Legal judgment prediction via optimized multi-task learning fusing similarity correlation

The use of computer-assisted legal judgment prediction (LJP) is a current research hotspot, driven by advances in artificial intelligence technology. Previous LJP methods have mainly relied on feature models and emphasized parameter sharing within the coding layer, ignoring progressive sequential relationships between LJP subtasks as well as potential similarity correlations between cases. These limitations have hindered the improvement of the accuracy of LJP methods. This article proposes an LJP algorithm, called MTL-LJP, based on optimised multi-task learning that fuses similarity correlations. MTL-LJP uses EnMo to encode cases and SiMa to compute similarity matrices between cases. MuTa, which is based on multi-task learning, is used to predict LJP subtasks. EnMo vectorises case facts from multiple perspectives using encoders based on CNN, Bi-GRU, Bi-GRU with attention mechanism and MMoE. SiMa computes centroids and distance vectors based on historical case labels and LJP subtask predictions, allowing the computation of similarity matrices for each subtask with low computational complexity. MuTa predicts subsequent subtasks by using the intermediate results of previous subtasks through the forward auxiliary mechanisms. MuTa modifies predictions by correlating the similarity of cases. Experimental results on several real case datasets show that MTL-LJP outperforms previous methods on LJP subtasks.