Introduction
While the overwhelming majority of Chomsky’s work in linguistics has centered on syntactic theory, from the earliest days he has been deeply concerned with phonology, semantics, and the connections between them and syntax. In this brief note, I will be concerned with the syntax-semantics interface and Chomsky’s proposals about the nature of that interface.
We find a clear indication of the importance Chomsky ascribed to the connection between syntax and semantics already in Chomsky (1957), where the notion ‘linguistic level [of representation]’ is crucial:
What we are suggesting is that the notion of ‘understanding a sentence’ be explained in part in terms of the notion of ‘linguistic level’. To understand a sentence, then, it is first necessary to reconstruct its analysis on each linguistic level; and we can test the adequacy of a given set of abstract linguistic levels by asking whether or not grammars formulated in terms of these levels enable us to provide a satisfactory analysis of the notion of ‘understanding’. (CHOMSKY, 1957, p. 87)
The question is exactly how levels of representation provide suitable input to semantic interpretation.
1. T-marker as locus of interpretation
In the early days of transformational generative grammar, the nature of the syntax- semantics interface was sometimes discussed in terms of the question of whether, or to what extent, transformations ‘preserve meaning’. Consider the following passage from Chomsky (1955):
It is fairly clear from all the examples that we have discussed in this chapter that in some sense meaning is preserved under transformation. Naturally we could not hold that transform and transformed string are synonymous, since a transformation may add or subtract morphemes. But we might have proposed that transform and pre-image differ in meaning only in the meanings of the morphemes dropped or added. For example, ‘John - is not here’ differs semantically from ‘John is here’ in the meaning of not. But ... this cannot be the case, because [there are cases] of two distinct sequences of transformations with exactly the same starting point in unambiguous kernel sentences, and with exactly the same end point, but with different meanings associated with this final transform. While transformations have semantic correlation, it is not obvious just how this is to be described. [CHOMSKY, 1955, p. 509 of 1975 version]
This was one factor that made the issue of syntax-semantics connection a tricky one. Another factor was the lack of recursion in the phrase structure component in this model. To the extent that a complex sentence could be said to have an underlying structure, that underlying structure was a forest of trees rather than a single tree. And a forest clearly had no determinate semantic interpretation. To take a simple example, the structures underlying (1) are those of the sentences in (2).
(1) Mary said John knows it.
(2) a. Mary said it.
b. John knows it.
A generalized transformation embeds (2)b into (2)a. However, the forest underlying (1) also is the source of (3).
(3) John knows Mary said it.
Thus, the interface with semantic interpretation must be the T-marker, the entire history of the transformational derivation. There was no Deep Structure in this model, as there was no recursion in the base, as mentioned above. And Surface Structure couldn’t determine thematic relations, as there were no traces.
The case Chomsky was actually discussing in the quoted passage concerns the ‘constructional homonymity’ [structural ambiguity] of ‘John kept the car in the garage’. Chomsky analyzes this example, on either interpretation, as having as its phrase structure basis the pair of kernel sentences (the car is in the garage, John kept it). Distinct transformational derivations ensue, with the PP winding up inside or outside the object NP. So while it is true that the alternative derivations of two sentences have “exactly the same end point” in terms of words, morphemes, and lower level structure, the final derived phrase markers are actually distinct. Thus, the problematic situation implied by Chomsky does not obviously arise in this specific instance. But the general issue persists.
