Learning the Taxonomy of Function Words for Parsing

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1886–1896, Dublin, Ireland, August 23-29 2014. Learning the Taxonomy of Function Words for Parsing Dongchen Li, Xiantao Zhang, Dingsheng Luo and Xihong Wu Key Laboratory of Machine Perception and Intelligence Speech and Hearing Research Center Peking University, Beijing, China {lidc,zhangxt,dsluo,wxh}@cis.pku.edu.cn Abstract Completely data-driven grammar training is prone to over-fitting. Human-defined word class knowledge is useful to address this issue. However, the manual word class taxonomy may be unreliable and irrational for statistical natural language processing, aside from its insufficient linguistic phenomena coverage and domain adaptivity. In this paper, a formalized representation of function word subcategorization is developed for parsing in an automatic manner. The function word classification representing intrinsic features of syntactic usages is used to supervise the grammar induction, and the structure of the taxonomy is learned simultaneously. The grammar learning process is no longer a unilaterally supervised training by hierarchical knowledge, but an interactive process between the knowledge structure learning and the grammar training. The established taxonomy implies the stochastic significance of the diversified syntactic features. The experiments on both Penn Chinese Treebank and Tsinghua Treebank show that the proposed method improves parsing performance by 1.6% and 7.6% respectively over the baseline. 1 Introduction Probabilistic context-free grammar (PCFG) is widely used in the fields of speech recognition, machine translation, information retrieval, etc. It takes the empirical rules and probabilities from a Treebank. However, due to the context-free assumption, PCFG does not always perform well (Klein and Man- ning, 2003). For instance, it assumes adverbs, including temporal adverbs, degree adverbs and negation adverbs, to share the same distribution, whereas the distinction would provide useful indication for dis- ambiguating the syntactic structure of the context. It arose that the manual word classification in linguistic research was used to enrich PCFG and im- prove the performance. However, from the point of view of statistical natural language processing, there are some drawbacks for manual classification. Firstly, Linguistic phenomena covered by the manual refinement may be limited by the linguistic observations of human. Secondly, the evidence of manual refinement is often based on a particular corpus or specific sources of knowledge acquisition. As a result, its adaptivity to different domains or genres may be insufficient. As for function words, due to the ambi- guity and complexity in syntactic grammar, it is more difficult to develop formalized representation than for content words. There are diversified standards for grammar refinement. Consequently, the word clas- sification or category refinement can be conducted in distinct manners, while each of them is reasonable in some sense. A delicate hierarchical classification inevitably involves in multiple dividing standards. However, the word sets under distinct dividing standards may be overlapping. The problems come up that how to choose the set of the multiple standards to cooperate to build the taxonomy, and how to de- cide the priority of each standard. Regarding that the manual method is hard to overcome critical issues, manual taxonomy for function words may …

Content truncated to 500 words