A Relation Learning Hierarchical Framework for Multi-label Charge Prediction

In legal field, multi-label charge prediction is a popular and foundational task to predict charges (labels) by a case description (a fact). From perspectives of content analysis and label decision, there are two major difficulties. One is content confusion that the case descriptions of some charges are almost identical. The other is dynamic label number that the numbers of labels (label number) of different cases may be different. In this paper, we propose a relation learning hierarchical framework for multi-label charge prediction with two models, i.e., dynamic merging attention (DMA) and number learning network (NLN). Specially, DMA can improve the charge prediction performance by dynamically learning the similarity relation between a fact and external knowledge (provisions) and the difference relation between different provisions, which alleviates the phenomenon of content confusion. NLN mitigates the dynamic label number by learning the co-occurring relation between labels. Moreover, we put the two models into a unified framework to enhance their effects. Conducted on a public large real-world law dataset, experimental results demonstrate that our framework with DMA and NLN outperforms well-known baselines by more than 3%–23%.