Research On The Prediction Method Of Prison Sentence In Theft Cases Based On Inverse Deductive Learning

With the development of the national intelligent justice system construction,legal judgment prediction(LJP)had become a hot issue in the judicial intelligence research field.As a pivotal part of LJP,sentencing prediction is of great research significance and important application value to promote the “ Smart Court ”construction.Theft sentencing prediction has directly related to the amount of money involved and case elements,such as theft behavior,identity information of defendant et,and the sentencing process has strong rule and logic.According to the sentencing logic of theft,this paper predicts theft sentencing time by extracting case elements in judgment documents and the amount involved in money.Lack of domain labeled data and difficulty in data labeling leads to poor model performance,which are the main problems in the judicial field.However,many unlabeled data and domain knowledge bases(KB)are not fully utilized.Given the above problems,this work combines the sentencing prediction task with the abductive learning method and makes full use of the sentencing logic rules.Combine the element extraction model with logical reasoning effectively.Abductive learning is based on the principle of maximum consistency of the KB and makes full use of unlabeled data to abductive reasoning,revised errors in the element extraction model,improve the performance of case element extraction,and improve sentencing prediction effect.The main research work and achievements of this paper are as follows:(1)This paper proposed a theft sentencing time prediction based on a semi-supervised abductive learning model,predict theft sentencing time by extracting theft case elements and total money involved.Construct domain knowledge base.Based on the principle of maximum consistency of KB,minimizes the difference between the predicted sentencing and the actual sentencing,the abductive learning method is used to abduce and revise sentencing elements pseudo-labels generated by the supervised learning model,and re-train the model with revised pseudo labels of unlabeled data generated by abductive learning,to improve the performance of theft elements extraction,solve the problem of model poor performance due to lack of domain labeled data and poor quality of labeled data,improve the accuracy of sentencing prediction.The experiments show that the MSE of the model decreased by5.35,the accuracy of 3-month increased by 3.83%.(2)When the case elements increase,the try-error search space and try-error cost of the abductive reasoning will increase exponentially.Abductive reasoning algorithm speed drops and the accuracy is lower,and the part of the sentencing penalty model parameters are inaccurate.This paper proposes a theft sentencing time prediction based on a constrained abductive learning model given the above problems.The model constructs the abductive reasoning method with certainty-factor constrained of case elements,based on the principle of maximum consistency of KB and certainty-factor constrained,improve abductive reasoning performance and reduce try-error cost.We also Construct a constraint-penalty model,the parameters of the penalty model are restricted to make them consistent with actual applications.The experimental results show that the model’s 1-month sentencing prediction accuracy has increased by 19.41%.The F-value of our model is at least increased by 7.76%,compared with existing LJP methods.