The Primacy of Grammar Part 14

Consider (104), paint. No doubt, it is generally true that painting the exterior is the default option for painting because the color of the exterior is privileged in appearance; in fact, in the typical cases of painting walls, doors, fences, and so on, the exterior is the only plausible option. As Chomsky observes elsewhere, this could extend to imaginary objects such as golden mountains, even to impossible objects such as a round square. On this basis, Chomsky suggests that the option is mind-directed ( innate) rather than based on world knowledge.

Words and Concepts 131.

My intuitions dier. I can think of apartment complexes in which the exterior of the buildings are maintained by the management while resi-dents have the freedom to paint the interior of their apartments the way they wish. In such circ.u.mstances, John painted his apartment green has the default meaning that the interior of the apartment has been painted.

Furthermore, suppose the distribution of expertise is such that laypersons are able to paint the interiors of their houses themselves; painting the exteriors requires professional help. If we know that John is not a professional, then John painted his house brown will typically mean the interiors.

Finally, Gricean maxims (Grice 1975) suggest that, other things being equal, the sentence John painted the entire house brown is likely to mean the interiors since people typically paint individual rooms dierently.

That is, since people typically do not paint the exteriors with dierent colors, the phrase the entire house violates the maxim of relevance if it is intended to mean the exteriors. In this light, the sentence John painted the entire town brown has the opposite eect. Default readings of paint vary as they are tied to world knowledge about houses and towns (see Stainton 2006, 931932).

Climb seems to pose a dierent set of problems. Chomsky (1988, 190191; 2000a, 75) observes that its meaning is ''very complicated'' such that human nature gives the concept CLIMB virtually ''for free.'' Jackendo 's a.n.a.lysis of climb brings out some of this complexity (1992, 4647). It is instructive to study his a.n.a.lysis to see if the semantic intuition reported in (105) can be satisfactorily explained as part of a theory of I-language.

Jackendo cites four examples to capture the ''feature system'' of the ''cl.u.s.ter'' concept CLIMB.

(107) Bill climbed (up) the mountain.

(108) Bill climbed down the mountain.

(109) The snake climbed (up) the tree.

(110) ?The snake climbed down the tree.

According to Jackendo, uses of climb involve two independent features: (A) an individual is traveling upward, and (B) the individual is moving with characteristic eortful grasping motions. Jackendo calls this manner of motion ''clambering,'' so climbing is clambering upward.

For Jackendo, (107) is salient since it satisfies both (A) and (B); (108) and (109) are partly deviant because they satisfy either (A) as in (109) or (B) as in (108), not both; (110) is fully deviant because it fails to satisfy 132

Chapter 4.

either (A) or (B). In particular, the up in (107) is optional since, by (A), climbing up is the default meaning.

There are a number of problems with both (A) and (B). Since, according to Jackendo, climb is defined as clambering upward, strictly we cannot say, without redundancy, John climbed up the mountain. Further, since up is part of the meaning of climb, strictly we cannot say, without contradiction, John climbed down the mountain. As a matter of fact, we can say either without redundancy or contradiction, suggesting that up cannot be central to the meaning of climb. If up is not central, what explains the default reading? As his a.n.a.lysis suggests, Jackendo is likely to say that up as in (A) is central, but because (A) needs to interact with (B), the deviant option (108) needs to be admitted, which in turn forces up as a (deletable) option in (107); in other words, unless up is an option in (107), (108) cannot be admitted at all. Why is (108) admitted? Because (107) and (108) share (B), the manner of motion.

Thus, the entire weight of the a.n.a.lysis s.h.i.+fts to the characterization of the manner of motion. Now the characterization ''eortful grasping motions'' is vague enough to admit of a large variety of motion. Walking through rocks or thickets or a dark room full of furniture requires eortful grasping motions. None of these are nondefault readings of climb!

Further, when we look specifically at the manners of motion involved in, say, climbing up and down a mountain, it is unclear if the manners of motion match at all in the way in which, say, the manners of motion of swimming up and down the river match. In fact, the manners of motion of a snake climbing up and down a tree are a lot more similar than those involved in humans climbing up and down a mountain. Yet Jackendo declares (110) to be fully deviant because, as far as I can see, (110) fails the ''feature system'' stipulated by him. Also, when leopards or mountain goats climb a mountain they do not seem to require eortful grasping motions; in fact, even humans do not require grasping motions when they climb stairs. It is unclear to me if Jackendo 's internalist a.n.a.lysis of climb with conceptual features explains the intuitions reported in Chomsky's example (105) even if particular uses of the verb currently seem to confirm those intuitions.

4.2.2.

