Scholarlink

<p>This study aims to address the challenges of cited text span recognition, a task that is crucial for understanding inter-document relationships, in scientific papers. However, existing methods have limitations in capturing the deep semantic of cited text spans. To overcome these gaps, we propose a novel cited text span&#xa0;recognition model,&#xa0;DACL-cite. DACL-cite first filters candidate cited text spans through supervised ranking. Then, it uses a&#xa0;Large Language Model (LLM) for data augmentation and optimizes the model through contrastive learning. Through this method, the model can learn feature vectors that are helpful for downstream tasks, thereby improving the stability and accuracy of cited text spans recognition. The experimental results show that DACL-cite performs well on public datasets, which proves the effectiveness and superiority of the method in the task of cited text span recognition. This not only verifies the rationality of the model design, but also provides a valuable reference for the research and application of LLMs in cited text span recognition.</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

A two-stage cited text span recognition model based on LLM-driven data augmentation and contrastive learning

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Mingxing Han,
  • Shiyan Ou,
  • Jiaxuan Li,
  • Kunhao Zhu,
  • Weimin Nie

摘要

This study aims to address the challenges of cited text span recognition, a task that is crucial for understanding inter-document relationships, in scientific papers. However, existing methods have limitations in capturing the deep semantic of cited text spans. To overcome these gaps, we propose a novel cited text span recognition model, DACL-cite. DACL-cite first filters candidate cited text spans through supervised ranking. Then, it uses a Large Language Model (LLM) for data augmentation and optimizes the model through contrastive learning. Through this method, the model can learn feature vectors that are helpful for downstream tasks, thereby improving the stability and accuracy of cited text spans recognition. The experimental results show that DACL-cite performs well on public datasets, which proves the effectiveness and superiority of the method in the task of cited text span recognition. This not only verifies the rationality of the model design, but also provides a valuable reference for the research and application of LLMs in cited text span recognition.