Open-World Knowledge Graph Completion Based On Bayesian Network

As the underlying technique of massive data,Knowledge Graph(KG)provides knowledge service for search engine,question answering,recommender system and other applications.A triple composing of entities and corresponding relations is the preliminary structure of KG.In order to enrich knowledge,traditional KG Completion(KGC)method embeds entities and relations of KG to vector space based on representation learning,and then constructs more triples via distance calculation between vectors.However,data in real world is constantly increasing and changing,which requires that KG can promptly reflect the changing new knowledge in real world.At the same time,knowledge contained in data can also be regarded as a new knowledge source,which is quite significant for KGC.The data not included in KG are called as open-world data.Open-world KGC method was proposed to obtain the new entities that are not included in the KG,and then use new entities to construct triples.However,existing open-world KGC methods can only be used to construct one triple for a new entity at a time,which limits the richness of new knowledge to a certain extent.In fact,there is interdependent relevance between relations among entities in KG,and more triples can be constructed by using new entities in open-world data based on the relevance.Bayesian Network(BN)is widely used in representing and reasoning of interdependent relevance and uncertain knowledge between variables.In order to complete KG,we discussed open-world KGC based on BN.We used BN as representation model for interdependent relevance between relations in KG.For new entities extracted from data,we acquired other relations related to new entities based on probabilistic reasoning of BN,and then constructed more triples containing new entities to complete KG.To sum up,the contents of this thesis are summarized as follows:(1)We constructed BN based on KG to represent interdependent relevance among relations,which was used as basis of open-world KGC.(2)We presented the method of constructing triples based on probability reasoning of BN,and obtained triples containing new entities from open-world data,which were taken as evidence,and then constructed more triples by the probability reasoning of BN to complete KG.(3)We used triple classification and link prediction tasks to test our method.The efficiency of the model construction method and effectiveness of the KGC method are verified.