UofA

Spring 2004

Ling/Phil 596D: Topics in Linguistics and Philosophy

Heidi Harley and Massimo Piattelli-Palmarini

Compositionality

Wednesday January 28

Handout 1 (M. Piattelli-Palmarini)

Fodor and compositionality

1 – From Chapter 5 of “Hume Variations”

A “reductionist” metaphysics of meaning is one that can provide sufficient conditions for meaning in a vocabulary that is not itself either semantic or intentional. The Empiricists’ resemblance and association is an instance, and so are causation, information (à la Fred Dretske, and Fodor himself) and evolution (à la Millikan and Dennett). Consensus is “most tenuous” in cognitive science on which metaphysics can really do the job of naturalizing meaning and representation.

Associationism fails because it’s unable “to distinguish the intentional relations among the contents of thoughts, from the causal relations among the thoughts themselves” (p. 115, emph. orig.) Take (as Hume does) probabilistic reasoning. Imagine a fair die with 3 sides, two of which have a triangle on them, and the third side has a circle. You roll it over and over. Suppose that, as a result, certain expectations do form in the mind. Here comes the big glitch, for the associationist: One thing is to explain why the thought “it will come up triangle” will occur twice as frequently as the thought “it will come up circle”. Quite another to explain how you can come to have the thought “it will come up triangle and it will come up circle in the ratio two to one”. The “two to one” bit has to be in the scope of the “think that” bit. But associationism has no way to get this. This is a most elegant example of the need to distinguish between causal relations among mental states from relations (logical or whatever) among their intentional objects. Hume tries to overcome this big glitch by means of the imagination (“fancy”). Fodor’s Chapter 5 is devoted to explaining why this does not work. Hume’s imagination is a blank check, a mystery. And Fodor strikes further heavy blows onto associationism, including (notably including) probabilistic associationism. You need Turing, rule following, innatism, and tacit knowledge, and syntacticity, to get out of the morass.

The thought “horse” followed (even invariably followed) by the thought “hoof” does not amount to the thought “hoofed horse” (horse with a hoof). (Remember ANTI-ANTI-MISSILE from last week). Succession is not the right kind of glue to hold complex concepts together. Contingencies in thought do not copy contingencies in experience. “Associationists have spent literally centuries in the fruitless search for a way out of this.” (p. 117)

The long and the short of it all is that “the least that the mind must be able to represent is the content of its experience together with whatever higher-order and relational properties of its experience determine the character of the associations it forms.” (p.129. Emph. orig.) (For a cogent defense of this thesis even in the domain of animal behavior and un-classical conditioning, see the work of Charles R. Gallistel over many years)

2 – Chapter 6, toward compositionality

To repeat: “Cognitive processes are constituted by causal interactions among mental representations, that is, among semantically evaluable mental particulars.” (p. 135). It’s this, or we are in total darkness. Three important, but distinct and independent, properties of propositional attitudes (PA): systematicity, productivity, compositionality. These are properties of the contents of the PAs.

Fodor runs a new argument against a dispositional account of mental causation. In a nutshell, a disposition must manifest itself here and now, in order for it to be the explanation of a behavior here and now. An event must cause its manifestation. So dispositions cannot be sufficient causes. “Dispositions manifest themselves only when something that’s not a disposition causes them to do so. It’s not sufficient for the vase to break that it’s fragile; something has to happen that causes its fragility to cause it to break.” (p. 140. Emph. orig.) If we want (as we do) a “robust” mental causation, we cannot have a dispositional account of mental states and mental processes. A mental link is sufficient to bring about its successor in the causal chain. “A thought is a bona fide mental particular and having a thought is a bona fide mental event.’ (ibid). Some mental events are causally sufficient for others. There is mind/mind causation.

[A bookmark, here, for the following weeks. Fodor’s long fight against the primacy of dispositions presently brings him in disagreement with Chomsky. Chomsky is adamant in insisting that I-language is an internal structure of the speaker, not any sort of “relation” between the speaker and something else external (or internal). Especially not between the speaker and his/her own (alleged) LOT.

Heidi’s comment: In fact, it seems to me that the ultimate LF representation, assuming that we ever determine what it is, just IS the LOT; distinguishing the LOT and the I-language in this regard, I suspect, will be impossible. (Unless LOT = (I-language + deictic-reference-resolution system) -- but most laypeople would doubtless want to include the deictic-reference resolution system as part of I-language, although of course Chomsky wouldn’t.

