Developing Language Resources and Advancing AI Techniques for Automatic Detection of Cyberbullying Causes
Year: 2025
Type: dissertation
Abstract: Cyberbullying, defined as repetitive online aggression intended to harm others through digital communication, has emerged as a critical societal issue due to its severe mental, physical, and social repercussions. The rapid advancements and widespread use of social media platforms, which serve as spaces for discussing and critiquing social issues, have exacerbated this phenomenon as these platforms provide perpetrators with opportunities to identify vulnerable individuals and communities. Despite numerous proposed detection and prevention methods, the prevalence of cyberbullying continues to rise, underscoring the critical requirement of more effective mitigation and prevention strategies. A comprehensive understanding of the underlying causes of cyberbullying and the motivations of perpetrators is essential, particularly through the application of automated techniques for detecting these causes. Although existing studies have sought to identify the causes of cyberbullying, they are fragmented and limited by a lack of computational methods and linguistic resources, which collectively hinder the progress of this field. To address these challenges, this thesis aims to systematically review and organise existing literature on the identification and detection of cyberbullying causes, identifying critical limitations and gaps. Building upon these insights, the research seeks to develop language resources and AI techniques to empower practitioners in designing tailored intervention programs, ultimately contributing to more effective cyberbullying prevention strategies. On achieving the aforementioned goal, this research systematically identified 56 cyberbullying causes from 212 academic publications to define a comprehensive taxonomy of cyberbullying causes, laying a foundation and presenting their complex interrelationship. Furthermore, it identified 14 computational techniques proposed for the automatic detection of cyberbullying causes from 19 academic studies, alongside five datasets and two lexicons employed in these techniques. An exploratory study conducted in this research identified ingroup bias-stemmed cyberbullying as the most prevalent underlying cause, highlighting common ingroup biases associated with cyberbullying through data collection and analysis from social media platforms. Then, by proposing an endto- end pipeline for constructing datasets labelled with cyberbullying causes utilising existing resources, two novel datasets labelled with cyberbullying causes, gold-standard and silver-standard datasets, consisting of 4,730 and 34,108 data records were developed. Next, through a comprehensive linguistic analysis, it was identified that frequent and contextually relevant words, toxicity scores, and sentiment values as linguistic features that can use to distinguish between different cyberbullying causes. Building upon these findings and resources, a set of baseline models were established to serve as benchmarks for future research aimed at developing automated techniques for detecting cyberbullying causes. Experimental results demonstrated that feature engineering (e.g., integration of toxicity values) and fine-tuning pre-trained language models outperformed the defined baseline models, demonstrating their potential to use in tools and applications design for automatic detection of cyberbullying causes. Finally, this research conducted an expert evaluation to identify limitations of the developed models and provide insights for future enhancements. In summary, this thesis systematises the existing knowledge on cyberbullying causes and identifies prominent causes on social media platforms. Notably, it offers language resources and key insights gained from developing techniques for the automatic detection of cyberbullying causes. These contributions aim to support and assist practitioners and researchers in developing tailored intervention and support programs to prevent cyberbullying and mitigate its consequences.
Institution: The University of Adelaide
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: cyberbullying, causes of cyberbullying, cyberbullying detection, in-group favoritism, social media, annotated dataset, sentiment, linguistic features, pre-trained language models, cyberbullying prevention, perpetrator motivation, language resources, vocabulary, feature engineering, benchmarking model
Open Access status: green
OpenAlex ID: W7216921570
