of utterance, the behaviour of other speakers etc.)""
(Higginbotham 1985, 550, cited in Hornstein 1995, 3). The child has to form judgments about semantic organization essentially by observing what people are doing with language, and what people are doing is subject to so many dierent interpretations that it supplies only vague cues about how a sound is to be interpreted. The data for the sound systems, in contrast, are somewhat more directly available to the child in that the ""perceptual apparatus just tunes in"" (Chomsky, in Hornstein 1995, 4; Karmilo-Smith 1992, 37; Ramus et al. 2000). In sum, the idea is that Linguistic Theory I
65.
the argument from the poverty of stimulus applies more strongly for semantics than for phonetics.
The issues of whether LF exists and whether LF is language-invariant are distinct, though the idea of language invariance bolsters the existence claim. In view of indirectedness, it is natural to postulate a level of representation where all ""grammatically determined information relevant to interpretation are consolidated"" (Higginbotham 1985, 549), where ""in-formation relevant to interpretation"" stands for the list of semantic issues drawn up above. Explicit empirical argumentation is then needed to determine if the postulation makes a dierence in explanation especially in terms of wider coverage of data, and whether the argument for language invariance is bolstered. There are a variety of interesting arguments to that end. I will barely scratch the surface below.9 A number of crosslinguistic studies conducted in recent years seem to supply the evidence needed. To take just one study that is widely mentioned in the literature, Huang (1982) suggested that languages such as Chinese and j.a.panese do not have overt wh-movement in syntax-that is, at s-structure. In other words, unlike English, WPs in Chinese and j.a.panese remain in situ in syntax. However, when the interpretations of sentences containing WPs are closely investigated, especially with respect to their scopal properties, it turns out that WPs in Chinese and j.a.panese covertly move to the clause boundary just as in English; there are other more complex considerations. Therefore, not only that interpretation is a.s.signed to a level of representation dierent but derived from s-structure, the structure to which the interpretation is a.s.signed conceals dierences between languages. There is good reason, then, to believe that there is a level of representation, to be called ""Logical Form"" (LF), that is largely language-invariant.
Recall a pending problem concerning (42), repeated below as (54), in which the second WP of the sentence who ate what could not be moved out of the direct Object position at s-structure. It is a problem since, despite being a nonthematic item in the sense suggested, a WP cannot fail to be a.s.signed a y-role in that position.
(54) [S-bar [Comp whoi][S ei Infl eat what]]
Following the example of covert wh-movement in Chinese, it is natural to think that the element what also moves covertly at LF. This results in the LF-structure (55).
(55) [S-bar [Comp whatj [Comp whoi]][S ei INFL eat ej]]
66.
Chapter 2.
In a similar fashion, we may now extend the a.n.a.lysis easily to various multiple wh-constructions such as those in (56).
(56) a. I wonder who gave the book to whom.
b. Who remembers where John read what?
There are other interests here that are widely discussed in the literature.
For example, some LF-representations of multiple wh-constructions seem to violate Subjacency. This suggests that Subjacency might hold only at s-structure and be violated at LF, just as Case theory holds at s-structure and is sometimes violated at d-structure. In English the wh-instance of the rule Move-a obeys Subjacency at s-structure, and ECP at LF. In the Minimalist Program, obviously, these issues of selective application disappear.
2.4.
Grammar and Scope Problem We can now appreciate how scope ambiguities of sentences containing quantified phrases (QPs), which were at the heart of Russell"s theory that brought us here, may now be handled in the general theory of LF. In fact, the theory easily extends uniformly to various subcla.s.ses of quantifiers, to interactions between them as well as to interactions between quantifiers and WPs. These include number quantifiers (two), ""strong"" quantifiers (every, most), ""indefinites"" (a) and ""definites"" (the), among others. A small sample of simple examples are given in (57).
(57) a. Most Indians speak two languages.
b. Every boy danced with a girl.
c. Who gave a gift to every child?
To simplify matters, let us a.s.sume that this instance of Move-a-that is, quantifier raising (QR)-has exactly the same properties as wh-movement except that a quantifier (controversially) raises to the IP node.10 Let us define ""quantifier scope"" in terms of asymmetric c-command: a QP b is in the scope of a QP a just in case a c-commands b but b does not c-command a. Move-a ( QR) generates multiple structures from a single set of lexical elements, creating traces along the way, if all of the structures are individually licensed. The first two sentences in (57) have, roughly, the LF-representations (58a) and (58b). I am skipping fuller a.n.a.lysis of these structures since their basic forms should be fairly obvious by now. I will a.s.sume that scopal ambiguities of each sentence have been structurally represented with full generality, as promised above (section 2.1).
