MVEA: Multi-View Graph Convolutional Networks for Entity Alignment in Cross-lingual Knowledge Graphs

Cross-lingual Entity Alignment (EA) aims to identify whether different entities in two Knowledge Graphs (KGs) constructed in various languages refer to the same real-world object.Current approaches predominantly focus on the structure aspect to address the heterogeneity between different KGs, however, they often overlook the semantic information of relations.Some methods attempt to incorporate attribute information to alleviate the sparsity issue present in some KGs.Nevertheless, these approaches fail to account for the dynamic nature of entities that are influenced by time, treating the attributes of entities and their relations as static.In fact, knowledge networks in the real world evolve over time, with dynamic changes in attributes and relations across different time periods.To fully exploit the rich information embedded in cross-lingual KGs, we propose Multi-View Graph Convolutional Networks for Entity Alignment (MVEA) which comprehensively consider entity representations from static relation and attribute views, as well as a dynamic time view.We then aggregate similarity matrices from both the static joint view and the dynamic time view for alignment learning.Our model has been tested on widely-used cross-lingual EA datasets and sparser datasets, demonstrating its superiority.