Research On Method Of Predicting Prison Term In Criminal Cases Based On Mixed Deep Learning Model

With the unceasing development of China’s economy and continuous improvement of the law,people have increasing demands for social fairness and justice.And as a result,judicial justice has now become the focus of public concern.However,due to differences in understandings and cognition of circumstances of crime by judicial workers,their different working experiences,and influences of culture or humanities in different regions,judges in China are relatively subjective in the standard of sentencing.This may lead to judgment results of great difference for similar circumstances of crime,which is regarded as legitimate "unfairness".But sentencing deviation will impose negative effects on credibility of judicial branch,and will weaken people’s trust in judicial justice to a certain extent.In order to further improve the impartiality of judicial trials,and promote the results of trial judgments towards the direction of "consistent judgments to similar cases",this paper aims at improving the difference score of prison term in judgment for criminal cases.After pretreatment of criminal judgment by data cleaning and word segmentation,it conducts vectorized representation of words using word2 vec.Then,features are extracted through deep learning model and the prison term is classified.The main research work of this paper is as follows.(1)The CNN(Convolutional Neural Networks)model based on word2 vec and the bidirectional LSTM(Long Short-Term Memory)model are used to classify the prison term in criminal cases.CNN can extract local correlated features of data,while LSTM can take advantage of context information when dealing with sequence problems,so they are mutually complementary.With regard to the long text and sparse information of written judgement,this paper utilized advantages of Attention mechanism to better identify the most influential keywords.For the above-mentioned reasons,a hybrid model of CNN-LSTM and Attention mechanism was applied to the prediction of prison term for the first time.Experiments in this paper proved that the CNN-LSTM-Attention hybrid model has improved the accuracy of classifying prison term and its difference score.(2)For the long text of written judgment,this paper combines text vector features constructed by doc2 vec to optimize the hybrid model,so as to preserve the overall information of the text.Because the distribution of prison term in criminal judgment papers is extremely uneven,the model is optimized by using weighted loss function according to the proportion of different classes of prison term.Through comparative analysis of the experimental results,it showed that doc2 vec can make full use of the context information,based on which the difference score of classifying prison term can be improved,and optimization of the model by weighted loss function is conducive to the application of prison term’s classification.