Fine-Grained Analysis of Cyberbullying Using Weakly-Supervised Topic Models

Abstract

The possibility of anonymity and lack of effective ways to identify inappropriate messages have resulted in a significant amount of online interaction data that attempt to harass, bully, or offend the recipient. In this work, we perform a fine-grained quantitative and qualitative linguistic analysis of messages exchanged using one such recent web/smartphone application—Sarahah, that allows friends to exchange messages anonymously. We first develop a weakly supervised hierarchical framework using seeded topic models to automatically categorize Sarahah messages into different coarse and fine-grained bullying categories. Our linguistic analysis reveals that a significant number of messages exchanged using this platform (~ 20%) include inappropriate, hurtful, or profane language intended to embarrass, offend, or bully the recipient. We then present a detailed analysis of the messages and corresponding users' responses to these messages in the different bullying categories by comparing them across different linguistic and psychological attributes such as sentiment and psycho-linguistic categories from Linguistic Inquiry Word Count (LIWC). Finally, we perform a comparative analysis of messages exchanged on Sarahah to an existing labeled cyberbullying dataset from the Formspring social network on the severity of bullying, coarse-grained bullying categories, and anonymity. Our analysis sheds light on the different categories of bullying and the effect each category has on the recipient and helps quantify the different types and amounts of negativity existing in online social media.