Linguistic Theory I
67.
(58) a. LF1: [IP [most Indians]i [ei speak two languages]]
LF2: [IP [two languages]j [IP [most Indians]i [ei speak ej]]]
b. LF1: [IP [every boy]i [ei danced with a girl]]
LF2: [IP [a girl]j [IP [every boy]i [ei danced with ej]]]
A natural way of initiating the discussion of the interpretation of these structures is to go all the way back to the pending problem of how to interpret the elements of the chain (wh-, e). Recall that a WP moves to the Comp position and a wh-trace is created at the original position. At that point, we just recorded that each of the elements of the chain (wh-, e) is interesting. Let us review some of the facts about the empty element e, a wh-trace. We know that e is an r-expression like John and the child since it is unlike an NP-trace in two respects: (i) e is A-free while an NP-trace is A-bound, and (ii) e is a.s.signed both Case and y-role while an NP-trace typically lacks Case. Thus, a WP-trace may be viewed as a ""place holder for a true referential expression"" (Chomsky 1988, 88).
However, e is also unlike John and the child in that e must be locally A0-bound either to a WP or to its trace; in other words, it must appear in a chain (wh-, e1, . . . , en). Turning to WPs, although a WP must head a chain containing wh-traces and, hence, is coindexed with them, coindexing here, unlike argument chains, does not amount to coreference since a WP is a nonreferring expression. In sum, a WP binds an element in its domain where the element is otherwise given full semantic interpretation.
These are exactly the properties of quantifiers and bound variables in standard quantification theory. Though a bound variable, like a name and a free variable, can be a.s.signed an independent interpretation (in a model under a system of a.s.signment), it can only occur, unlike names and free variables, in a.s.sociation with a quantifier whose task is to bind the variable. So a WP and its trace are naturally viewed as a quantified phrase and a variable respectively. Let us examine this crucial point from a slightly dierent angle.
It is well known that standard QP-constructions of English, such as all men are mortal, are naturally viewed as having a quantifier-variable structure in the sense just outlined. This explains why, despite serious problems noted above (section 2.1), canonical representations of natural-language sentences in the notation of first-order logic look so natural. Adoption of this notation helps explain the telling semantic dierences between, say, all men are mortal and Socrates is mortal. All men are mortal expresses a general proposition-that is, a proposition that is not about anyone in particular. The proposition concerns, at best, a cla.s.s whose members are 68
Chapter 2.
referred to, as Quine puts it, ""with studied ambiguity."" Socrates is mortal, on the other hand, expresses a singular proposition since it is plainly about Socrates. This dierence is traceable to dierences between all men and Socrates and the dierential relationship they bear to the predicate terms. All of this is clearly brought out in the notation of (restricted) quantification as follows: (59) All men are mortal : (all x: man x)(x is mortal) Socrates is mortal : (Mortal)Socrates Therefore, the idea that the chain (wh-, e) has a quantifier-variable structure will be bolstered if we can show explicitly that WPs are more naturally aligned with QPs than with names. Following Chomsky 1980, 160162, consider (60) and (61).
(60) Johni betrayed the woman hei loved (61) The woman hei loved betrayed Johni The coindexing shows that in either case we may take the p.r.o.noun he to refer to John. Now, let us replace John in either case by a QP everyone to yield (62) and (63).
(62).
Everyonei betrayed the woman hei loved (63) *The woman hei loved betrayed everyonei This time coindexing fails for the ""quantified"" version of (61). It is natural to explain these results in the following idiom. (62) allows the interpretation that for each person x, x betrayed the woman x loved. This interpretation is not allowed in (63); however, if he and everyone are not coindexed, (63) will allow the interpretation that for each person x, the woman he loved betrayed x, where he refers to someone else. Suppose we ocially adopt this interpretive idiom and a.s.sign it to LF-representations.
I have not given the full LF-representations for (62) and (63) because the essential forms can be easily constructed. In the LF-representation for (62), the QP everyone moves from the Subject position of the main clause to adjoin to the IP, leaving a coindexed trace behind; this creates the LF-structure [everyonei ei . . . hei . . . ]. In (63), the QP moves from the Object position of the embedded clause to the sentence front, and the coindexed trace appears at the sentence end; this creates the LF-structure [everyonei . . . hei . . . ei]. (62) and (63) may thus be said to have the logical forms (64) and (65) after the idiom is a.s.signed.
Linguistic Theory I