ConMask-GNN: Leveraging Graph Neural Networks for Enhanced Knowledge Graph Completion in Static Contexts

With the rapid development of knowledge graphs (KG), enhancing the performance of knowledge graph completion has become a critical research challenge. Traditional methods for KG completion rely heavily on entity and relation embeddings, often neglecting the synergistic interaction between graph structures and textual information. To address this issue, this paper proposes a novel approach that combines Graph Neural Networks (GNN) with the ConMask model to improve the effectiveness of static knowledge graph completion tasks. Specifically, the GNN is employed to generate entity embeddings based on graph structures, while ConMask extracts key information from text using a relation-dependent content masking mechanism. The model further integrates the embeddings through a multi-head attention mechanism at the embedding layer, fusing the GNN-based entity embeddings with text embeddings. Additionally, the loss function incorporates contrastive learning, which enhances the representational capacity of the embeddings. Experiments are conducted on standard datasets FB15k-237 and WN18RR, evaluated with metrics such as Mean Rank (MR), Mean Reciprocal Rank (MRR), Hits@1, Hits@3, and Hits@10. The results demonstrate that the proposed model significantly outperforms existing baseline methods in prediction accuracy. Finally, ablation studies validate the critical role of combining GNN and ConMask in improving model performance.