融入法律条款的可解释罪名预测模型
Interpretable Model for Predicting Charges with Legal Provisions Incorporated
陈俊霖 1刘群 1李能1
作者信息
- 1. 重庆邮电大学计算智能重庆市重点实验室,重庆 400065
- 折叠
摘要
罪名预测旨在根据案件的相关事实和证据,对犯罪嫌疑人可能被指控的罪名进行推断和预测.随着人工智能技术的发展,利用深度学习方法进行罪名预测的模型已经展现了优秀的性能.然而,在司法领域,依靠深度学习实现的罪名预测模型,由于其背后的黑盒性质,容易受到法律专家的质疑,从而影响了预测结果的可靠性和可信度.本文提出了一种融入法律条款的可解释罪名预测模型.首先,通过文本匹配任务寻找犯罪事实适用的法律条款对刑事法律文书进行表示学习,该表示过程将事实描述和法律条款映射到相同的空间进行特征对齐,使隐空间中的犯罪事实围绕相关法条聚集.进一步将事实描述的表示进行聚类形成描述不同类型罪名的概念,并嵌入到预测模型的隐空间作为决策依据,指导训练过程.在罪名预测时,计算待测样本表示与概念之间的相似度,将相似度最大的概念作为决策依据生成预测结果.在CAIL2018数据集上的两组不同类型实验表明,本文模型预测准确率达到了 99.70％和93.60％,在分类准确率和模型解释性上均优于对比模型.
Abstract
Charge prediction aims to infer and predict the crimes that a suspect may be charged with based on the relevant facts and ev-idence of the case.With the development of artificial intelligence technology,charge prediction models using deep learning methods have demonstrated excellent performance.However,in the legal field,the charge prediction models based on deep learning are easily challenged by legal experts due to the black-box nature behind them,which affects the reliability and credibility of the prediction re-sults.This paper proposes an interpretable charge prediction model with legal provisions incorporated.First,the representation learning of criminal legal instruments is performed by matching the legal provisions with the crime facts through a text-matching task that maps the factual descriptions and the legal provisions to the same space for feature alignment,so that the crime facts in the hidden space are clustered around the relevant legal provisions.The representation of factual descriptions is then further clustered to form concepts de-scribing different types of crimes and embedded into the hidden space of the prediction model as a decision base to guide the training process.During the prediction,the similarity between the sample representation to be tested and the concepts is computed,and the con-cept with the greatest similarity is used as the decision base to generate the prediction results.Two different types of experiments on the CAIL2018 dataset shows that the model prediction accuracy reaches 99.70％and 93.60％,outperforming the comparison model in terms of classification accuracy and model interpretability.
关键词
法律文书表示学习/罪名预测/可解释性/文本匹配Key words
representation learning of legal documents/charge prediction/interpretability/text matching引用本文复制引用
出版年
2025
NETL
NSTL
万方数据