Refining Grammars for Parsing with Hierarchical Semantic Knowledge

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1298–1307, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Refining Grammars for Parsing with Hierarchical Semantic Knowledge Xiaojun Lin, Yang Fan, Meng Zhang, Xihong Wu∗, Huisheng Chi Speech and Hearing Research Center Key Laboratory of Machine Perception (Ministry of Education) School of Electronics Engineering and Computer Science Peking University, Beijing, 100871, China {linxj, fanyang, zhangm, wxh}@cis.pku.edu.cn, chi@pku.edu.cn Abstract This paper proposes a novel method to refine the grammars in parsing by utiliz- ing semantic knowledge from HowNet. Based on the hierarchical state-split ap- proach, which can refine grammars au- tomatically in a data-driven manner, this study introduces semantic knowledge into the splitting process at two steps. Firstly, each part-of-speech node will be anno- tated with a semantic tag of its termi- nal word. These new tags generated in this step are semantic-related, which can provide a good start for splitting. Sec- ondly, a knowledge-based criterion is used to supervise the hierarchical splitting of these semantic-related tags, which can al- leviate overfitting. The experiments are carried out on both Chinese and English Penn Treebank show that the refined gram- mars with semantic knowledge can im- prove parsing performance significantly. Especially with respect to Chinese, our parser achieves an F1 score of 87.5%, which is the best published result we are aware of. 1 Introduction At present, most high-performance parsers are based on probabilistic context-free grammars (PCFGs) in one way or another (Collins, 1999; Charniak and Johnson, 2005; Petrov and Klein, 2007). However, restricted by the strong context- free assumptions, the original PCFG model which simply takes the grammars and probabilities off a treebank, does not perform well. Therefore, a va- riety of techniques have been developed to enrich and generalize the original grammar, ranging from lexicalization to symbol annotation. ∗Corresponding author: Xihong Wu. Lexicalized PCFGs use the structural features on the lexical head of phrasal node in a tree, and get significant improvements for parsing (Collins, 1997; Charniak, 1997; Collins, 1999; Charniak, 2000). However, they suffer from the problem of fundamental sparseness of the lexical dependency information. (Klein and Manning, 2003). In order to deal with this limitation, a variety of unlexicalized parsing techniques have been pro- posed. Johnson (1998) annotates each node by its parent category in a tree, and gets significant improvements compared with the original PCFGs on the Penn Treebank. Then, some manual and automatic symbol splitting methods are presented, which get comparable performance with lexical- ized parsers (Klein and Manning, 2003; Matsuzaki et al., 2005). Recently, Petrov et al. (2006) in- troduces an automatic hierarchical state-split ap- proach to refine the grammars, which can alter- nately split and merge the basic nonterminals by the Expectation-Maximization (EM) algorithm. In this method, the nonterminals are split to differ- ent degrees, as appropriate to the actual complex- ity in the data. The grammars refined in this way are proved to be much more accurate and compact than previous work on automatic annotation. This data-driven method still suffers …

Content truncated to 500 words