Multi-scale Heterogeneous Graph Attention Network for Prison Term Prediction

Legal judgment prediction (LJP) is one of the most critical tasks in legal artificial intelligence. One of the most important subtasks is the prison term prediction, which aims to predict the term of the penalty based on the documents of the criminal case. Existing methods consider the problem of the prison term prediction either as a pure text classification problem or a pure text regression problem. The drawback of such approaches is that the accuracy of the prediction is largely affected by the imbalanced data which is due to the complicated distribution of the prison terms. Moreover, most existing methods fail to use the prior legal knowledge or fail to have a better representation of it that may enhance the interpretability of the methods. To address these issues, we construct a heterogeneous graph attention network (HGAT) with a multi-scale structure for the prison term prediction. Our method first takes sentencing elements as the input, which can be considered as the summaries of the information of the legal documents by legal experts. We then use the graph network as the basic module to represent the sentencing elements and to obtain the relationships between the groups of the sentencing elements. Most importantly, we propose a multi-scale structure, which essentially separates the prison term prediction problem into a classification problem of the time intervals and a regression problem of the accurate term. Since this structure decomposes the original complicated distribution of the prison terms into simple distributions, it effectively promotes the learning ability of the regression models and consequently improves the overall performance of the prediction. We conduct a series of experiments on the dataset of the crime of fraud in mainland China. The results demonstrate that our model can outperform most of the baselines and generate good robustness. In addition, in the classification problem, we find a good legal interpretation for the labels of the intervals when clustering the prison terms. This may enhance the interpretability of our method in the eyes of legal practitioners.