COTSAE: co-training of structure and attribute embeddings for entity alignment