Nature of Lexical Inquiry To say that lexical intuitions are uncertain is not to say that we do not have pretty firm intuitions of nongrammatical anomaly in some cases.

Asking of certain questions is moot-for example, Can you hear me? We are likely to be puzzled by strings such as Bill admires the Pope's wife, Words and Concepts 133.

John's subordinate is John's boss, The mat sat on the cat, and the like.

Any popular piece on newspaper bloopers, political speeches, student papers, and so forth, contains dozens of examples of this sort. But these are not usually cited in scholarly treatises on semantics for a very good reason. Examples from language use, unless they are properly controlled, admit of any number of factors not all of which pertain directly to a theory of language; clearly, the anomaly of Bill admires the Pope's wife has nothing to do with English (or Italian, for that matter). When such popular, and often hilarious, examples are closely scrutinized, it turns out that they involve grammatical failure, performance failure, pragmatic failure, and insucient world knowledge, among many other things (Pinker 1995a). I am suggesting, however, that when it comes to data that is supposed to be theoretically interesting, the quality of the data leaves much to be desired insofar as nongrammatical ( for now, conceptual) aspects of meaning are concerned. So the chances are that problems with the data may lead up to problems with the theories they sp.a.w.n.

I attempted to raise two points simultaneously: an explicit point and an implicit one. The explicit point was to cast some preliminary doubt as to whether the alleged data for lexical semantics is salient enough for us to feel interested in the search for deep theory. Because we are looking for a theory in this domain from LF-up, as it were, we expect to find some initial data that has some promise of throwing further light on the human makeup: that is the only notion of an explanatory (cognitive) theory under discussion in this work. I am personally not convinced that the data cited in this regard hold the promise.

The implicit point was to show that what is called ''semantics'' in this area essentially amounts to giving at least a systematic account of the organization of a vast network of concepts used by humans. In each case, the data demanded that we know more about individual concepts and their relations to one another, as they attach to words, whether a noun or a verb. Apparently, then, the two points are opposed to one another: the second urges a research program that the first casts doubt on.

Yet they are not really in opposition. If we could pursue the program with some degree of success, then it is quite possible that new data will show up, and doubts about the quality of the initial data will gradually subside.

Perhaps we can already see that the resulting enterprise will have a very dierent flavor than that of grammatical research, in which a general research program took o from very specific (and incontrovertible) data of 134

Chapter 4.

linguistic behavior. The same, it could be argued, is the case with physics.

It is hard to imagine that physics could have reached the level it soon achieved if some ancient scientists had proclaimed, ''Let us try to understand the physical universe.'' It would not have been clear what was specifically there to understand. Even though the general query might well have been entertained in cla.s.sical philosophy as a quest for ''being,'' what led to physical theory were some sharp and deep facts that demanded sustained explanation: day and night, tides, eclipses, motion of pendulums and projectiles, angle of shadows cast by the sun, and so on.

Such facts abound in grammatical theory at various levels of generality, as we saw: rapid acquisition of language, the ambiguity of flying planes can be dangerous, the missing lexical Object in John is eager to please, to mention a few. I am not convinced that this sort of motivation exists for lexical semantics; all we have is some general motivation, not unlike the proclamation cited above. For example, a contemporary work on lexical semantics begins with the following project: ''It would be at least useful to investigate our semantic competence, that is, to wonder what kind of knowledge and abilities we possess that make it possible for us to understand language'' (Marconi 1996, 1).

4.3.

Lexical Decomposition The general task for lexical semantics is to come up with some account of how individual concepts are organized to lend, so to speak, meanings to individual words. Following poverty-of-stimulus arguments, it is natural to think that most concepts are somehow built out of primitive concepts which must be just a handful. In this picture, meanings of most lexical items, perhaps all, will be captured ''decompositionally''; in other words, the total meaning of an item will be broken down until the conceptual primitives are reached. Once the system is made available to the child, some of the nodes of the system-that is, the individual concepts-will be a.s.sociated with the sounds that the child hears: each sound-node a.s.sociation will count as a word. This is a fairly obvious and standard a.s.sumption: ''Two things are involved in knowing the meaning of a word-having the concept and mapping the concept onto the right form. This is the sense of 'knowing the meaning of a word' implicit in most discussions of language development, both scientific and informal'' (Bloom 2000, 17).

The child has to learn these a.s.sociations individually in any case. I said the project looks ''natural,'' although it is far from clear how it is going to be executed.

Words and Concepts 135.