Even Chosmky’s own expression “knowledge of language” needs a caveat: It is not the case that there is the speaker’s mind and, distinct from it, a “language” which the speaker “knows”. All we have, according to Chomsky, is a complex internal (computational-derivational) structure of the speaker. This has consequences for compositionality as well. Chomsky finds it perplexing that Fodor may admit the compositionality of LOT, but not of natural languages. Though, of course, Chomsky’s account is not at all “dispositional” (in anything remotely resembling Wittgenstein and Ryle, or Travis), the notion of “internal structure” without a representationally separate LOT is different from Fodor’s unrepentant representational theory. Fodor, however, gladly admits that many representations are unconscious and that there is a huge heterogeneity in the kinds of (innate) mental representations we have. Notice too that in Minimalism, there are no representations any more. All is derivational (we will go back to this next week, or the week after next).]

The decomposition of complex concepts into “syntactic constituents” is what differentiates them one from the other (notably those that have co-extensive applications). Their canonical decomposition is intrinsic, not contingent

Heidi’s comment: a point rather like what I was making on the first day: linguistic (I-language)) expressions just are structured; the very idea that they might not be is an incoherent one which I suspect we can blame on the extreme incompleteness of their written representation, and to a certain extent their spoken representation. There is no expression of I-language without constituent structure.

Concepts are “semantically evaluable, causally active, mental particulars” (p.144). They are Modes of Presentation (MOPs) “only, psychologized”. The same extension can be presented to the mind in lots of different ways (Chomsky’s term is “points of view”).

Rule-following versus “being in accord with” a rule. How can we trace the difference? Fodor (rightly, in my view) dismisses awareness as a criterion (contra, for instance, Searle’s position, whom he does not cite). The key is causal power. MOPs distinguish the causal power of otherwise extensionally identical concepts. How things (or rules) are represented in the mind is what decides.

“Hard wired” rules may be extensionally equivalent to bona fide rules, but they are intentionally different. (Intentional, with a ‘t’, subsumes intensional, with an ‘s’).

Heidi’s comment: Is this reiterating that language is an automatically triggered computational device, not a set of rules (understood as ‘instructions’ or similar) followed by a homunculus who represents them independently?

Again, a cogent defense of atomism (see Fodor’s previous work). Data on language acquisition make it very unlikely that the child has to acquire PHYSICAL OBJECT or PARENT before she acquires MAMA. These data, together with further arguments against making inferential power constitutive of meaning (and against the holism that inevitably ensues) strongly recommend to espouse atomism.

3 - On compositionality proper

(The very last 5-6 pages of the book). Systematicity and productivity are explained by compositionality. Nay, only compositionality can explain them. Either thought (LOT, Mentalese) or language have to be compositional. Or both. If you can show that one of them isn’t, then you have shown ipso facto that the other is. Well, then, if you show that English is not compositional, then you have shown that Mentalese is, and you have shown that TOI (or RTM) is true.

Argument 1: When you learn English, you learn how to translate from Mentalese to English and vice versa. You think in Mentalese but you communicate in English. English may well borrow systematicity, productivity and compositionality from Mentalese. Mentalese is compositional. Is English too? This is an empirical issue (see infra), and there are no knockdown arguments. But surely English does not “look to be frightfully compositional” (p.153). Everyone agrees that natural languages are not compositional at surface form (see my first week’s handout). So, really the whole burden of proof falls onto Logical Form (LF). If you undermine the legitimacy of LF as being the semantic level, then you undermine the compositionality of natural languages.

Argument 2: Ambiguous sentences are supposed to map onto two or more un-ambiguous LFs. These, in turn, are supposed to map univocally onto un-ambiguous thoughts. Do LFs represent sentences, or do they rather represent thoughts? (The thoughts that the sentences are used to express). Mentalese “points in two different directions: on the one hand, towards thought; on the other hand, towards language” (p. 155). That’s how (sometimes at least) you manage to say what you think. What grants that sentences are un-ambiguous at some level of representation (LF)? “..There is no obvious reason why sentences should be ambiguity-free at any level of representation; in other words, there is no obvious reason why there shouldn’t be really ambiguous sentences. […] Perhaps some sentences of L [natural languages] are ambiguous at every level of representation that the grammar of L recognizes, or perhaps none are. If we want to choose, we need an argument” (ibid, emph orig). Not much (no really cogent argument) justifies the claim that there is a level of representation of sentences (LF) that is ambiguity-free.

