Scholarlink

This paper proposes SGCoT, an innovative framework for scientific data discipline classification that addresses critical challenges in multilingual support and cross-domain adaptability. Unlike traditional methods constrained by structured data requirements or annotation dependency, SGCoT integrates two key points: (1) A self-generated Chain-of-Thought mechanism that autonomously constructs reasoning processes through few-shot prompting, eliminating manual annotation; (2) Knowledge distillation that reduces inference costs by a factor of 10 while maintaining 66.4% accuracy comparable to supervised models. Experimental results on the CSLDCP dataset demonstrate superior performance over existing approaches, achieving 62.12% accuracy with GPT-4o-mini and 66.41% with GPT-4o, while significantly lowering annotation costs for 36M scientific records in practical deployments.

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

SGCoT: Self-generating Chain-Of-Thought for Discipline Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Peng Yu,
  • Faren Yan,
  • Xin Chen

摘要

This paper proposes SGCoT, an innovative framework for scientific data discipline classification that addresses critical challenges in multilingual support and cross-domain adaptability. Unlike traditional methods constrained by structured data requirements or annotation dependency, SGCoT integrates two key points: (1) A self-generated Chain-of-Thought mechanism that autonomously constructs reasoning processes through few-shot prompting, eliminating manual annotation; (2) Knowledge distillation that reduces inference costs by a factor of 10 while maintaining 66.4% accuracy comparable to supervised models. Experimental results on the CSLDCP dataset demonstrate superior performance over existing approaches, achieving 62.12% accuracy with GPT-4o-mini and 66.41% with GPT-4o, while significantly lowering annotation costs for 36M scientific records in practical deployments.