Automatic tweets classification under an intelligent agents framework

Twitter is a microblogging platform that allows users to share opinions with a restricted amount of characters. Given the social characteristic of Twitter, it is a potential source for sentiment analysis. For this reason, opinions of certain subjects such as people or brands can change in short periods of time. A traditional approach of a sentiment classifier implementation performs poorly since it depends on how it is trained. We propose a novel method for tackling this problem, with the implementation of a multi-agent system for classifying and corpus analysis mechanism for retraining the classifier. This mechanism consists of a critic agent which compares the trained corpus of the classifier agent with new collections of documents from future time steps, using primarily two methods: hypothesis test analysis and histogram differences. A Naïve-Bayes based classifier was implemented with this mechanism with multiple configurations. The results of experimental data show that the mechanism boosts its performance, when compared to a pure Naïve Bayes classifier.