On the Unsupervised Induction of Phrase-Structure Grammars

Abstract

Researchers investigating the acquisition of phrase-structure grammars from raw text have had only mixed success. In particular, unsupervised learning techniques, such as the inside-outside algorithm (Baker, 1979) for estimating the parameters of stochastic context-free grammars (SCFGs), tend to produce grammars that structure text in ways contrary to our linguistic intuitions. One effective way around this problem is to use hand-structured text like the Penn Treebank (Marcus, 1991) to constrain the learner: (Pereira and Schabes, 1992) demonstrate that the inside-outside algorithm can learn grammars effectively given such constraint, and currently the best performing parsers are trained on treebanks (Black et al., 1992; Magerman, 1995).