Knowledge Graph Completion Based On Learning And Deduction Of Rules

Knowledge Graph(KG)has been widely used in many tasks and fields.Most of KGs are still developing rapidly,and the relations between the entities in KG need to fully explore.In recent years,some researchers have embedded KG into different vector spaces,and calculate the vector for the potential relations in the KG.Some researchers have proposed to use neural network to predict the missing relations or entities in KG.The traditional methods require a large amount of training data for model training in the early stage to ensure the accuracy of representation and completion.However,the new knowledge obtained from the real world is not uniform in scale.How to complete the missing relations or entities in the KG with a small amount of training data has become an important problem in the current research.The completion task is often called Knowledge Graph Completion(KGC).KG is established on the basis of semantic network,and the knowledge contained in KG can be directly and effectively expressed by using the KG ontology rules.In the process of learning the ontology rules,the structure and content of the rules need to be learned.Structure learning is complex and time-consuming,and content learning usually uses the full traversal method to ensure the coverage of the rules learning.However,this method greatly increases the learning time.After obtaining the rules related to the KG,an effective strategy is necessary to further explore the knowledge that can facilitate the task of KGC.In conclusion,the contents of this thesis are summarized as follows:(1)By extending the Horn rules and the logical rules deduction,we define the ontology rules to express the semantics of KG.Then,we propose a framework for KGC based on the rules deduction named RuleB.(2)Aiming at the problem that the rules learning takes a lot of time to traverse the KG,we propose an algorithm that combining the random walk algorithm and K-sized traverse(RWK)to search the triples in KG.In order to solve the problem that RWK cannot guarantee the rules coverage,we propose a method based on Bloom Filter that can reduce the learning time and ensure the coverage of the rule learning.(3)Using the rules learned from KG,we propose a strategy to generate new rules through the rules deduction and use the new rules to obtain new knowledge.Finally,the acquired new knowledge is used to complete the KG.(4)Upon real databases,we test the efficiency of rule learning and the precision,recall rate and F1 value of KGC task based on RuleB.Experimental results show that the method proposed in this thesis can effectively improve the precision of KGC with a small scale of training data.