Detecting Antisocial Behavior on Social Media During COVID-19 Lockdown

Abstract

The widespread availability of the internet has rendered the engagement with social media an integral component of contemporary society. Platforms such as Facebook, Twitter/X, YouTube, among others, are designed to facilitate extensive, efficient, and sustained user participation, offering both anonymity and opportunities for positive engagement. However, these platforms have also become arenas for antisocial behaviors, including disregard for others’ rights, lack of empathy, trolling, and aggression, leading to significant negative psychological impacts on affected individuals. These impacts range from anxiety and emotional trauma to depression, psychological disorders, self-isolation, diminished self-esteem, and even suicidal thoughts. This study focuses on antisocial behavior (ASB) manifested in tweets from Ghana during the 21-day COVID-19 lockdown. We develop a gold-standard annotated ASB corpus from collected and pre-processed data. We then assess the performance of different baseline classifiers against three transformer models-BERT, RoBERTa, and ELECTRA-in a binary classification task designed to detect ASB. Each model demonstrated varying degrees of success; however, the RoBERTa model, upon fine-tuning, exhibited superior performance, achieving an accuracy rate of 95.59% and an F1 score of 94.99%, thereby outperforming the other models.