CONTEXT-BOOSTED CYBERBULLYING DETECTION IN SOCIAL MEDIA USING ATTENTION

Abstract

The detection of cyberbullying is a growing area of study because of its significant effects on social
media users, particularly children and teenagers. Although there has been significant advancement in the use of
effective machine learning and natural language processing techniques to solve this issue, contemporary
approaches have not adequately addressed contextualizing the textual information to the greatest degree possible.
Using an attention-based technique that can concentrate on more significant parts of the text can be highly
relevant because social media postings and comments typically contain lengthy, loud material that is cluttered
with numerous unrelated tokens and characters. Furthermore, social media data is typically multimodal in nature
and may include a variety of contextual information and metadata that can improve the cyberbullying prediction
system. In this study, we present a novel machine learning technique that: (i) refines a deep attention-based
language model, a variant of BERT, which can identify patterns in lengthy and noisy text; (ii) extracts contextual
information from various sources, such as images, meta-data, and even external knowledge sources, and utilizes
these features to enhance the learner model; and (iii) effectively combines textual and contextual features using
boosting and a wide-and-deep architecture. We contrast our suggested strategy with cutting-edge techniques and
demonstrate how, in the majority of circumstances, our method produces outcomes that are noticeably better
than those of those techniques.