JurisNexus Two robots face each other, one blue and one yellow, each with a speech bubble containing a question mark. The blue robot has a square head with antennae and a neutral expression, while the yellow robot has a round head and a smiling face. This illustration highlights a conversation or interaction between two different types of robots, possibly representing a question or uncertainty in communication between them. The contrasting colors and expressions emphasize their distinct identities and the theme of dialogue or inquiry. : Enhancing Legal Judgment Prediction via Cross-Reasoning-Chain Representation Learning Mechanism

Legal Judgment Prediction (LJP) methods seek to automatically make lawful judgments by analyzing and learning criminal features hidden in the fact description. However, accurately identifying and representing legal nuances in similar confusing crimes, e.g., fraud and robbery criminal charges, remains challenging. Inspired by the jury system, an equitable judicial process should harmonize the jury’s factual analogical-inductive retrieval among similar criminal cases, the plaintiff lawyer’s factual-reconstructive reasoning for ascertaining legal clues, and the veteran judges’ evidentiary-deductive reasoning. Therefore, we propose a legal representation learning framework (JurisNexus), which designs a novel cross-reasoning-chain mechanism to establish a specialized legal semantic space for better performance. Our cross-reasoning-chain mechanism involves: 1) Similar Case Reasoning Chain, which makes judgments based on analogies with similar precedents; 2) Legal Clue Reasoning Chain, which can help the general-domain language model to recognize and embed legal clues. 3) Legal Judgment Reasoning Chain, which can enhance the judicial association between criminal cases and legal judgment generation. Our pre-training strategy boosts JurisNexus’s accuracy by 4.6% on average, reflecting notable advances across various criminal scenarios. Experimental results indicate that JurisNexus effectively learns uniform and discriminative fact representations, resulting in more precise predictions of confusing crimes and a significant minimization of uncertainty. The implementation details will be open on GitHub.