On the Robustness of Semi-supervised Learning for Cyberbullying Detection in Social Media

Cyberbullying has become an usual form of harassment nowadays because most of the time we use digital technologies to communicate with others. This type of bullying can affect our mental, emotional, and also physical health. Also, the significant impact of cyberbullying is that it can spread easily and quickly around the world. In most of the cases, a cyberbullying attack is discovered too late, after all the negative effects have already affected the assaulted person. Researchers have proposed several solutions to detect cyberbullying attacks on social media mainly using machine learning techniques. While Artificial Intelligence provides a solution to discover cyberbullying, Human-Computer Interaction principles can be integrated with machine learning to offer the user a reliable feature to protect them from cyberbulling. The majority of the proposed solutions are supervised approaches that leverage the representation power of deep neural networks. In our recent days, the amount of unlabeled data heavily outnumbers the amount of annotated data, as it is easier and more convenient to obtain it. Labeling samples usually implies manual annotation by human experts, a process which is expensive and prone to error. The aim of this chapter is to leverage the large amount of unlabeled data that can be easily collected and use it alongside with semi-supervised learning approaches to enhance cyberbullying detection.