基于迁移学习的适用法条推荐模型

A model fro recommendation of applicable law articles based on transfer learning

摘要:

本文提出了一个基于法律事实的适用法条推荐模型.作为应用,本文选取了刑事诈骗罪和民事离婚纠纷两个案由的裁判文书数据集,使用FastText模型，并基于迁移学习方法从预训练的通用词向量出发训练法律词向量，再以此为基础进行文本分类.结果显示,进过迁移学习后,无论是诈骗罪案件还是离婚纠纷案件模型均能做到对案情描述文本全面准确地推荐适用法条,特别是具有针对性的法规、司法解释等.随着迁移学习模式的不断完善，本文研究的方法还应该可以进一步用于证据推送、量刑预测等.

Abstract:

In this paper, we propose a model for the recommendation of applicable law articles. As an application, the judgment documents data sets of fraud and divorce dispute are selected from the criminal and civil cases. Based on transfer learning, the legal word vectors are trained from the pre-trained general word vectors by using the FastText model. Then, the text is classified according to the well trained vectors. The simulation results show that for both the fraud and the divorce dispute, after the transfer learning, the applicable law can be recommended comprehensively and accurately for the case discription text, especially for the targeted regulations and judicial interpretations. With the continuous improvement of the transfer learning mode, our model is expected to be further applied to the evidence pushing and sentencing prediction.