Scholarlink

<p>This paper focuses on deep learning–based methods for the Author Name Disambiguation (AND) task. It introduces a feature attribution method to explain AND models by identifying the most influential features in distinguishing name-ambiguous entities (papers). Based on these explanations, we propose a novel technique to improve model accuracy in the AND task. This technique optimizes model training using explanations in two key areas: training set selection and entity attribution prior. Experiments were conducted on the benchmark dataset S2AND, which includes six mainstream AND datasets such as Aminer, Arnetminer, and PubMed. The results show that applying explanations during AND model training improves accuracy by an average of 4.2% across the six datasets compared to basic neural networks. An ablation study further validates the performance improvements at different stages of model training.</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

Improving author name disambiguation with explanations of deep learning models

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • ZhiJian Fang,
  • Yue Zhuo,
  • Liming Tu,
  • Lili He,
  • Huang Chen,
  • Jinying Xu,
  • Ganwei Shi,
  • Yu Cheng,
  • Yiqun Zheng,
  • HuaXiong Zhang

摘要

This paper focuses on deep learning–based methods for the Author Name Disambiguation (AND) task. It introduces a feature attribution method to explain AND models by identifying the most influential features in distinguishing name-ambiguous entities (papers). Based on these explanations, we propose a novel technique to improve model accuracy in the AND task. This technique optimizes model training using explanations in two key areas: training set selection and entity attribution prior. Experiments were conducted on the benchmark dataset S2AND, which includes six mainstream AND datasets such as Aminer, Arnetminer, and PubMed. The results show that applying explanations during AND model training improves accuracy by an average of 4.2% across the six datasets compared to basic neural networks. An ablation study further validates the performance improvements at different stages of model training.