<![CDATA[Sentiment classification in Indonesian-language social media text remains a challenge because informal spelling, code-mixing, culture-specific expressions, and class imbalance can reduce model reliability. Using a dataset containing 1,336 labeled X posts related to the 2024 Indonesian Presidential Election, this research compares two transformer workflows for three-class sentiment classification, which are fine-tuning IndoBERT on Indonesian text and a translation-based workflow, where NLLB-200 translates the same text into English before the RoBERTa training process. The dataset was divided via stratified sampling into 1,069 training examples, 133 validation examples, and 134 test examples. Both models were trained for five epochs with identical optimization settings and evaluated using accuracy, weighted precision, weighted recall, weighted F1, classwise scores, and confusion matrices. IndoBERT achieved 82.09% accuracy and 80.31% weighted F1, compared with 80.60% and 75.69% for the NLLB-200 plus RoBERTa pipeline. The largest difference occurred in the neutral class, for which IndoBERT obtained 0.26 recall and RoBERTa only 0.05. Error analysis indicates that translation artifacts and majority-class bias jointly reduced sensitivity to neutral and context-dependent expressions. Direct monolingual fine-tuning was more reliable for this dataset, although translation-based transfer remained competitive for the dominant positive class. Future work should use larger independently annotated datasets, repeated runs, translation-quality analysis, and class-aware training objectives.]]>
Copyrights © 2026
