TST-IOC: A Text Style Transfer-based Approach to Automatic Intervention of Online Offensive Content on Social Media to Improve Online Safety

Social media platforms such as Facebook, TikTok and Instagram have witnessed increasing use of offensive language by online users, which can be harmful to other users. Recently the continuance of the pandemic has propelled the propagation of offensive content associated with Covid-19 on social media. Some researchers begin to develop effective methods for detecting online offensive language from social media content automatically, yet automatic intervention of offensive language after it is detected remains largely understudied. To address the gaps, this dissertation develops an effective text style transfer-based approach, TST-IOC, for automatic offensive intervention tasks. The promising outcome suggests that our proposed method shows significant potential and could be a preferred choice among users for offensive intervention tasks. This dissertation provides some contributions. First, it contributes significantly to the field of offensive language research by introducing a novel text style transfer-based approach, which has been rarely explored in existing intervention studies. This approach shows a step forward in the development of an automatic offensive intervention system, addressing the limitations of current filtering systems deployed by social media platforms. Second, existing research has mainly focused on using performance metrics for evaluating offensive intervention methods quantitatively. However, this study goes beyond by proposing a comprehensive automatic evaluation paradigm. By exploring both quantitative and qualitative aspects of automatic intervention assessment, it fills a crucial gap in the current offensive language research landscape. Finally, it recognizes the scarcity of studies comparing human evaluation with automatic evaluation in automatic intervention systems. To bridge this gap, we conduct a user study, which allows for an investigation of user acceptance of the proposed automatic intervention approach in real-world scenarios. The insights gained from this user study not only guide the design of more comprehensive automatic intervention systems but also hold the potential to shape the development of human-centric automatic intervention systems in the future.