Scholarlink

<p>Legal document summarization presents a critical challenge due to the dense, complex nature and extreme length of legal texts. This paper proposes and evaluates two distinct methods for this task. The first is an optimized hybrid model that combines TextRank for extractive summarization with a fine-tuned Facebook-BART-Large-CNN model for abstractive summarization. We then apply Bayesian Optimization to enhance the coherence and fluency of the final summary. Our second approach involves fine-tuning the LLaMA-2 (7B) model using Low-Rank Adaptation (LoRA) on a custom dataset of 7036 Indian Supreme Court judgments. We evaluated the summaries using ROUGE metrics and human evaluation by a legal expert on criteria of consistency, coherence, relevance, and fluency. The optimized hybrid approach achieved a ROUGE-1 F1-Score of 0.76, significantly outperforming the fine-tuned LLaMA-2 model (ROUGE-1 F1-Score: 0.11) and strong baselines. Human evaluation confirmed that the hybrid model produced more relevant and consistent summaries. Our work demonstrates that a carefully engineered hybrid pipeline can be more effective than a directly fine-tuned large language model for the specialized domain of legal text summarization.</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

Legal text summarization with optimized hybrid models and fine-tuned LLaMA-2

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月17日

查看全文
  • Dharil Patel,
  • Shruti Patil,
  • Deepali Vora,
  • Aniket K. Shahade

摘要

Legal document summarization presents a critical challenge due to the dense, complex nature and extreme length of legal texts. This paper proposes and evaluates two distinct methods for this task. The first is an optimized hybrid model that combines TextRank for extractive summarization with a fine-tuned Facebook-BART-Large-CNN model for abstractive summarization. We then apply Bayesian Optimization to enhance the coherence and fluency of the final summary. Our second approach involves fine-tuning the LLaMA-2 (7B) model using Low-Rank Adaptation (LoRA) on a custom dataset of 7036 Indian Supreme Court judgments. We evaluated the summaries using ROUGE metrics and human evaluation by a legal expert on criteria of consistency, coherence, relevance, and fluency. The optimized hybrid approach achieved a ROUGE-1 F1-Score of 0.76, significantly outperforming the fine-tuned LLaMA-2 model (ROUGE-1 F1-Score: 0.11) and strong baselines. Human evaluation confirmed that the hybrid model produced more relevant and consistent summaries. Our work demonstrates that a carefully engineered hybrid pipeline can be more effective than a directly fine-tuned large language model for the specialized domain of legal text summarization.