Sentiment Analysis and Political Party Classification in 2016 U.S. President Debates in Twitter

We introduce a framework of combining tweet sentiment analysis with available default user profiles to classify political party of users who posted tweets in 2016 U.S. president debates. The main works focus on extracting event-related information in short event period instead of collecting tweets in a long-time period as most previous works do. Our framework is not limited in debate event, it can be used by researchers to build rationale of other events study. In sentiment analysis, we show that all three Naïve Bayes classifiers with different distributions obtain accuracy above 75% and the results reveal positive tweets most likely follow Gaussian or Multinomial distributions while negative tweets most likely follow Bernoulli distribution in our training data. We also show that under unbalanced sparse term document setting, instead of using “Add-1” parameter, tuning Laplace smoothing parameter to adjust the weights of new terms in a tweet can help improve the classifier’s performance in targeted direction. Finally, we show sentiment might help classifying political party.