Scholarlink

<p>Knowledge graph completion (KGC) is used to infer new facts from existing facts. Embedding-based KGC methods efficiently predict new facts by computing similarities among embeddings, whereas rule-based KGC methods achieve accuracy by applying logical rules. Both methods are combined in studies that involve evaluating rules on the basis of the matching degree between relation embeddings and scoring functions and using triplets derived from high-quality rules to enhance embedding learning. Despite notable improvements, these rule evaluation methods are limited to simple models, and potential noise in high-quality rules is frequently overlooked. To overcome these limitations, we propose the NAAER framework, a noise-aware and model-agnostic framework that iteratively refines embeddings and rules. Our key innovation lies in the implementation of a generalizable rule quality assessment that is conducted by measuring the absolute difference between the scores of the rule body and the head triplets, coupled with adaptive noise reduction that is achieved via performance-guided rule filtering. Improvements in the link prediction performance of the NAAER framework were demonstrated using standard and sparse datasets, illustrating the broad applicability of the rule evaluation method.</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

Knowledge graph completion based on iteratively learning embeddings and noise-aware rules

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月25日

查看全文
  • Jinglin Zhang,
  • Bo Shen

摘要

Knowledge graph completion (KGC) is used to infer new facts from existing facts. Embedding-based KGC methods efficiently predict new facts by computing similarities among embeddings, whereas rule-based KGC methods achieve accuracy by applying logical rules. Both methods are combined in studies that involve evaluating rules on the basis of the matching degree between relation embeddings and scoring functions and using triplets derived from high-quality rules to enhance embedding learning. Despite notable improvements, these rule evaluation methods are limited to simple models, and potential noise in high-quality rules is frequently overlooked. To overcome these limitations, we propose the NAAER framework, a noise-aware and model-agnostic framework that iteratively refines embeddings and rules. Our key innovation lies in the implementation of a generalizable rule quality assessment that is conducted by measuring the absolute difference between the scores of the rule body and the head triplets, coupled with adaptive noise reduction that is achieved via performance-guided rule filtering. Improvements in the link prediction performance of the NAAER framework were demonstrated using standard and sparse datasets, illustrating the broad applicability of the rule evaluation method.