基于深度条件依赖网络的裁判文书信息抽取

Information extraction of judgment documents based on deep conditional dependency network

裁判文书信息抽取旨在对裁判文书中包含的信息进行结构化处理,是法律人工智能研究的重要任务.目前的研究多是把裁判文书的信息抽取任务看成一个多标签文本分类任务,而没有考虑标签之间的依赖性.为此,提出了基于深度条件依赖网络的裁判文书信息抽取框架,用于裁判文书的案件要素信息抽取和争议焦点信息抽取.该框架由两部分组成,一是框架前端的特征提取网络用来挖掘裁判文书的文本特征,二是框架后端的标签关系网络来构建多个标签间的依赖性关系.在两个真实数据集上的实验结果表明,该模型在基线上都取得了显著且一致的改进,并且具有很好的扩展性,在该框架下可以获得多标签文本分类任务效果的提升.

As an important task for legal artificial intelligence,judgment document information extraction aims to structurally process the information contained in the judgment document.Most of the current research regards the task of judgment document information extraction as a multi-label text classification work,without considering the dependency between labels.To solve this problem,we propose a framework for information extraction of judgment documents based on deep conditional dependency network(DCDN),which is used for the extraction of case-element information and dispute-focus information of judgment documents.The framework is composed of two parts described below.First,the feature extraction network in the front of the framework is used to mine the text features of judgment documents.Second,the label relationship network in the back of the framework is used to construct the dependency relationship between multiple labels.We conduct experiments on two real datasets.Experimental results show that our model achieves significant and consistent improvements over baselines and has good scalability,thus improving the performance of multi-label text classification task under this framework.