Dangerous Speech Classification on Twitter Using Multilabel Aspects and Structural Features

Dangerous speech on social media—particularly content inciting violence against specific groups—has become a growing concern due to its role in fueling conflict and polarization. Prior research on dangerous speech classification mainly focuses on tweet text and features such as user influence level, emotions, or user attributes. However, many studies overlook essential aspects, including social and historical context, dehumanization, accusation in a mirror, attacks on children or women, questioning group loyalty, and threats to group integrity. The co-occurrence of two or more aspects within a single tweet indicates that it falls under a multilabel aspect. This study proposes a multimodal classification framework integrating textual features, multilabel dangerous speech aspects, and structural features—namely engagement metrics and graph-based user influence levels. The dataset comprises 3,254 tweets collected in 2019–2022, a period of peak political polarization after the 2019 Presidential Election, when social media was laden with dangerous narratives. We evaluate three models: a traditional machine learning model (TF-DangerML), a transformer-based hybrid model (BERT-DangerML), and a neural multimodal architecture (Multi-DangerBERT). Results show that aspect-based features substantially improve model performance. TF-DangerML with XGBoost achieves the best results (96.3% accuracy, 95.9% macro F1-score) when combining text and aspect features. Multi-DangerBERT with concat pooling achieves 95.4% accuracy, 94.8% macro F1-score, and 0.983 macro ROC-AUC, showing balanced and discriminative performance across classes. These findings highlight the complementary strengths of aspect-based modeling: while traditional models like TF-DangerML achieve the highest overall accuracy, transformer-based multimodal models such as Multi-DangerBERT offer more stable and consistently balanced predictions.