Speaker: Richard Blythe (Edinburgh)

Title: Statistical mechanics of lexicon learning in an uncertain and nonuniform world

Abstract: Children learn words at a rapid pace (about 10 a day) despite the fact that every time a word is uttered, it is not certain which meaning the speaker intended. I will describe simple models of word learning that are inspired by psychology experiments and discuss the conditions under which words can be learnt efficiently. Of particular interest is an interplay between nonuniformity in the space of meanings (i.e., the fact that some meanings may suggest themselves more frequently than others) which acts to make the learning problem harder, and a mutual exclusivity constraint (i.e., an assumption that no two words can have the same meaning) that speeds up learning. In fact, we show, by means of a statistical mechanical analysis of a word-learning model, that two simple word-learning strategies interact in such a way to generate the maximum possible learning rate up to a critical difficulty threshold, above which the learning rate begins to decrease.