Multi-modal Graph Convolutional Network for Knowledge Graph Entity Alignment

Entity Alignment (EA) plays a crucial role in the integration of multiple knowledge graphs (KGs). With the blooming of KGs, the auxiliary multi-modal data, such as attributions and images, are widely used to enhance alignment performance. However, most existing techniques for multi-modal knowledge exploitation separately pre-train uni-modal features and heuristically merge these features, failing to adequately consider the interplay between different modalities. To tackle this problem, we propose a novel model entitled MGCEA (Multi-modal Graph Convolutional network for knowledge graph Entity Alignment), which considers the guidance of neighborhood structure in cross-modal embedding enhancement. Specifically, MGCEA pre-trains multiple modal features to initialize their corresponding embeddings. Then a multi-modal embedding enhancement mechanism, which consists of a multi-modal graph convolution network and an attention network, is developed to achieve cross-modal enhancement guided by the neighborhood structure and learn an effective joint embedding. Moreover, a joint loss based on contrast learning is introduced to optimize model parameters by considering intra-modal relationships and cross-modal interactions. The extensive experiments conducted on two benchmarks demonstrate that MGCEA significantly outperforms the state-of-the-art multi-modal knowledge graph entity alignment baselines.