Sentiment analysis reads opinion out of text. On Twitter that job is harder than it looks, because posts are short, informal, and full of sarcasm. A study from Universitas Brawijaya asked a simple question: if the text is not enough on its own, what else on the platform can help?
The team split the problem in two. Text was classified with Naive Bayes, a long standing workhorse for this kind of task. People also react to a post in ways the words do not capture, so the researchers treated likes and retweets as a second set of features, measured rather than read.
Not every count is worth keeping. Fisher Score was used to test which of those engagement features actually carried information about sentiment, so that weak signals could be filtered out before they added noise.
Text and engagement were combined with weights named alpha and beta. Choosing those weights is the whole design decision, because they set how much the model trusts the words and how much it trusts the crowd's reaction.
The best configuration put slightly more weight on the text than on the engagement features, and the combined model reached an F1 score of 0.838. The result is a useful reminder that engagement data is a supplement, not a replacement. Words still carry most of the meaning, and reactions sharpen the picture around the edges.
For teams building anything on social data, the takeaway is that a platform gives you more than the sentences it stores. The surrounding behaviour is cheap to collect and, with the right weighting, it earns its place in the model.
Rizal Setya Perdana and Aryo Pinandito (2018). Combining Likes-Retweet Analysis and Naive Bayes Classifier within Twitter for Sentiment Analysis. Journal of Telecommunication, Electronic and Computer Engineering.
