Addressing Antisocial Behavior in Multi-Party Dialogs Through Multimodal Representation Learning

Antisocial behavior (ASB) on social media – including hate speech, harassment, and cyberbullying – poses growing risks to platform safety and societal well-being. Prior research has focused largely on networks such as X and Reddit, while $\mathit{\text{multi-party}}\mathit{\text{conversational}}\mathit{\text{settings}}$ remain underexplored due to limited data. To address this gap, we use $\mathit{\text{CyberAgressionAdo-Large}}$, a French open-access dataset simulating ASB in multi-party conversations, and evaluate three tasks: $\mathit{\text{abuse}}\mathit{\text{detection}}$, $\mathit{\text{bullying}}\mathit{\text{behavior}}\mathit{\text{analysis}}$, and $\mathit{\text{bullying}}\mathit{\text{peer-group}}\mathit{\text{identification}}$. We benchmark six text-based and eight graph-based $\mathit{\text{representation-learning}}\mathit{\text{methods}}$, analyzing lexical cues, interactional dynamics, and their multimodal fusion. Results show that multimodal models outperform unimodal baselines. The late fusion model $\mathtt{\text{mBERT}}\mathtt{\text{+}}\mathtt{\text{WD-SGCN}}$ achieves the best overall results, with top performance on abuse detection (0.718) and competitive scores on peer-group identification (0.286) and bullying analysis (0.606). Error analysis highlights its effectiveness in handling nuanced ASB phenomena such as implicit aggression, role transitions, and context-dependent hostility.