Learning the Taxonomy of Function Words for Parsing

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.