Knowledge Graph Completion Based on Entity Description Information and Graph Attention

Knowledge graphs, as a structured knowledge representation, are widely used in fields such as recommender systems and intelligent Q&A. However, existing knowledge graphs usually suffer from the problem of incomplete information, especially in large-scale entity and relationship scenarios, and lack efficient complementation methods. To this end, this paper proposes TNGAT, a knowledge graph complementation model based on entity descriptions and neighborhood information, which improves the complementation performance by combining BERT to generate semantic representations of entities with Graph Attention Networks (GAT) to weightedly aggregate neighbor information. In addition, TNGAT employs ConvR decoder to optimize relational inference. Experimental results show that TNGAT significantly outperforms existing methods on the FB15K237 and WN18RR datasets, validating its superiority.