Research On Judicial Punishment Based On Deep Learning

Judicial workers need to read a lot of legal documents,and constantly update the legal knowledge system to ensure the accurate judgment of the case.However,there are still many problems in the judicial industry,such as more cases but less staff,different judgments for the same case,which lead to the slow progress of judicial penalty.In order to improve the efficiency of judicial penalty and reduce the burden of judicial workers to read a large number of legal documents,the judicial penalty model were constructed,and deep learning technology were used for the judicial penalty to assist judicial workers to complete the task of accusation prediction and recommendation of the article of the law in judicial penalty.Firstly,the data of the judicial penalty model is processed.The text data of legal documents comes from the CAIL(China judicial Artificial Intelligence challenge of Law research cup).Firstly,through the analysis of the given data set,it is concluded that the data set is unevenly distributed and there are many confusing accusation.By oversampling small categories of charges,the distribution of charges in the dataset is adjusted,and key features of element dimensions are extracted to distinguish easily confused accusation.Using Regular Expression technology to extract the key contents of legal documents,such as fact description,defendant's accusation,relevant laws of article and regulations,and establish JSON(Java Script Object Notation)format documents;Using stuttering segmentation and stop words list to stop words;Word2Vec(Word to Vector)algorithm is used to express the text as vector.Secondly,establish the prediction model and optimization model of judicial penalty.Through the comparative analysis of a variety of text classification models,the text classification effect of Tex CNN(Text Convolutional Neural Networks)model is excellent,so the judicial penalty prediction model based on Text CNN is established.After the evaluation,the accuracy rate of the prediction model is 77.16%,the precision rate is 77%,the recall rate is 77%,and the F1 value is 76%.Based on the prediction model,the optimization model is established by adding Gru-Attention network,and the loss function of focal loss was added to the output layer of the optimization model.The weight of accusation is adjusted by the loss function to solve the problem of uneven distribution of accusation at the model level.The experimental results show that the performance of the model is improved by 1.82%,0.45% and1.62% compared with the cross entropy loss,The final accuracy rate of the optimized model is 84.78%,the precision rate is 87%,the recall rate is 85%,and the F1 value is85%.Finally,a judicial penalty system was established based on deep learning.The system takes the optimization model as the core,designs and implements the core function modules of accusation classification,accusation classification document generation and so on.Through the functional test of the system,the availability of the system is verified.The system is expected to assist judicial workers to classify the accusation of judicial documents,and reduce the burden of judicial workers to read a large number of legal documents.