Joint Models for Chinese POS Tagging and Dependency Parsing

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1180–1191, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Joint Models for Chinese POS Tagging and Dependency Parsing Zhenghua Li†, Min Zhang‡, Wanxiang Che†, Ting Liu†, Wenliang Chen‡ and Haizhou Li‡ †Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {lzh,car,tliu}@ir.hit.edu.cn ‡Institute for Infocomm Research, Singapore {mzhang,wechen,hli}@i2r.a-star.edu.sg Abstract Part-of-speech (POS) is an indispensable fea- ture in dependency parsing. Current research usually models POS tagging and dependency parsing independently. This may suffer from error propagation problem. Our experiments show that parsing accuracy drops by about 6% when using automatic POS tags instead of gold ones. To solve this issue, this pa- per proposes a solution by jointly optimiz- ing POS tagging and dependency parsing in a unique model. We design several joint models and their corresponding decoding algorithms to incorporate different feature sets. We fur- ther present an effective pruning strategy to re- duce the search space of candidate POS tags, leading to significant improvement of parsing speed. Experimental results on Chinese Penn Treebank 5 show that our joint models sig- nificantly improve the state-of-the-art parsing accuracy by about 1.5%. Detailed analysis shows that the joint method is able to choose such POS tags that are more helpful and dis- criminative from parsing viewpoint. This is the fundamental reason of parsing accuracy improvement. 1 Introduction In dependency parsing, features consisting of part- of-speech (POS) tags are very effective, since pure lexical features lead to severe data sparseness prob- lem. Typically, POS tagging and dependency pars- ing are modeled in a pipelined way. However, the pipelined method is prone to error propagation, es- pecially for Chinese. Due to the lack of morpholog- ical features, Chinese POS tagging is even harder than other languages such as English. The state-of- the-art accuracy of Chinese POS tagging is about 93.5%, which is much lower than that of English (about 97% (Collins, 2002)). Our experimental re- sults show that parsing accuracy decreases by about 6% on Chinese when using automatic POS tagging results instead of gold ones (see Table 3 in Section 5). Recent research on dependency parsing usually overlooks this issue by simply adopting gold POS tags for Chinese data (Duan et al., 2007; Zhang and Clark, 2008b; Huang and Sagae, 2010). In this pa- per, we address this issue by jointly optimizing POS tagging and dependency parsing. Joint modeling has been a popular and effec- tive approach to simultaneously solve related tasks. Recently, many successful joint models have been proposed, such as joint tokenization and POS tag- ging (Zhang and Clark, 2008a; Jiang et al., 2008; Kruengkrai et al., 2009), joint lemmatization and POS tagging (Toutanova and Cherry, 2009), joint tokenization and parsing (Cohen and Smith, 2007; Goldberg and Tsarfaty, 2008), joint named en- tity recognition and parsing (Finkel and Manning, 2009), joint parsing and semantic role labeling (SRL) (Li et al., 2010), joint word sense disambigua- tion and SRL (Che and Liu, 2010), …

Content truncated to 500 words