Towards Contextualised Deep Learning Based Cyberbullying Detection

Cyberbullying is a growing problem in social media, and this makes the task of automatic cyberbullying detection and protecting its victims quite crucial and of immense importance. Cyberbullying or hate speech is the act of posting hateful content on social media, which is found abusive due to its repetitive and profane nature. The automatic detection of cyberbullying is of immense importance as it impacts vulnerable cohorts like teenagers, youngsters, and queerer people. Social media platforms are pretty complex as there are various moving parts and angles to a social media platform, such as the enormous graph of social relationships, the massive load of content generated by social media users, and the multimodal nature of the content. The complexity of the data in social media could be further highlighted by taking into account the extensive amount of unstructured data, including text, images, and videos that are rapidly produced in huge volumes. The problem’s significance and complexity require employing Big Data analysis and machine learning techniques to tackle this use case efficiently. Therefore, it has made cyberbullying a popular research topic and motivated many researchers to address it. Since cyberbullying primarily occurs through natural language expression, its automatic detection is also based on the best practices of natural language processing (NLP). However, major challenges arise in the process of cyberbullying detection using NLP. These challenges are mainly rooted in the inability of conventional NLP methods to contextualise the textual inputs and make sense of them for cyberbullying detection. Secondly, the naive approach of conventional NLP-based cyberbullying methods, which focus on textual features only and neglect the invaluable features beyond text, also plays an important role. Finally, it is worthwhile to mention the challenges of dealing with heterogeneous features from various modalities and combining them using an efficient mechanism to perform a task like cyberbullying detection. This dissertation aims to highlight the challenges of cyberbullying detection and tackle them effectively using the best practices of natural language processing and deep learning. More specifically, we focus on the challenges of textual content complexity and the multimodal nature of social media data. We will propose methods that significantly enhance the process of automatic cyberbullying detection.