Knowledge Graph Completion Based on Graph Attention Networks and Text Information

In knowledge graphs (KGs), there exist some unsolved problems such as incomplete data, hidden information with incomplete mining and so on. In the most completion models, the information of the triples in the KG is generally utilized, but the neighborhood information and rich entity description information are not included in the triples. In this paper, the knowledge graph completion (KGC) method is improved based on graph attention networks (GATs) with text information by using the neighborhood information of aggregated triples and entity description information. And the embedding capability of semantic information is enhanced in KGs. First, the feature vector of entity description information is extracted by the Bi-LSTM model and concatenated with the entity embedding in the triples. Then the joint vectors are trained by GATs to aggregate the neighborhood information. Next, the KGC task is realized by a decoder. Finally, the effectiveness of the proposed method is verified by the link prediction experiments in the public datasets FB15K-237 and WNISRR and comparison is investigated with several other existing methods. The test results show that most of the indicators in the two datasets are improved. Furthermore, it is proved that the model combined with multi-source information has better representation ability for entities, which can further improve the accuracy and comprehensive performance of KGC tasks.