Scholarlink

As a structured knowledge representation, Knowledge Graphs (KGs) provide abundant semantic information and contextual understanding for a variety of downstream tasks. However, their inherent incompleteness problem significantly impacts the effectiveness of knowledge graph applications. Knowledge graph completion (KGC) models are designed to predict the missing triples to address this issue. Traditional KGC methods primarily leverage structural information, which presents certain limitations. Introducing diverse multimodal data can supplement richer external information for KGC task. Existing multimodal methods mainly focus on optimizing the fusion mechanism. However, aligning external heterogeneous multimodal data with KGs introduces inevitable redundancies and may cause noise interference. Therefore, we introduce an innovative inference framework for multimodal knowledge graph completion, featuring Multi-Granularity Multimodal Information Interaction, abbreviated as MI2KGC. Our framework alleviates the effects of modal redundancy and heterogeneity, and further capturing the semantic commonality across different multimodal representations. It significantly achieves competitive performance. We evaluated the models on three benchmarks. Thorough experiments confirm the performance and efficacy of our MI2KGC framework in addressing the MKGC task.

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 Multimodal Information Interaction for Knowledge Graph Completion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Fengyu Li,
  • Xinning Zhu,
  • Zheng Hu

摘要

As a structured knowledge representation, Knowledge Graphs (KGs) provide abundant semantic information and contextual understanding for a variety of downstream tasks. However, their inherent incompleteness problem significantly impacts the effectiveness of knowledge graph applications. Knowledge graph completion (KGC) models are designed to predict the missing triples to address this issue. Traditional KGC methods primarily leverage structural information, which presents certain limitations. Introducing diverse multimodal data can supplement richer external information for KGC task. Existing multimodal methods mainly focus on optimizing the fusion mechanism. However, aligning external heterogeneous multimodal data with KGs introduces inevitable redundancies and may cause noise interference. Therefore, we introduce an innovative inference framework for multimodal knowledge graph completion, featuring Multi-Granularity Multimodal Information Interaction, abbreviated as MI2KGC. Our framework alleviates the effects of modal redundancy and heterogeneity, and further capturing the semantic commonality across different multimodal representations. It significantly achieves competitive performance. We evaluated the models on three benchmarks. Thorough experiments confirm the performance and efficacy of our MI2KGC framework in addressing the MKGC task.