Research On Key Technologies Of Knowledge Graph Semantic Representation Learning Based On Curvature Space

Knowledge graph(KG)representation learning aims to encode both entities and relations into a continuous low-dimensional vector space.However,Most existing methods only concentrate on learning representations from structural triples in Euclidean space,which cannot well exploit the rich semantic information with hierarchical structure in KGs.This paper will combine the curvature space and semantic information to propose four knowledge graph semantic representation learning models.The attribute values of data types are introduced in this paper,and the data types are refined into five primitive modalities,including integer,double,Boolean,temporal,and textual.By designing general and dedicated encoders,this paper proposes the DTEGCN model based on the Euclidean space and the DT-SGCN model based on the spherical space,to efficiently integrate data types information into the knowledge graph representation learning,and explore the representation learning ability of the zero and positive curvature space.This paper introduce the hyperbolic geometry,and define the basic operators in hyperbolic space,including space mapping function,M ¨obius addition,M ¨obius matrixvector multiplication,Hadamard product,and hyperbolic activation function.Furthermore,this paper defines a unified space based on Euclidean,spherical and hyperbolic space,which has continuous curvature and can combine the advantages of three different spaces.By introducing the above ideas into the DT-EGCN and DT-SGCN model,this paper proposes the DT-HGCN model based on the hyperbolic space and the DTUGCN model based on the unified space,to explore the representation learning ability of the negative and continuous curvature space.Extensive experiments on both synthetic and real-world datasets are conducted to evaluate the performance of the proposed models on the tasks of node classification and link prediction.The experimental results show that compared with the optimal baseline model in Euclidean space,the average performance of DT-EGCN and DTSGCN are improved by 17.42% and 14.34%.Compared with the optimal baseline model in hyperbolic space,the average performance of DT-EGCN and DT-SGCN are improved by 10.58% and 13.57%,which demonstrate the advantages of embedding data types information and leveraging the unified space.