Data-driven, PCFG-based and Pseudo-PCFG-based Models for Chinese Dependency Parsing

Data-driven, PCFG-based and Pseudo-PCFG-based Models for Chinese Dependency Parsing Weiwei Sun and Xiaojun Wan Institute of Computer Science and Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {ws,wanxiaojun}@pku.edu.cn Abstract We present a comparative study of transition-, graph- and PCFG-based models aimed at il- luminating more precisely the likely contri- bution of CFGs in improving Chinese depen- dency parsing accuracy, especially by com- bining heterogeneous models. Inspired by the impact of a constituency grammar on de- pendency parsing, we propose several strate- gies to acquire pseudo CFGs only from de- pendency annotations. Compared to linguistic grammars learned from rich phrase-structure treebanks, well designed pseudo grammars achieve similar parsing accuracy and have equivalent contributions to parser ensemble. Moreover, pseudo grammars increase the di- versity of base models; therefore, together with all other models, further improve sys- tem combination. Based on automatic POS tagging, our final model achieves a UAS of 87.23%, resulting in a significant improve- ment of the state of the art. 1 Introduction Popular approaches to dependency parsing can be divided into two classes: grammar-free and grammar-based. Data-driven, grammar-free ap- proaches make essential use of machine learning from linguistic annotations in order to parse new sentences. Such approaches, e.g. transition-based (Nivre, 2008) and graph-based (McDonald, 2006; Torres Martins et al., 2009) have attracted the most attention in recent years. In contrast, grammar- based approaches rely on linguistic grammars (in either dependency or constituency formalisms) to shape the search space for possible syntactic anal- ysis. In particular, CFG-based dependency parsing exploits a mapping between dependency and con- stituency representations and reuses parsing algo- rithms developed for CFG to produce dependency structures. In previous work, data-driven, discrim- inative approaches have been widely discussed for Chinese dependency parsing. On the other hand, various PCFG-based constituent parsing methods have been applied to obtain phrase-structures as well. With rich linguistic rules, phrase-structures of Chinese sentences can be well transformed to their corresponding dependency structures (Xue, 2007). Therefore, PCFG parsers with such conversion rules can be taken as another type of dependency parser. We call them PCFG-based parsers, in this paper. Explicitly defining linguistic rules to express precisely generic grammatical regularities, a con- stituency grammar can be applied to arrange sen- tences into a hierarchy of nested phrases, which de- termines constructions between larger phrases and their smaller component phrases. This type of infor- mation is different from, but highly related to, the information captured by a dependency representa- tion. A constituency grammar, thus, has great possi- ble contributions to dependency parsing. In order to pave the way for new and better methods, we study the impact of CFGs on Chinese dependency parsing. A series of empirical analysis of state-of- the-art graph-, transition-and PCFG-based parsers is presented to illuminate more precisely the properties of heterogeneous models. We show that CFGs have a great impact on dependency parsing and PCFG- based models have complementary predictive pow- ers to data-driven models. System ensemble is an effective and important technique to build more accurate parsers …

Content truncated to 500 words