LePREC: A Neuro-Symbolic Framework to Solve Legal Problem Recognition Challenges, Improving Accuracy by 30-40%

LePREC is a neuro-symbolic reasoning framework inspired by legal professionals. It combines large language model (LLM) generation of structured analytical factors with sparse linear model classification. Through interpretable feature weight learning, it significantly improves the accuracy of legal problem recognition, outperforming advanced baselines like GPT-4o and Claude with a 30-40% accuracy increase. This framework addresses the accuracy deficit caused by the "black box" nature of existing end-to-end neural networks, providing a new path for legal AI to move toward trustworthiness.

Legal problem recognition is the starting point of legal analysis and crucial for AI systems. However, the accuracy of existing top LLMs (e.g., GPT-4o) is only 62%, with high error rates. The root causes include: the "black box" nature of end-to-end neural networks, which prevents explicit verification of reasoning processes; LLMs generate candidate problems with "breadth but lack of precision", mixing irrelevant or incorrect content; and the complex hierarchy and cross relationships of legal concepts further exacerbate the challenge. Amid the global civil justice demand dilemma, AI needs to break through this "last mile".

LePREC consists of neural and symbolic components. The neural component uses multi-turn prompts to guide LLMs to generate question-answer pairs (structured analytical factors) for case facts, making the reasoning process explicit; it adopts multi-sample generation and aggregation strategies to ensure comprehensiveness. The symbolic component applies sparse logistic regression classification, learning explicit feature weights via L1 regularization, retaining only predictive features to enhance interpretability and generalization ability.

On the Malaysian contract law dataset, LePREC's accuracy is 30-40% higher than GPT-4o, with precision increasing from 62% to over 85% and F1 score from 0.71 to 0.89. The improvement comes from structured factor extraction enhancing recall quality and sparse classification filtering noise. Feature weights align with the logic of human legal judgment, verifying interpretability. Its stability on cross-year test sets is better than pure neural network methods.

The success of LePREC proves that neuro-symbolic fusion is the key for legal AI to break through bottlenecks. Its advantages include interpretability (explicit weights support decision transparency), data efficiency (good generalization with small samples), and debuggability (modular design makes error localization easy). This technology is expected to democratize access to legal services, drive legal AI to leap from "usable" to "trustworthy", and assist human professionals in improving efficiency.

LePREC is currently mainly applicable to the field of contract law; its applicability to other legal fields needs to be verified. Factor extraction relies on the quality of LLM prompts, so more robust methods or human-machine collaboration need to be explored. It does not involve multi-case comparison and case retrieval, so the end-to-end workflow needs to be expanded. Decisions are based on statistical correlation, so legal deductive rules need to be integrated to enhance rigor.