Sentiment Analysis of Twitter Posts About news
G. Gebremeskel Gebrekirstos · 2011
The thesis set out to solve a practical problem of sentiment analysis of Twitter posts about news. The thesis has made contributions is data collection of tweets about news, empirical study of the role of context in sentiment analysis of tweets about news, and best feature selection. Test data was collected by a bootstrapping approach where some tweets that contain some number of words from the news headline were used to extract links and to obtain more tweets about news from Twitter. The test data was manually inspected and annotated for its sentiment to see if context plays role in determining sentiment of a tweet. Uni-gram+bi-gram was selected as a feature that captures two important features of the data: uni-grams provides better coverage of the data, and bi-grams capture sentiment expression patterns. The thesis has shown that tweets about news can be automatically collected and successfully analyzed for their sentiment. Multinomial Naive Bayes classifier using uni-gram+bi-gram presence was found to give the highest accuracy, an accuracy of 87.78% for a three-classed classification and an accuracy of 90.79% for the two-classed (subjective, objective) classifier derived from the three-classed classifier. The accuracies of the classifier on both three-classed and two-classed classification is impressive and can be applied for practical applications dealing with sentiment analysis of tweets.