Chomsky summarizes the state of affairs with respect to semantic interpretation as follows:
More generally, it appears that the notion of ‘understanding a sentence’ must be partially analyzed in grammatical terms. To understand a sentence it is necessary (though not, of course, sufficient) to reconstruct its representation on each level, including the transformational level where the kernel sentences underlying a given sentence can be thought of, in a sense, as the ‘elementary content elements’ out of which this sentence is constructed. In other words, one result of the formal study of grammatical structure is that a syntactic framework is brought to light which can support semantic analysis. (CHOMSKY, 1957, p. 108)
2. Deep Structure as locus of interpretation
Chomsky (1965) introduced a significantly revised syntactic model, and a correspondingly revised model of the syntax-semantics interface. He observes of the earlier model:
... the only contribution of transformations to semantic interpretation is that they interrelate Phrase-markers (i.e., combine semantic interpretations of already interpreted Phrase-markers in a fixed way). It follows, then, that transformations cannot introduce meaning-bearing elements (nor can they delete lexical items unrecoverably ... (CHOMSKY, 1965, p. 132)
In part for this reason he proposed eliminating generalized transformations in favor of recursion in the base, thus introducing the level of deep structure.1 We thus have a new view of the syntax-semantics interface:
It is clear, as Katz and Fodor have emphasized, that the meaning of a sentence is based on the meaning of its elementary parts and the manner of their combination. It is also clear that the manner of combination provided by the surface (immediate constituent) structure is in general almost totally irrelevant to semantic interpretation, whereas the grammatical relations expressed in the abstract deep structure are, in many cases, just those that determine the meaning of the sentence. (CHOMSKY, 1965, pp. 161-162)
He summarizes the resulting model this way:
... one major function of the transformational rules is to convert an abstract deep structure that expresses the content of a sentence into a fairly concrete surface structure that indicates its form .... we are, in effect, assuming that the semantic interpretation of a sentence depends only on its lexical items and the grammatical functions and relations represented in the underlying structures in which they appear. (CHOMSKY, 1965, p. 136)
Chomsky called this revised approach the ‘Standard Theory’. It introduced recursion in the base, thus, the level of Deep Structure.
The new model is elegantly simple with respect to the syntax-semantics interface. BUT, there is a problem:
As it stands, this claim [that deep structure determines meaning] seems to me somewhat too strong ... For example, it seems clear that the order of ‘quantifiers’ in surface structures sometimes plays a role in semantic interpretation. Thus for many speakers – in particular, for me – the sentences ‘everyone in the room knows at least two languages’ and ‘at least two languages are known by everyone in the room’ are not synonymous. (CHOMSKY, 1965, p. 224)
In fact, this problem had already been noted in Chomsky (1957) with respect to the question, mentioned above, of what the contribution of transformations to semantic interpretation is:
... we can describe circumstances in which a ‘quantificational’ sentence such as ‘everyone in the room knows at least two languages’ may be true, while the corresponding passive ‘at least two languages are known by everyone in the room’ is false, under the normal interpretation of these sentences – e.g., if one person in the room knows only French and German, and another only Spanish and Italian. This indicates that not even the weakest semantic relation (factual equivalence) holds in general between active and passive. (CHOMSKY, 1957, pp. 100-101)
3. Deep Structure and Surface Structure as the locus of interpretation
A while later, Chomsky returns to this phenomenon, and a closely related one:
It has been frequently noted that order of quantifiers in surface structure determines the preferred interpretation of a sentence ... Similarly, consider the interpretation of negations such as:
I didn't want to argue with Bill about money
I didn't give the book to Bill
I don't approve of John's actions wholeheartedly
I didn't want to argue about money with Bill
I didn't give Bill the book
I don't approve wholeheartedly of John's actions
In all such cases, under normal intonation (i.e., nuclear stress and the associated pitch center on the final phrase) there is, it seems to me, a preferred interpretation, with the nuclear stressed phrase being the one that is ‘negated’. The same is true, as Lakoff has pointed out, in such sentences as ‘I didn't see John in England’, which suggests (under non-contrastive, normal intonation) that I saw him somewhere else; etc. Since the ordering in these cases is determined by fairly late transformations, it seems that once again surface structure is involved in determining semantic content or at least, preferred interpretation. (CHOMSKY, 1970b, pp. 258-259)
Ray Jackendoff explored this problem with interpretation of quantifiers and negation in detail. Here is one of his representative examples:
(4) Many of the arrows didn't hit the target. ≠ in truth conditions to
(5) The target wasn't hit by many of the arrows.
