I just heard this story, probably well-known to many of you.

For many years, the leading automatic translation software worked on a Chomsky-ian theory that to translate between two languages, you parse the grammar/syntax of the portion in language A, look up the words in a dictionary, then apply B’s rules to them. The results have been mixed at best.

Then IBM began the Candide Project. It took hundreds of thousands of pages of Hansard, the bilingual record of the Canadian Parliament. The project did nothing but associate words and phrases by position in the French and English versions. It had no dictionary and no rules of syntax. And it did better than the rule-based technique. (In this case “better” means that human readers gave the IBM project’s translations a higher score.)

This appeals to me because I’ve always resisted the idea that humans understand things by interiorizing rules and maps. On the other hand, this makes the argument against AI harder, for if computers and human brains are now both working associatively, we’re forced to argue about what probably can’t be argued: whether thought is necessarily an organic function, something that living flesh does, whether you have to be alive to think/experience.

(See here for a history of computational linguistics.)

Categories: Uncategorized dw