Character-Level Chinese Dependency Parsing

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1326–1336, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Character-Level Chinese Dependency Parsing Meishan Zhang†, Yue Zhang‡ , Wanxiang Che†, Ting Liu†∗ †Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {mszhang, car, tliu}@ir.hit.edu.cn ‡Singapore University of Technology and Design yue zhang@sutd.edu.sg Abstract Recent work on Chinese analysis has led to large-scale annotations of the internal structures of words, enabling character- level analysis of Chinese syntactic struc- tures. In this paper, we investigate the problem of character-level Chinese depen- dency parsing, building dependency trees over characters. Character-level infor- mation can benefit downstream applica- tions by offering flexible granularities for word segmentation while improving word- level dependency parsing accuracies. We present novel adaptations of two ma- jor shift-reduce dependency parsing algo- rithms to character-level parsing. Exper- imental results on the Chinese Treebank demonstrate improved performances over word-based parsing methods. 1 Introduction As a light-weight formalism offering syntactic information to downstream applications such as SMT, the dependency grammar has received in- creasing interest in the syntax parsing commu- nity (McDonald et al., 2005; Nivre and Nilsson, 2005; Carreras et al., 2006; Duan et al., 2007; Koo and Collins, 2010; Zhang and Clark, 2008; Nivre, 2008; Bohnet, 2010; Zhang and Nivre, 2011; Choi and McCallum, 2013). Chinese dependency trees were conventionally defined over words (Chang et al., 2009; Li et al., 2012), requiring word segmen- tation and POS-tagging as pre-processing steps. Recent work on Chinese analysis has embarked on investigating the syntactic roles of characters, leading to large-scale annotations of word internal structures (Li, 2011; Zhang et al., 2013). Such an- notations enable dependency parsing on the char- acter level, building dependency trees over Chi- nese characters. Figure 1(c) shows an example of ∗Corresponding author. 林业局 副局长 会 上 发言 forestry administration deputy director meeting in make a speech (a) a word-based dependency tree 林 业 局 副 局 长 会 上 发 言 woods industry office deputy office manager meeting in make speech (b) a character-level dependency tree by Zhao (2009) with real intra-word and pseudo inter-word dependencies 林 业 局 副 局 长 会 上 发 言 woods industry office deputy office manager meeting in make speech (c) a character-level dependency tree investigated in this pa- per with both real intra- and inter-word dependencies Figure 1: An example character-level dependency tree. “林业局副局长在大会上发言 (The deputy director of forestry administration make a speech in the meeting)”. a character-level dependency tree, where the leaf nodes are Chinese characters. Character-level dependency parsing is interest- ing in at least two aspects. First, character-level trees circumvent the issue that no universal stan- dard exists for Chinese word segmentation. In the well-known Chinese word segmentation bakeoff tasks, for example, different segmentation stan- dards have been used by different data sets (Emer- son, 2005). On the other hand, most disagreement on segmentation standards boils down to disagree- ment on segmentation granularity. As demon- strated by Zhao (2009), one can …

Content truncated to 500 words