Scholarlink

<p>Few-shot knowledge graph completion aims to address the long-tail distribution problem in knowledge graphs. Existing methods primarily aggregate neighborhood information to learn representations of task entity pairs. However, these methods struggle to accurately distinguish the contributions of neighbors under multiple-relations, leading to insufficiently discriminative representations of task entity pairs. Additionally, in relation embedding learning, these methods overly rely on semantic information while neglecting the mapping properties of triples, which makes them difficult to effectively model complex task relations. In this paper, we propose a relation learning with neighborhood fusion interaction method for few-shot knowledge graph completion. It uses a multi-layer perceptron to fuse features of relations and entities during neighborhood aggregation, and employs a cross-attention network to capture interactions among the fused embeddings, dynamically allocating more accurate weights to neighbors under both complex relations and multiple relations, thereby enhancing the discriminability of task entity pair representations. During the embedding learning phase, background knowledge is leveraged to extract the mapping properties of triples and combined with the semantic information to optimize the update of relation meta-representations, thereby improving the modeling capacity for complex task relations. Experimental results on the public datasets NELL-One and Wiki-One show that our method outperforms the existing methods. Compared to the state-of-the-art baseline, our method achieves 5-shot performance improvements of 2.1% and 0.9% on the two datasets, respectively, by the metric MRR. The code can be found in <a href="https://github.com/wenxuanliu666/NFI-FKGC.">https://github.com/wenxuanliu666/NFI-FKGC.</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

Relation learning with neighborhood fusion interaction for few-shot knowledge graph completion

  • 期刊论文

  • 发表日期：

    2025年10月24日

查看全文
  • Bin Yang,
  • Shuai Liu,
  • Mengqi Shi,
  • Mengxiang Peng

摘要

Few-shot knowledge graph completion aims to address the long-tail distribution problem in knowledge graphs. Existing methods primarily aggregate neighborhood information to learn representations of task entity pairs. However, these methods struggle to accurately distinguish the contributions of neighbors under multiple-relations, leading to insufficiently discriminative representations of task entity pairs. Additionally, in relation embedding learning, these methods overly rely on semantic information while neglecting the mapping properties of triples, which makes them difficult to effectively model complex task relations. In this paper, we propose a relation learning with neighborhood fusion interaction method for few-shot knowledge graph completion. It uses a multi-layer perceptron to fuse features of relations and entities during neighborhood aggregation, and employs a cross-attention network to capture interactions among the fused embeddings, dynamically allocating more accurate weights to neighbors under both complex relations and multiple relations, thereby enhancing the discriminability of task entity pair representations. During the embedding learning phase, background knowledge is leveraged to extract the mapping properties of triples and combined with the semantic information to optimize the update of relation meta-representations, thereby improving the modeling capacity for complex task relations. Experimental results on the public datasets NELL-One and Wiki-One show that our method outperforms the existing methods. Compared to the state-of-the-art baseline, our method achieves 5-shot performance improvements of 2.1% and 0.9% on the two datasets, respectively, by the metric MRR. The code can be found in https://github.com/wenxuanliu666/NFI-FKGC.