Deep Learning Techniques for Cyber Bullying Detection in Social Networks

Cyberbullying detection (CBD) is a technology that is both popular and strong, thus many natural language processing (NLP) experts are intrigued in it. Whereas the latter need considerable and complex engineering, traditional approaches mostly depend on natural language processing (NLP) software developed by a third party and features defined manually. Furthermore, aggravating the situation is the way the several subtasks are used to divide a task into a lot of smaller ones on CBD procedure, which does not follow the usual approach. By using a combination approach depending on Capsule Network (CapsNet) to complete a job, this research is able to bypass these constraints. After feature representations are extracted from input corpora using CapsNet, the next stage is the reconstruction of events from the RNN output using the combination approach. Using several annotated corpora in the case that an event occurs on more than one level of a social media network, this approach is possible to extract the tasks from a CBD. We assess the proposed model against models regarded to be state-of-the-art using a broad spectrum of text corpora datasets. Regarding cases of cyberbullying directed against cancer sufferers, the results show that the CapsNet classification approach has a higher accuracy than the approaches now in use.