Research On Sentencing Deviation Identification Method Of Trial Supervision

In December 2018,the Supreme People’s Court issued the "Implementation Opinions on Further Comprehensively Implementing the System of Judicial Responsibility",proposing to improve the supervision mechanism and punishment system,effectively implement the principle of "judges taking responsibility".Judicial supervision relates to the evaluation of the effectiveness of judicial operation,which is of great significance to ensure the smooth operation of society and maintain judicial authority.In the face of the digital transformation of social governance,promoting digital social governance through digital technology,has become an important means of the rule of law in the new situation.It is urgent to strengthen and improve criminal judicial supervision.Trial supervision and sentencing deviation identification method integrates legal logic and algorithm,does not take the existing judgment result as the fitting object,and minimizes the prediction interference brought by unfair cases.This method makes use of crime facts and sentencing circumstances,finds similar cases and constructs sentencing labels with the help of Lawk-means algorithm,and further identifies abnormal sentencing cases with the help of deep learning algorithm.The method of sentencing deviation identification in trial supervision explores the boundary between discretion and case justice on the basis of case retrieval,and provides reference for trial supervision.The main research work around the sentencing deviation identification thesis is as follows.1.Corpus construction for criminal case documents.First of all,split and re-spliced the judgment data,regular expressions are used to extract the elements for the experiment,namely,the contents of the paragraphs "charges by the procuratorate","confirmation by trial","Court’s opinion" and "Judgment is as follows".Then,the data is integrated and processed by removing redundancy and cleaning it up.Secondly,the integration of data to eliminate redundancy and impurity,cleaning processing.Further,determining the starting point of sentencing according to the facts of the crime,and dividing the data into different data sets.2.Label construction of similar cases based on "crime result-sentencing circumstances".Under different starting points of sentencing,in this part,the word vector of sentencing circumstances trained by Word2vec is used as input,adds the weighting matrix of sentencing circumstances for Lawk-means clustering as model,and outputs the labels of similar cases.3.Discovery of abnormal cases in sentencing based on text classification.In this part,crime facts and sentencing circumstances are taken as model input,and the similar case labels generated by the model are taken as text classification label values.The text classification models were trained by the traditional neural network model and the BERT-LSTM ensemble model,to output the prediction accuracy of the model and the abnormal sentencing cases.4.Analysis of case with abnormal sentencing.The purpose of legal analysis of cases with abnormal sentencing is not only to judge whether the model results are accurate or not,but more importantly,to provide reference for trial supervision through jurisprudential analysis,as well as to further improve the identification method of sentencing deviation in trial supervision.