Abstract

The purpose of this paper is to propose an algorithm for automatic text classification, as an alternative for those traditionally used for this task. The proposed classifier considers dependence between predictor variables (words or terms), an approach ignored by traditional classifiers. The dependence between predictor variables is captured as links of co-ocurrent words networks, objects that are used for training the classifier and also estimate the category of an unknown text. The results obtained from the automatic sentiment classification of more than 1,000 Twitter messages in positive, negative or neutral categories, and considering different context (topics), show that the proposed classifier, besides being a novel proposal, performs well compared to other algorithms traditionally used in automatic text classification such as Support Vector Machines or algorithms based in Bayesian statistic.