Dual Context Representation Learning Framework for Entity Alignment

Entity alignment, which aims to identify entities with the same meaning in different Knowledge Graphs (KGs), is a key step in knowledge integration. Despite the promising results achieved by existing methods, they often fail to fully leverage the structure information of KGs for entity alignment. Therefore, our goal is to thoroughly explore the features of entity neighbors and relationships to obtain better entity embeddings. In this work, we propose DCEA, an effective dual-context representation learning framework for entity alignment. Specifically, the neighbor-level embedding module introduces relation information to more accurately aggregate neighbor context. The relation-level embedding module utilizes neighbor context to enhance relation-level embeddings. To eliminate semantic gaps between neighbor-level and relation-level embeddings, and fully exploit their complementarity, we design a hybrid embedding fusion model that adaptively performs embedding fusion to obtain powerful joint entity embeddings. We also jointly optimize the contrastive loss of multi-level embeddings, enhancing their mutual reinforcement while preserving the characteristics of neighbor and relation embeddings. Additionally, the decision fusion module combines the similarity scores calculated between entities based on embeddings at different levels to make the final alignment decision. Extensive experimental results on public datasets indicate that our DCEA performs better than state-of-the-art baselines.