Research On Multi-Granularity Relational Linking In Knowledge Graph

Knowledge Graph(KG),known as a semantic network that stores a large amount of structured and semi-structured knowledge,has been widely used in various fields in Natural Language Processing.To better understand the plain text with knowledge in KGs,establishing a “bridge” between knowledge and natural language texts becomes increasingly urgent.The building process,to be specific,could be implemented by a relation linking method.It might be conducive to the Knowledge Graphs Completion,and still benefit various applications related to KGs such as knowledge extraction,automatic question answering and intelligent search engine.The knowledge in KGs mainly includes the entities and the relationship between entities.And,in order to associate the relationships with the corresponding natural language expression in the texts,a Multi-Granularity Relation Linking System for Wikidata(MGRLSW)is proposed.MGRLSW establishes the effective relation linking by mapping the words or sequences,which could present entities' properties in the plain texts,into corresponding attributes in KGs.For this relation linking model,this paper's contributions are as follows:(1)The description of a relationship always consists of various expressions in plain texts.Therefore,to distinguish these expressions,we need to separate them into different categories,and gather resemble representations together.The most common approach is cluster.Using Location Sensitive Word Mover's Distance(LSWMD)algorithm to calculate the similarity between sentences,later clustering the sentences by DPC algorithm,eventually we could obtain these clusters,which only indicate one relationship's similar expressions in one cluster.(2)By analyzing the distribution of these words,which could represent internal relations,it could be conveniently observed that they tend to follow certain rules.Thus,we try to use Beta distribution,Gauss distribution and Gaussion Mixtrue Model(GMM)distribution to fit the positions of the words,and construct corresponding BoD-beta templates,BoD-gauss templates and BoD-GMM template.Among these,the BoD-GMM template transforms a Gauss distribution of one single granularity into multiple granularity by the multi-granularity idea,and extracts feature from each granularity level to realize the transformation at multiple granularity levels.Finally,we apply MGRLSW to the task of relation classification and verify the validity of system in comparison with two state-of-the-art baselines.(3)Besides,we built a presentation system to provide users with more information,rather than merely entity information,to demonstrate the main functions of MGRLSW.The experimental results present that MGRLSW is effective.We can also apply MGRLSW to the relation prediction,automatic question answering and intelligent search engine.By building the relation linking system,a complete KGs can be further constructed so as to provide a better support for artificial intelligence.