AbstractThe Internet and social media have undoubtedly improved our abilities to keep in touch with friends and loved ones. Additionally, it has opened up new avenues for journalism, activism, commerce and entertainment. The unbridled ubiquity of social media is, however, not without negative consequences and one such effect is the increased prevalence of cyberbullying and online abuse. While cyberbullying was previously restricted to electronic mail, online forums and text messages, social media has propelled it across the breadth of the Internet, establishing it as one of the main dangers associated with online interactions. Recent advances in deep learning algorithms have progressed the state of the art in natural language processing considerably, and it is now possible to develop Machine Learning (ML) models with an in-depth understanding of written language and utilise them to detect cyberbullying and online abuse. Despite these advances, there is a conspicuous lack of real-world applications for cyberbullying detection and prevention. Scalability; responsiveness; obsolescence; and acceptability are challenges that researchers must overcome to develop robust cyberbullying detection and prevention systems.
This research addressed these challenges by developing a novel mobile-based application system for the detection and prevention of cyberbullying and online abuse. The application mitigates obsolescence by using different ML models in a “plug and play” manner, thus providing a mean to incorporate future classifiers. It uses ground truth provided by the enduser to create a personalised ML model for each user. A new large-scale cyberbullying dataset of over 62K tweets annotated using a taxonomy of different cyberbullying types was created to facilitate the training of the ML models. Additionally, the design incorporated facilities to initiate appropriate actions on behalf of the user when cyberbullying events are detected.
To improve the app’s acceptability to the target audience, user-centred design methods were used to discover stakeholders’ requirements and collaboratively design the mobile app with young people. Overall, the research showed that (a) the cyberbullying dataset sufficiently captures different forms of online abuse to allow the detection of cyberbullying and online abuse; (b) the developed cyberbullying prevention application is highly scalable and responsive and can cope with the demands of modern social media platforms (b) the use of user-centred and participatory design approaches improved the app’s acceptability amongst the target audience.
|Date of Award||2021|
|Supervisor||Jo Lumsden (Supervisor) & Yulan He (Supervisor)|
- cyberbullying detection
- cyberbullying prevention
- deep learning
- participatory design
- mobile application