Key Technology Research And Application Of Knowledge Graph Relation Reasoning Based On Contrastive Learning

Relational reasoning is a crucial research area in the field of knowledge graphs.In recent years,effectively utilizing existing triple knowledge to infer relational facts from knowledge graphs and making knowledge graphs more complete has become a research hotspot.However,the staple relational reasoning methods often only consider the structured information of the knowledge graph.It also ignores the rich unstructured multisource information and semantic similarity between related entities.To address these issues,this thesis proposes two knowledge representation relational reasoning algorithms.Firstly,this thesis fuses multi-source information to construct entities and relational representations.It also combines contrastive learning to model similar entities in different triples and improve the accuracy of relational reasoning.The main contributions of this thesis include:1.This thesis proposes a multiple-source knowledge representation learning algorithm based on Transformer(MKRLT)that fuses multi-source information,which aims to utilize multi-source information effectively.Firstly,MKRLT constructs three multi-source information encoders of text relation,entity description,and entity hierarchy type to obtain the unstructured multi-source representation of triples.Secondly,a mixed item is added to the energy function to assist in learning the triple structure representation.Finally,MKRLT uses entities and relationship’ vector representations for relational reasoning.Experimental results show that this method significantly affects the two subtasks of relational reasoning compared with existing knowledge representation methods.2.This thesis proposes a multiple-source knowledge representation learning algorithm based on Transformer with contrastive learning(MKRLT-CL).It enhances the accuracy of relational reasoning by using contrastive learning methods to model similar entities in different tuples.First,the model designs three strategies for selecting positive and negative samples based on the similar structures,entity descriptions,and hierarchy types among different entities.Then,based on these strategies,the model proposes three relational reasoning models includes MKRLT-CL(S,D,Z).The model jointly learns the contrast loss based on positive and negative samples of entities,and the maximum semantic spacing loss based on negative samples of tuples.This enables it to draw the similar entities closer together in vector space.Experimental results demonstrate that the method effectively models similar entities and improves the accuracy of knowledge representation in relational reasoning tasks.3.To meet the demand of automatic relational reasoning in the field of knowledge graphs,this thesis has designed and implemented a prototype system.This system aims at relational automatic reasoning in knowledge graphs by using the MKRLT and MKRLT-CL algorithms.The implemented knowledge graph relational reasoning system includes modules for data upload and download,knowledge graph display,relational reasoning,and knowledge graph completion.To demonstrate the algorithmic efficacy of the system,this thesis crawls web data to build a movie knowledge graph and uses the system to perform relational inference and reasoning completion.