<![CDATA[Rising geopolitical tensions in the Middle East have triggered public concerns in Indonesia regarding fuel subsidy policies and fuel availability. This study aims to compare the performance of IndoBERT and NusaBERT in classifying Indonesian public sentiment on social media related to fuel subsidy policies. Data were collected from X (Twitter) and Instagram comments between October and November 2025 using keywords such as “BBM”, “Pertalite”, “fuel subsidy”, and “Middle East conflict”. After filtering duplicate, spam, and irrelevant content, a total of 1,500 opinion texts were manually annotated into positive, neutral, and negative sentiment classes and divided using an 80:20 train-test split configuration. The preprocessing stage included case folding, text cleansing, slang word normalization, emoji removal, duplicate filtering, and tokenization. Both Transformer models were fine-tuned using the AdamW optimizer with a learning rate of 2e-5, batch size of 16, and 3 training epochs. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. Experimental results show that NusaBERT achieved better performance than IndoBERT, obtaining an accuracy of 96.3% and weighted F1-score of 96.3%, while IndoBERT achieved an accuracy of 92.5%. Additional evaluation through confusion matrix and error analysis indicates that both models still face challenges in handling sarcasm, ambiguous expressions, and mixed-context sentences commonly found in informal Indonesian social media text. The findings suggest that NusaBERT is more effective for Indonesian social media sentiment classification due to its stronger adaptation to informal language patterns.]]>
