Graph Attention Networks Fusing Semantic Information for Knowledge Graph Completion

Knowledge graph embedding aims to embed triples into low-dimensional vector spaces. Through the attention mechanism, models based on the graph attention network can learn the information of the neighboring node, improving knowledge graph completion and enriching the structural information. On the other hand, because of the shallow graph attention network, these models can only learn local structural information and cannot capture the global semantic properties of entities and relations. They also only use linear matrices to update relations, ignoring the impact of neighboring information on relations. We propose a fused semantic information graph attention network (FSI-GAT) to address the above challenges for knowledge graph completion. FSI-GAT consists of three modules: structural information learning module, semantic information learning module, and scoring module. The structural information learning module adopts the improved graph attention network to learn the local structural information of entities and relations; the semantic information learning module encodes the descriptive statements of entities and relations through the BERT pretrained language model to obtain the global semantic information; and the scoring module uses the ConvE to score the candidate tail entities. The experimental results show that FSI-GAT significantly outperforms the existing models on the link prediction task and can effectively improve the link prediction performance.