A Weakly Supervised Approach for Adaptive Detection of Cyberbullying Roles

Abstract

As human communication increasingly occurs through online social networks, online harassment and cyberbullying are becoming a serious social health threat. In this talk, I will introduce an automated, data-driven method for adaptive detection of bullying roles. The model represents whether social media users are instigating bullying, whether they are victimized by bullying, and how indicative certain key phrases are of bullying behavior. The algorithm simultaneously estimates these variables by extrapolating from social media graph data and an expert-provided set of highly indicative bullying phrases. This weak supervision provides training input without requiring excessive effort from human annotators. I will describe quantitative and qualitative experiments on three social media datasets from networks with frequent incidences of cyberbullying: Twitter, Ask.fm, and Instagram. I will conclude by describing a vision of a future where technology helps prevent and mitigate the harm of cyberviolence, placing detection algorithms like the main topic of this talk in context.