Exploiting Lexical Semantic Information For Chinese Relation Extraction

Lexical semantic information plays an important role in semantic relation extractionbetween named entities. There exist many kinds of Chinese lexical semantic resources,such as TongYiCi CiLin and HowNet, which have ever been exploited in feature-basedrelation extraction. As kernel-based methods can explore structural features of discreteobjects, they have gained wide popularity in many areas of natural language processing.This paper proposes two methods that incorporate lexical semantic information into treekernel-based Chinese relation extraction to improve its performance. The maincontributions lie in:1. Proposing a method to incorporate semantic information, based on TongYiCiCiLin and HowNet, into structural information; comparing the impact of these two kinds ofsemantic resources on Chinese relation extraction; analyzing the influence of differentlevels of semantic information and polysemy phenomenon in TongYiCi CiLin, andinvestigating the redundancy and complementarity between lexical semantic informationand entity type information.2. Investigating the effects of thesaurus-based and corpus-based lexical semanticsimilarity scores on Chinese relation extraction; Incorporating lexical semantic similarityinto the standard tree kernel to constitute the Semantic Convolution Tree Kernel (SCTK);Using the above two similarity measures to investigate the impacts on Chinese relationextraction; analyzing the impact of lexical semantic similarity on the extractionperformance with different sizes of training set.The experiments of relation extraction on the ACE2005Chinese corpus shows thatsemantic information can significantly improve the extraction performance without entitytypes, while in the case of known entity types, semantic information can also noticeablyenhance the extraction performance for some relation types. Meanwhile, semantic convolution tree kernel can also facilitate the Chinese relation extraction on a certaindegree, particularly the fine-grained subtype extraction. While only with appropriate sizeof the training set, can lexical semantic similarity excert its biggest effect.