Detecting harassment and defamation in cyberbullying with emotion-adaptive training

TL;DR

Cyberbullying detection has largely focused on direct forms, leaving indirect forms and celebrity-targeted abuse underexplored. This paper introduces HDCyberbullying, a celebrity-focused dataset with harassment and defamation, and presents Emotion Adaptive Training (EAT), a domain-adaptation framework that transfers knowledge from emotion detection to cyberbullying detection under low-resource conditions. Across nine transformer models, EAT yields about 20% improvements in macro-F1, precision, and recall, especially enhancing detection of indirect cyberbullying such as defamation. The work demonstrates the value of emotion signals for cross-domain transfer, offers a practical dataset for celebrity cyberbullying research, and points to richer emotion indicators and broader forms as fruitful directions for future work.

Abstract

Existing research on detecting cyberbullying incidents on social media has primarily concentrated on harassment and is typically approached as a binary classification task. However, cyberbullying encompasses various forms, such as denigration and harassment, which celebrities frequently face. Furthermore, suitable training data for these diverse forms of cyberbullying remains scarce. In this study, we first develop a celebrity cyberbullying dataset that encompasses two distinct types of incidents: harassment and defamation. We investigate various types of transformer-based models, namely masked (RoBERTa, Bert and DistilBert), replacing(Electra), autoregressive (XLnet), masked&permuted (Mpnet), text-text (T5) and large language models (Llama2 and Llama3) under low source settings. We find that they perform competitively on explicit harassment binary detection. However, their performance is substantially lower on harassment and denigration multi-classification tasks. Therefore, we propose an emotion-adaptive training framework (EAT) that helps transfer knowledge from the domain of emotion detection to the domain of cyberbullying detection to help detect indirect cyberbullying events. EAT consistently improves the average macro F1, precision and recall by 20% in cyberbullying detection tasks across nine transformer-based models under low-resource settings. Our claims are supported by intuitive theoretical insights and extensive experiments.

Figures (10)

• Figure 1: Examples of celebrity cyberbullying, with key terms highlighted and names anonymised as ##.
• Figure 2: Architecture of EAT
• Figure 3: Domain similarity. Orange: Emotion data; Blue: Cyberbullying data.
• Figure 4: Mapping between Emotion and Cyberbullying domain.
• Figure 5: Performance of EAT in the defamation detection task.