Research On Entity Alignment Method Based-on Representation Learning

Knowledge Graphs(KGs)provide data support for various knowledge-based intelligent applications,and play an increasingly important role in fields like data-mining,artificial intelligence.However,it is difficult for a single KG to satisfy all the knowledge required by various applications,so the interconnection of KGs becomes an inevitable trend.In recent years,representation learning-based entity alignment algorithms have emerged and became the main method of linking knowledge graph.However,the existing methods have many limitations.First,they use either only semantic translation relations or only topology structure features of entities to align entities.Second,when learning the characteristics of entity topology structure,the GNN-based method only uses entity nodes,neglecting or can't utilize effectively the characteristics of entities' relationship and attribute features.Moreover,it does not distinguish the importance of nodes while learning the embedding of entities,but uses the same weight for all nodes.Third,there is less research on iterative alignment,but relying on many priori alignment entities.To solve the above problems,this paper proposes a method that combines graph attention network(GAT)with translation model and iteratively aligning entities.The method uses GAT to learn embedded vectors of entities based on the topology of KG,and assigns different weights to the nodes according to their characteristics.Meanwhile,in addition to the entity-entity knowledge graph,the entity-relation structure graph and the entity-attribute structure graph are constructed to learn the embedded vectors of the entity based on the relation structure and attribute structure.To alleviate the problem of error propagation during iteration,a weight decay method combined with bi-directory matching is proposed.Experimental results on five real world datasets show that the proposed method is better than the baselines.