Case Study of Criminal Law Based on Multi-task Learning

摘要

With the development of natural language processing technology and the advancement of judicial intelligence, the analysis of legal documents and the prediction of case decisions are attracting more and more attention. In this paper, the analysis of criminal case judgment is divided into four subtasks: recommendation of relevant laws, prediction of charges, prediction of sentences and query of similar cases. Based on the characteristics of the civil law system, we recommend relevant laws as the core subtask, use the method of multi-task learning to conduct joint training with the task of crime prediction, and integrate the results into the task of sentences prediction and similar case query. At the same time, in order to better understand the description of the case and enhance the interpretability of the results, we extract and integrate the sentencing rules according to the relevant laws. The experimental results show that the model that incorporates rules and shares specific generic knowledge for training has better performance than the single subtask model.