<![CDATA[The escalation of cyberbullying on social media necessitates automated content moderation systems capable of deciphering the characteristics of non-standard digital language. This study compares the effectiveness of contextual word embeddings (IndoBERT) and static subword embeddings (FastText) in multi-class text classification (Safe, Mild, Moderate, Severe, and Very Dangerous) for Indonesian-language social media comments. The research focuses on analyzing embedding architectures: IndoBERT is integrated with a Dense Layer, while FastText is evaluated using Logistic Regression and Random Forest (15 trees). A dataset of 16,643 samples was extracted from social media platforms and processed through text cleaning and dictionary-based slang normalization stages. Test results indicate that the bidirectional contextual architecture of IndoBERT achieved the highest performance, with an accuracy of 82.15% and a Macro F1-Score of 74.97%. Regarding FastText representations, the linear Logistic Regression approach (78.70% accuracy) proved more stable than the Random Forest model (65.50% accuracy). These findings demonstrate that contextual embeddings are more adaptive than static embeddings at distinguishing sarcasm and mutated lexicons in digital text. As a research contribution, this study not only offers a precise five-level severity classification architecture for informal Indonesian but also provides a prototype automated moderation system (CyberGuard AI) ready for implementation in real-time risk mitigation.]]>
