Heterogeneous Knowledge Graph Fusion System

Knowledge graph is a large-scale semantic network,attracting wide concern from a-cademia and industry in recent years.As more and more knowledge graphs are constructed,how to conduct the heterogeneous knowledge graph fusion automatically has become a pop-ular research topic.It aims to increase the coverage of domain knowledge and achieve better results on applications based on knowledge graph.In addition to ontology alignment on which the semantic web was focused in early years,two sub-tasks of heterogeneous knowledge graph fusion are:entity alignment and attribute alignment,respectively corresponding to the fusion of entity layer and pattern layer in knowledge graph.However,previous research works mainly focus on entity alignment,while less attention is paid to attribute alignment.Moreover,the correlation between these two subtasks is often ignored.But in fact,not only the alignment of entities can promote the alignment of attributes,but the alignment of attributes can also help align entities better.Based on the research and analysis above,this paper studies and proposes a knowl-edge graph fusion system,conducting entity alignment and attribute alignment interactively.Based on this algorithm,a prototype system to do heterogeneous knowledge graph fusion is developed.More details are shown as follows:(1)In the entity alignment module,an unsupervised entity alignment algorithm that comprehensively uses attribute triples and relation triples is designed.This algorithm iter-atively performs entity alignment and attribute alignment,and combines with the relation embedding model,which improves the accuracy of entity alignment by about 5%compared with existing methods.(2)In the attribute alignment module,the similarity of attributes is calculated based on the related entities.Attribute pairs,whose similarities are above a certain threshold,are regarded as the same and can be merged.This method greatly reduces attribute redundancy of the fused knowledge graph.(3)We develop a system for the fusion of heterogeneous knowledge graphs,which makes the two tasks conduct interactively.Under the condition of a unified concept layer,the system can combine entity alignment and attribute alignment and make them promote each other,to truly finish the final fusion of two knowledge graphs.We prove the effectiveness of the algorithm proposed in this paper on real datasets.The results show that the entity alignment algorithm proposed in this paper greatly improves accuracy compared with the previous mainstream algorithms.In addition,the attribute align-ment algorithm based on the related entities is superior to traditional methods in both accu-racy and recall.