A Charge Prediction Model Based on Category Frequency Division

Charge prediction aims to charge a defendant through the case description. A fusion model to predict charges based on category frequency partitioning is proposed to address the serious imbalance of categories in charge prediction and the limited capture of text semantic features by existing models. A sample of CAIL2018(China AI and Law Challenge) was classified as low, medium-low, medium, and high frequency based on the frequency of the crime category, and a weighted stratified sampling method was adopted to construct three datasets. Dynamic vectors characterizing the semantics of the text using the pretrained RoBERTa model were used for each of the three datasets to capture the global contextual information. TextCNN, a convolutional neural network with a multi-window sliding mechanism, was used to extract the local semantic features of the text. Additionally, BiGRU, a bidirectional gated recurrent unit network, was used at the same time to capture the information of the textual sequences, combined with Attention to highlight the textual keyword information. The learning of feature representations for the three datasets was completed through multi-channel feature extraction. The outputs of the three parts were further fused by a feedforward neural network to complete the charge prediction. In experiments, the accuracy of the proposed model was 91.65%; the accuracy and Macro_F1 values were improved by 6.84% and 34.88%, respectively, compared with a BERT baseline; and the Macro_F1 value was improved by 1.49% compared with the suboptimal model. The proposed model improved the prediction accuracy of low-, medium-low-, medium-, and high-frequency offenses by 3.02%, 1.16%, 1.06%, and 0.5%, respectively, compared with the suboptimal model, effectively learning the semantic representations and improving the classification accuracy of low-frequency offenses, and providing references to achieve judicially assisted sentencing.