Identification of cyberbullying: a deep learning based multimodal approach

Cyberbullying can be delineated as a purposive and recurrent act, which is aggressive in nature, done via different social media platforms such as Facebook, Twitter, Instagram and others. While existing approaches for detecting cyberbullying concentrate on unimodal approaches, e.g., text or visual based methods, we proposed a deep learning based early identification framework which is a multimodal (textual and visual) approach (inspired by the informal nature of social media data) and performed a broad analysis on vine dataset. Early identification framework predicts a post or a media session as bully or non-bully as early as possible as we have processed information for each of the modalities (both independently and fusion-based) chronologically. Our multimodal feature-fusion based experimental analysis achieved 0.75 F-measure using ResidualBiLSTM-RCNN architecture, which clearly reflects the effectiveness of our proposed framework. All the codes of this study are made publicly available on paper’s companion repository.