Cross-Platform Sentiment Analysis of Indonesian Public Policies 2024-2025 Using IndoBERT on X and YouTube

Public policy discourse in Indonesia is increasingly shifting to digital platforms such as X and YouTube, where citizens express support, criticism, and evaluations of government programs. This study evaluates IndoBERT for sentiment classification across eight Indonesian public policy topics during 2024-2025 using manually annotated datasets from both platforms. Sentiment labels are categorized as Positive, Negative, and Neutral, with performance assessed using Accuracy, Balanced Accuracy, and Macro-F1. IndoBERT achieves average Accuracy, Balanced Accuracy, and Macro-F1 scores of 0.78, 0.74, and 0.73 on X, respectively, compared with 0.72, 0.65, and 0.66 on YouTube. Compared with a Multinomial Naïve Bayes baseline, IndoBERT achieves higher average performance across all three metrics on both platforms. Inter-annotator agreement, measured using Cohen’s Kappa, yields average values of 0.77 for X and 0.66 for YouTube. The results show that IndoBERT performs more consistently on X, while YouTube presents greater classification challenges due to contextual complexity, implicit sentiment, and class imbalance. Overall, the findings demonstrate the suitability of IndoBERT for Indonesian public policy sentiment classification while highlighting the influence of platform-specific characteristics and class distribution on model performance, as well as the need to consider class imbalance when developing sentiment classification models.