Research On Prison Term Prediction Method Based On Multi-Neural Network Models

Prison term prediction as one of the key techniques in the prediction of legal judgements,aims to predict the prison term based on the charge committed by the defendant and the severity of the sentencing circumstances.At present,most of the research on prison term prediction uses a single neural network to extract features of the case description text,ignoring the external factors that affect the prison term,reducing the accuracy of the prediction.At the same time,there is no convincing basis for the prediction process and prediction results,which reduce interpretability of predictions.This thesis conducts in-depth research on prison term prediction technology from the aspects of legal judgment prediction technology and neural network model,and explores the judge’s sentencing process in reality.It aims to analyze the case description text from different angles by using a variety of neural network models and provide a basis for the judgment result,thereby effectively improving the accuracy and interpretability of prison term prediction.Based on the analysis of legal judgment prediction technology,legal document vectorization,neural network model and other technologies,combined with the actual sentencing process,this thesis conducts in-depth research on the prison term prediction method,and proposes a prison term prediction framework based on multi-neural network models,which focus on the case feature extraction based on multi-neural network models and the prison term prediction based on case features and fully connected neural networks.First,the case description is encoded by the gated recurrent unit,and the case description features are obtained.Then the short fact summary representation in the case description is extracted by the reinforcement learning method,and we obtain the case fact summary features,which can provide explanations for the prediction results and improve interpretability of prison term prediction results.At the same time,based on the charge of the case and relevant legal provisions,the prison term is divided into the sentencing benchmark interval,which is in line with the actual sentencing process,improves the interpretability of the prison term prediction process.And we introduce the relevant information of the charge as an external knowledge,the attention mechanism is used to combine it and the case description features to predict the sentencing benchmark interval and obtain the sentencing benchmark interval features,which initially narrows the scope of the prison term,improves the accuracy of the prison term prediction.Next the case fact summary feature and the sentencing benchmark interval feature are spliced and integrated into a one-dimensional vector,which is input into the fully connected neural network to predict the prison term.Finally,experimental verification and comparative analysis of our prison term prediction method.Experimental results show that the method in this thesis can better capture the details of the case,and the prediction results have obvious advantages in terms of accuracy and interpretability.