Scholarlink

<p>Multi-modal knowledge graph completion and fusion aims to enrich relational information, address missing data and uni-modal limitations in traditional graphs, and provide high-quality graphs for downstream tasks. However, existing models often treat all entities equally, ignoring semantic distances between entities. Besides, they often encode triples using transformers alone, overlooking the original graph topology, which weakens structural dependency modeling and relational reasoning. To bridge this gap, we propose a novel model incorporating an entity weight regulation (WR) module and a multi-layer graph fusion network (MGFN). The WR module assigns weights to entity vectors based on semantic distance, reducing the noise from weakly related entities. To effectively model structural patterns, features are processed through a multi-layer convolutional attention network before entity content encoding. Extensive experiments on public datasets demonstrate the effectiveness of our model and show notable gains over baselines, especially in mean reciprocal rank (MRR) and Hit@1.</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

MGFN-WR: multi-modal knowledge graph completion with multi-layer graph fusion network based on weight regulation

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Min Zhong,
  • Ziyang Fu,
  • Hong Liang,
  • Xiaofeng Zhu,
  • Siting Le

摘要

Multi-modal knowledge graph completion and fusion aims to enrich relational information, address missing data and uni-modal limitations in traditional graphs, and provide high-quality graphs for downstream tasks. However, existing models often treat all entities equally, ignoring semantic distances between entities. Besides, they often encode triples using transformers alone, overlooking the original graph topology, which weakens structural dependency modeling and relational reasoning. To bridge this gap, we propose a novel model incorporating an entity weight regulation (WR) module and a multi-layer graph fusion network (MGFN). The WR module assigns weights to entity vectors based on semantic distance, reducing the noise from weakly related entities. To effectively model structural patterns, features are processed through a multi-layer convolutional attention network before entity content encoding. Extensive experiments on public datasets demonstrate the effectiveness of our model and show notable gains over baselines, especially in mean reciprocal rank (MRR) and Hit@1.