Scholarlink

<p>Knowledge graph completion (KGC) is a task for predicting missing fact triples in an incomplete knowledge graph (KG), which is crucial for enhancing the comprehensiveness and utility of KGs. In recent years, multilingual KGs bridged with the same entities and relations have been introduced into KGC to alleviate the poor-resource issue in an individual KG, with popular multilingual KG completion (MKGC) methods demonstrating that multilingual KG alignment (MKGA) and KGC can be mutually beneficial. However, existing methods do not sufficiently address the relation alignment (RA) within MKGA. Even though some methods consider the importance of RA, the unbalanced training issue between EA and RA may harm MKGC. Motivated by efficient contrastive learning, we propose a contrastive learning framework based on relation alignment for MKGC, named SimRMKGC, which aims to improve both MKGC and MKGA and solve the unbalanced issue between the EA and the RA. Specifically, the RA is regarded as a central component in MKGA, employing a supervised contrastive learning optimization method to align well-designed positive relation pairs effectively. In addition, we introduce two types of negatives: in-batch sampling negative and generated area-wise mixup hard negatives to improve the learning effect. Experiments on DBPedia-5L and E-PKG demonstrate state-of-the-art results: For MKGC, SimRMKGC achieves an improvement of 3.5%∼7.0% Hits@1 and 4.0% ∼7.1% MRR on DBPedia-5L, and achieves an improvement of 0.5%∼1.3% Hits@1 and 0.5% ∼1.0% MRR on E-PKG.</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

SimRMKGC: Simple relational contrastive learning on multilingual knowledge graph completion

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Xiaohan Fang,
  • Qian Zang,
  • Jibing Gong,
  • Tao Fang,
  • Yili Xu,
  • Xi Wu

摘要

Knowledge graph completion (KGC) is a task for predicting missing fact triples in an incomplete knowledge graph (KG), which is crucial for enhancing the comprehensiveness and utility of KGs. In recent years, multilingual KGs bridged with the same entities and relations have been introduced into KGC to alleviate the poor-resource issue in an individual KG, with popular multilingual KG completion (MKGC) methods demonstrating that multilingual KG alignment (MKGA) and KGC can be mutually beneficial. However, existing methods do not sufficiently address the relation alignment (RA) within MKGA. Even though some methods consider the importance of RA, the unbalanced training issue between EA and RA may harm MKGC. Motivated by efficient contrastive learning, we propose a contrastive learning framework based on relation alignment for MKGC, named SimRMKGC, which aims to improve both MKGC and MKGA and solve the unbalanced issue between the EA and the RA. Specifically, the RA is regarded as a central component in MKGA, employing a supervised contrastive learning optimization method to align well-designed positive relation pairs effectively. In addition, we introduce two types of negatives: in-batch sampling negative and generated area-wise mixup hard negatives to improve the learning effect. Experiments on DBPedia-5L and E-PKG demonstrate state-of-the-art results: For MKGC, SimRMKGC achieves an improvement of 3.5%∼7.0% Hits@1 and 4.0% ∼7.1% MRR on DBPedia-5L, and achieves an improvement of 0.5%∼1.3% Hits@1 and 0.5% ∼1.0% MRR on E-PKG.