Research On Crime Prediction Method For Case Documents

With the development of artificial intelligence technology and the disclosure of judicial data,the research and application of artificial intelligence in the judicial field have attracted extensive attention.As an important sub-task in the prediction of legal judgment,accusation prediction is an important part of judicial intelligent auxiliary system.The task of charge prediction predicts the charges the accused will be convicted of according to the description and facts of the case.Crime prediction is usually regarded as a text classification problem in the judicial field.Most of the existing methods use deep neural network to construct the crime prediction model,which has achieved good results in predicting common crimes.However,the existing researches pay less attention to the data imbalance in the crime prediction task,which leads to the poor prediction effect of low-frequency and easily confused crimes.In this paper,a crime prediction method based on the judgment document data is studied to improve the crime prediction method from two aspects of metric learning text classification and data enhancement.The main research work of this paper includes:(1)On charges of crime prediction of single case,this paper introduces the mean prototype network in the field of judicial charges to predict a particular task,this paper proposes a using mean prototype of network prediction method,the method combining measurement study and Bert build charges prediction model,the classification of the integration of all charges by moving average prototype vector.Compared with the baseline model,this method improved the predicted F1 value by 5.4%.(2)In order to solve the problem of category imbalance in crime prediction,this paper proposes a Mixup data enhancement strategy that integrates the priori information of category,which effectively improves the prediction performance of low-frequency and easily confused crimes.This method firstly using Bi-LSTM and structured the attention mechanism for case description and the fact of the text vector said,on this basis,through the Mixup data enhancement strategy in text vector representation space synthesis of pseudo samples,and use the category prior to synthesis the label samples to low-frequency charges category,to amplify the low-frequency charges the training sample.The experimental results show that,compared with the existing methods,the accuracy,macro accuracy,macro recall rate and macro F1 value of the proposed method are greatly improved,and the macro F1 value of low-frequency crime prediction is increased by 13.5%.(3)In this paper,a prototype system of accusation prediction is constructed based on the proposed prediction model.On the basis of the proposed crime prediction model,combined with the SANIC Web development framework and the Vue.js front-end framework,I developed a prototype system of crime prediction based on B/S framework.