In learning their first language, children must learn the meanings of words and the syntactic mechanism by which these word meanings can be composed to give the meanings of sentences and phrases. The cognitive theory of language acquisition holds that both of these are learnt through the process of mapping the words heard in an utterance onto some contextually afforded interpretation of what that utterance may mean.
In this talk I present a probabilistic model of language acquisition that uses a psycholinguistically plausible learning algorithm to learn word meanings and syntax from child-directed utterances annotated with logical approximations of the context in which they appear. I show that the approach is able to learn an accurate parsing model for the target language even in the face of ambiguous training data. Furthermore, I show that both word-meanings and syntactic rules are learnt in a manner that correlates with observations of language learning in children, overcoming criticisms of previous statistical models of language acquisition.
Tom Kwiatkowski is a post-doctoral researcher working at the University of Washington on building computational systems capable of natural language understanding. He received a PhD for his thesis titled Probabilistic Grammar Induction from Sentences and Structured Meanings from the University of Edinburgh in 2012.