摘要：

案情要素识别旨在根据案情描述文本识别该案件涉及的重要情节及事实。现有工作针对任务案情要素之间存在标签关系的问题，采用融合标签信息的多标签文本分类方法和基于反绎学习(ABL)的方法予以解决，但前者存在监督信息不足导致模型学习不充分的问题，后者存在先验知识的局限性，并且标签关系为某些案情要素提供的监督信息，较为片面。为此，提出融合标签关系与法条逻辑的案情要素识别方法。该方法从数据样本和领域知识两个来源出发，分别挖掘和利用标签关系和法条逻辑两类外部知识，为识别过程提供约束和引导信息，使模型能更加准确地提取案情描述文本语义特征，从而提升识别结果性能。具体来说，该方法首先利用基于Transformer的双向编码器表示(BERT)作为编码器，提取案情描述文本表征得到上下文编码，再通过全连接(FC)层得到原始预测结果；然后设计标签关系增强网络(LRE-Net)学习和利用案情要素之间的标签关系，对原始预测结果进行修正和增强，以突破标签关系作为先验知识的局限性；最后引入一阶逻辑对法条逻辑进行形式化表达，并根据一种一阶逻辑表达到连续数值计算的映射关系，构建了法条逻辑约束下的损失函数，为模型提供额外的监督信息。实验结果表明，在标签关系与法条逻辑的共同约束与相互增益下，该方法在F1值上相较于基线模型提升了2.41百分点，相较于最优对比模型提升了1.21百分点。

关键词: 案情要素, 多标签文本分类, 反绎学习, 一阶逻辑, 法条逻辑

Abstract:

Case element recognition aims to recognize the important sequence of events and facts involved in a case based on the case description text. Considering the task characteristic that label relations exist between case elements, existing studies use either multilabel text classification methods that fuse label information or methods based on Abductive Learning (ABL) for recognition; however, the former suffer from insufficient supervised information, leading to inadequate model learning, whereas the latter suffer from the limitation of priori knowledge. Additionally, the supervised information that label relations provide for some elements is one-sided. To address these issues, a case element recognition method integrated with label relations and logic of legal articles is proposed. This method mines and utilizes two types of external knowledge, namely, label relations and logic of legal articles, from two sources, data samples and domain knowledge, to provide constraints and guidance information for the recognition process, enabling the model to extract semantic features from the case description text more accurately. This improves the recognition result performance. Specifically, the method utilizes Bidirectional Encoder Representations from Transformers (BERT) to extract the textual representation of case descriptions and then obtains the original prediction results using a Fully Connected (FC) layer. Subsequently, to overcome the limitations of prior knowledge, a Label Relation Enhancement Network (LRE-Net) is designed to learn and utilize the label relations for correcting and enhancing the original prediction results. To provide additional supervised information for the model, first-order logic is introduced to formalize the expression of logic of legal articles. Additionally, the loss function is constructed under the constraints of this logic according to the mapping from first-order logic expressions to continuous numerical computation. Finally, under the joint constraints and mutual benefits of label relations and legal logic, the proposed method improves the F1 value by 2.41 and 1.21 percentage points compared to those of the baseline model and the optimal comparison model, respectively.

Key words: case element, multi-label text classification, Abductive Learning (ABL), first-order logic, logic of legal articles