The growing importance of social networks in our society leads governments and public institutions to privilege these networks, not only in communicating with their citizens, but also in the perception of citizens' opinion and degree of satisfaction with the services provided. By using text mining techniques, more specifically, of sentiment analysis or opinion mining, it is possible to extract useful information from social networks that allows the identification and monitoring of citizens' opinions. In this sense, we present in this article, the work developed, from tweets of the social network Twitter related to the Portuguese Tax Authority, with the purpose of exploring algorithms of text classification, that allow to identify the opinions, positive and negative, of the citizens. As a main contribution of this study, we highlight the discovery of a new approach, which consists in applying the Naive Bayes algorithm to a dataset generated from the use of the Lexicon-PT library on the original dataset, which was named LexiNB. This new approach showed a significant increase in the Kappa statistic and Balanced Accuracy.