Persona and Contextual Semantic Embeddings for Entity Alignment

Entity alignment is to identify entities from different KGs that are equivalent in the real world. The complexity of neighborhood relations and structural heterogeneity remain major obstacles to entity alignment. To tackle the challenges, our proposed PCSE utilizes three components to form a new framework that projects persona embedding of the entity on the hyperplane based on the translation model and graph attention network. It employs BERT to obtain the name and description semantics of the entity and applies a message propagation mechanism to aggregate the context information. This strategy can produce a cognitive expression of the entity, which abstracts the whole picture of it. The experimental results on three datasets demonstrate that our proposed model outperforms existing baselines.