See No Evil, Hear No Evil

Emerging multimedia communication apps are allowing for more natural communication and richer user engagement. At the same time, they can be abused to engage in cyberbullying, which can cause significant psychological harm to those affected. Thus, with the growth in multimodal communication platforms, there is an urgent need to devise multimodal methods for cyberbullying detection and prevention. However, there are no existing approaches that use automated audio and video analysis to complement textual analysis. Based on the analysis of a human-labeled cyberbullying data-set of Vine "media sessions' (six-second videos, with audio, and corresponding text comments), we report that: 1) multiple audio and visual features are significantly associated with the occurrence of cyberbullying, and 2) audio and video features complement textual features for more accurate and earlier cyberbullying detection. These results pave the way for more effective cyberbullying detection in emerging multimodal (audio, visual, virtual reality) social interaction spaces.