Leveraging Task Dependencies and Label Constraints for Legal Judgment Prediction

Legal Judgment Prediction aims to automatically predict the judgment results of multiple subtasks, such as legal article, charge, and term of penalty prediction. Practically, there exist dependency and constraint relationships among the above three subtasks. Existing work treats the three subtasks as sequential relationships but ignores mutual dependencies and consistency constraints between tasks, which leads to unreliable prediction results. Moreover, the result of term of penalty is closely related to the seriousness of the circumstance. However, existing research fails to model the impact of the seriousness of the circumstance on term of penalty. To address these problems, we present a sequential multi-task learning framework with task Dependencies and label Constraints for Legal Judgment Prediction (DCLJP). The DCLJP captures the multi-task dependency by the prediction forward-propagate mechanism over a directed acyclic graph and designs a novel loss function for obtaining the constraint relationships between the multi-task labels. In addition, we further adopt a reasoning method based on first-order predicate logic. It can drive the model to pay more attention to the correct term of penalty corresponding to the seriousness of the circumstance. The experimental results on two real-world datasets demonstrate that our proposed framework outperforms state-of-the-art methods on benchmarks.