Legal Judgement Prediction via Contrastive Learning Based-Retrieval and Semantic Embedding

Legal judgment prediction is a significant task in Legal Artificial Intelligence, aiming to predict judgment results based on fact descriptions automatically. The judgment output may include law articles, charges, prison terms, or combinations. Relevant legal cases with similar backgrounds or events should likewise be considered essential references for judgment, as they play an indispensable role in evaluating the prediction task. However, when applied to Chinese legal data, current methodologies often neglect the specialized preprocessing required for unique semantic attributes of Chinese, which affects their prediction performance. This paper proposes a novel contrastive learning- based model for legal judgment prediction named CLR-LJP to improve prediction performance. According to the legal judgment process in real life, law article prediction is converted as a multi- label classification task that includes both relative law cases and law articles. The CLR-LJP firstly converts fact descriptions, law articles, and charges into vectors by embedding Chinese semantics. Then, contrastive learning is introduced in retrieval training to distinguish the similar content of different label definitions. Finally, the role embedding method is applied to learn the interactions among multiple law articles to predict prison terms. Experiment results indicate that our CLR-LJP model achieves the best performance compared to the state-of-the-art models, especially in the task of prison term prediction.