Attention Based Context Boosted Cyberbullying Detection in Social Media

Abstract

Cyberbullying detection is a rising research topic due to its paramount impact on social media users, especially youngsters and adolescents. While there has been an enormous amount of progress in utilising efficient machine learning and NLP techniques for tackling this task, recent methods have not fully addressed contextualizing the textual content to the highest possible extent. The textual content of social media posts and comments is normally long, noisy and mixed with lots of irrelevant tokens and characters, and therefore utilizing an attention-based approach that can focus on more relevant parts of the text can be quite pertinent. Moreover, social media information is normally multi-modal in nature and may contain various metadata and contextual information that can contribute to enhancing the Cyberbullying prediction system. In this research, we propose a novel machine learning method that, (i) fine tunes a variant of BERT, a deep attention-based language model, which is capable of detecting patterns in long and noisy bodies of text; (ii)~extracts contextual information from multiple sources including metadata information, images and even external knowledge sources and uses these features to complement the learner model; and (iii) efficiently combines textual and contextual features using boosting and a wide-and-deep architecture. We compare our proposed method with state-of-the-art methods and highlight how our approach significantly outperforming the quality of results compared to those methods in most cases.