(JACKENDOFF, 1969b, p. 167; JACKENDOFF, 1969a, pp. 223-224)
As Jackendoff observes, “Under the assumption that transformations do not change meaning, [this fact] will be very difficult to account for.” (Jackendoff, 1969b, p. 168; Jackendoff, 1969a, p. 225). Jackendoff therefore argues against the semantic interpretive architecture of Chomsky (1965):
If we give up the assumption that transformations do not change meaning and that all semantic information is represented in deep structure, it immediately becomes apparent how to go about explaining the interpretation of ... negation. We simply need a way to relate the understood order of quantifiers and negation to their position in the derived structure. We can do this informally by saying that the understood order is the same as the order in the surface structure. (JACKENDOFF, 1969b, p. 172; JACKENDOFF, 1969a, p. 228)
In passing I point out a very interesting remark of Jackendoff’s concerning the structural counterpart of scope: “If the preceding discussion is correct, we have found that the structural notion ‘in construction with neg’ ... has the semantic correlate ‘within the scope of neg’.” (Jackendoff, 1969b, p. 176; Jackendoff, 1969a, p. 230). This essentially follows Klima (1964), and anticipates Reinhart (1976). Klima’s ‘in construction with’ was identical to Reinhart’s first version of c-command. Jackendoff then has a straightforward conclusion, and a novel conjecture: “We conclude then that the scope of negation must be determined at surface structure (or perhaps at end-of-cycle structure ...)” (Jackendoff, 1972, p. 303). The conjecture in parentheses is an obvious early forerunner of minimalist multiple spell-out. Citing Jackendoff, Chomskysays:
Since it is the notion ‘surface subject’ that is involved in determining sameness or difference of meaning [in passive sentences], the principle is inconsistent with the standard theory.
Furthermore, the principle of interpretation of surface structures seems clear, and, in addition, the transformations that form passives can be left in a simple form (though they will drastically change meaning, if they change the order of quantifier and negation). These facts, then, provide strong support for the hypothesis that surface structure determines (in part, at least) the scope of logical elements, and serve as strong counter-evidence to the standard theory in its most general form. (Chomsky 1970a, p. 81)
Chomsky continues, stating that one of the core ideas of the standard theory remains:
Jackendoff's arguments ... leave unaffected the hypothesis that the grammatical relations defined in the deep structure are those that determine semantic interpretation. If we modify the standard theory, restricting in this way the contribution of the base to semantic interpretation, we can take account of the fact that many aspects of surface structure appear to play a role in determining semantic interpretation; correspondingly, insofar as some development in syntactic theory is motivated by the demand that these aspects of semantic interpretation be expressed in deep structure, it will have lost its justification. (CHOMSKY, 1970a, pp. 81-82)
Barbara Partee pursued this issue as well:
Among the important issues in the area of semantics and its relation to the rest of the grammar, the idea that transformations might be meaning preserving is one that has an interesting history and one whose fate is still far from clear.... I will talk about some of the more interesting phenomena with respect to which the meaning-preservingness of transformations seems to be called into question. (Partee, 1971, p. 1)
... in talk about transformations preserving or changing meaning, the most common examples are transformations which can informally be thought of as relating sentences to other sentences. But transformations in fact operate not on sentences but on abstract phrase-markers, and it is not obvious that we have any direct semantic intuitions about these abstract structures, in particular any notion of synonymy between them. (Partee, 1971, pp. 4-5)
Partee goes on to make an important conceptual point:
For obligatory rules, ... the question of meaning-preservingness does not even make sense, for the input to the rule is an abstract structure with which we have no independent acquaintance. ... The question of whether transformations change meaning can therefore be meaningfully asked only of optional transformations. The clearest case is that in which two sentences are derived from the same deep structure, their derivations differing only in the application versus nonapplication of a certain optional rule. If the two sentences are synonymous, and if the same is true of all pairs related by the given rule, the rule is meaning- preserving; otherwise it is not. (Partee, 1971, p. 5)
And to the extent that transformations are meaning preserving, deep structure determines meaning.
