WLEDD: Legal judgment prediction with legal feature word subgraph label-embedding and dual-knowledge distillation

Abstract

Legal judgment prediction(LJP) has achieved remarkable results. However, existing methods still face problems such as difficulties in obtaining key feature words for charges, which impose limitations on the improvement of prediction results. To this end, we propose a legal judgment prediction model with legal feature Word subgraph Label-Embedding and Dual-knowledge Distillation(WLEDD). Compared with traditional methods, our method has two contributions: (1) To mitigate the impact of overly sparse tail class data and high similarity text representations, we capture the critical features related to the charges by fusing LDA and legal feature word subgraphs. Then we encode them as label information to obtain highly distinguished representations of legal documents. (2) To solve the problem of high difficulty in some subtasks in LJP, we perform subtask-oriented compression of models to construct a student model with lower complexity and higher accuracy through dual knowledge distillation. Moreover, we exploit the logical association between the subtasks to constrain the labels of articles by charge prediction results. It greatly reduces the difficulty of article prediction. Experimental results on four datasets show that our approach significantly outperforms the baseline models. Compared with the state-of-art method, the F1 value of WLEDD for charge prediction has increased by an average of 2.57% . For article prediction, the F1 value has increased by an average of 1.09% . In addition, we demonstrate its effectiveness through ablation experiments and analytical experiments.