Multi-information embedding based entity alignment

Entity alignment refers to discovering two entities in different knowledge bases that represent the same thing in reality. Existing methods generally only adopt TransE or TransE-like knowledge graph representation learning models, which usually assume that there are enough training triples for each entity, and entities appearing in few triples are easily misaligned. In this paper, we propose a multi-information embedding based entity alignment method (MEEA), which utilizes embedding methods based on multi-information, including triple embedding and neighbor information embedding, to obtain the vector representations of each entity, which are then used for aligning entities. In addition, we propose a weighted neighbor information encoding method to make the neighbor information based vector representation suitable for entity alignment, which measures the effects of different neighbors on entity alignment from three aspects (i.e., the mapping cardinality of the neighbor, the relation association of the neighbor, and the attention mechanism) and gives corresponding weights. Experiments are conducted on two cross-lingual knowledge bases, and the experimental results show that MEEA is able to yield a better performance compared to the state-of-the-art methods.