Like Chomsky and Jackendoff, Partee examined sentences with quantificational expressions, and with the same concern: “A good bit of the current interest in quantifiers stems from the fact that there are a number of transformations which, as traditionally formulated, preserve meaning except when quantifiers are involved.” (Partee, 1971, p. 9). She gives the following pair of examples related by raising, observing that they are “sharply nonsynonymous” (Partee, 1971, p. 18):
(6)
a. Nobody is (absolutely) certain to pass the test.
b. It is (absolutely) certain that nobody will pass the test.
Crucially, she shows that (6) does indeed instantiate raising, using a test that was already standard, and remains so. “Certain must allow subject-raising because of sentences like [(7)].”
(7) There is certain to be an argument over that.
An expletive in matrix subject position excludes the possibility of a control analysis, thus implicating raising. Yet...
...under a common kind of grading system, [(6)b] [is] false, while [(6)a] is true.... A logician would have no difficulty in representing the differences in the sentences of [(6)] in terms of differing ‘scopes’ of three elements: negation, an existential quantifier, and a modal operator ‘certain’ ... (PARTEE, 1971, p. 18)
4. Logical Form (LF) as the syntactic level relevant to quantifier scope
May (1977) had extensive discussion of the syntax-semantics interface, arguing in detail that LF is the syntactic level that inputs semantic interpretation. One kind of example he examined involved syntactic raising, of the sort Partee had considered, but with apparently diverging properties. Here is one of his examples:
(8) Some politician is likely to address John's constituency.
May reports that
[(8)] may be taken as asserting either (i) that there is a politician, e.g., Rockefeller, who is likely to address John’s constituency, or (ii) that it is likely that there is some politician (or other) who will address John’s constituency. ... the reason that sentences like [(8)], containing raising predicates, are ambiguous is that they can be associated with two distinct logical forms ... (MAY, 1977, p. 189)
(9) [S [some politician]a [S a is likely [S t to address John's constituency] ] ] ]
This logical form arises from a standard raising instance of May's Quantifier Rule (QR).
(10) [S a is likely [S [some politician]a [S t to address John's constituency] ] ] ]
We get this logical form from a lowering instance of QR, as QR “applies freely”. And, interestingly, it is the sort of reading that Partee showed didn’t arise in her example. More on that below. May summarizes the situation as follows: “In [(9)], the quantifier 'some’ has scope wider than the matrix predicate ‘likely’; it corresponds to the (i) reading of this sentence as described above. In [(10)], on the other hand, the quantifier has scope narrower than 'likely'; this logical form corresponds to the (ii) reading above.” (MAY, 1977, p. 192)
May also claims that, in effect, a in (10) can be ignored as it is not an argument position. A popular more recent instantiation of May’s basic idea assumes, with Chomsky (1995b), that movement is a copying operation and that there is (sometimes) a choice as to which copy to use for semantic interpretation. The ‘lowering’ phenomenon (which I will call QL) arises from using a lower copy for semantic purposes.
Sloan and Uriagereka (1988) suggested an interestingly different account of the Q- lowering phenomenon. Rather than having a lowering operation largely undoing the effects of raising, or calling upon activating a lower copy, they proposed, in the spirit of Lasnik (1972), and the conjecture of Jackendoff (1972) cited above, that quantifier scope is determined cyclically, but as part of the syntactic cycle, not as part of a later LF cycle. This is an obvious precursor of the multiple spell-out of Uriagereka (1999), which shortly led to Minimalist single-cycle syntax.
5. An empirical difficulty: QL is much less general than it is expected to be
None of these accounts predict any limitation on lowered interpretations. Whenever there is a raising construction, ‘lowered’ readings should be possible. But, as first observed by Partee (1971), cited above, and as I discussed in Lasnik (1998) and Lasnik (2010), among other places, such readings are very often unavailable, as indicated by the standard test of paraphrase by the it ... [finite clause] alternant. Below, I provide a sampling of examples from those works, as well as some new ones, and some brief discussion. (I use ≠ to indicate that the first example cannot be paraphrased by the second.)
