Data science supporting a question classifier model
Rafael Jardim, Carla Delgado, Daniel Schneider
Abstract
Deep Learning has been successful in developing e-learning applications, including engineering question classifiers. However, we are not aware of any classification model to date that recognizes the level of difficulty of questions in the Portuguese language. To overcome this challenge, this paper proposes a data-driven classifier model. A corpus was built with exam questions from the “Exame Nacional do Ensino Médio” in Brazil, whose hit level was validated by millions of students and a difficulty label was assigned to them. Then, a pipeline was designed, in which a Convolutional Neural Network (CNN) received the matrix of the question vectors and predicted the probability of the classes of difficulty, in the output layer. Besides, its performance was measured. As a result, a data-driven classifier was obtained that identifies the difficulty of questions.