Heidi’s comment: Well, this is true but in an uninteresting way, seems to me. It’s true if you separate out the deixis-resolution device from L, which is probably not too contentious. An expression like he or tomorrow will be ambiguous all the way through to LF. What you can guarantee, however, is that once you anchor the reference of such expressions to the context, the meaning computed by an LF representation will be unambiguous...

Argument 3: Thoughts cannot be ambiguous (see the case of Lincoln You can fool some of the people…). You do not open your mouth and close your eyes, and think ambiguous thoughts. It’s open whether there are ambiguous sentences, qua sentences, but not whether there are ambiguous thoughts. ‘Thoughts are where the buck stops.” What could disambiguate them? The impossibility of entertaining ambiguous thoughts (for thoughts that have propositional content) does not depend on presupposing that LF represents them. Sentences are (sometimes at least) ambiguous. They become ambiguity-free only under the LF hypothesis. Now, thoughts have causal roles, but not sentences (nor propositions, being abstracta). Therefore (out of a simple Occamian principle of minimizing the number of mysteries) let’s conclude that LF represents thoughts, rather than sentences. English does not look terribly compositional because it isn’t! (p.157) But, if it isn’t, then Mentalese has to be.

These arguments are presented rather tentatively. The crux is that whatever LF represents is ipso facto compositional. Linguists, Fodor says (correctly), agree on that.

Coda. Notes from a conversation with Jerry (November 16, 2003)

JF: Take Phonological Form (PF.) That is not, again, at first sight (i.e. phonetically) compositional. You have all sorts of contiguity phenomena, vowel-shifts, absorptions, rounding etc. non-articulated phonemes etc. The canonical compositional analysis (features etc.) is compositional, but that’s a theoretical construct, a complex algorithm, not a property of surface pronunciation. [See Halle’s Æ morphemes]. A lot of gap-filling must go on at EVERY level. Why not at the semantic level too? Why should LF need no gap-filling and have EVERYTHING made explicit?

Basically, the semantics of natural languages is a chapter of a theory of communication. There is quite a lot that we must be able to compute anyway reconstructing each other’s communicative intentions, reconstructing each other’s THOUGHTS. We know how to do it, basically, in virtue of the fact that we are all members of the same species. We know how to fill the gaps (in the main, in a huge variety of situations)

Take It’s raining. (And all deictics in general). We obviously understand that it means It’s raining here & now. Why should we suppose that the specifications here and now are somehow tacitly present in the language, at some level of representation (allegedly, at LF)? Why not suppose that we fill those gaps in virtue of what we know about the situation, the speaker’s intentions and so on, and the standard way of expressing the thought IT’S RAINING HERE AND NOW in English? Something like Grice’s conversational maxims.

MPP: In Italian, even the it is omitted. But we have every reason to believe that it’s there. Not in virtue of Grice’s maxims, or any knowledge of the world, but rather in virtue of our knowledge of language. (JF agrees). Why admit it for it, but not for here and now?

JF These are empirical questions, and I have no strong objection, but it NEEDS ARGUMENTS AND DATA, it cannot be just taken for granted. It cannot just be an assumption that L is compositional, and that there is a strictly compositional level of representation (LF) of L, such that everything is made explicit at that level. This is a very strong hypothesis, that forces one to hypothesize hugely complicated underlying structures, with a huge amount of silent components (functional categories of all sorts, projections of all sorts) And he has an inherent problem with that? just because the theory imposes that Ls are strictly compositional. A theory that imposes so many diversified and complex posits ought to reconsider the basic assumption that make these inevitable, i.e. the assumption of compositionality.

Heidi’s comment: Basically, it seems to me that the primary reason that the generative program results in productive research is because it has assumed compositionality for L. Linguists see an ambiguity or an underspecification (Mary saw the man with the telescope; John wants to go), assume that there must be something there that disambiguates or specifies it (constituent structure in the first case, big PRO + the theory of control in the second), and then go out and discover it, with corroborating evidence, predictive power, and so on. Why on earth should we give up this extremely productive line of inquiry on the off chance that L isn’t compositional? Jeepers! Seems soundly unscientific.

