Inference for PCFGs and Adaptor Grammars
Mark Johnson (Dept. of Computing - Macquarie University)
NICTA SML SEMINARDATE: 2013-05-23
TIME: 11:15:00 - 12:15:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
This talk describes the procedures we've developed for adaptor grammar inference. Adaptor grammars are a non-parametric extension to Probabilistic Context-Free Grammars (PCFGs) that can be used to describe a variety of language learning and information extraction problems. We start by reviewing MCMC samplers for PCFGs, and then introduce Adaptor Grammars, which are a non-parametric extension of PCFGs. We explain how samples from a PCFG whose rules depend on the other sampled trees can be used as a proposal distribution in an MCMC procedure for estimating Adaptor Grammars, and describe several optimizations that dramatically speed inference of complex adaptor grammars.
BIO:
Mark Johnson is a Professor of Language Sciences in Macquarie University's Department of Computing, and Director of Macquarie's Centre for Language Sciences (CLaS). His work focuses on computational models of language acquisition and parsing, and recently he has been exploring information extraction applications. He was President of the Association for Computational Linguistics (ACL) in 2003 and will be President of ACL's SIGDAT in 2015.
