Abstract: Legal rules are behavior norms formulated by competent authorities with binding legal force and effectiveness,essential to maintaining social order.As the preliminary of Legal AI,numerous studies have attempted to convert natural-language legal texts into machine-readable rule sets,however the results remain unsatisfactory.To address the challenges of formal representation and extraction of legal rules,this paper proposes a systematic legal rule extraction paradigm and introduces a negative-aware approach to boosting generative rule extraction with large language models(LLMs).The paradigm defines a legal rule schema by decomposing a legal rule into a tetrad of subject,object,conditions and consequences,thus clarifying its applicability,targets and effects.Building on this,this paper proposes a generative legal rule extraction enhancement method leveraging LLMs,which incorporates the concept of “learning from errors” by constructing a negative-aware training framework to improve the model’s ability to recognize hard negative cases and mitigate hallucination issues in generative rule extraction.Experimental results show that the rule extraction model based on Mistral-Small-24B(a mid-size LLM) outperforms the general-purpose LLM(Deepseek-r1) by 18.23% and even surpasses human-annotated performance by 1.5%,demonstrating that the negative-aware training framework significantly enhances the rule extraction capability of the model.

Key words: Rule extraction, Negative-aware training, Large language model, Corpus construction, Supervised fine-tuning