基于深度学习的司法判罚研究  

Research on judicial penalty based on deep learning

作　　者：高珊 何安娜 肖清泉 GAO Shan;HE Anna;XIAO Qingquan(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵州贵阳550025

出　　处：《电子设计工程》2022年第17期23-27,共5页Electronic Design Engineering

基　　金：国家自然科学基金(61264004);贵州省留学回国人员科技活动择优资助项目([2018]09);贵州省高层次创新型人才培养项目([2015]4015)。

摘　　要：基于深度学习的司法判罚研究有助于缓解司法行业案多人少以及同案不同判的问题,提高司法判罚效率。文章基于深度学习技术建立司法判罚模型,主要实现罪名预测功能。对司法文本数据进行分析以及处理,通过提取要素维度关键特征的方式区分易混淆罪名。建立基于CNN+GRU-Attention(Gated Recurrent Unit-Attention)的司法判罚模型,在模型输出层加入分组focal loss损失函数,解决罪名分布不均衡问题,根据模型评估指标进行模型验证以及对比分析。建立基于深度学习的司法判罚系统,系统以模型为核心,主要实现罪名分类功能,通过对系统的测试,验证了其有效性。The research on judicial penalty based on deep learning can alleviate the problems of many cases and different judgments in the same case and improve the efficiency of judicial judgment. Based on the deep learning technology,a judicial penalty model was established,which mainly can realize the function of crime prediction. The judicial text data were analyzed and processed,distinguishing the easily confused charges by extracting the key features of the elements dimension. The optimization model of judicial penalty based on CNN+GRU-Attention(Gated Recurrent Unit-Attention) was proposed. The group focal loss function was added to the output layer of the model to solve the problem of imbalance of crime distribution. The model was verified and compared with the evaluation index of the model. The judicial penalty system based on deep learning was established. The system takes the model as the core,and mainly realizes the function of charge classification. Through the test of the system,its effectiveness is verified.

关 键 词：深度学习 司法判罚 卷积神经网络 GRU-Attention网络 

分 类 号：TN915.03[电子电信—通信与信息系统]