(11)
a. Nobody is (absolutely) certain to pass the test ≠
b. It is (absolutely) certain that nobody will pass the test (Partee, 1971)
(12)
a. No large Mersenne number was proven to be prime ≠
b. It was proven that no large Mersenne number is prime
(13)
a. Noone is certain to solve the problem ≠
b. It is certain that noone will solve the problem
(14)
a. Every coin is 50% likely to land heads ≠
b. It is 50% likely that every coin will land heads
(15)
a. Every coin is 3% likely to land heads ≠
b. It is 3% likely that every coin will land heads
In (15), if we have 5 coins, the b. reading (i.e., the lowered one) would be far more plausible. Yet it is still unavailable. However, Boeckx (2001) argues that unlike likely, 3% likely is not a raising predicate (using the Partee test above). Then the only source for (15)a would be control, and lowering is not expected with a control structure. Boeckx provides example (16).
(16) *There is 30% [sic] likely to be a man in the garden (Boeckx, 2001, p. 541)
In fact, a number of my consultants agree with Boeckx’s judgment on (16). But not all of them do. Yet none of them get the lowered reading for (15)a.
Following are some examples like (15), but where many speakers (in fact, nearly all of my consultants) clearly allow raising, by standard tests:
(17)
a. Every student is quite likely to pass the exam ≠
b. It is quite likely that every student will pass the exam
(18)
a. There is quite likely to be an investigation
b. The cat is quite likely to be out of the bag
(19)
a. Everyone is fairly certain to pass the exam ≠
b. It is fairly certain that everyone will pass the exam
(20)
a. There is fairly certain to be a storm today
b. The shit is fairly certain to hit the fan (when this news breaks)
Another possible failure of ‘lowering’ comes from an observation about scope that Zubizarreta (1982) attributes to Chomsky, and that is discussed again by Chomsky (1995a):
(21)
a. (it seems that) everyone isn't there yet
b. everyone seems [t not to be there yet]
Chomsky (p. 327) argues as follows: “Negation can have wide scope over the Q in [(21)a]... but not in [(21)b]”, concluding that “...reconstruction in the A-chain does not take place, so it appears.” The scope facts seen in this kind of example might still be compatible with May style literal lowering (which would adjoin the universal quantifier to the lower clause), as in fact Chomsky suggests, but would still be incompatible with activation of a lower copy in an A-movement chain (since a lower copy in lower subject position should be able to scope under lower clausal negation, just as everyone can scope under negation in (21)a), a point that Chomsky makes.
6. Towards a theory?
A theory of what?
(22) QL exists, as it would under any of the accounts above. Then we need an account of why it is so often blocked (basically, with anything except indefinites.
OR
(23) QL doesn't exist. Then we need an account of why it doesn't. And we also need an
account of why it looks like it does with indefinites.
Suppose it doesn’t exist. Well, there are at least three standard powerful arguments against this point of view. First, though in many lowering situations the truth conditional differences between the claimed two readings are not very easy to pinpoint, Fox (1999) presented a case with sharper differences:
(24) Someone from New York is very likely t to win the lottery
Fox explicates the situation as follows.
One interpretation results when the quantifier takes scope in the final landing site. For the sentence to be true under this interpretation, there must be a person from New York who is very likely to win the lottery (e.g., a person who bought enough tickets to make winning a likely outcome). Under the second interpretation, in which the quantifier has scope in the position of t, the truth conditions are much less demanding; they merely require that there be enough ticket buyers from New York to make it likely that the city would yield a winner. (FOX, 1999, p. 160)
Let’s now strengthen the situation:
(25) Someone from New York is certain t to win the lottery.
Imagine that a particular New Yorker bought all the tickets. Or, alternatively, imagine that only New Yorkers bought tickets. (25) is a good description of either situation, parallel to what was seen with the Fox example. And, as in that example, it is tempting to implement the second reading via lowering, with Someone from New York below certain at LF.
