Jackendoff talk: semantics must be generative

On Friday (4th) I heard Ray Jackendoff give the keynote lecture at a conference organised by the UCL Centre for Human Communication which my department (UCL Phonetics and Linguistics) is part of (in some way I don't understand).

What he said may not be news to anyone else, but I hadn't heard it, not having read any of his recent stuff, except the bits about music.

Broadly, he thinks that mainstream - ie Chomskyan - linguistics is on the wrong track by supposing that syntax is the only generative component needed in the grammar, so that phonology and semantics need only interpret the output from syntax.

On the phonology side, he thinks that auto-segmental phonology is on the right track, since it is a separate generative system with a 'dirty connection' to syntax, in the sense that there is no one-to-one mapping between phonological and syntactic structure.

Jackendoff wants to treat semantics in the same way, as an independent generative component, linked with syntax, and - separately - with phonology, by constraints, not isomorphic mappings.

This has a number of interesting consequences, one of which is that words become constraints across components of the language faculty, which seems like a neat idea. Another is that in the evolution of language, syntax might have come last to mediate a previously direct link between rich conceptual structure and phonology, rather than first as Hauser, Chomsky and Fitch (2002) seem to propose.

I'm with Jackendoff on the last point - it doesn't make sense to me that recursion would develop first in syntax to link a conceptual system with a sound system, because a conceptual system without recursion wouldn't be worth linking with anything. If you couldn't think, you wouldn't have anything to say.

But I don't think agreeing with this commits anyone to Jackendoff's position. It seems to me that he assumes that anything (or, presumably, anything propositional) that you can think must be representable in linguistic semantics. But that doesn't seem right. All that you can really say is that anything (propositional) you can think must be representable in the Language of Thought (assuming there is such a thing). The semantics of sentences, on the other hand, are notoriously underspecified relative to the Language of Thought, and if Sperber and Wilson's recent work is right, the same goes for the semantics of lexical items. (Since the contribution a lexical item ends up making to the propostion expressed by an utterance will generally be an ad hoc concept reached by narrowing and/or broadening the concept encoded by the lexical item.)

So it's perfectly consistent to think that we need to have a richly structured conceptual system for propositional thought, and that some parts of this must predate a structurally rich linguistic syntax (or there would have been nothing for the syntax to express) - without thinking that linguistic semantics needs to be generative in its own right. It could just be read off LF, which seems the simplest assumption.

I'll give one example to try to make this clearer. Jackendoff mentioned Pustejofsky's work on lexical semantics. The basic idea is something like this: You interpret 'begin' differently in 'I began the book' and 'I began the beer' or, say, 'The goat began the book'.
It's clear (I think) that something like this is the case at the level of the thought formed in interpreting utterances of these sentences, but it's a huge (and apparently unjustified) step from there to say that those differences in the meaning of 'begin' should be encoded in the lexical semantics.