The Science of Language

by Noam Chomsky

  1 Pietroski puts the basic point in a different way in his (2002) and elsewhere. He construes semantic features as instructions to build concepts. “Instructions to build concepts” might appear to some less worrisome than a concept/MOP that has alternate features. But the concept so built raises the same issues: lexical + morphosyntactic instructions to build a concept at a sentence's semantic interface or other ‘transfer’ place still builds a (complex sentential) polity-type concept displaying a France that is hexagonal and a republic. It also raises puzzles: what is an “instruction to build a concept”? To whom or what is it directed? Not a person, surely; is it then a module? If so, what is the nature of that module: how does it operate, what does it take as input, how does it produce its output? Does it produce its output by duplicating morphosyntactic processing? If so, why not have the sentential MOP and its constituents at a semantic interface serve as the complex concept itself? I prefer to place both semantic features in a cluster of semantic features that constitute a lexical MOP that then becomes (a part of) a sentential MOP. This allows also for a different and I think preferable view of interpretation: features do not give instructions, but (help) configure understanding and experience. See below, Appendix XII.

  That said, I am aware that Pietroski has strong reasons internal to his well-developed and impressively defended semantic theory (2005) and a complementary account of the lexicon (forthcoming) to adopt the ‘instruction giving’ account that he does. I would like to find a way to adopt both his account of concepts and mine.

  2 They are not appearance properties, of course, if that means properties such as colors and sounds that might themselves constitute an animal's DOGmop. They are human-specific, and include features such as NATURAL OBJECT and presumably some kind of function-for-us. That said, like colors conceived of as ways in which the visual system configures visual experience, the basic assumption is that these features partially configure experience (here, of dogs). That is just what MOPs ‘do’. See Appendix XII.

  3 Cases such as dog-pictures are not very interesting: that is part of what PICTURE does. Dog-toys are more interesting cases. Children do often invest their toys with psychic continuity, as well as other properties of living creatures. As I mentioned: interpretation is probably beyond the reach of science. But the potential contributions of various cognitive systems are not.

  4 Perhaps there is a measurable probability that people will greet each other in English by uttering “hello.” But there are few other such cases. Their view is on the face of it absurd. But it has attracted many others. Showing it is absurd in a way that will convince these others may be impossible, but perhaps it is worth the effort.

  5 These statistical measures indicate again that their view is not driven by any actual assessment of probabilities, but by dogma.

  6 I add an ‘s’ to ‘paper’ because – apparently – Chomsky was wrong to think that the view expressed is in a single paper. The relevant papers include Weckerly and Elman (1992), Elman (1990, 1991, 1993), and Elman and Lewis (2001).

  7 It is difficult to believe that their efforts really do have anything to do with truth except for something like ‘truth in a model’ (a very different thing). Chomsky remarks that their approach works as well for brains in a vat.

  Appendix VII: Hierarchy, structure, domination, c-command, etc.

  The sentences of natural languages all display hierarchical structure. Structure is often represented in a tree; a tree for a simple sentence such as Harry saw a house drawn in a way that reflects a fairly recent version of generative syntax might look like this:

  The labels on the diagram are explained below.

  I briefly address two questions: (i) where do linguistic hierarchical structures ‘come from,’ and (ii) what is their significance for understanding the way that language ‘works’?

  On the first question, it is important to recognize that it is difficult to avoid hierarchy. The successor function for generating the natural numbers introduces it, and in fact any finite state grammar introduces hierarchy: each additional element added in generating a string of words or other elements yields a larger set of which the prior-generated set is a subset. If one adds the operation of associativity, one can make the hierarchy disappear and end up with a string, but that requires an additional operation – and justification for introducing it. I do not mention a finite state grammar because it introduces the right hierarchy for natural languages. Chomsky demonstrated long ago that it did not (1957; see also Chomsky & Miller 1963). The issue is not how some kind of hierarchy is introduced, but rather how the ‘right’ kind is – the kind of hierarchy found in natural languages. The ‘right’ kinds of grammar yield this.

  Answers to this question have changed over the last sixty years or so, demonstrating progress in developing the study of language into a natural science. To illustrate, consider two stages on the way to the minimalist program approach to explaining structure, the earliest (1950s on) ‘phrase structure grammar’ one and the later (beginning about 1970) X-bar theory approach. Historically speaking, these are not distinct stages; it is better to think in terms of evolving views. However, for purposes of explication, I compress and idealize. During the first stage, linguistic structure in the derivation of a sentence was attributed primarily to phrase structure rewrite rules that included some like this:

  S → NP + VP (‘Sentences’ [more carefully, abstract descriptions of structures and LIs] consist of Noun Phrases followed by Verb Phrases.)

