Legal judgment prediction via dynamic task graph and keyword-aware contrastive learning

Legal judgment prediction (LJP) aims to predict judgment results based on the factual descriptions of criminal cases and is gradually becoming a popular research topic in the legal field. Case factual descriptions contain a large amount of keyword information, and there are complex dependencies among the subtasks. For example, law article prediction can guide charge prediction and term of penalty prediction. However, most of the previous methods often failed to take advantage of the effective keywords that are commonly present in case descriptions, or failed to consider the interrelationships between sub-tasks. To address the shortcomings, we propose a multi-task legal judgment prediction framework based on Dynamic Task Graph and Keyword-aware Contrastive Learning, termed DTG-KCL. Specifically, we design a sample-level method of dynamic task relationship graph convolution to encode task relationships from the perspective of each input case description. Subsequently, we enhance the ability to distinguish confusing charges by developing a more profound and novel keyword fusion mechanism based on soft keyword representation and an approach of keyword-aware contrastive learning. Our model effectively improved the performance of the sub-tasks. Experimental results on two real legal datasets demonstrated that our model had a significant advantage over the five methods in terms of prediction performance. Code can be obtained from: https://github.com/YiLisiForever/DTG-KCL.