MAKE:Knowledge Graph Embedding via Multi-Attention neural network

Knowledge graph embedding is a popular method to solve the incompleteness of knowledge graph. At present, research on knowledge graph embedding based on neural network has achieved remarkable results, but most models ignore the influence of the correlation among with subject entity and relation and object entity within triple. Existing attention-based models take into account the effect of correlation, but perform moderately. In this paper, we propose a multi-attention neural network-based embedding model, named MAKE, which utilizes a novel multi-attention mechanism to generate feature maps of triples by computing correlations within triple. To fully exploit the performance of the multi-attention mechanism, MAKE uses a trainable batch normalization method and a novel composite loss function to improve the model learning ability. Evaluation results on FB15K-237 and WN18RR standard datasets show that our MAKE achieves better performance than previous state-of-the-art knowledge graph embedding models.