Scholarlink

<p>Multi-modal knowledge graph completion (MKGC) has garnered substantial research interest, particularly in applying multi-modal knowledge graphs (MMKGs). Previous studies have proposed various approaches to address the MKGC problem, including a variety of knowledge graph embedding techniques and approaches based on large language models. However, these methods face two limitations. First, many models embed entities and relations in a regular embedding space, failing to capture multi-modal components’ hierarchical semantics. Second, existing methods struggle to learn hybrid relations between entities of different modalities, which constrains the performance of MKGC models. In this paper, we propose a novel model, <b>H</b>ybrid <b>R</b>elations-guided MKGC with <b>H</b>ierarchical <b>E</b>mbeddings (HRHE), to address the aforementioned limitations. Specifically, HRHE projects MMKGs into a polar coordinate system to capture hierarchical semantics and facilitate intra-inter-class interactions. Subsequently, HRHE introduces a general hybrid relations learning approach to discover implicit relations among entities of different modalities. Noting that our work can flexibly employ various neural network architectures, including convolutional neural networks, multi-layer perceptrons, and linear graph networks, to infer the implicit associations of MMKGs. Finally, an objective function is defined to leverage the inferred hybrid relations in guiding the modeling process of MMKGs. We investigate HRHE using two widely evaluated benchmarks and show that HRHE achieves better results than state-of-the-art MKGC baselines. Extensive evaluations demonstrate that constructing hierarchical semantic space and utilizing implicit associations among entities of different modalities can advance MKGC.</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

HRHE: hybrid relations-guided multi-modal knowledge graph completion using hierarchical embeddings

  • 期刊论文

  • 发表日期：

    2025年9月2日

查看全文
  • Xinyu Lu,
  • Hao Li,
  • Wei Lu,
  • Yanshuo Chang,
  • Feng Xue

摘要

Multi-modal knowledge graph completion (MKGC) has garnered substantial research interest, particularly in applying multi-modal knowledge graphs (MMKGs). Previous studies have proposed various approaches to address the MKGC problem, including a variety of knowledge graph embedding techniques and approaches based on large language models. However, these methods face two limitations. First, many models embed entities and relations in a regular embedding space, failing to capture multi-modal components’ hierarchical semantics. Second, existing methods struggle to learn hybrid relations between entities of different modalities, which constrains the performance of MKGC models. In this paper, we propose a novel model, Hybrid Relations-guided MKGC with Hierarchical Embeddings (HRHE), to address the aforementioned limitations. Specifically, HRHE projects MMKGs into a polar coordinate system to capture hierarchical semantics and facilitate intra-inter-class interactions. Subsequently, HRHE introduces a general hybrid relations learning approach to discover implicit relations among entities of different modalities. Noting that our work can flexibly employ various neural network architectures, including convolutional neural networks, multi-layer perceptrons, and linear graph networks, to infer the implicit associations of MMKGs. Finally, an objective function is defined to leverage the inferred hybrid relations in guiding the modeling process of MMKGs. We investigate HRHE using two widely evaluated benchmarks and show that HRHE achieves better results than state-of-the-art MKGC baselines. Extensive evaluations demonstrate that constructing hierarchical semantic space and utilizing implicit associations among entities of different modalities can advance MKGC.