This paper highlights that the semantic context surrounding a triplet (h, r, t) provides crucial clues for prediction in Knowledge Graph Completion (KGC). Existing node-based message passing mechanisms indiscriminately aggregate information from all adjacent edges, leading to noise, information dilution, or excessive smoothing. To address this issue, this paper proposes a semantically-aware relational message passing approach. The key innovation is the introduction of a semantically-aware Top-K neighbor selection strategy. This strategy evaluates the semantic relevance between a central node and its connected edges within a shared latent space and selects only the most relevant Top-K edges. Then, a multi-head attention aggregator is used to effectively fuse the information from the selected edges with the central node's unique representation to generate a semantically-driven node message. Therefore, our model not only leverages the structure and characteristics of edges within the knowledge graph but also more accurately captures and propagates the most relevant contextual information for a specific link prediction task, effectively mitigating interference from irrelevant information. Extensive experiments demonstrate that the proposed approach outperforms existing methods on several existing benchmarks.
