MANN: A Multichannel Attentive Neural Network for Legal Judgment Prediction

Recent years have witnessed an opportunity to improve trial efficiency and quality by predictive analysis of massive judgment documents. A practical legal judgment prediction (LJP) system should provide a judge with feasible judgment suggestions, including the charges, applicable law articles, and prison term, whereas most existing works focus on only part of the LJP task. Inspired by the impressive success of deep neural networks in a wide range of application scenarios, we propose a multichannel attentive neural network model, MANN, which learns from previous judgment documents and performs the integrated LJP task in a unified framework. In general, MANN takes the textual description of a criminal case as the input for attention-based neural networks to learn its latent feature representations oriented to the case fact, the defendant persona, and relevant law articles. Moreover, we adopt a two-tier structure to empower attentive sequence encoders to hierarchically model the semantic interactions from different parts of case description at both the word and sentence levels. The experiments are conducted on four real-world datasets of criminal cases in mainland China. The experimental results demonstrate that MANN achieves state-of-the-art LJP performance on all evaluation metrics.