Defesa de Mestrado – Thiago Raulino Dal Pont – 28/7/2021
13/07/2021 21:43
| Defesa de Dissertação de Mestrado | |
| Aluno | Thiago Raulino Dal Pont |
| Orientador
Coorientador |
Prof. Jomi Fred Hübner, Dr. – DAS/UFSC
Prof. Aires José Rover, Dr. – CCJ/UFSC |
| Data | 28/7/2021 (quarta-feira) – 14h
Videoconferência (https://meet.google.com/xgu-jvoj-kct) |
| Banca | Prof. Jomi Fred Hübner, Dr. – DAS/UFSC (presidente);
Prof. Fabiano Hartmann Peixoto, Dr. – UnB; Prof. Eric Aislan Antonelo, Dr. – DAS/UFSC; Prof. Marcelo Ricardo Stemmer, Dr. – DAS/UFSC. |
| Título | Representation, Classification and Regresstion Techniques applied to Legal Judgments about Immaterial Damage due to Failures in Air Transport Services |
| Abstract: According to the last report Justiça em Números, annually published by the National Council of Justice, 77.1 million processes were waiting for a solution in the Brazilian judiciary, 5.2 million in the Special Civel Courts (JECs). Those numbers have been growing year after year, indicating the need of creating mechanisms to speed up the Brazilian Judiciary. Thus, this work aims to contribute to the improvement of the efficiency of the judiciary by applying Machine Learning (ML) and Text Mining (TM) techniques to the prediction of the results of the legal judgments from the JEC located at the Federal University of Santa Catarina, which relate to failures in air transport services. To do so, we divided the problem into three parts: representation, classification, and regression. In the first part, we evaluate whether the size and specificity of the corpora used to train word embeddings, impact the performance of text classification. We, thus, trained embeddings based on judgments in Portuguese. As a result, we discovered that size and specificity matter, however size influences the results until a certain point. In the second part, we evaluate whether Deep Learning (DL) techniques perform better than Classical ML techniques in the classification of the judgments’ results from JEC. Thus, we trained several DL and Classical ML techniques using two types of the dataset, one containing the full text of the judgments and another with the result part removed. In the former, the DL techniques performed better, implying that they can better assimilate the parts of the texts that explicitly indicate the result. In the latter, classical techniques performed better, indicating that without the explicit result part, those techniques can better learn from the other parts. In the third part, we focus on predicting the compensation for immaterial damage, using regression techniques. Based on several pipelines and on a legal expert’s evaluation, we noticed that the prediction quality achieved in such a task is acceptable and it can be helpful in the legal domain. Thus, we concluded that it was possible to accurately predict the results of the judgments from JEC and the compensation for immaterial damage using the proposed pipelines. | |
