Each human language contains an unbounded number of different sentences. How can something so large and complex possibly be learnt? Over the past decade and a half we've figured out how to define probability distributions over grammars and the linguistic structures they generate, opening up the possibility of Bayesian models of language acquisition. Bayesian approaches are particularly attractive because they can exploit "prior" (e.g., innate) knowledge as well as statistical generalizations from the input. Recently a variety of new non-parametric Bayesian methods have been developed that let us identify the relevant parameters as well as their values. This talk describes a specific class of non-parametric models called Adaptor Grammars (AGs), which generalise over the infinite sets of subtrees defined by a CFG. We explain how AGs can be applied to morphology induction, unsupervised word segmentation and topic modelling.
Joint work with Tom Griffiths (Berkeley) and Sharon Goldwater (Edinburgh)