Valence Induction with a Head-Lexicalized PCFG

TL;DR

This paper introduces a method for learning verb valences from a large corpus using a lexicalized PCFG and a modified EM algorithm, achieving highly accurate subcategorization frames.

Contribution

It presents a novel approach combining lexicalized PCFGs with a modified EM algorithm to induce valence frames from large text data.

Findings

High accuracy in induced frame distributions

Effective comparison with dictionary data

Entropy measures confirm model quality

Abstract

This paper presents an experiment in learning valences (subcategorization frames) from a 50 million word text corpus, based on a lexicalized probabilistic context free grammar. Distributions are estimated using a modified EM algorithm. We evaluate the acquired lexicon both by comparison with a dictionary and by entropy measures. Results show that our model produces highly accurate frame distributions.