MPP Take ellipsis, something like John believes that Mary is smart but Bill doesn’t.

A syntactic treatment of ellipsis is surely a possibility (several interesting ideas are on the market). This gap-filling does not depend on conversational maxims, on guessing communicative intentions, or the like. It’s allowed (sometimes imposed) by our knowledge of language, of syntax. Punkt.

JF Yes, but these are cases of deletion under identity. It’s quite understandable that you may succeed in developing an algorithmic account of deletion under identity. But take all the other ubiquitous cases of ellipsis (in a more general sense), where there is no identity. They require, quite obviously, a gap-filling based on other stuff (presuppositions, understanding of communicative intentions, etc.) and these are non-algorithmic. It may well be the case that bona fide semantics, the individuation of thoughts from surface sentences of L, turns out to be algorithmic, but I doubt it. A lot more is required that is very probably non-algorithmic.

MPP In “Hume Variations” (HV), as well as in many other papers and books, you pay homage to Turing, and say that a healthy injection of Turing-computationalism, bringing to an end associationism and Wittgensteinianism (the “use” theory of meaning), is arguably the best thing that ever happened to cognitive science. But, if so much is non-algorithmic, then the role of Turing-computationalism becomes rather minor.

JF But thought IS Turing-computable. We manage to reconstruct the underlying thoughts and then thoughts are causally related one to the other in (roughly speaking) syntax-sensitive computational ways, à la Turing. The fundamental kind of causality is not associations, but Turing-computability. THAT was the crucial move.

On propositions

MPP In a footnote in HV you say that the (hypothetical) end-product of the algorithm, the entry into LOT (what really corresponds to the thought that the sentence is uttered to convey) cannot be a proposition, because propositions are abstract entities and, as such, cannot have any causal power. But thoughts do have causal powers, so it cannot be propositions. Many years ago, in your paper on propositional attitudes, you also said that if propositions are somehow to be brought into the picture, they must be something that possesses a syntactic structure. (JF confirms). Well, then, how can it be that the “end product” must possess some syntactic structure, and LF does possess a syntactic structure, but LF is not the ‘entry” into the corresponding thought?

JF The very existence of LF needs arguments and data. There is hardly any doubt that there are thoughts, and that they are compositional (the systematicity, productivity and fine-graininess of thought guarantee that). Why not have just thoughts? Why add LF?

MPP Is this a kind of Ockam’s razor consideration?

JF Well, something like that, but not just that. Why not just have an algorithm (syntax) that is compositional and operates on surface English, computing sentence types out of sentence tokens, and a bunch of communicative operations (Gricean maxims, knowledge of the situation, intuitions about the speaker’s intentions etc.) that fill the gaps, and indeed reconstruct the thought that it is the speaker’s intention to convey? We really do not need anything else, unless there are convincing arguments and empirical evidence that we must have something else (an LF level of representation that is compositional). I (JF) see none being offered that is convincing.

Heidi’s comment: Does he (JF) argree that the wh-movement is interpreted via an operator-variable construction? Further, does he agree that there are wh-in-situ languages? Do such languages have a different algorithm for mapping questions to thoughts than wh-movement languages do (i.e. not an operator-variable structure?) Seems to me that without LF the learnability problem rears its ugly head again. How can a child reconstruct the algorithms that map the structures of his/er language to LOT if those algorithms are (wildly) different from language to language? If they are not (wildly) different from language to language, but rather parametrically different in constrained ways, then I submit that whatever the algorithms accomplish is equivalent to what LF is intended to (and is) accomplishing. That is, LF is a concrete proposal for how the mapping to LOT could happen. Until there’s a convincing alternative that does as well as LF does, and doesn’t open the Pandora’s box of learnability considerations, I plan to stick with LF as the only concrete proposal that comes close to doing the job (perhaps supplemented by a set of context-mapping principles that fix deictic reference).

