Semantic Enhancement Based Knowledge Graph Completion for Graph Convolutional Neural Networks

Knowledge Graph Completion (KGC) is a task that aims to predict missing links in a knowledge graph based on known triples. Recent studies have demonstrated outstanding performance in KGC employing models grounded on Graph Convolutional Networks (GCN). Nevertheless, prevailing GCN-based models solely utilize neighborhood information of entities to reason, disregarding the textual semantic information of entities and relationships in the knowledge graph. Existing GCN models suffer from poor prediction performance when dealing with tail entities due to limitations. Additionally, these models still have shortcomings in the semantic feature interaction between entities and relations. This paper proposes a Semantic-Enhanced Graph Convolutional Network (SEGCN) for knowledge graph completion. The SEGCN leverages textual descriptions of entities and relations to obtain better entity and relation embeddings using a language model. Additionally, a new Attention-Convolutions Network (ACN) has been developed to enhance the semantic interaction among entities and relations. Based on experimental findings, SEGCN outperforms the state-of-the-art GCN-based model, CompGCN, by showing 0.4%, 0.6%, 0.1 %, and 0.2% improvements in MRR, Hits@l, Hits@3, and Hits@10 on the FB15k-237 dataset, and 1.9%, 1.0%, 2.5%, and 2.9% improvements on the WN18RR dataset, respectively. These findings demonstrate that SEGCN displays improved generalization and accuracy.