Research On Judicial Intelligence Based On Deep Learning

The subject is the study of judicial intelligence based on deep learning,which mainly includes automatic sentencing,predicting related laws and recommending similar cases in the field of justice.The task solves the problem of judicial intelligence by the technology of deep learning.In the research process,the main data is criminal case which is single suspect and criminal offense.Automatic sentencing refers to predict the charge,prison term and penalty for given description of law case.In the experiment,Bag-of-Words,fast Text and convolution neural network model are used respectively,and predictive value transformation and digital discretization are compared to use in the task of prison term and penalty.The results show that the convolution neural network model is the best in predicting the charge,and the accuracy rate is 96.22%.The best result for prison term is the mean absolute error of 5.42 months,the mean absolute percentage error of 36.60% and the consistency rate of 43.04%.And the best result for penalty is the mean absolute error of 5199 yuan,the mean absolute percentage error of 52.36% and the consistency rate of 34.06%.Predicting relevant laws is find the related provision info for given description of law case.In the experiment,a variety of experimental ideas were tried,such as comparing the text of law,multi-label classification and predicting by similar cases.We also attempt the model of merging more information,such as the result of automatic sentencing in charge and elements of cases.Among them,the best result is multi-label classification model with more information,which is the mean coverage at 5 of 92.34%,the macro precision rate of 89.43%,the macro recall rate of 87.02%,the macro F1-measure of 88.21%,the micro precision rate of 88.08%,the micro recall rate of 84.23% and the micro F1-measure of 86.11%.Similar case recommendation refers to recommend parts of similar cases in existing case library by calculating the text similarity for given description of law case.In the study,we try the methods of term frequency-inverted document frequency,doc2 vec and the combination of term frequency-inverted document frequency and word2 vec,and the combination of term frequency-inverted document frequency and word2 vec is the best.For evaluation,avg-DCG@5 is used as the evaluation index.The best result is 18.51.