GIRMSF-global information reconstruction and multi-scale feature sharpening framework for knowledge graph embedding

In contemporary research, Knowledge Graph Embedding (KGE) is recognized as an effective technique for prediction tasks in structured data. However, current KGE models face three major limitations in global–local modeling: (1) high complexity of information extraction models, (2) inadequate utilization of relation embeddings, and (3) absence of a feature sharpening mechanism to enhance global–local features. To address these issues, this paper proposes a global information reconstruction and multi-scale feature sharpening framework (GIRMSF). First, GIRMSF introduces an entity-relation semantic reconstruction module (GSR), which reconstructs a semantic information matrix composed of entities and relations using multi-layer dilated convolution, thereby capturing features with contextual and logical association. Second, a multi-scale feature capture module (MSC) is designed to extract semantic information from multiple perspectives. Finally, the framework proposes a group normalization-based information sharpening module (GNS), which standardizes and thresholds multi-scale features to salient information while suppressing redundancy. Extensive experiments on 7 benchmark datasets of varying scales demonstrate that GIRMSF significantly outperforms existing methods and exhibits strong generalization capabilities across diverse scenarios.