Research On Knowledge Graph Entity Alignment Based On Graph Neural Network

Entity alignment is the task of identifying equivalent entities from different knowledge graphs,and then complete the knowledge graphs fusion.The current entity alignment methods embed the entities in the knowledge graph into low-dimensional vectors and find equivalent entities by calculating the vector similarity.The graph convolutional network is a graph embedding model,which obtains vector representation of nodes by aggregating their neighborhood information,but it ignores the edge information in the graph.Existing graph convolutional network-based entity alignment methods learn edge information in knowledge graphs by aggregating relations around entities.However,due to the heterogeneity of knowledge graphs,equivalent entities do not have identical relations around them,entities will cause noise when aggregating relations around them,which will reduce the effectiveness of the model.Meanwhile,triple is the basic composition of the knowledge graph,the graph convolutional network-based approaches are also difficult to learn the triple information in knowledge graphs.To address the above problems,this paper gives a corresponding solution based on graph convolutional networks,the main work is as follows:(1)To alleviate the problem regarding inadequate utilization of relation information in entity alignment,a novel model,joint Unsupervised Relation Alignment for Entity Alignment(URAEA),is proposed.The model first proposes a method of calculating relation embeddings to obtain accurate relation vectors;after that,by calculating the similarity of the relationship vectors,the model generates unsupervised sets of equivalent relations as the training set for relation alignment;finally,by performing relation alignment along with entity alignment,the model learns the relation information in the sets of equivalent relations.Furthermore,aiming at the problem of small training set in entity alignment,the entity training set expansion strategy is introduced.Meanwhile,a relation-aware training set expansion strategy is proposed to further utilize the relation information.The strategy considers the relations around entities when expanding the training set based on the set of equivalent relation pairs and then obtains more accurate prediction results for the training set.Experiments are conducted on the publicly available cross-lingual dataset DBP15 K,compared to the benchmark model RNM in this paper,URAEA improved the Hits@1 metric on the three datasets by 4.0%、2.9%、2.5%,which validates the effectiveness of the proposed model.(2)In order to alleviate the problem regarding inadequate utilization of triple information in entity alignment,an entity alignment model with joint triple embedding is proposed.The model calculates the triple embedding for each entity,and then uses this triple embedding for entity alignment to utilize the triple information in knowledge graphs.In addition,considering that the relations in the knowledge graphs have different types,a relation type-aware calculation method of triple embedding is proposed;meanwhile,constraints based on relation types are added to this model,to learn the potential mapping properties of relations.Experiments are conducted on the publicly available cross-lingual dataset DBP15 K,the Hits@1 metrics on the three datasets are 84.7%、87.6%、94.8%,the experimental data outperform the benchmark model RNM in this paper,which validates the effectiveness of the proposed model.