UofA

Spring 2004

Ling/Phil 596D: Topics in Linguistics and Philosophy

Heidi Harley and Massimo Piattelli-Palmarini

Compositionality

Wednesday January 21st14

Handout 1 (M. Piattelli-Palmarini)

I - Novelties in “Hume Variations”

Three Fodorian novelties, in this crisp essay:

- The stress on canonical decompositions (as opposed to non-interpretable “parts”);

- Admitting that simple concepts may well have “thematic roles”, but remain atomic nonetheless;

- Denying compositionality for natural languages, while maintaining it for thoughts (sentences in

LOT) and for syntax.

II - Hume being both right and wrong

Right about:

A cognitive psychology entirely based on the Theory of Ideas (TOI). With minor variations this is very similar to the Representational Theory of Mind (RTM).

The compositionality of Complex Ideas (Complex Concepts in modern terminology).

The Cartesian thesis that possession of concept C amounts to the capacity to think about the property C; Representing in thought the property of being C, nay, the property of being C as such, is primary. Uses and licensed inferences are derived.

Possessing concepts = there being certain items (particulars) in the mind; These items “making their entrances into minds that originally lack them” (Barry Stroud, 1977, quoted in HV).

Running arguments both about the genetics (entrances) and about the etiology (causality) of these entities.

The difference between simple concepts and composite concepts. The latter being composed of the former in certain ways.

The natural “interconversion” between what concepts are and what having them amounts to. Not to be confused with the problem of knowing what to do with them (pace Wittgenstein, Ryle and pragmatists like Dummett and Co.);

Postulating entities to explain empirical data, and then “letting later investigation make their precise nature clear”. This is the characteristic strategy of rational theory construction.

Concepts being the natural constituents of judgments, without confusing this thesis with the thesis that “accepting certain of the judgments in which a concept occurs is constitutive of grasping it” (HV, p.12)

“…in much the same way that sentences are constructed out of words….. mental representations that express propositions … are constructed out of mental representations that express concepts” (p. 13. emphasis in the original). “Maybe only a mind that can judge could conceptualize. It does not follow that concepts are constituted by the judgments they are deployed in.” (ibid.)

The prevailing doctrine that complexes are primary and components derived (sentences versus words, languages versus sentences) “is very dark indeed.”

Our uses of concepts make them “worth the bother of having or of studying” but it is not constitutive of having them.

Inferring must be licensed by some principle. And this must somehow be internally represented.

Sorting must be carried out depending on representing items as being, or not-being, this or that. So, neither uses, nor inferences, nor sorting capacities, can be primary with respect to concept possession.

Hume was wrong about:

The iconic nature of ideas. Ideas resembling sensory impressions.

The thesis that ideas “copy” the content of impressions.

Empiricism (the thesis that all mental contents derive their origins from sensory impressions).

Mixing psychology with epistemology (how things are versus our justification(s) for believing that they are so).

The crucial role of associations.

The nature of abstract ideas (a puzzle that Hume acknowledges). The source of non-perceptual ideas (UNICORN). The power of the “imagination” in producing them remains a mystery.

The difference between sensations and ideas being “the degree of force and liveliness with which they strike upon the mind”.

The mystery of the “knowability” of the structure of the world, given that the world is (allegedly, due to the empiricist conception) itself built out of experience alone (i.e. from sensations) (Kant will give a better solution)

III - Fodor on canonical decomposition

Unlike sentences of natural languages (see infra), complex concepts decompose univocally. They decompose into canonical constituents. BROWN and COW are canonical constituents of BROWN COW, but AND A is not a canonical constituents of A BOY AND A GIRL. Nor is BOY GIRL. (Remember last week’s point about non-constituents?). Complex concepts cannot (pace Hume) be copies of complex impressions, because complex impressions do not decompose. Surely do not decompose in any canonical way. “They are complex in different ways” (p. 35)

Criteria for canonicity:

Substitutions allowed under some recursive scheme (again, remember last week)

Parts vs constituents: “Each constituent is ipso facto among the parts of a composite concept, but it is not the case that each part is ipso facto among its constituents”. (Footnote 6, p. 34)

Interpretability: Only canonical constituents are guaranteed to be semantically interpretable.

(Something like) participating in a logical form (Thematic roles, or argument structure, page 61. This remains a bit mysterious – to me at least)

I (MPP) think that all this comes very close to syntactic criteria. Moreover, as a way of exemplifying it, mono-morphemic words are cited. But beware! A mono-morphemic word in one language may well have a precise non-mono-morphemic translation into another language. This is stressed contra the objection to atomism based on translation (in, say, Turkish transitive MOVE is translated as cause-move). That’s irrelevant (see infra)

Responsibility to counterfactuals. An old Fodorian horse. We go beyond sensations and impressions in building up the structure of the world (objects that would occlude each other, if I moved around, borders that would remain invariant, if I changed the luminance etc.)

