Dependency Tree Based Chinese Relation Extraction over Web Data

Abstract

A new semi-supervised approach for Chinese relation extraction (RE) over constantly growing and edgeless web data is introduced in this paper. Existing semi-supervised approaches have the better improvement potential while lacking syntactic structure and semantic meaning of a sentence and unsuitable to loosely structured Chinese sentences. To follow their basic procedures as well as covering their remaining shortages, a dependency tree (DT) including both structure and semantic information is drawn in. Based on DTs, a new kind of pattern, called DT-based pattern, is proposed to extract new triples. Later patterns are optimized according to the characteristics of Chinese and typed dependency trees. Finally, extensive experiments show the higher precision and more efficiency of the proposed approach against DIPRE.