Abstract

Social media holds the freedom to express anyone as they are. Still, people fail to follow community standards and cross the boundary of self-limit, hurting other people, sometimes leading to cyberbullying. Social media mining is the frontier where researchers contend with ensuring safe cyberspace with the help of robust information retrieval and data mining techniques. In this paper, we are aiming towards achieving such a goal for Bangla language-spoken people. We have created a corpus that contains 11006 Bangla comments from Facebook, analyzed them demographically, annotated them to create robust classifiers to classify these comments as positive, negative, and neutral polarity. We have decomposed these polarities to further sentiments based on contents of the text varying from wishful thinking to gender-based hate speech. Our multiclass classification algorithm, consisting of TF-IDF vectorizer alongside uni-gram, bi-gram, and tri-gram followed by MNB, MNB, and KNN, gives 82.60%, 82.33%, and 79.63% accuracy, respectively.