基于BERT与改进BP神经网络的盗窃案刑期预测方法研究
来源:用户上传
作者:郭彬彬
文章编号:2096-1472(2022)-02-06-03
DOI:10.19644/j.cnki.issn2096-1472.2022.002.002
摘 要:为了应对智慧法院项目中刑期预测任务的实际需求,提出了基于BERT与改进BP神经网络的刑期预测模型。以盗窃案为切入点,剖析相关案情要素,介绍刑期预测的整体框架和具体过程。基于大量真实案件数据,结合法官的审理流程,首先使用BERT识别裁判文书中的案情要素,然后基于规则抽取对应的涉案金额,最后使用改进的BP神经网络预测刑期,并与传统模型对比。实验证明,提出的模型刑期预测的平均误差小于2.5 个月,优于进行对比的传统模型。
关键词:神经网络;刑期预测;盗窃案件;BERT
中图分类号:TP39 文献标识码:A
Research on Prediction Model of Sentence for Theft based onBERT and Improved BP Neural Network
GUO Binbin1,2
(1. College of Computer Science and Technology, Guizhou University, Guiyang 550025, China;
2.State Key Laboratory of Public Big Data, Guiyang 550025, China)
gbb96@qq.com
Abstract: In order to meet the actual needs of sentence prediction task in the smart court project, this paper proposes a sentence prediction model based on BERT (Bidirectional Encoder Representation from Transformers) and improved BP neural network. Starting the theft cases, relevant case elements are analyzed, and the overall framework and specific process of sentence prediction are introduced. Based on a large amount of real case data and the judge's trial process, BERT is used to identify the case elements in the judgment documents. Then the amount of money involved based on the rules is extracted. Finally, the improved BP neural network is used to predict the sentence period and the proposed model is compared with the traditional one. Experiments show that the average error of the proposed sentence prediction model is less than 2.5 months, which is better than the traditional model used for comparison.
Keywords: neural network; sentence prediction; theft case; BERT
1 引言(Introduction)
智慧法院是我国2016 年提出的一项任务,主要目的是提高案件受理、判、执行、监督等环节的信息化水平,推动司法信息公开,促进司法公平正义。在当前的“智慧法院”建设中,刑期预测是其中一项重要任务,其核心目标是通过分析裁判文书中案情描述部分,自动预测出案件的判决刑期。该任务的实现结果可以作为补充来提高法官的审判效率,且促进同案同判。
早在20 世纪,自动法律判决预测就已经引起研究者的关注,这个时期主要是利用数学模型和统计学原理对法律案件进行分析。这种方法对专业性知识要求较高,且效率低下。随着人工智能的发展,研究人员逐渐将AI应用到法律方面。KATZ使用随机森林(Random Forest)从案情描述中提取有效特征对美国最高法院的判决结果进行预测。王文广等将层次注意力网络(Hybrid Attention Network, HAN)应用到刑期预测模型中,提出混合注意力和卷积神经网络模型(Hybrid Attention and CNN model, HAC)。谭红叶等采用多模型投票方法结合量刑属性进行刑期预测。ZHONG等认为法律审判的多个子任务之间存在依赖关系,提出了多任务拓扑依赖学习模型TOPJUDGE。YANG等基于多个子任务之间的拓扑依赖关系,引入词与词之间的组合语义关系,提出了多视角双向反馈网络MPBFN。以上这些模型均是基于分类的方法进行刑期预测,并没有给出最终的预测刑期;且不同类型的案件存在一定的差异,缺乏对某一类型案件的针对性。
转载注明来源:https://www.xzbu.com/1/view-15425116.htm