Parse reranking with WordNet using a hidden variable model

Abstract

We present a new parse reranking algorithm that extends work in (Michael Collins and Terry Koo 2004) by incorporating WordNet (Miller et al. 1993) word senses. Instead of attempting explicit word sense disambiguation, we retain word sense ambiguity in a hidden variable model. We define a probability distribution over candidate parses and word sense assignments with a feature-based log-linear model, and we employ belief propagation to obtain an efficient implementation. Our main results are a relative improvement of [approximately] 0.97% over the baseline parser in development testing, which translated into a [approximately] 0.5% improvement in final testing. We also performed experiments in which our reranker was appended to the (Michael Collins and Terry Koo 2004) boosting reranker. The cascaded system achieved a development set improvement of [approximately] 0.15% over the boosting reranker by itself, but this gain did not carry over into final testing.