The Research On Legal Judgment Prediction Based On Deep Learning

In recent years,in order to improve judicial efficiency and promote judicial fairness,China has vigorously promoted the construction of smart justice.Legal judgment prediction is an important application of artificial intelligence technology in the field of judiciary,which usually includes three subtasks: crime prediction,law article prediction,and sentence prediction,among which crime prediction and law article prediction belong to multi-label classification task.Existing research mostly focuses on predicting judgment results based on case description information,without fully utilizing the existing criminal constitution information in the law.This is obviously different from the thinking process of human judges.In reality,human judges not only understand the case description information but also refer to relevant legal texts,especially the criminal constitution information,during the process of hearing cases.Furthermore,the current research on task relationship modeling is relatively scarce,and these works often model the tasks in the order of predicting crime,then law article,and finally sentence.In practice,human judges determine the crime and law article involved in a case based on the case description and criminal constitution information,and then determine the sentence for the defendant.Lastly,the legal judgment prediction task mainly encodes the case description.The case description text contains a large number of specialized terms in the field of judiciary and has an obvious hierarchical structure.How to extract relevant information effectively from the case description directly affects the accuracy of prediction.To address these issues,two new multi-task legal judgment prediction frameworks are proposed:(1)Considering the relationship between subtasks and criminal constitution information,a multi-task judgment prediction model fusing crime constitution(MJP-FCC)is proposed.Considering the close relationship between the crime prediction task and the law article prediction task,they are regarded as a whole prediction task and jointly modeled with shared parameters.To fully utilize the criminal constitution information to assist the task of crime and law article prediction,keywords are extracted from the criminal constitution and incorporated into the model.In order to learn the dependency features of tasks and improve the performance of sentence prediction,the output of the overall prediction task is input into the sentence prediction task.Experimental results show that incorporating criminal constitution keywords into the model can effectively improve the performance of crime and law article prediction.When the outputs of predicting crime and law article are added in the sentence prediction,it can effectively improve the prediction accuracy of sentence.(2)Considering the presence of judicial specialized terms and text hierarchical structure in the case description,a multi-task judgment prediction based on the Lawformer pre-trained model(MJP-Law)is proposed.Lawformer is used to embed the case description text,which can transfer learned legal prior knowledge into the model and generate dynamic word vectors with semantic information.Case description text is a long text that contains multiple types of information elements and has a typical hierarchical structure.Therefore,a hierarchical attention network(HAN)encoder is adopted to encode the case description representation from word level to sentence level and learn the representation with multiple hierarchical information.To assist the judgment prediction,the crime constitution information related to the case description is encoded and fused into the model.To emulate the judicial trial process,the outputs of predicting crime and law article are separately used as inputs for the sentence prediction task to assist in sentence prediction.Through experimental comparison,MJP-Law model outperforms all baseline models in crime and law article prediction,and is second only to the MJP-FCC model proposed in this paper in sentence prediction,which verifies the effectiveness of the model.