汉语交叉依存类非投射性现象 = Edge-crossing Non-projective Phenomenon in Chinese Language

Abstract

Most of Dependency-based parsing is only restricted in projective dependency. In order to improve parsing precision, it is inevitable to involve non-projective dependency in dependency parsing. Therefore, it is also an important topic for dependency research.

We choose edge-crossing dependency, one of the non-projective phenomena, as our research topic. By analyzing language data involved in U-penn Chinese Treebank, we found nine kinds of edge-crossing phenomena.

We proposed two principles to explain the edge-crossing phenomena. First, it is the semantic attribute harmony principle, which determines the syntactic relation among words. Second is the ring-shaped field principle of word order, which determines the linear order distribution of words. These two principles jointly restrict the formation of edge-crossing dependency.

According to these two principles, we proposed new standards for dependency-graph bank annotation. We use graph-form to present sentences. The dependency-graph bank can be used as data learning and evaluation platform for statistic-based parsers.