  VP → V + NP

  V → V + AUX

  V → {leave, choose, want, drink, fall. . .}

  . . . and so on, in very considerable detail, tailored to specific natural languages. These rules, it was thought, and a few “obligatory” transformations, yielded Deep Structures. The Deep Structures then could be subject to “optional” transformations that would (for instance) move elements around and perhaps add or subtract some in order to (for instance) change a declarative structure into a passive one. Basically, though, transformations depended on the basic structure established by the phrase structure grammar. So to the question, where does structure come from, the answer is: the phrase structure component of a grammar, modified where appropriate and needed by a transformational component. Notice that this kind of answer is not very different from answering the question of where the structures of molecules come from by saying “this is the way atoms combine” with no further explanation of why atoms do so in the ways they do. In effect, phrase structure grammars are descriptive, not explanatory in any illuminating sense. Segmenting linguistic derivations into a phrase structure component and a transformational one simplifies things a bit, and phrase structure grammars greatly reduce the number of rules that a child is conceived of as required to learn to acquire a language. However, that is a long way from offering a workable answer to how the child manages to acquire a language in the way the poverty of stimulus observations indicate. So phrase structure grammars do not do much more than begin to speak to the issue of acquisition: because structure differs from language to language (a prime example is the verbal auxiliary system), rewrite rules are far from universally applicable, making it difficult to understand how specific grammars could be acquired under poverty of stimulus conditions. They also leave wide open the question of how language came to have the hierarchical structure it does. They offer no way to trace language's structure to biology and/or other sciences, and do not speak at all to the question of why only humans seem to have the structure available to them – structure that is, furthermore, usable in cognition in general, including thought and speculation. Indeed, because so many variants are required for different languages, they make it even more difficult to imagine how they could have a biological basis.

  The second, X-bar approach made some progress. There are a couple of ways of looking at X-bar theory. One is this: instead of appearing as in phrase structure grammars to try to build structure top-down, move instead from lexical items up. Assume that lexical items
come already placed in a small set of possible categories, where each item in each category has a set of features that in effect say what the relevant lexical item can do in a derivation/computation. Thus, as a lexical item ‘projects’ up through the structure, it carries its features with it, and these determine how they can combine, where, and how they will be read. The categories are N (for noun), V (verb), A (adjective/adverb) and – with some dispute concerning its fundamental character – P (post/preposition). The projected structures are reasonably close to being common to all languages. The structures that are built consist of three ‘bar levels’ – so-called because the original way of representing the level of the structure took an N, say, and represented the first or “zero” level of structure as the lexical item itself (N), the next level as an ‘N’ with a single bar written on top, and the third an ‘N’ with two bars. A more convenient notation looks like this: N0, N1, N2 or this: N, N′, N″. The number of primes corresponds to the number of ‘levels” to be found in a hierarchical structure that looks like this, for a noun:

  Phrases of other types can stem from the N′ position; when they are found there, they are “adjuncts.” A hierarchical drawing for the green house with the adjective (A) green as an adjunct would place a branch of the tree stemming from the diagram's N′ down to the left (for English) to an AP, which would then go to an A′ and in turn to an A, and finally to green. ‘Det’ is short for “determiner”; determiners for noun phrases include the and a (in English). The position occupied by ‘Det’ in the diagram is the “specifier” position. X′s have specifiers; a specifier for a sentence (now an “inflectional phrase” to capture the idea that “inflections” such as tense need to be added to verbs to get sentences as we normally think of them) would normally be an NP.

  Another way to look at X-bar structure is to conceive of it as required to conform to a set of three schemata. The schemata are: 1. Specifier schema: XP → (YP) – X′

  2. Adjunct schema: X′ → (YP) – X′

  3. Complement schema: X′ → X – (YP)

  X = any of N, V, A, P

  P = phrase

  (. . .) = optional

  − = order either way

  Details concerning X-bar theory and structure are available in many places; my aim is to speak to what its introduction accomplished. X-bar theory, however one conceives of it, simplifies the structure that phrase structure grammar once was used to describe a great deal. The schemata or projection principles capture all the possible phrasal structures of all languages, thereby making a case as linguistic universals; and joined to the early stages of the principles and parameters framework as it was after 1980, X-bar could accommodate at least the headedness parameter. To the extent that it reduced multiple rule systems into a relatively simple set of schemata or projection principles, it aided – at least to a degree – the task of speaking to the acquisition issue and the matter of accommodating the theory of language to biology (the genome). It still left a lot unexplained: why this form, why three ‘bar levels,’ and where did this structure come from? However, it did speak to the matter of simplicity, and also to another important issue that Chomsky emphasizes in comments on an earlier draft of this appendix. In phrase structure grammar, even more was left unexplained. Phrase structure grammar simply stipulates abstract structures as NP, VP, S and stipulates too the rules in which they figure. Because of this, it offered no principled reason for the rule V → V NP rather than V → CP N. X-bar theory eliminated the wrong options and the stipulated formal technology in a more principled way than did phrase structure grammar. Again, that represents progress.

