コンピューティング科学研究領域のNGUYEN, Minh Le 教授のチームが、The 12th Competition on Legal Information Extraction and Entailment (COLIEE 2025)のTask 1（判例検索）、Task 3（法令検索）、Pilot Task（司法判断予測）のサブタスクであるTort Prediction（判決予測）の3部門において、最高成績を達成しました。

　COLIEEは、判例や法令などの法的文書を対象に、AIによる情報検索や法的推論の精度を競う国際コンペティションで、司法手続きの効率化や、法制度へのアクセス向上を目的としています。第12回となる2025年大会は、令和7年6月20日に、アメリカ・シカゴのノースウェスタン大学（Northwestern University）にて開催されました。
　COLIEE 2025では、判例法に基づくTask 1・2、成文法に基づくTask 3・4に加え、日本の不法行為（民法第709条）を題材としたPilot Task（LJPJT25）の計5つの課題が出題されました。Pilot Taskは、日本の裁判所で実際に扱われた不法行為に関する判例データを活用し、研究者が新たな手法を開発・検証できる実験環境（サンドボックス）として、今回初めて導入されたものです。
　COLIEE 2025は、AIと法律分野における主要な国際会議であるICAIL 2025（International Conference on Artificial Intelligence and Law）と併催されました。

※参考：COLIEE 2025
　　　　ICAIL 2025

■受賞年月日
　令和7年6月20日

【Task 1・3】
■チーム名・メンバー
JNLP：Hai Nguyen, Hiep Nguyen, Trang Pham, Minh Nguyen, An Trieu, Dinh-Truong Do, Nguyen-Khang Le, Le-Minh Nguyen

■論文のタイトル
Hybrid Large Language Model-based Framework for Legal Information Retrieval and Entailment

■手法の概要等
This paper presents the JNLP team's approaches for the COLIEE 2025 competition, addressing all four legal information processing tasks: case law retrieval, case law entailment, statute law retrieval, and statute law entailment. Our systems leverage a hybrid framework that synergistically combines classical information retrieval (IR) pipelines, fine-tuned Transformer-based models, and instruction-tuned Large Language Models (LLMs) for deep legal reasoning. For case law retrieval (Task 1), we enhance a proposition-based ranking model by integrating lexical and structural-semantic features. For case law entailment (Task 2), we adopt a two-stage pipeline: we fine-tune re-rankers with hard-negative sampling and refine predictions using few-shot prompted LLMs. In statute law retrieval (Task 3), we implement a three-stage pipeline consisting of embedding-based pre-retrieval, LoRA/QLoRA-based fine-tuning, and model ensembling. For statute law entailment (Task 4), we explore zero-shot, few-shot, and reasoning ensemble prompting using models like Qwen2-72B to generate well-justified yes/no answers. Experimental results show that our methods achieve top-tier performance across multiple tasks in the official COLIEE 2025 evaluation. Our findings highlight the practicality and effectiveness of integrating lightweight IR models with large-scale LLMs for high-stakes legal NLP applications.

　　　

　　　

【Pilot Task】
■チーム名・メンバー
CAPTAIN：Dat Nguyen, Minh-Phuong Nguyen, Quang-Huy Chu, Son T. Luu, Nguyen-Hoang Chu, Trung Vo, Le-Minh Nguyen

■論文のタイトル
Enhancing Legal Text Processing and Structural Analysis with Large Language Models

■手法の概要等
The legal domain poses unique challenges for information extraction and reasoning due to the intricate structure and domain-specific language of legal texts. To address these challenges, our team, CAPTAIN, leverages recent advances in Large Language Models (LLMs) to enhance legal information processing within the scope of the COLIEE 2025 competition. We participate in four tasks: Legal Case Entailment (Task 2), Statute Law Retrieval (Task 3), Legal Textual Entailment (Task 4), and Legal Judgment Prediction for Japanese Tort Law (Pilot Task). Our approach harnesses the interpretive power of LLMs to analyze and summarize complex legal documents, uncover semantic relationships between legal cases and relevant statutes, and perform contextual reasoning. By leveraging diverse prompting techniques, our approach effectively uncovers implicit relationships between legal cases and their corresponding statutes, thereby enhancing both interpretability and accuracy. Experimental results demonstrate the strength of our method: it achieved first place in the Tort Prediction sub-task of the Pilot Task, and second place in both the Legal Statute Law Retrieval and Rationale Extraction sub-tasks, confirming the potential of LLM-based approaches in legal AI.


■受賞にあたって一言
COLIEE 2025 is a prestigious international competition in the field of Legal AI. This year, Nguyen Lab achieved the highest overall performance, securing first place in three out of five tasks.
These outstanding results reflect the dedication and talent of our students, whose efforts continue to advance our contributions to the Legal AI domain. We are deeply honored to receive this recognition and sincerely thank JAIST for providing an exceptional research environment that supports our work.
This achievement serves as a strong source of encouragement for our students and all members of the lab, inspiring us to pursue further research and strive for even greater accomplishments in the future.

令和7年8月20日