Ruminati : modeling the detection of textual cyber-bullying

Abstract

The scourge of cyber-bullying has received widespread attention at all levels of society including parents, educators, adolescents, social scientists, psychiatrists and policy makers at the highest echelons of power. Cyber-bullying and it's complex intermingling with traditional bullying has been shown to have a deeply negative impact on both the bully as well as the victim. We hypothesize that tackling cyber-bullying entails two parts - detection and user-interaction strategies for effective mitigation. In this thesis, we investigate the problem of detecting textual cyber-bullying. A companion thesis by Birago Jones will investigate use-interaction strategies. In this thesis, we explore mechanisms to tackle the problem of textual cyber-bullying using computational empathy - a combination of detection and intervention techniques informed by scoping the social parameters that underlie the problem as well as a socio-linguistic treatment of the underlying socially mediated communication on the web. We begin by presenting a qualitative analysis of textual cyber-bullying based on data gathered from two major social networking websites and decompose the problem of detection into sub-problems. I then present Ruminati - a society of models of models involving supervised learning, commonsense reasoning and probabilistic topic modeling to tackle each sub-problem.