Convolutional Neural Networks and Long Short-Term Memory Networks for Textual Classification of Information Access Requests

The e-sic system aims to centralize requests for access to information addressed to the Brazilian Federal Executive. However, the volume of requests received can be an impediment to responses to those requests. The purpose of this article is to create an automatic classifier for these requests. For that, they were analyzed as architectures of the Convolutional Neural Network (CNN) and of Long Short Term Memory (LSTM), as well as a combination of these two architectures in order to identify the best architectures to this problem. The metrics used to evaluate the results were the area under curve roc and accuracy, and the error function used was cross entropy. The study concluded that the CNN network performed the best. Thus, the main contribution of this article is the identification of the most appropriate network architecture for classifying texts of interaction between citizens and government written in Portuguese.