The Challenges of Parsing Chinese with Combinatory Categorial Grammar

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 295–304, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics The Challenges of Parsing Chinese with Combinatory Categorial Grammar Daniel Tse and James R. Curran School of Information Technologies University of Sydney Australia {dtse6695,james}@it.usyd.edu.au Abstract We apply Combinatory Categorial Grammar to wide-coverage parsing in Chinese with the new Chinese CCGbank, bringing a formalism capable of transparently recovering non-local dependencies to a language in which they are particularly frequent. We train two state-of-the-art English  parsers: the parser of Petrov and Klein (P&K), and the Clark and Curran (C&C) parser, uncov- ering a surprising performance gap between them not observed in English — 72.73 (P&K) and 67.09 (C&C) F-score on  6. We explore the challenges of Chinese  parsing through three novel ideas: develop- ing corpus variants rather than treating the cor- pus as fixed; controlling noun/verb and other  ambiguities; and quantifying the impact of constructions like pro-drop. 1 Introduction Automatic corpus conversions from the Penn Tree- bank (Marcus et al., 1994) have driven research in lexicalised grammar formalisms, such as  (Xia, 1999),  (Miyao et al., 2004) and  (Hock- enmaier and Steedman, 2007), producing the lexical resources key to wide-coverage statistical parsing. The Chinese Penn Treebank (; Xue et al., 2005) has filled a comparable niche, enabling the development of a Chinese  (Xia et al., 2000), a wide-coverage  parser (Yu et al., 2011), and recently Chinese CCGbank (Tse and Curran, 2010), a 750 000-word corpus of Combinatory Categorial Grammar (; Steedman, 2000) derivations. We train two  parsers, Clark and Curran (C&C; 2007), and the Petrov and Klein (P&K; 2007)  parser, on Chinese CCGbank. We follow Fowler and Penn (2010), who treat the English CCGbank (Hock- enmaier and Steedman, 2007) grammar as a  and train and evaluate the P&K parser directly on it. We obtain the first Chinese  parsing results: F-scores of 72.73 (P&K) and 67.09 (C&C) on la- belled dependencies computed over the  6 test set. While the state-of-the-art in Chinese syntactic parsing has always lagged behind English, this large gap is surprising, given that Fowler and Penn (2010) found only a small margin separated the two parsers on English CCGbank (86.0 versus 85.8). Levy and Manning (2003) established that prop- erties of Chinese such as noun/verb ambiguity con- tribute to the difficulty of Chinese parsing. We focus on two factors within our control: annotation deci- sions and parser architecture. Existing research has varied parsers whilst keep- ing the corpus fixed. We vary the corpus whilst keep- ing the parsers fixed by exploring multiple design choices for particular constructions. By exploiting the fully automatic CCGbank extraction process, we can immediately implement these choices and assess their impact on parsing performance. Secondly, we contrast the performance of C&C, with its tagging/parsing pipeline, with P&K, a parser which performs joint tagging and parsing, and estab- lish that P&K is less sensitive to the …

Content truncated to 500 words