Construção e análise comparativa de métodos de classificação˜ para categorização das mensagens direcionadas ao ˜ “Fale Conosco” da SEFAZ-PB

Abstract

Every day people come up with questions or requests for organiza tions where they benefit from some type of service. Therefore, a service that connects the population with organizations is essential. This is the case of ser vices called “Contact Us”. It is extremely important that such services provide good service to the public, providing intelligent options and good user experien ces. In this work, the use of the “Contact Us” from the State of Para´ıba Finance Department was emphasized and seeking to improve the experience of using the tool, classify messages according to the topics related to it. Thus, this work analyzes and compares the Naive Bayes, decision tree and logistic regression classification methods in conjunction with message vectorization using Term Frequency-Inverse Document Frequency. In addition, a qualitative analysis of the model with the best performance was carried out against an unobserved data set. The method with the best performance analyzed was Naive Bayes with a balanced accuracy of 51.0% and based on this, it is possible to see that the mo dels are able to identify patterns and make good distinctions between the classes worked on, but there is still room for improvement in relation to the predictions made.