Scholarlink

<p>Recent studies reveal that innovative designs combining graph structures with soft prompts enable pre-trained language models (PLMs) to achieve deep fusion of knowledge graph structural and textual information. Although PLMs’ multilayer interaction mechanisms effectively capture fine-grained textual features, these works still exhibit notable limitations in perceiving coarse-grained structural patterns. This results in feature representations overly focused on fine-grained semantic units and local structural patterns, thereby hindering the effective modeling of global hierarchical correlations inherent in knowledge graphs. To tackle this issue, we propose multi-granularity information fusion for knowledge graph completion (MGIF-KGC) that systematically integrates multi-level semantics from entity, relation, and graph structure perspectives. The proposed architecture enhances structured data comprehension and improves performance on complex tasks. Experiments validate MGIF-KGC on three static knowledge graph completion and two temporal knowledge graph completion benchmarks. MGIF-KGC outperforms competitive baselines and achieves state-of-the-art results on all benchmarks. For example, on the ICEWS05-15 dataset, it achieves an MRR of 0.646, representing a 2.87% relative improvement over the strong CSProm-KG baseline (0.628). On the FB15k-237 dataset, our method also demonstrates stable improvements, with relative gains of 1.68% in MRR and 2.04% in Hits@10. Our code is available at <a href="https://github.com/kotosatsuki/MGIF-KGC">https://github.com/kotosatsuki/MGIF-KGC</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-granularity information fusion for knowledge graph completion

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Yongjin Zhang,
  • Ji Xu,
  • Honghao Liu

摘要

Recent studies reveal that innovative designs combining graph structures with soft prompts enable pre-trained language models (PLMs) to achieve deep fusion of knowledge graph structural and textual information. Although PLMs’ multilayer interaction mechanisms effectively capture fine-grained textual features, these works still exhibit notable limitations in perceiving coarse-grained structural patterns. This results in feature representations overly focused on fine-grained semantic units and local structural patterns, thereby hindering the effective modeling of global hierarchical correlations inherent in knowledge graphs. To tackle this issue, we propose multi-granularity information fusion for knowledge graph completion (MGIF-KGC) that systematically integrates multi-level semantics from entity, relation, and graph structure perspectives. The proposed architecture enhances structured data comprehension and improves performance on complex tasks. Experiments validate MGIF-KGC on three static knowledge graph completion and two temporal knowledge graph completion benchmarks. MGIF-KGC outperforms competitive baselines and achieves state-of-the-art results on all benchmarks. For example, on the ICEWS05-15 dataset, it achieves an MRR of 0.646, representing a 2.87% relative improvement over the strong CSProm-KG baseline (0.628). On the FB15k-237 dataset, our method also demonstrates stable improvements, with relative gains of 1.68% in MRR and 2.04% in Hits@10. Our code is available at https://github.com/kotosatsuki/MGIF-KGC.