K-Best Combination of Syntactic Parsers

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1552–1560, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP K-Best Combination of Syntactic Parsers Hui Zhang1, 2 Min Zhang1 Chew Lim Tan2 Haizhou Li1 1Institute for Infocomm Research 2National University of Singapore zhangh1982@gmail.com {mzhang, hli}@i2r.a-star.edu.sg tancl@comp.nus.edu.sg Abstract In this paper, we propose a linear model-based general framework to combine k-best parse outputs from multiple parsers. The proposed framework leverages on the strengths of pre- vious system combination and re-ranking techniques in parsing by integrating them into a linear model. As a result, it is able to fully utilize both the logarithm of the probability of each k-best parse tree from each individual parser and any additional useful features. For feature weight tuning, we compare the simu- lated-annealing algorithm and the perceptron algorithm. Our experiments are carried out on both the Chinese and English Penn Treebank syntactic parsing task by combining two state- of-the-art parsing models, a head-driven lexi- calized model and a latent-annotation-based un-lexicalized model. Experimental results show that our F-Scores of 85.45 on Chinese and 92.62 on English outperform the previ- ously best-reported systems by 1.21 and 0.52, respectively. 1 Introduction Statistical models have achieved great success in language parsing and obtained the state-of-the- art results in a variety of languages. In general, they can be divided into two major categories, namely lexicalized models (Collins 1997, 1999; Charniak 1997, 2000) and un-lexicalized models (Klein and Manning 2003; Matsuzaki et al. 2005; Petrov et al. 2006; Petrov and Klein 2007). In lexicalized models, word information play a key role in modeling grammar rule generation, while un-lexicalized models usually utilize latent in- formation derived from the parse structure diver- sity. Although the two models are different from each other in essence, both have achieved state- of-the-art results in a variety of languages and are complementary to each other (this will be empirically verified later in this paper). There- fore, it is natural to combine the two models for better parsing performance. Besides individual parsing models, many sys- tem combination methods for parsing have been proposed (Henderson and Brill 1999; Zeman and Žabokrtský 2005; Sagae and Lavie 2006) and promising performance improvements have been reported. In addition, parsing re-ranking (Collins 2000; Riezler et al. 2002; Charniak and Johnson 2005; Huang 2008) has also been shown to be another effective technique to improve parsing performance. This technique utilizes a bunch of linguistic features to re-rank the k-best (Huang and Chiang 2005) output on the forest level or tree level. In prior work, system combination was applied on multiple parsers while re-ranking was applied on the k-best outputs of individual parsers. In this paper, we propose a linear model-based general framework for multiple parsers combina- tion. The proposed framework leverages on the strengths of previous system combination and re- ranking methods and is open to any type of fea- tures. In particular, it is capable of utilizing the logarithm of the parse tree probability from each individual parser while previous combination …

Content truncated to 500 words