摘要
法律判决预测指的是在给定案情描述的情况下,对案件的罪名和刑期进行预测。当前罪名预测主要使用深度神经网络模型,刑期预测主要使用混合深度神经网络模型。现有研究只关注局部特征或全局特征,没有考虑到将二者结合。因此本文使用双向门控循环单元提取上下文特征,并结合注意力机制学习文本中词的重要性,使用胶囊网络克服卷积神经网络丢失空间信息的缺点,学习文本局部与全局之间的关系特征。同时由于刑期分类时分类粒度较大,导致辅助决策结果不够理想。为实现更加理想的分类效果,将刑期按年进行更加细粒度的分类,分为28类。实验结果表明,该混合模型比基线模型效果更好,在罪名预测和28类刑期预测准确率分别为98.88%和74.32%。
Legal judgment prediction refers to predicting accusations and prison terms of criminal cases on the basis of case descriptions.At present,accusation prediction mainly uses deep neural network models to extract features,and prison term prediction is grounded on hybrid deep neural network models.Existing works either focus on global or local features,not considering their combination.Therefore,this paper proposes Bidirectional Gated Recurrent Unit(BiGRU)to extract context features and merge attention mechanism(Attention)to learn the importance of words in text.Then it combines with CapsNet,which can overcome the disadvantage that Convolution Neural Network(CNN)will lose spatial information.The combination makes it possible to learn the relationship between local and global text features.At the same time,due to the large classification granularity of sentence classification,the result of auxiliary decision is not ideal.In order to achieve a more ideal classification effect,the sentence is classified into 28 categories by year.The experimental results show that the mixed model is better than the baseline model,and the accuracy of the prediction is 98.88%and 74.32%respectively.
作者
刘璐瑶
李实
LIU Luyao;LI Shi(College of Information and Computer Engineering,Northeast Forestry University,Harbin 150006,China)
出处
《智能计算机与应用》
2021年第8期42-46,共5页
Intelligent Computer and Applications
基金
中央高校基本科研业务费资助项目(2572019BH03)。
