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 13th Competition on Legal Information Extraction and Entailment（COLIEE 2026）のPilot Task 1 - Tort Predictionにおいて、1位を獲得しました。

　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す。法律情報の抽出やテキスト含意認識など、AIと法律に関するさまざまな課題が設けられ、研究成果や手法が共有されています。第13回となる2026年大会は、令和8年6月12日に、シンガポール・マネージメント大学（Singapore Management University：SMU）にて開催されました。
　今回、NGUYEN教授のチーム（JNLP）は、Pilot Task 1 - Tort Predictionで1位を獲得したほか、Pilot Task 2 - Rationale Extraction、Task 1 - Legal Case Retrieval、Task 3 - Statute Retrieval and Entailment、Task 4 - Legal Textual Entailmentでそれぞれ2位を獲得し、さらにTask 2 - Legal Case Entailmentで3位となるなど、各タスクにおいて上位の成績を収めました。
　COLIEE 2026は、AIと法律分野における主要な国際会議であるICAIL 2026（21st International Conference on Artificial Intelligence and Law）と併催されました。

※参考：COLIEE 2026
　　　　ICAIL 2026

■受賞年月日 
　令和8年6月12日

■チーム名・メンバー
JNLP：Dang Le、Cuong Hoang、Duy-Minh Nguyen-Tran、Nguyen Khac Vu Hiep、Tan-Minh Nguyen、Do Minh Duc、Son T. Luu、Trung Vo、An Trieu、Nguyen Ngoc Minh、Dat Nguyen、The-Hai Nguyen、Trang Pham、Nguyen-Khang Le、Dinh-Truong Do、Vu Tran、Le-Minh Nguyen

■研究題目、論文タイトル等
JNLP at COLIEE 2026: Multi-Stage Hybrid LLM Framework for Case Law Retrieval, Statute Entailment, and Tort Prediction

■概要等
This paper presents the JNLP team's approach to addressing these challenges in the Competition on Legal Information Extraction and Entailment (COLIEE) 2026. We address all four main tasks: case law retrieval, case law entailment, statute law retrieval, and statute law entailment, alongside the pilot tasks. For Legal Case Retrieval (Task 1), we improve upon previous approaches by reducing the search space via BM25 and utilizing alternative classification models. For Legal Case Entailment (Task 2), we propose a two-stage architecture that combines dense retrieval with LLM-based reasoning to accurately identify entailing paragraphs. For Statute Retrieval and Entailment (Task 3), we employ a three-stage retrieval pipeline integrated with LLMs, incorporating dynamic few-shot prompting and a multi-agent debate framework. For Legal Textual Entailment (Task 4), we utilize Japanese-specific LLMs with few-shot prompting and self-consistency through majority voting to assess legal validity. Finally, we formulate both Tort Prediction and Rationale Extraction as binary classification tasks, leveraging LLMs to analyze the legal context and predict the target labels. Experimental results highlight the efficiency of our proposed approaches. Notably, we achieved the highest ranking on the pilot task, third place in Task 2, and second place across all other tasks.

■受賞にあたって一言
COLIEE 2026 is one of the world's premier competitions in Legal AI, bringing together leading research teams to advance intelligent legal reasoning and analysis. We are deeply honored that the Nguyen Lab won first place in Pilot Task 1 - Tort Prediction at COLIEE 2026. We sincerely thank JAIST for providing an outstanding research environment and continuous support for our research. We are also grateful to our students and collaborators for their dedication, hard work, and enthusiasm. This recognition inspires us to continue pursuing cutting-edge research and developing trustworthy AI technologies that contribute to the future of Legal AI.

令和8年9月10日