It is obvious that the project requires some notational scheme to mark the concepts themselves, independently of how they are marked by the sound systems of particular languages. The overall organization of concepts is thus described in terms of a system of semantic or conceptual markers. The form of explanation is called ''markerese explanation.'' Almost everyone who looks at the universal bases of semantics adopts markerese explanation in one form or another, even if there is good deal of resistance in recent years to use discredited labels like ''decomposition''

and ''markerese explanation.'' This is particularly true of empirical research across languages. I will restrict the discussion to only those approaches which at least profess to adopt the broad explanatory goal of biolinguistics, namely, to solve Plato's problem in this domain.

Before I proceed I must note that the notion of lexical decomposition is also used in some recent work in distributed morphology (Harley and Noyer 1999; Embick and Marantz 2006, section 2) which apparently does not take recourse to markerese explanation.6 The distributed morphology approach proposes that lexical information is not located in one store called the ''lexicon,'' but is distributed/decomposed in at least three components. The formal morphosyntactic items, but not the phonological ones, enter into syntactic computation to LF and are then mapped onto the meaning interface. The insertion of ''vocabulary items,'' such as count, animate and phonological properties among others, is delayed and fed directly into the meaning interface via phonological computation.

Encyclopaedic information such as canine are fed into the meaning interface directly. In this scheme, then, there is no ''sound'' in narrow syntax ( computation to LF) and LF is viewed as a level of representation which exhibits certain meaning-related structural relations, such as quantifier scope (Harley and Noyer 1999).

Insofar as the second set of items do not enter into syntactic computation to LF, the scheme bypa.s.ses the vexing issue, noted above, regarding the role of these ''semantic'' features, such as artifact, during syntactic computation; according to the proposed scheme these features do not enter narrow syntax at all. For the purposes of this work, I certainly welcome as frugal a conception of LF ( output of narrow syntax) as possible. In that sense, the postulation, if valid, of ''late insertion'' of vocabulary items (on which most of the work in distribured morphology has been done, not surprisingly) strengthens the perspective developed in this work. As for the third-encyclopedia-component, it is hard to see how the scheme avoids markerese explanation in one form or other when a systematic attempt is made to describe the general organization of this 136

Chapter 4.

component. Notice that the organization of the post-LF component of meaning is exactly the topic currently under discussion.

4.3.1.

Initial Objections Some authors working on language dismiss the very project of markerese explanation. One of these objections runs as follows. As Norbert Hornstein (1989, 36) puts it: ''how does saying that in understanding a sentence we map them into mentalese help explain anything unless we understand what it is for someone to understand mentalese? Postulating semantic markers only pushed the problem back one step. Instead of explaining what it was for a word to have a content, the problem was to explain what it was for a marker in mentalese to have a content.''

There are many domains of inquiry where pus.h.i.+ng ''the problem back one step'' signals significant progress. The theory of evolution is a case in point. Arguably, the task for the theory of evolution is to explain how species, any species, evolved. But the explanations actually oered typically fall short of the global task; usually it is said that species X evolved from species Y, Y from Z, and so on, without really hitting a nonspecies end until very recently. Similar remarks apply to cosmological theories in physics. It does not follow that hundreds of years of evolutionary and cosmological explorations have been a waste of time. It seems then that the force of Hornstein's objection depends entirely on the specific nature of ''pus.h.i.+ng back,'' not on the general fact alone.

A related objection to the very project is reported by Ray Jackendo (1990, 4), a leading exponent of markerese semantics himself: ''How do you know your putative semantic primitives really are primitive?

Mightn't there be an infinite regress?'' Again, the demand that the ''real''

primitives ought to be available all at once, and at the very beginning of the enterprise, is unfair. As Jackendo rightly notes, by these standards even the search for the fundamental const.i.tuents of the physical universe (the ''primitives'' of physics) will appear to be dubious since physics, at no point in its history, can legitimately claim to have reached rock bottom. There are fairly standard ways in which progress can still be judged from incremental steps. The question of course is whether markerese semantics meets these standards. It is hard to see that this question has an answer in advance of the enterprise.

Perhaps this is the right place to examine a set of more general objections that aim to beset the very idea of language research, including especially the study of word meanings currently under discussion. Recall that, in a decompositional a.n.a.lysis, concepts are linked in a complex network.

Words and Concepts 137.

Each link in that network represents a conceptual relation, which, in turn, represents a part of the meaning of related words, if the (nodes of the) links are verbalized at all. For example, COW will form a link with ANIMAL such that, when verbalized, the conceptual relation COW !

ANIMAL will represent a part of the meaning of cow (and animal ). Since this inferential link represents a part of the meaning of cow, the link must hold come what may, that is, whenever cow is used; in other words, the enterprise requires that there are a.n.a.lytical inferences of the sort just sketched (Fodor and Lepore 1994).

