Combining constituent parsers

Abstract

Combining the 1-best output of multiple parsers via parse selection or parse hybridization improves f-score over the best individual parser (Henderson and Brill, 1999; Sagae and Lavie, 2006). We propose three ways to improve upon existing methods for parser combination. First, we propose a method of parse hybridization that recombines context-free productions instead of constituents, thereby preserving the structure of the output of the individual parsers to a greater extent. Second, we propose an efficient linear-time algorithm for computing expected f-score using Minimum Bayes Risk parse selection. Third, we extend these parser combination methods from multiple 1-best outputs to multiple n-best outputs. We present results on WSJ section 23 and also on the English side of a Chinese-English parallel corpus.