Improving Dependency Parsing with Fined-Grained Features

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 228–236, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Improving Dependency Parsing with Fined-Grained Features Guangyou Zhou, Li Cai, Kang Liu, and Jun Zhao National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences 95 Zhongguancun East Road, Beijing 100190, China {gyzhou,lcai,kliu,jzhao}@nlpr.ia.ac.cn Abstract In this paper, we present a simple and effective fine-grained feature generation scheme for dependency parsing. We focus on the problem of grammar representa- tion, introducing fine-grained features by splitting various POS tags to different de- grees using HowNet hierarchical semantic knowledge. To prevent the oversplitting, we adopt a threshold-constrained bottom- up strategy to merge the derived subcat- egories. We conduct the experiments on the Penn Chinese Treebank. The results show that, with the fine-grained features, we can improve the dependency parsing accuracies by 0.52% (absolute) for the un- labeled first-order parser, and in the case of second-order parser, we can improve the dependency parsing accuracies by 0.61% (absolute). 1 Introduction In natural language parsing, part-of-speech (POS) information is seen as crucial to resolving ambigu- ous relationships, yet POS tags are usually too general to encapsulate a word’s syntactic behav- ior. It is therefore attractive to consider intermedi- ate entities which exist at a finer level than the POS tags, and the relationship between specific words and their syntactic contexts may be best modeled. In this paper, we introduce the fine-grained fea- tures by splitting various POS tags to different de- grees. First, we split the POS tags of each word in the Treebank using HowNet hypernym-hyponymy hierarchical semantic knowledge (Dong and Dong, 2000). Then we adopt a threshold- constrained bottom-up strategy to merge the semantic-related subcategories which are plagued by the oversplitting problems. Finally, we use $ ROOT OBJ NMOD NMOD SBJ NMOD NN NN VV NN JJ NN �� �� �� �� �� ��� foreign capital important growth enterprise become foreign trade Figure 1: An example of a labeled dependency tree. The tree contains a special token ”$” which is always the root of the tree. Each arc is directed from head to modifier and has a label describing the function of the attachment. the generated sub-categories to construct a new fine-grained feature mapping for a discriminative learner. We are thus relying on the ability of dis- criminative learning methods to identify and ex- ploit informative features. To demonstrate the effectiveness of our ap- proach, we conduct the dependency parsing exper- iments on the Penn Chinese Treebank (CTB) (Xue et al., 2005). The results show that, with the fine-grained features, we can obtain mildly signifi- cant improvements both for first-order and second- order parsing (e.g., the absolute improvements are 0.52% and 0.61%, respectively) (see Section 6). The remainder of this paper is organized as fol- lows. Section 2 introduces the Motivation. Section 3 gives background on dependency parsing and HowNet hierarchical semantic knowledge. Sec- tion 4 describes the fine-grained feature generation scheme. Section 5 presents fine-grained features. Experimental evaluation …

Content truncated to 500 words