CrsKGC-RSAN: Conflict Resolution Strategy for Knowledge Graph Completion

Knowledge Graph Completion (KGC) is a crucial task aimed at enhancing the completeness and accuracy of knowledge graphs (KGs) by filling in missing information. However, existing knowledge graphs face two significant challenges: 1) the negative sampling issues, where false negative samples contain potential true facts, while easy negative samples limit the generalization capacity of KGC models; and 2) the entity conflict issues, many candidate entities exhibit similarities in their explicit structural features, making it difficult for models to make accurate decisions. To address the above challenges, we propose a novel method, CrsKGC-RSAN. For the negative sampling issues, we propose a Relative Self Adversarial Negative Sampling (RSAN) module. RSAN allows for a smaller margin between positive and negative samples compared to traditional self-adversarial sampling, while assigning higher training weights to hard negative samples to address false negative and esay negative sample issues. For the entity conflict issues, we present CrsKGC module, which, in addition to learning explicit structural features of KGs, mines implicit structural features, enhancing the comprehensiveness of feature learning and effectively resolving entity conflicts. Extensive experiments on the FB15K-237 and WN18RR datasets demonstrate that our method achieves state-of-the-art results and exhibits compatibility.