Basis is also explanation: Interpretable Legal Judgment Reasoning prompted by multi-source knowledge

Abstract

The task of Legal Judgment Prediction (LJP) aims to forecast case outcomes by analyzing fact descriptions, playing a pivotal role in enhancing judicial system efficiency and fairness. Existing LJP methods primarily focus on improving representations of fact descriptions to enhance judgment performance. However, these methods typically depend on the superficial case information and neglect the underlying legal basis, resulting in a lack of in-depth reasoning and interpretability in the judgment process of long-tail or confusing cases. Recognizing that the basis for judgments in real-world legal contexts encompasses both factual logic and related legal knowledge, we introduce the interpretable legal judgment reasoning framework with multi-source knowledge prompted. The essence of this framework is to transform the implicit factual logic of cases and external legal knowledge into explicit basis for judgment, aiming to enhance not only the accuracy of judgment predictions but also the interpretability of the reasoning process. Specifically, we design a chain prompt reasoning module that guides a large language model to elucidate factual logic basis through incremental reasoning, aligning the model prior knowledge with task-oriented knowledge in the process. To match the above fact-based information with legal knowledge basis, we propose a contrastive knowledge fusing module to inject external statutes knowledge into the fact description embedding. It pushes away the distance of similar knowledge in the semantic space during the encoding of external knowledge base without manual annotation, thus improving the judgment prediction performance of long-tail and confusing cases. Experimental results on two real datasets indicate that our framework significantly outperforms existing LJP baseline methods in accuracy and interpretability, achieving new state-of-the-art performance. In addition, tests on specially constructed long-tail and confusing case datasets demonstrate that the proposed framework possesses improved generalization abilities for predicting these complex cases.