Fine-tuning Large Language Models for Improving Factuality in Legal Question Answering

2025-01-11

AI Summary

To address severe hallucination and frequent factual errors in large language models (LLMs) for legal question answering, this paper introduces LegalHalBench—the first domain-specific benchmark for evaluating legal hallucinations—and proposes Hard-sample-aware Iterative Direct Preference Optimization (HIPO), the first framework integrating behavior cloning, hard-sample identification, and iterative DPO to enhance semantic alignment with legal texts and enable automatic hallucination pattern detection. Compared to strong baselines, HIPO significantly improves non-hallucinatory statute recall, statute relevance, and factual accuracy of legal claims. It achieves state-of-the-art performance across METEOR, BERTScore, ROUGE-L, and human preference win rates. Our core contributions are: (1) LegalHalBench, the first dedicated legal hallucination evaluation benchmark; and (2) HIPO, the first fine-tuning paradigm explicitly designed for factual consistency in legal LLMs.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Hallucination, or the generation of incorrect or fabricated information, remains a critical challenge in large language models (LLMs), particularly in high-stake domains such as legal question answering (QA). In order to mitigate the hallucination rate in legal QA, we first introduce a benchmark called LegalHalBench and three automatic metrics to evaluate the common hallucinations when LLMs answer legal questions. We then propose a hallucination mitigation method that integrates behavior cloning and a novel Hard Sample-aware Iterative Direct Preference Optimization (HIPO). We conduct extensive real-data experiments to validate the effectiveness of our approach. Our results demonstrate remarkable improvements in various metrics, including the newly proposed Non-Hallucinated Statute Rate, Statute Relevance Rate, Legal Claim Truthfulness, as well as traditional metrics such as METEOR, BERTScore, ROUGE-L, and win rates.

Problem

• Large Language Models
• Legal Question Accuracy
• Information Fabrication

Innovation

• Imitation Learning
• HIPO Method
• Legal Information Accuracy