Beyond Hostility: Detecting Subtle and Overt Forms of Online Conflict with Multi-Objective Learning

Social networks have become a dominant and influential aspect of modern society, with numerous users engaging in observing, creating, and distributing content. The growth of content has led to user conflicts that include bullying, aggression, harassment, and threats. Consequently, recent research has been aimed at identifying and addressing these openly hostile forms of social conflict. However, in current studies, the less overtly hostile yet equally damaging types of conflict, including teasing, criticism, and sarcasm, have been overlooked. We introduce a comprehensive multi-class conflict dataset and develop a robust multi-objective classification model to capture the full spectrum of conflict; from subtle tensions to open hostility, significantly advancing conflict detection capabilities. This innovative approach leverages class-based reward functions to improve model performance and is implemented by fine-tuning pre-trained transformer language models within a decision transformer framework. We also propose and evaluate multiple knowledge distillation strategies that compress large LLM teachers into efficient student models. These methods result in statistically significant increases in classification performances, leading to further evaluation of performance-cost trade-off. Our experiments on three datasets demonstrate superior recall, precision, F1-score, and accuracy compared to traditional state-of-the-art deep learning classifiers. Furthermore, we analyse class ambiguity and its impact on model performance as well as conducting thematic analysis on model misclassifications.