Stochastic HPSG Parse Selection using the Redwoods Corpus

This article details our experiments on HPSG parse disambiguation, based on the Redwoods treebank. Using existing and novel stochastic models, we evaluate the usefulness of different information sources for disambiguation – lexical, syntactic, and semantic. We perform careful comparisons of generative and discrimintative models using equivalent features and show the consistent advantage of discriminatively trained models. Our best system performs at 76% sentence exact march accuracy.