Abstract: Entity alignment is one of the main tasks in the stage of knowledge fusion. Representation learning is the main research direction of entity alignment. Firstly, after a thorough study of the current representative entity alignment techniques, the characteristics and architecture of these methods are described, and a framework to capture the key features of these techniques is proposed. Then, based on the knowledge representation technologies they use, they are divided into two categories: Trans-based and GNN-based. Two currently widely used datasets are summarized, and 11 representative models of the above two categories are built. These models run on three datasets of the DBP15k cross-language dataset in the comparative experiments. Finally, this paper evaluates the alignment effect of mainstream models and models with different side information such as attributes and words, and provides a reference for future large-scale single-mode and even multi-modal knowledge map entity alignment studies.

Key words: knowledge graph, entity alignment, knowledge representation, profile information, similarity