On Classifying the Political Sentiment of Tweets

For this project, we attempted to classify the political sentiment of tweets containing the case-insensitive string ‘Obama’ in an effort to automatically gauge the public opinion of US President Barack Obama. To accomplish this goal we investigated rule-based, supervised, and semi-supervised learning methods. Our main approach involved bootstrapping an ngram-feature-based maximum entropy classifier using label propagation to exploit relationships between ngrams, tweets, and users. To evaluate each of our approaches to this problem we measured the percent accuracy on a test set of roughly 2, 500 labeled tweets. We also did cross-correlation analysis of time series produced by our sentiment predictions, from a larger test set of roughly 550, 000 unlabeled tweets, and Gallup poll’s data on President’s job approval.