Contra iconicity: Parts of a photograph of X are photographs of parts of X, no matter how you slice it. NOT so for complex concepts (MR JAMES’S TAIL). (And, I add, not so for sentences)

It all depends of how you slice them. Having canonical components is the crucial dividing line.

“Pictures don’t have canonical decompositions even if they are pictures of things that do” (p. 38), But thoughts have their constituent structure essentially.

Spectrograms of speech versus phonological decompositions.

IV - Fodor against the decomposition of simple concepts (transitive MOVE)

Reiterations and refinements of why one should not try to derive KILL from “CAUSE TO BECOME NOT ALIVE”.

Plus data from language acquisition. Plus the huge problem of unifying forms of causation.

New argument: MOVE has argument structure, but so has CAUSE. Moreover, they have the same argument structure. In X MOVES Y and X CAUSES Y (TO) MOVE, X and Y are in an argument relation to one another such that the concept transitive MOVE expresses. “Whether a concept has argument structure is independent of the question whether it has constituent structure”(ftn. 3, p. 63)

New argument: Contra Susan Carey’s decomposition of CAR and HOUSE (versus HURRICANE and WAR) into ontological primitives such as HUMAN ARTIFACT (and, respectively, EVENT). These are not constituents of CAR, HOUSE, HURRICANE etc. simply because there is similarity in their extensions. Similarity of things is no evidence for similarity of the concepts of those things. If the two things A and B have, in reality, a common component qua things, this in no way entails that the concept of A and the concept of B have a common component qua concepts. No structural conceptual similarity. The similarity of their extensions will make certain inferences come out right, but that’s not because the concepts are similar (or possess a common component).

Antecedently intuitively plausible complexity is no proof of reasonable independent evidence for internal structure.

Simple concepts are boring, though the fact that they are is not. Fodor develops at length arguments against the idea that simple concepts share, qua simple concepts, interesting common properties. Being simple is not as such an interesting semantic property. You have to claim it is only if you are an empiricist (because they are - allegedly - all derived only from sensations). Then you touch upon the issue of the existence and the allowable heterogeneity of innate ideas. Rationalist have no problem with either (there are many, many, many innate ideas, and there are no limits to their degree of heterogeneity, and they do not all have to connect with sensory data).

New argument: There is a sufficiency claim and a necessity claim. If (allegedly) MOVE_T decomposes as CAUSE MOVE_I , then the concept CAUSE MOVE_I suffices to give MOVE_T (This is

sufficiency). If (allegedly) MOVE_T decomposes as CAUSE MOVE_I , then possession of the concept MOVE_T entails possession of the concept CAUSE and of the concept MOVE_I (necessity).

But, the sufficiency claim leads us into the notorious predicament of +X. And no one has come up with any remotely plausible candidate for X. (Desperate attempts: CAUSE is really something like IMMEDIATE CAUSE. CAUSE does not really mean cause. CAUSE is not effable. All bad suggestions). This is old stuff, no need to hammer the nail in again.

The necessity claim deserves some attention. “The crucial point is that concept possession is closed under constituency, but not under entailment” (p.69, emph. orig.). It goes both ways. Entailment does not entail constituency. Entailment is not closed under concept possession. Entailment is the “necessity of the hypothetical”. Can we single out kinds of necessitation? That seems doubtful.

It’s a truism that you have to have all the parts to have the whole. But it’s far (very far) from obvious that, if concept A entails concept B, then if you have A, you must also have B. And that B is a constituent of A. In fact, not only it’s not obvious, it’s squarely false. The case of 2 and IS A PRIME. (Again, Fodor stresses the distinction between mere parts, and bona fide constituents).

“Although X moves Y is a causal relation, it does not follow that the concept X MOVES Y presents that relation to the mind by invoking the concept CAUSE” …. “To think of a causal relation is one thing, to think of a relation as causal is quite another”. (p. 78, emph. orig).

Reliable inferences are orthogonal to concept constitutivity.

The productivity of thought and of complex concepts is again (see last week) an argument for compositionality. Thought is prior to the content of language (p.87). Language expresses thought.

“The productivity of content (‘semantic productivity’) demands something more than the productivity of structure” (ibid). You need compositional content and an arrangement. MISSILE, ANTI-MISSILE, ANTI-ANTI-MISSILE. Two simple concepts, arranged in different ways. Once you add arrangement, then compositionality does explain semantic productivity.

Determining versus exhausting the content of complex concepts (see last week). Empiricists have to claim that there is such exhaustion and that experience suffices to grant it. Moreover, that association is the engine.