But now imagine the exact same pair of situations concerning ticket purchasers (and the speaker's knowledge thereof). And suppose the drawing has already taken place, but the winner has not yet been announced. Suppose the speaker were to utter (26):
(26) Someone from New York won the lottery.
(26) seems to be an accurate and felicitous report of either situation, just as in the case of Fox's (24) or my modification in (25). But (26) is a completely transparent extensional context. If the “lowered reading” is to be instantiated by lowering of Someone from New York, what operator does that expression lower below?
Second, the classic ‘trapping’ effects (as in May (1985) and Lebeaux (1998)), where the need to bind something in the matrix clause evidently precludes the low interpretation for the quantifier:
(27)
a. No agenti was believed by hisi superior to be a spy for the other side ≠
b. *It was believed by hisi superior that no agenti was a spy for the other side (May, 1985)
Note, though, that there are two interfering factors here: Negatives don't lower in the first place, as seen above; and, even controlling for that, what can we really conclude from the fact that a particular sentence cannot be paraphrased by an ungrammatical sentence (one violating weak crossover)?
(28)
a. Some agenti was believed by hisi superior to be a spy for the other side ≠
b. *It was believed by hisi superior that some agenti was a spy for the other side
Third, 3-way scope interactions:
(29) Someone seems [ t to love everyone ] " can scope over $ (Aoun and Li, 1993)
(30) Someone wants [ PRO to kiss everyone] ?" cannot scope over $
Note that for the issue at hand, it does not suffice to show that " can scope over $. Rather, it must also be true that seem scopes over both of them. I am willing to believe that that is true, but I am not certain. The following example is relevant:
(31) Two women seem to each other to be expected to dance with every senator (Lebeaux, 1998)
Two women must be high (to license each other). " cannot scope over 2. Thus, it is reasonable to conclude that in (29) also, the scope of " is limited to the embedded clause, and hence that the scope of $ is the lower clause.
Another pretty clear instance of apparent distinct interpretations:
(32)
Two of the students in the room are (almost/virtually) certain to have the same birthday
[Suppose there are 60 students in the room. The low reading has truth conditions clearly distinct, and they are satisfied.]
7. A new approach (one excluding lowering in general, but allowing genuine low scope where it plausibly exists)
Recall the Sloan and Uriagereka (1988) approach to lowering phenomena, one that fits neatly into single cycle syntax. And suppose, following May (1977), that scope is generally achieved via QR, but where QR is always an raising operation. Then, to get embedded scope, QR would have to operate on the embedded cycle. But then subsequent raising to matrix subject position would constitute an instance of ‘improper movement’ from an -position to an A-position (a constraint first discussed by Chomsky (1973)). Matrix scope would cause no such problem. Its derivation would involve perfectly standard A-movement (raising) followed by -movement (QR), similar to what happens in:
(33) Who [t is likely [t to solve the problem]]
But what of the low readings of indefinites in raising constructions? For these we can rely on a special property of indefinites: that they are, or can be, variables rather than quantifiers, an idea developed by Heim (1982). These variables are then provided with binders by existential closure. Following Reinhart (1997) among others, I would take existential closure to be available in all clausal domains, not just the matrix. If closure is introduced in the lower clause in the examples at issue, we get low scope. And no constraint on improper movement would prevent subsequent A-movement of the indefinite up to subject position of the higher clause (though the raising would be semantically vacuous, at least with respect to scope). An alternative derivation would have raising, then closure in the higher clause. This gives high scope. The Lebeaux examples with high binding and no possibility of low scope fall out, as low scope implicates low closure. But then the high pronoun or variable could not be in the scope of the introduced existential.
One remaining task (maybe a hard one): Develop a principled theory of improper movement that would be effective here. As far as I know, noone has until now claimed that improper movement might block raising of the subject of a raising complement.
8. Acknowledgments
I would like to thank Aaron Doliana and Rodrigo Ranero for helpful suggestions and for indispensable assistance with the manuscript.
XML
