Scholarlink

Existing large language models(LLMs) in the legal domain typically only handle text within a limited context window size, making it challenging to fully meet the requirements of long legal texts. In this paper, we propose to achieve long-text information extraction for Chinese judgement documents by training on short-text models with low fine-tuning costs. Through training on short-text data and improving the rotational positional encoding method, we achieve a model capable of directly inferring on long texts with relatively low fine-tuning costs. Overall experimental results demonstrate that the our model for long judgement documents information extraction outperforms all compared LLMs.

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

Empowering LLMs for Long-Text Information Extraction in Chinese Legal Documents

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Chenchen Shen,
  • Chengwei Ji,
  • Shengbin Yue,
  • Xiaoyu Shen,
  • Yun Song,
  • Xuanjing Huang,
  • Zhongyu Wei

摘要

Existing large language models(LLMs) in the legal domain typically only handle text within a limited context window size, making it challenging to fully meet the requirements of long legal texts. In this paper, we propose to achieve long-text information extraction for Chinese judgement documents by training on short-text models with low fine-tuning costs. Through training on short-text data and improving the rotational positional encoding method, we achieve a model capable of directly inferring on long texts with relatively low fine-tuning costs. Overall experimental results demonstrate that the our model for long judgement documents information extraction outperforms all compared LLMs.