V - Fodor contra associationism

“The content that’s before the mind when you first think concept A and then, by association, think B is the very same content that is before the mind when you first think A and then just happen to think B.”(p. 91, e.o.). The mere sequence of representation should determine and exhaust (if associationism were right, but it isn’t) the compositionality of representations. Association is semantically transparent, it cannot “outrun the content of [the] constituents”. (ibid). But then semantic transparency fails to explain semantic productivity. A huge problem. (It’s Hume’s dilemma. He invokes the imagination, but that is totally unsatisfactory – see Chapter 5). Plainly, you do not get the composite concept ANTI-(ANTI-MISSILE) by thinking ANTI twice, and then thinking MISSILE. (Kant and Frege made a much better move, abandoning empiricism and associationism). “The right way to proceed is to assume that mental representations have syntactic constituent structure” (ftn 11, p. 93). This buys us the right kind of constituency. (Reminder: Fodor has insisted in several previous papers that the syntactic nature of representations as possessing causal powers, due to Turing, is just about the best thing that ever happened to cognitive science).

In a nutshell: (Also) associations are causal relations between representations (and this much is OK), but they are only sensitive to proximity and contiguity, unlike the causal relations that are sensitive to syntax (and therefore to constitutivity). Ultimately, associations are not (nor can possibly be) sensitive to truth (as Kant was right to point out). But thought is. So associationism has to go away (pace Connectionists & Co). Constitutivity and compositionality (rationalism) offer (after Frege, Turing and Tarski) precisely the sensitivity to truth that we need.

VI - Fodor contra Wittgenstein (contra context-dependency)

The standard neo-Wittgensteinian view is that “all content is content-in-context, the content of simple concepts (and of mono-morphemic words) is too…. But either content isn’t inherently contextualized, or there is something wrong with the compositionality thesis” (p. 96. My emphasis)

We do not want to invert the direction of compositionality, having simple representations inherit their content from complex ones. We do not want a metaphysical priority, and we do not want a semantic priority, of the complex over the simple.

What we want to grant, nonetheless, is “externalism” (the content of representations is somehow supervenient on their mode of being in the world. BROWN COW has the content it has because of its being related to how things are in the world, i.e. to brown cows). “No cow is a text. You can milk a cow, but you cannot milk a text” (p. 97)

Externalism is no threat to compositionality. Semantic contextualism is. Fodor here reiterates (see his book with Lepore on holism, and Fodor and Lepore 2002) that we must not conflate epistemology (how we, or the child, actually end up knowing about cows etc.) with the metaphysics of concepts and of meanings. Concept acquisition may well reverse compositionality. Usually the child does not hear words in isolation, in the absence of suitable situations, sentences etc. Acquisition may well work mostly top-down. This does not impugn the thesis that semantics proceeds bottom-up. Enough of the content of the constituents must be preserved in the composite to allow the child to grasp it from the composite.

Lexical equivocation (BANK, TRUNK, PAGE etc.) may well require a context for disambiguation. No big problem here. But, in the end, the disambiguated content is what it is. Thoughts cannot be ambiguous (see infra). Fodor argues against radical contextualism (the thesis that there is an essential in-eliminability of ambiguity, save by taking in the whole text). The epistemology of communication (interpreting the speaker) must not be confused with metaphysics of meaning.

Reiterating here a consideration that Fodor made in his old paper (1978) on propositional attitudes, the distinction must be made between inferences licensed by the contents of one’s beliefs, and the consequences of having those beliefs.

The Groucho Marx joke I shot an elephant in my pajamas. (I shot (an elephant in my pajamas) versus (I in my pajamas) (shot (an elephant)). Different constituents, different thoughts. There is no ambiguity in thoughts, only in what is said. Only univocality disambiguates, and thoughts are univocal. “The disambiguation of a statement supervenes on what thought the speaker intended to express in making it” (p.108). Constitutivity uni-vocates (is this LF? See infra). Fodor’s argument (contra C. Travis 2000) is to the effect that even discourse effects have to be accountable (are at least constrained by) context-independent semantic content.

A reflection that touches on what will occupy us next week:

The content of a thought supervenes on its relational properties (some nomic connection between representations and the world). The disambiguation of language by thought requires some (metaphysical) story about what bestows content on thoughts. Thoughts are not objects for interpretation in anything like the way that utterances are. “Even if the content of statements is not strictu dictu compositional, the content of thoughts may well be” (p. 108). The choice between the content of thoughts being intrinsic or context-relative (interpretation-relative) is not disjunctive. You cannot rule out one, and therefore conclude that it’s the other. As we will see next week, Fodor now things that both are required. That’s why he claims that natural languages are not compositional, though thoughts are, and syntax is. Stay tuned.