Classificação de textos de decisões judiciais

Abstract

The great amount of legal documents produced in law courts makes it very inefficient and slow to be reviewed by humans. For example, courts produce millions of legal decisions per year in Brazil. In this paper, we propose an approach to automate the classification of court decisions. Specifically, we address the classification of sentences that contains judge decisions. We use machine learning and natural language processing (NLP) techniques to classify appeal decisions as granted, not granted, and partially granted. As a result, legal professionals do not need to read the case decisions – or even their summary (Ementa) – to know its decision. The data set used in this paper has 1,596 legal decisions from the State Supreme Court of Alagoas (Tribunal de Justiça de Alagoas). Moreover, we assessed the performance of five supervised machine learning models: Gaussian Naive Bayes, Decision Tree, Support Vector Machine, Random Forest, and XGBoost. The Decision Tree, Support Vector Machine, and XGBoost models achieved standout results with approximately an F1-score of 99%.