Exploring the interpretability of legal terms in tasks of classification of final decisions in administrative procedures

Nodaways, diverse artificial intelligence techniques have been applied to analyse datasets in the legal domain. Precisely, several studies aim at predicting the decision to help the competent authority resolve a specific legal process. However, AI-based prediction algorithms are usually black-box, and explaining why the algorithm predicted a label remains challenging. Therefore, this paper proposes a 5-step methodology for analysing legal documents from the agency responsible for resolving administrative sanction procedures related to consumer protection. Our methodology starts with corpus collection, pre-processing, and TF vectorisation. Later, fifteen machine and deep learning algorithms were tested, and the best-performing one was selected based on quality metrics. Interpretability is emphasised, with the SHAP scores used to explain predictions. The results show that our methodology contributes to the understanding the decisive influence of legal terms and their connection to the decision made by the competent authority. By providing tools for legal professionals to make more informed decisions, develop effective legal strategies, and ensure fairness and transparency in the legal decision-making process, this methodology has broad implications for various legal areas beyond disputes, including administrative procedures like bankruptcies and unfair competition.