Following the work of Willard Quine (1953, 1960), the very idea of a.n.a.lytic relations.h.i.+ps between linguistic expressions has been widely questioned. Skipping details of Quine's argument (Mukherji 1983), his basic observation is that knowledge of linguistic meaning cannot be interestingly separated from extralinguistic knowledge. It follows that, according to Quine, since extralinguistic knowledge is in principle revisable, so are relations.h.i.+ps of meaning. Hence, there are no a.n.a.lytical inferences that display the (invariant) meaning-relations.h.i.+p between expressions.

It is dicult to estimate the scope of this generalization. For example, it is indisputable that there are a.n.a.lytical relations.h.i.+ps between, say, an active sentence and its pa.s.sive: John likes Mary strictly entails Mary is liked by John, but does not entail Mary likes John. Turning to words, we saw that there are aspects of the meanings of persuade and intend such that John persuaded Mary to buy the book entails Mary intends to buy the book, but does not entail anything about John's intentions. We also saw that in each case, general grammatical explanations are available.

Active-pa.s.sive entailment is predicted by the interaction of y-theory and Case theory with the theory of movement; the entailment between persuade and intend is captured in the general syntactic frames for causatives.

It follows that there are aspects of the concept of a.n.a.lyticity that are satisfactorily covered within grammatical theory; if there are other problematic aspects of the concept, then they are likely to belong to the nongrammatical aspects of linguistic expressions. Grammatical theory thus looks insulated from Quinean charges. The question arises as to whether the lessons from grammatical theory can be extended to the study of specific word meanings.

It is unclear if the question can be meaningfully addressed from a priori philosophical grounds alone, as Quine advocated (Uriagereka 1998).

After all, Quinean charges could have been generally advanced for the relations.h.i.+p between persuade and intend before the causative frames were unearthed. Once grammatical understanding was reached, the charges 138

Chapter 4.

lost their general eect. Maybe, if we begin with such facts, we will be able to reach a point of abstraction where some allegedly a.n.a.lytical inferences will not look so bad in that they will engender a suciently attractive semantic picture; maybe we will not even call them ''a.n.a.lytical inferences''

anymore. Or, maybe, we will locate specific reasons why these charges continue to hold for word meanings.

The strategy of deferment possibly extends to a more intractable problem raised by Quine (1960). Imagine an English-speaking linguist visiting an alien culture to learn its language. Suppose a rabbit scurries by and a native utters gavagai. Should the linguist write down that gavagai means rabbit? As Quine (1960, 52) observes, ''Point to a rabbit and you have pointed to a stage of a rabbit, to an integral part of a rabbit, to the rabbit fusion, and to where rabbithood is manifested.'' In eect, the linguist has no empirical control on the (specific) meaning of gavagai, and, thus, on the concept GAVAGAI. Decades after it was proposed, many authors continue to view the problem as fundamental for any coherent theory of word meaning, semantics, and learning of words (Larson and Segal 1995, 1617; Bloom 2000, 34; Jackendo 2002, 88; among others). As Chomsky (1980, 14) notes, Quine intended the problem to arise not just for ''problems of meaning but to any theoretical move in linguistics.''

Again, the scope of this generalization is unclear. As Chomsky (1980) suggested, Quine needs to show that the gavagai-problem appears in the same form for grammatical theory as well. For example, grammatical theory holds that the sentence the man you met read the book I wrote contains two grammatically significant units (noun phrases) the man you met and the book I wrote, but met read the is no phrase at all. Quine needs to hold that there is no fact of the matter here. Recall that the ''indeterminacy'' arose in the gavagai-case because the English-speaking linguist was faced with at least two ''equivalent descriptions'': (111) gavagai means (whole enduring) rabbit (112) gavagai means (undetached) rabbit-stage This can only happen if both RABBIT and RABBIT-STAGE are simultaneously available to the linguist as resources for describing the native's utterance of gavagai. In eect, (111) and (112) are equivalent descriptions of the linguist's intuition of what gavagai means to the native.

In order for this case to extend to grammar, then, the linguist should be able to formulate two equivalent descriptions of the intuition underlying the grammaticality of the man you met read the book I wrote: one in terms of noun phrases, the other in terms of cl.u.s.ters like met read the. It is Words and Concepts 139.

totally unclear what it means to formulate a description of the latter kind.

In any case, there are tons of evidence showing why a phrase-structure description of linguistic streams is salient (Chomsky 1957). Jenny Saran (2002) shows that structures such as (The) ( professor graded the) (students) are not ''predictable dependencies,'' that is, these structures are simply not learned from the linguistic environment even though the words are presented serially in the stimuli (see also Chomsky 2000d, 121).

