Detecting the correlation between sentiment and user-level as well as text-level meta-data from benchmark corpora

Do tweets from users with similar Twitter characteristics have similar sentiments? What meta-data features of tweets and users correlate with tweet sentiment? In this paper, we address these two questions by analyzing six popular benchmark datasets where tweets are annotated with sentiment labels. We consider user-level as well as tweet-level meta-data features, and identify patterns and correlations of these feature with the log-odds for sentiment classes. We further strengthen our analysis by replicating this set of experiments on recent tweets from users present in our datasets; finding that most of the patterns are consistent across our analysis. Finally, we use our identified meta-data features as features for a sentiment classification algorithm, which results in around 2% increase in F1 score for sentiment classification, compared to text-only classifiers, along with a significant drop in KL-divergence. These results have potential to improve sentiment analysis applications on social media data.