  Minimalism is a program for advanced research in linguistics. It is not a theory. It is a program that aims to answer questions such as the ones left unanswered by earlier attempts at dealing with linguistic structure. It became possible because linguists around the beginning of the 1990s could – because of a lot of evidence in its favor – rely on the principles and parameters framework, and particularly on parameters. They could afford, then, to set aside the explanatory issue that dominated research until that time, that of speaking to Plato's Problem. They could – as the discussion of the main text emphasizes in several ways – go “beyond explanation” (as the title of one of Chomsky's papers reads), meaning by that not by any means that they could consider the task of explanation complete, but that they could speak to other, deeper explanatory issues, such as those raised by evolution and the fact that language is unique to human beings, and by ‘Where does structure comes from?’, ‘Why this structure?’, and ‘What does the structure accomplish?’ For obvious reasons, Merge – its nature, its introduction to the human species, and what it can do – and third factor considerations came as a result to be placed at the forefront of research. Since the main text and other appendices speak to what this new focus provided (the structure introduced by Merge and the variants in structure offered, possibly, by third factor considerations alone), I do not repeat it here. I just emphasize that even coming to the point where minimalism could come to be taken seriously as a scientific program of research represents great (but definitely not finished) progress on the way to making linguistics a naturalistic science. Only once it is possible to initiate it can one begin to see how the theory of language and of linguistic structure could be accommodated to other sciences.

  The second issue concerning structure is what it is for – what linguistic structure does. In what follows, keep in mind again that any grammar introduces structure and hierarchy; the crucial issue is what is the right structure and grammar (and getting principled answers to this question, ultimately by appeal to biology). In the same vein, one should keep in mind that what is commonly called “weak generative capacity” (the capacity of a grammar to yield a set of unstructured strings) is a less primitive operation than strong generation (yielding strings with structures – in the case of language, stated in terms of structural descriptions or structural specifications). Weak generative capacity involves a second operation – perhaps associativity, as indicated above.

  As for the (connected) issue of what structures do: there have been several answers, but also a degree of agreement. Structures constitute phrases, and phrases are domains that play central roles in determining what can move where, and what the various elements of a phrase ‘do’ – what kind of theta role a noun phrase might have, for example. I illustrate in a very small way by looking at the notion of c-command – how it is defined over hierarchical structures and some of the principles in which it is involved. For ‘c,’ read ‘constituent,’ hence “constituent command.”

  To understand c-command and a bit of its roles, consider the following hierarchical linguistic structures:

  Now introduce the notion of domination, which is intuitively defined over a hierarchical structure of the sort illustrated. In the first figure above, B immediately dominates A and C, B dominates A in the second. Given this intuitive notion, we can now define c-command this way: X c-commands Y if X does not immediately dominate Y or vice versa and the first branching node that dominates X also dominates Y. Thus for the first structure above, A c-commands C, D, and E; B does not c-command anything; C c-commands A; D c-commands E; E c-commands D. And for the second, A c-commands D, E, and F.

  C-command figures in several central linguistic principles, principles that govern the ways in which sentences and the expressions within them can/must be interpreted/understood. One example is found in “binding theory,” a set of three principles that for all natural languages describes the syntactically determined binding/co-indexing properties of referential expressions, pronouns, and anaphors. An example of an anaphor in English is a reflexive pronoun, such as herself . . . Harriet washed herself. Clearly, herself must be coreferential with Harriet, unlike Harriet washed her, where her clearly cannot be used/understood to corefer with Harriet. An informal way to express this is to say that anaphors must be bound within the minimal (smallest) domain containing a subject: Harriet watched Mary wash herse
lf is fine, but herself must be used to corefer with Mary, not Harriet. A more formal and theoretically useful way – more useful because c-command plays a central role elsewhere too, and in capturing a linguistic universal allows the theoretician to consolidate – is this: an anaphor's antecedent must c-command it. This is “Condition A” of binding theory.

  The Minimalist Program illuminates why c-command plays an important role, and how it does so. The minimalist picture of sentence derivation/computation, at least on certain assumptions that are rather technical and that cannot be discussed here, offers an elegant picture of c-command and how it comes to have the properties it does. Minimalism treats a derivation as a matter of Merge – either internal or external. Consider an internal merge; it places an element that is already a part of a derivation “at an edge” – in effect, at the front of a set in which it leaves a copy [or a copy of A is put at the edge . . .]. For example,{A {B. . . {A. . .}}}. This portrays c-command in a nutshell: the ‘fronted’ element c-commands all the other elements, on the assumption – one that is well motivated – that the copy/copies of A will not be pronounced. C-command becomes precedence – being a ‘later’ Merge. For some fairly non-technical discussion, see Boeckx (2006).

  Minimalist efforts to eliminate ‘artifacts’ from theories of grammar suggest that one should not take c-command as a primitive of linguistic theory. It is a useful descriptive tool but – as the paragraph above suggests – it might be eliminable in favor of features of internal and external Merge. As Chomsky (2008) argues, Condition C and Condition A of binding theory can be dealt with in this way. One point of the Minimalist Program is, of course, to eliminate items that cannot be justified by “principled explanation” and to aim toward conceiving of the language system as perfect. The discussion above illustrates some progress in that regard.