Exploiting Lexical Semantic Resource for Tree Kernel-Based Chinese Relation Extraction

摘要

Lexical semantic resources play an important role in semantic relation extraction between named entities.This paper exploits lexical semantic information based on HowNet to convolution tree kernels via two methods: incorporating lexical semantic similarity and embedding lexical sememes,and systematically investigates its effects on Chinese relation extraction.The experimental results on the ACE 2005 Chinese corpus show that the incorporation of lexical semantic similarity can significantly improve the performance whether entity-related information is known or not,while embedding lexical sememes can also improve the performance,but only when entity types are unknown.This demonstrates the effectiveness of lexical resources for Chinese relation extraction.In addition,the experiments also suggest that lexical semantic similarity facilitates the relation extraction,particularly the fine-grained subtype extraction,more than that of relation detection.