Classification of Special Civil Court judgments using machine learning

Abstract

The paper exposes a case study conducted at the Special Civil Court of the Federal University of Santa Catarina, aiming to employ different Machine Learning techniques to classify judgments on failures in the air transport service (Consumer Law) in four labels: “well founded”, “partly founded”, “not well founded” and “without prejudice”. Two experiments were performed, one with the full text of the judgments and other removing the dispositive part, the text that represents the procedural result (the label). The accuracy obtained by the classifiers in the second experiment was minimally reduced. In general, the models obtained by Logistic Regression, ANN and Random Forest achieved higher performance for the “well founded”, “partly founded” and “not well founded” labels, while for the “without prejudice” label, whose sample is smaller, the highest performance was achieved by SVM with RBF Kernel. Orange, version 3.22, an open source software running on Python, was used as tool.