Classificação de textos escolares com aprendizado de máquina e lógica Fuzzy

Abstract

Nowadays, during the teaching-learning process, it is prevalent for students to seek information about the subjects to complement studies or execute school work. For this, it is possible to use search engines on the Internet, but the texts do not always have the complexity compatible with the student’s level of education. A solution for this purpose is the automatic filtering of texts, where computational models are created for textual classification according to their contexts. Many studies have presented solutions based on Machine Learning or Fuzzy Logic techniques, but there is no comparison nor an analysis considering the possibility of combining both methods. This work explores this scenario, implementing two solutions: one based only on traditional machine learning algorithms and another combining the execution of a Decision Tree with the automation of a Fuzzy Logic system. In both solutions, Natural Language Processing techniques are applied in order to select characteristics of the textual contents and introduce them into the classification pipeline. Experiments were conducted and the results obtained suggest that, for the scenario explored, the Machine Learning algorithms have a superior result, with class-balanced accuracy of around 0.89 in the test data. Furthermore, the use of Decision Tree models to create the rules of a Fuzzy System could add more interpretability and understanding to the classification of school texts, but with a balanced accuracy of approximately 0.5, still superior to all the baselines chosen for the experiment of this work.