Friday, December 29, 2006

Dummett on Davidson on translation

In his "What is a theory of meaning? (I)", Dummett attributes an argument to Davidson about the inadequacy of a translation manual for meaning. The argument goes that one could have a complete translation manual from a language one doesn't understand into a language one doesn't understand. One could use this manual to get a perfectly adequate translation without understanding either the source or the target sentences. Since a theory of meaning is supposed to double as a theory of understanding for a language, it follows that translation manuals can't be all there is to meaning. This argument is used by Dummett against Davidson when he claims that (a) Davidson wants a theory of meaning to be modest (in the sense that you must already have the concepts in question to use it) and (b) that a modest theory of meaning does nothing over and above a translation manual. Premiss (a) is true. Most of the relevant argumentation goes to supporting (b), since convincing us of that will get Dummett to the conclusion that a theory of meaning should be full-blooded, as this is the only other option.

I don't know where this argument comes from in Davidson's work. I'd like to know. Dummett doesn't have a citation. It sounds like something that could be there. It seems like the thrust of the argument can be put another way. Translation manuals are solely syntactic (in the logician's sense). They map strings to strings (or if we have a fancy manual, phoneme sequences to phoneme sequences, or syntactic structures to syntactic structures). At no point in the use of a translation manual does meaning come into consideration. Of course, the use of a translation manual would require that source sentences be disambiguated and parsed properly. This job might require recourse to meanings, but apart from this possible presupposition there is no mention of meanings. So, Davidson's argument comes down to saying that syntax by itself can't take care of meaning cause syntax can't explain understanding. I fear there is some misinterpretation going on because I've just made Davidson sound like Chinese Room Searle. But, this might not be too far off base since Davidson's writings on truth emphasize the indispensability of the semantic in interpretation and in our theories of the world. See his "Material Mind" and "Mental Events" for examples of this.


Aidan said...

It's worth noting that Dummett takes back the attribution of modesty to Davidson in 'The Logical Basis of Metaphysics' (see 108-10).

Shawn said...

That is interesting. Do you know if Davidson used the argument that Dummett attributes to him and, if so, where he makes it?

Aidan said...

Sorry, nope. I don't know Davidson nearly as well as I should - something I intend to fix this coming semester.

Brendan said...

Ooh, that sounds Chinese room-ish! So I'll repeat my normal objection -- a phoneme => phoneme translation manual would be hideously complex. In fact, to even create such a table, you'd have to introduce all these complex dependencies whose structure would necessarily have to include syntax. (In the case of the Chinese room, note that the guy has to answer things like "How is your family? What did they teach you of the Cultural Revolution in school?" so that those allegedly mere symbol tables he has would have to encode huge amounts of real-world knowledge. Thought experiments hide important computational complexity considerations.)

Anyway, a syntactic parse => syntactic parse translation manual might be kind of feasible. A syntax => common-cross-linguistic-meaning => syntax translation manual would be much more manageable. It's just a matter of computational complexity for your pathway of language1 <=> language2. I claim that the pathway that centralizes on meaning:

phonemes1 <=> syntax1 <=> meaning! <=> syntax2 <=> phonemes2

is a much more compact and natural way to describe what's going on. A real-world example of this approach in machine translation is documented here.

Shawn said...

The complexity issue is a good point, but it doesn't really address the question here, I think. From an NLP standpoint it looks on target, but the question here is more of an "in principle" one abstracting from all computational issues. The thought experiment does sweep computational issues under the rug, mainly because they are not the issue.

The question answering directed at the Chinese room doesn't really hit the Davidson point since answering questions isn't really the concern. (I suppose using the translation manual one could translate into one's home language, formulate an answer, then translate that back to the target.) Although, I feel like Davidson would agree with the idea since he thinks the syntactic can't be used to exhaust the semantic.

The idea of going from phoneme to syntactic form to meaning to syntactic form to phoneme is something I had problems with when I took the NLP classes. Mainly the step from syntactic to meaning back to syntactic form. Whether meaning of that form is isolatable seems like it is still an open question (Searle says yes, Quine says no, e.g.). I had written up a post on how the translation manual idea isn't feasible, but then I remembered hearing something about NLP systems that operate in much the same way. Thanks for the link, since that will be useful in the future.

Also, it seems like there are a lot of forms that could be used to encode any particular meaning (if you had one). I suppose that last point would count against both sides.

Brad said...

Davidson makes something like the point you attribute to him in Radical Interpretation (and in nearly all of his discussions about Radical Translation vs. Radical Interpretation, because he wants to say why he was unsatisfied with radical translation and felt the need to move to radical interpretation).

But he goes over it again when he smokes Horwich in his responses in the Interpreting Davidson volume:

"I have criticized such accounts on the ground that someone who knows and understands a translation manual may not learn the meaning of a single word or sentence of a translated language. The reason is simple: a translation manual tells us how one language, say Italian, is to be translated into another, perhaps Japanese, but since the manual may be in English, it will not help anyone to understand either Italian or Japanese unless that person already knows one of them. As Horwich says, one needs to konw that 'Martie' in Italian means (or, better, refers to [oh snap...] Mars, which is a semantic, not a translational, fact. One may confuse the two if one assumes that one's own language is both the metalanguage and one of the languages for which we are given a translation manual."