Research On Classification And Prediction Of Judgment Documents Based On Legal Elements

The classification of judgment documents is to deal with the prediction task of law,crime and punishment by classification,according to the case description in the judgment documents.As a part of intelligent justice,the classification and prediction of judgment documents plays an important role in improving the efficiency of legal decision-making and helping people understand the trial mechanism.The difficulty of this study lies in the particularity of the case and the label itself.For example,a case may involve more than one crime and law,there is a dependency between the law and the crime,and the punishment is highly related to the crime,law and other elements corresponding to the case.In addition,considering the principle of "fairness,impartiality and openness" of case trial,the classification prediction model of judgment documents should also be understandable.In view of the above problems and characteristics,the processed CAIL2018 dataset is used to conduct research in this thesis.The data is the judgment documents of public criminal cases,involving 53 articles of law and 70 charges.Based on this dataset,we propose two models in this thesis.Firstly,based on the neural network framework and attention mechanism,the hierarchical multi-label classification prediction model for articles and crime(HMLC-AC)is constructed,by introducing the criminal law original text to mine label information.In this model,both law prediction and crime prediction are regarded as multi-label classification tasks,and carried out in combination.Secondly,in order to explore the specific relationship between the punishment and the elements of cases,we extract the elements of cases with the help of Kmeans clustering method,then constructs the trial model based on the proposed elements(TC-E)with Light GBM,and carries out the structural analysis of sentencing decision with the help of SHAP value.In order to verify the feasibility of the models,classification indicators such as accuracy,Hamming loss and clustering indicators such as Silhouette Coefficient and Calinski-Harabaz Index are used to test the effect of the models.The results show that model HMLC-AC performs well in the two tasks of law article prediction and crime prediction,effectively extracts the key information from the original law,and excavates the corresponding relationship between law article and crime.At the same time,the trial model,constructed in this thesis,quantifies the impact of various elements on the final sentencing,on the basis of maintaining a good classification effect,so as to make the trial mechanism more clear.