Abstract

I will introduce the Bayesian approach to grammar induction within an evolutionary and cognitive framework for thinking about language acquisition by first describing a simple though linguistically-inadequate account of the acquisition of a limitedclass of probabalistic context-free grammars. I will then introduce a Bayesian Incremental Parameter Setting (BIPS) algorithm for learning Generalized Categorial Grammars (GCG), which Iwill argue is cognitively feasible, linguistically-adequate, and offers plausible explanations for some putative exceptionless and statistical (typological) linguistic universals when embedded in an evolutionary model of language development and change.

Surprisingly, the BIPS -GCG theory is most naturally treated as a non-parametric Bayesian process, in which `hidden’ variables can be estimated deterministically (unlike, say, in Latent Dirichlet Processes), though the resulting theory bears similarities to recent accounts of parameter setting emerging from the ReCoS project at DTAL .

This talk is presented by the Cambridge Linguistics Forum.