Scholarlink

<p>Scientific entity recognition serves as a pivotal module for a multitude of downstream applications, including knowledge organization and technology landscape analysis. This paper addresses the intricate challenge of scientific entity recognition by introducing a novel, annotation-free self-distillation framework that leverages the capabilities of large language models (LLMs) through an unsupervised distillation pipeline. Since semantic interpretations of words are often contextually constrained and inferable, their comprehension tends to converge when learned from extensive text corpora. Our framework utilizes the inherent reading comprehension of LLMs to process vast amounts of scientific literature. It employs a credibility evaluation mechanism to automatically identify and distill high-confidence entity extractions. These refined outputs are fed back into the model through a fine-tuning process and further optimized via an adaptive fine-tuning strategy that uses in-context knowledge editing to learn how to incorporate lexical suggestions. This methodology enables the development of a task-specific model with enhanced accuracy, significantly reducing the requirement for manual annotation. Experimental evaluations demonstrate the effectiveness of our approach, showcasing substantial performance improvements compared to generic LLMs.</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

Enhancing large language models for scientific entity recognition via fully automated self-distillation

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Maodi Hu,
  • Donghuan Song,
  • Xiumin Liu,
  • Xi Sun,
  • Yuanzhe Zhang

摘要

Scientific entity recognition serves as a pivotal module for a multitude of downstream applications, including knowledge organization and technology landscape analysis. This paper addresses the intricate challenge of scientific entity recognition by introducing a novel, annotation-free self-distillation framework that leverages the capabilities of large language models (LLMs) through an unsupervised distillation pipeline. Since semantic interpretations of words are often contextually constrained and inferable, their comprehension tends to converge when learned from extensive text corpora. Our framework utilizes the inherent reading comprehension of LLMs to process vast amounts of scientific literature. It employs a credibility evaluation mechanism to automatically identify and distill high-confidence entity extractions. These refined outputs are fed back into the model through a fine-tuning process and further optimized via an adaptive fine-tuning strategy that uses in-context knowledge editing to learn how to incorporate lexical suggestions. This methodology enables the development of a task-specific model with enhanced accuracy, significantly reducing the requirement for manual annotation. Experimental evaluations demonstrate the effectiveness of our approach, showcasing substantial performance improvements compared to generic LLMs.