Multi-view Chinese Treebanking

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 257–268, Dublin, Ireland, August 23-29 2014. Multi-view Chinese Treebanking Likun Qiu1,2,3, Yue Zhang1, Peng Jin4 and Houfeng Wang2 1Singapore University of Technology and Design, Singapore 2Institute of Computational Linguistics, Peking University, China 3School of Chinese Language and Literature, Ludong University, China 4Lab of Intelligent Information Processing and Application, Leshan Normal University, China {qiulikun,jandp,wanghf}@pku.edu.cn, yue zhang@sutd.edu.sg Abstract We present a multi-view annotation framework for Chinese treebanking, which uses dependen- cy structures as the base view and supports conversion into phrase structures with minimal loss of information. A multi-view Chinese treebank was built under the proposed framework, and the first release (PMT 1.0) containing 14,463 sentences is be made freely available. To verify the effectiveness of the multi-view framework, we implemented an arc-standard transition-based dependency parser and added phrase structure features produced by the phrase structure view. Experimental results show the effectiveness of additional features for dependency parsing. Fur- ther, experiments on dependency-to-string machine translation show that our treebank and parser could achieve similar results compared to the Stanford Parser trained on CTB 7.0. 1 Introduction Phrase structures (PS) and dependency structures (DS) are two of the most popular grammar formalisms for statistical parsing (Collins, 2003; Charniak, 2000; McDonald et al., 2005; Nivre, 2006; Petrov and Klein, 2007; Zhang and Clark, 2008). While DS trees emphasize the grammatical relation between heads and dependents, PS trees stress the hierarchical constituent structures of sentences. Several researchers have explored DS and PS simultaneously to enhance the quality of syntactic parsing (Wang and Zong, 2010; Farkas and Bohnet, 2012; Sun and Wan, 2013) and tree-to-string machine translation (Meng et al., 2013), showing that the two types of information complement each other for NLP tasks. Most existing Chinese and English treebanks fall into the phrase structure category, and much work has been done to convert PS into DS (Magerman, 1994; Collins et al., 1999; Collins, 2003; Sun and Jurafsky, 2004; Johansson and Nugues, 2007; Duan et al., 2007; Zhang and Clark, 2008). Research on statistical dependency parsing has frequently used dependency treebanks converted from phrase structure treebanks, such as the Penn Treebank (PTB) (Marcus et al., 1993) and Penn Chinese Treebank (CTB) (Xue et al., 2000). However, previous research shows that dependency categories in converted tree- banks are simplified (Johansson and Nugues, 2007), and the widely used head-table PS to DS conversion approach encounters ambiguities and uncertainty, especially for complex coordination structures (Xue, 2007). The main reason is that the PS treebanks were designed without consideration of DS conversion, leading to inherent ambiguities in the mapping, and loss of information in the resulting DS treebanks. To minimize information loss during treebank conversions, a treebank could be designed by considering PS and DS information simultaneously; such treebanks have been proposed as multi-view treebanks (Xia et al., 2009). We develop a multi-view treebank for Chinese, which treats PS and DS as different views of the same internal structures of a sentence. We choose the DS …

Content truncated to 500 words