Equivalent descriptions are a fact of science since theories are necessarily underdetermined by the evidence they cover; from equivalent descriptions, we choose the best theory we can basically from nonevidential considerations (Chomsky 1980, 22). Quine's problem thus looks rather benign once again when raised for grammatical theory: ''indeterminacy''

is nothing other than familiar underdetermination (George 1986; Hornstein 1991 for more).

Does this conclusion extend to the gavagai-case? Chomsky (1980) thinks so. According to him, (111) and (112) are just equivalent descriptions familiar in science. A geneticist works with fruit flies, and he a.s.sumes that two fruit flies are the same in relevant respects. There is no conclusive empirical control on this a.s.sumption, yet he holds on to this as long as it works. Similarly, when a linguist works on an alien language, he a.s.sumes that the speakers of that language are like us in relevant respects such that a native is likely to mean rabbit rather than rabbit stage in uttering gavagai, just as we do in uttering rabbit. No special problem arises when we s.h.i.+ft from the study of fruit flies to the study of words.

Chomsky's response may not be as decisive as it looks. No doubt the geneticist simply a.s.sumes that, other things being equal, the two fruit-flies he is working with are alike in relevant respects. When other things are not equal, there are ways to find that out. Again, each such finding will be underdetermined by whatever evidence is relied on by the geneticist at that point. Thus suppose one of the fruit flies is in fact an engineered object slipped into the sample by a jealous colleague. We can imagine this object to be as close to a real fruit fly in its appearance as we wish, except that it diers in one minute feature caused by a planted gene or something; the feature is so minute that it escapes standard methods of observation, but it takes the inquiry totally o the track. We wish these things do not happen, but there are ways of finding out when they do, including a confession from the remorseful colleague.

It is unclear if a parallel is available for the gavagai case. The way we posed the problem, it does not easily disappear with the a.s.sumption that the native is suciently like us. In fact, if the native does resemble us, 140

Chapter 4.

then he is all the more likely to entertain either (111) and (112) without our ability to figure out which one, for, ultimately, it is we, namely the linguists, who formulated the alternative hypotheses on the native's behalf. The native just uttered gavagai, and in doing so he might have been following some of our respectable philosophical traditions.

Furthermore, each of the hypotheses is itself underdetermined by the nonlinguistic evidence-the glimpsed rabbit-that give rise to such hypotheses. When we are ascribing two of these to the native, the problem is not just that they are individually underdetermined, but, from the evidence we have about the native-namely, his gestures at a rabbit-it is underdetermined which of the (two) underdetermined schemes is at work.

This, as far as I can see, is the point about ''additional'' indeterminacy mentioned by Quine (1969a, 303). Finally, unlike the fruit-fly case, it is not easy to think of some ''discriminating'' evidence that will temporarily settle the problem, for the same problem is likely to be built in to any conceivable additional evidence: will that be an evidence about rabbits or rabbit stages? Contrasted to the problem of a.n.a.lyticity, indeterminacy is a wonderfully deep problem.

Nevertheless, it is also true that we typically think of whole enduring rabbits when uttering rabbit. If the native is suciently like us, why should he typically think otherwise? In that sense, Quine could be making just ''a philosophical point'' (Quine 1969b, 34). It could be that the precondition for forming (112) is (111), that is, there could be some explanation that acquisition of RABBIT necessarily precedes that of RABBIT-STAGE. On this view, unless other conditions so warrant, rabbit means rabbit by default; so does gavagai. This will meet Quine's challenge provided we are able to study the structure of RABBIT that enters into our uses of rabbit without begging any of Quine's questions. If, in other words, there is a suciently abstract general theory of concrete nouns, we will be in a position to examine how the specific meaning of rabbit plays out.

A similar perspective obtains for Saul Kripke's influential remarks on rule following (Kripke 1982). It is well known that the concept of rule following permeates much of the research in the cognitive sciences, including biolinguistics. Tracing the brief history of contemporary approaches to mental phenomena, Chomsky (2005, 2) suggests that in many cases, ''The best available explanatory theories attribute to the organism computational systems and what is called 'rule following' in informal usage.'' As Chomsky points out, the usage applies not only to straightforward rule-governed phenomena such as language, but also to the functioning of the visual system, insect navigation, and much else.

Words and Concepts 141.

Following some remarks by Wittgenstein (1953), Kripke argues that the very concept of rule following is suspect. Kripke explains some of Wittgenstein's ''elusive'' remarks in terms of the quus-puzzle where two alternative definitions, (113) and (114), are given for computing the arithmetical sum of two numbers. (113) could be viewed as the ''standard''