A Learning Based Model for Chinese Co-reference Resolution by Mining Contextual Evidence

This paper presents a learning based model for Chinese co-reference resolution, in which diverse contextual features are explored inspired by related linguistic theory. Our main motivation is to try to boost the co-reference resolution performance only by leveraging multiple shallow syntactic and semantic features, which can escape from tough problems such as deep syntactic and semantic structural analysis. Also, reconstruction of surface features based on contextual semantic similarity is conducted to approximate the syntactic and semantic parallel preferences in resolution linguistic theories. Furthermore, we consider two classifiers in the machine learning framework for the co-reference resolution, and performance comparison and combination between them are conducted and investigated. We experimentally evaluate our approaches on standard ACE (Automatic Content Extraction) corpus with promising results.