An Explainable Hierarchical Class Attention Model for Legal Appeal Automation

Abstract

Judicial systems worldwide are overburdened due to the limited number of legal professionals. The digitization of legal processes has resulted in abundant legal data, paving the way for the development of legal automation systems that can assist the public and legal professionals. Legal Appeal Automation, a key problem in the legal domain, aims to automate the filing of legal appeals by using machine learning techniques to predict allegedly violated articles and provide supporting explanations based on the facts and provisions presented in the articles. Machine understanding of legal documents is challenging as they are typically lengthy, and effectively analyzing them is difficult. Further, providing explanations to justify the model predictions is complex yet crucial to building user confidence and trust in the model. Although solution approaches for predicting allegedly violated articles in legal cases have been proposed in the literature, to the best of our knowledge, no solution provides explanations justifying predictions. This absence of explanation generation is mainly due to the lack of datasets. To this end, we curate a new legal appeal automation dataset containing 9.8k instances of case- violated article pairs with explanations for each violated article. Using this dataset, we propose a novel neural architecture, Hierarchical Class Attention for Legal Appeal Automation, that efficiently handles long legal documents, predicts the allegedly violated articles and generates explanations justifying the predictions. We also introduce a baseline model for the new dataset and demonstrate that the proposed model outperforms the baseline. Using different multi- label classification datasets in the legal domain, we show that the proposed approach achieves state-of-the-art performance.