Scholarlink

<p>Legal case entailment is a fundamental task in legal natural language processing that focuses on identifying entailment relationships between textual fragments extracted from court cases. This task poses particular challenges, involving lengthy and complex documents, cross-paragraph references, and a specialized linguistic style. These factors limit the performance of general-purpose textual entailment methods in the legal domain. To address these challenges, we propose <span>LexEntail</span>, a four-stage re-ranking framework that meticulously integrates the advantages of various advanced techniques, such as two-tower architecture, query-document interaction, late fusion strategies, LLM-as-reranker, and different voting approaches. Extensive experiments conducted on the COLIEE 2025 benchmark demonstrate the effectiveness and strong performance of our proposed framework, achieving state-of-the-art results in legal case entailment. Furthermore, we provide comprehensive analyses of each module within the framework and identify potential directions for further improvement. The full implementation and pretrained checkpoints will be publicly released to facilitate reproducibility and future research. The full implementation and pretrained checkpoints will be publicly released to facilitate reproducibility and future research.</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

LexEntail: A Multi-Stage Reranking Pipeline with LLM-Based Reasoning for Legal Case Entailment

  • 期刊论文

  • 发表日期：

    2026年3月11日

查看全文
  • Hoang-Trung Nguyen,
  • Tan-Minh Nguyen,
  • Thuong-Hieu Ngo,
  • Ha-Thanh Nguyen,
  • Thi-Hai-Yen Vuong

摘要

Legal case entailment is a fundamental task in legal natural language processing that focuses on identifying entailment relationships between textual fragments extracted from court cases. This task poses particular challenges, involving lengthy and complex documents, cross-paragraph references, and a specialized linguistic style. These factors limit the performance of general-purpose textual entailment methods in the legal domain. To address these challenges, we propose LexEntail, a four-stage re-ranking framework that meticulously integrates the advantages of various advanced techniques, such as two-tower architecture, query-document interaction, late fusion strategies, LLM-as-reranker, and different voting approaches. Extensive experiments conducted on the COLIEE 2025 benchmark demonstrate the effectiveness and strong performance of our proposed framework, achieving state-of-the-art results in legal case entailment. Furthermore, we provide comprehensive analyses of each module within the framework and identify potential directions for further improvement. The full implementation and pretrained checkpoints will be publicly released to facilitate reproducibility and future research. The full implementation and pretrained checkpoints will be publicly released to facilitate reproducibility and future research.