On ambiguous sentences

MPP: Let’s take the famous sentence (Chomsky, then Higginbotham) I almost had my wallet stolen. It’s instantly recognized as ambiguous between two meanings, but there is a third one that emerges, slowly, under guidance or careful reflection. Now, it’s a property of L (of I_L) that this surface form can receive three distinct LFs, not a property of our communicational skills. It really seems that something like LF (the computational apparatus that derives these forms) is needed, in order to access the associated thoughts.

JF The apparatus computes types. This token sentence can be mapped onto three different types, this is algorithmic, but then something else is needed (some filling of the gaps, as usual) to decide WHICH thought the speaker who uttered that token actually intended to convey. That’s where the buck stops.

Heidi’s comment: Sure -- but that doesn’t amount to saying that the surface form is underdetermined. The hearer can compute 3 LFs given the evidence of the written sentence; deciding which one is correct is the same as the problem of deciding what with the telescope is intended to modify in She saw the man with the telescope. Surely, surely that problem doesn’t call into question the compositionality of the 3 LFs that could match the surface string. The string can’t mean just anything!! And the speaker didn’t have an ambiguity; the speaker had one of the three LFs in mind on utterance. The problem for the hearer is the same as deciding which lexical item to identify when the string /bQNk/ is uttered. Does the existence of that problem mean that language isn’t compositional? What about the homophony of kills (3sgV), kills (plN), kill’s (genitiveN)? Are we not justified in positing homophony for -s here, or does it all come down to Gricean maxims, hearer-driven heuristics that the speaker knows the hearer will use? This all seems very odd to me...

Modes of presentation (MOPS)

JF Think about the acquisition of the lexicon. It requires a theory of mind, understanding other people’s intentions, understanding the relevant aspects of the situation etc. Quite a lot of gap-filling. It is causally dependent on what I have called MOPS (“modes of presentation”). Why should sentences be any different? There are sentential MOPS too. Ways of representing situations and communicative intentions and so forth. Maybe what we really need are MOPS, rather than LFs.

MPP Well, the classic experiments by Lila Gleitman show that syntactic frames are also essential for acquiring lexical meanings (X is gorping Y, versus X and Y are gorping).

JF The data are not so clear as they appeared initially. Lila is putting strong brakes on those conclusions, these days. She is seriously reconsidering that whole story. But, look, it may well be. I am not excluding that it may be the case, though, it seems to me, better arguments and better data are needed to affirm that syntactic frames are so frequently so essential for lexical acquisition.

Further considerations (not discussed with JF yet)

“Compositionality is, par excellence, a property of representations” (HV, p. 157, emphasis in the original). Sure, but it’s also a property of derivations (i.e. of operators), and in minimalism, these days, there are no representations (arguably), only derivations. And these are compositional all the way, up to and including LF, but not just at LF. Merge is, by its nature, compositional. And so are “phases” (modulo a better specification of what they are), and so are the derivations by phases. It really sounds a bit extravagant (in need of arguments and evidence, exactly as JF says, but in the opposite direction) to assert that:

(1) Thoughts (sentences in the LOT) are compositional

(2) Syntax is compositional Syntactic derivations are (to put it mildly) relevant to the individuation of the corresponding thought

BUT

(3) The interface between NS and LOT (something like LF) is not itself compositional.

This amounts, in my (MPP) opinion, to the insertion of a huge puzzle where none needs to be.

Heidi: Yes -- or regressing to ‘I don’t like the LF solution to this problem, so let’s throw it out and wait for something better to come along’....

In HV, as we saw last week. Jerry says that composite concepts are not “just” decomposable into basic components, but decomposable into canonical components. The composite concept A BOY AND A GIRL has A BOY and has A GIRL as canonical components, but not “AND A” as a component (surely not as a canonical component). This is quite Ok, but it sounds to me as syntactically based. A BOY is a DP, and so is A GIRL, but “AND A” is not a syntactic constituent. There seems to be a clear correspondence between syntactic constituents and conceptual (i.e. semantic) constituents. If L is not compositional, how can that be? This is, I (MPP) think, an old puzzle in Fodor’s theory. In his book on concepts, Jerry says that BROWN COW is a composite concept (unlike BROWN and COW) because its is composed of two mono-morphemic units, i.e. words (sic). How any such consideration (with which I fully agree) can be offered while maintaining that natural languages are not compositional, I fail to see.

To be continued.