Research On Tree Kernel-based Semantic Relation Extraction Between Chinese Named Entities

Semantic relation extraction is a subtask of information extraction,which aims to find various predefined semantic relations between pairs of entities in text.The current research on English relation extraction has achieved better performance,with F score greater than 75%.Nevertheless,there still exists a big gap between research and practical application. In contrast with English relation extraction,the research on Chinese relation extraction is still in its initial phase.Many researchers have applied feature-based methods Chinese relation extraction,which lead to disappointing performances.On the other hand, kernel-based methods have obtained better performance in English relation extraction, because it effectively captures structural information in relation instances,while it needs extensive research for Chinese relation extractionThis paper proposes a tree kernel-based method for Chinese semantic relation extraction,including the following objectives:1.Build a effective Chinese relation extraction prototype system,and compare the difference on relation and entity types between Chinese and English.In addition,we address the technical problems emerging during the course of Chinese corpus preprocessing.2.Propose a convolution tree kernel-based method for Chinese semantic relation, and refine the structural representation for relation instances.We also compare and analyze the difference of various relation types between Chinese and English.3.Investigate the effect of entity semantic information on Chinese semantic relation extraction.Furthermore,these pieces of semantic information are combined with syntactic structural information to form a Chinese entity semantic relation tree,which can effectively capture both the structural information and entity semantic information.Experiment results indicate that kernel-based method achieves reasonable performance on Chinese relation extraction as it did on English.Moreover,the addition of entity semantic information into the structural syntactic information leads to significant improvement in the performance of tree kernel-based relation extraction,with the final F score equals 64.2,thus lays a fairly good foundation for further research.