Beyond trolling: Fine-grained detection of antisocial behavior in social media during the pandemic

Abstract

Antisocial behavior (ASB), including trolling and aggression, undermines constructive discourse and escalates during periods of societal stress, such as the COVID-19 pandemic. This study aimed to examine ASB on social media during the COVID-19 pandemic by leveraging a novel annotated dataset and state-of-the-art transformer models for detection and classification of ASB categories. Specifically, this study examined ASB within a gold-standard corpus of tweets collected from Ghana during a 21-day lockdown. Each tweet was meticulously annotated into ASB categories or non-ASB, enabling a comprehensive analysis of online behaviors. We employed three state-of-the-art transformer-based language models (BERT, RoBERTa, and ELECTRA) and compared their performance against traditional machine learning models. The results demonstrate that the transformer-based approaches substantially outperformed the baseline models, achieving a high detection accuracy across both binary and multiclass classification tasks. RoBERTa excelled in binary ASB detection, attaining a 95.59% accuracy and an F1-score of 94.99%, while BERT led in multiclass classification, with a 94.38% accuracy and an F1-score of 93.92%. Trolling emerged as the most prevalent ASB type, reflecting the polarizing nature of online interactions during the lockdown. This study highlights the potential of transformer-based models in detecting diverse online behaviors and emphasizes the societal implications of ASB during crises. The findings provide a foundation for enhancing moderation tools and fostering healthier online environments.