Cross-lingual Entity Alignment based on Multi-dimensional Graph Convolutional Networks and Neighborhood Interaction Network

Learning the embedding representation of knowledge graphs through graph neural networks and measuring the similarity between the obtained entity embeddings is the conventional method for achieving entity alignment in knowledge graphs, but many methods do not take the entity neighborhood information and interaction properties between entities into account. To address the above problems, an entity alignment method based on multidimensional attention mechanism and neighborhood interaction, namely MGNI, is proposed. Firstly, Bi-LSTM is utilized to construct the initial feature representation of entities and relations. Next, similar attention (SA) and heterogeneous attention (HA) mechanisms are used to learn entity structure features and interaction features, and the entities are embedded into a unified spatial vector. Finally, entity alignment is performed by integrating the information of neighboring entity nodes. The method is validated using the DBP15K dataset, and the results reveal that all Hits@1 values are above 70%, Hits@10 values are above 91%, and Mean Reciprocal Rank (MRR) values are above 76%. Compared to other traditional entity alignment methods, the performance of each index of the proposed method is superior and now achieves the greatest level. Experiments including the deletion of each module in the proposed method demonstrate that each module has distinct impacts on entity alignment and that the proposed method can effectively increase the accuracy of entity matching.