Detecting Explicit Hate Comment and Sarcasm in Multilingual Indonesian-English Tweets Using mDeBERTaV3

Abstract

Detecting hate comment on Indonesian social media is challenging due to rapid digitalization and code-switching texts. Traditional machine learning and standard transformer models often struggle because they rely on word counts without understanding context, or entangle word meaning with positional context, often miscalssify sarcastic contents. This study proposes mDeBERTaV3 to overcome these limitations. By utilizing its disentangled attention, mDeBERTaV3 processes a word’s semantic meaning and its relative positional context separately. This also handles sarcastic contents effectively, validated through ablation study using an Indonesian sarcasm dataset. The model’s robustness is then evaluated against synthetic social media noise, like informal slangs, emojis, and misspellings. Using the Indonesian Abusive and Hate Speech Twitter Dataset with over 13,000 tweets annotated by humans, mDeBERTaV3 is compared against three baselines, which are TF-IDF with SVM, standard multilingual BERT, and XLM-RoBERTa. The dataset also went through a stratified split of 70/15/15 to prevent majority class bias, and the model were evaluated using macro F1-Score and inference latency. Results show that mDeBERTaV3 achieves a macro F1-score of 0.8676 and inference latency of 2.2179 ms/sentence, outperforming all baselines and effectively balances the trade-offs that cause baselines to overlook actual hate or trigger false alarms. Furthermore, it achieves 0.7636 macro F1-score on the sarcasm task. Additionally, the robustness test reveals that sarcasm detection is highly fragile to lexical perturbations while hate speech detection remains stable under noises. This result suggests mDeBERTaV3 offers a reliable and deployable solution for real-time detection of multilingual hate comment in social media moderation.