Building Semantic Dependency Knowledge Graph Based on HowNet

This paper introduces a method of constructing a semantic dependency knowledge graph (SDKG) by using the rich semantic knowledge in HowNet. The establishment of SDKG depends on correspondence between the lexical dependency labels in semantic dependence bank of BLCU-HIT and the event roles in HowNet. For words with few event roles or those which are not included in the knowledge graph, sememes are recommended to them based on SPWE and SPASE algorithms to extend the SDKG. The paper demonstrates that the experiments achieve an accuracy of 86% when the sememe recommendation is conducted. Considering the establishment of the dependency relationship, a correspondence table in this paper including 87 pieces of data of event role labels mapping to dependency labels is designed. The constructed SDKG has nearly 500000 nodes that contains rich dependency information, which can be used to assist the analysis of the Semantic Dependency Parser. Besides, the results of Semantic Dependency Analysis can be drawn on to supplement the SDKG.