Exploring Public Opinion on the 'Makan Bergizi Gratis' Program on X: A Comparative Analysis of IndoBERT-Large and NusaBERT-Large Models

Abstract

Program Makan Bergizi Gratis (MBG) has triggered extensive discourse on social media platform X, which serves as a primary space for public expression of opinions toward government policies. This study aims to analyze public sentiment toward the MBG program while simultaneously comparing the performance of two prominent Transformer-based models, namely IndoBERT-Large and NusaBERT-Large. This research adopts a quantitative approach employing supervised learning on 10,201 Indonesian-language posts (tweets) collected through web scraping from February 2024 to September 2025. A total of 2,000 samples were manually annotated as ground truth, achieving a high level of inter-annotator reliability (Cohen’s Kappa, κ = 0.81). The experimental results indicate that IndoBERT-Large outperforms NusaBERT-Large, achieving an accuracy of 83.00%, while NusaBERT-Large demonstrates competitive performance with an accuracy of 80.50%. Substantively, public discourse is dominated by negative sentiment, accounting for nearly 50% of the total data, reflecting public concerns regarding budgetary constraints and technical implementation issues. Positive sentiment ranges between 33% and 36%, indicating sustained and substantial public support for the program. These findings confirm the effectiveness of Transformer-based models in accurately capturing the dynamics of public opinion toward government policies using social media data.Program Makan Bergizi Gratis (MBG) has triggered extensive discourse on social media platform X, which serves as a primary space for public expression of opinions toward government policies. This study aims to analyze public sentiment toward the MBG program while simultaneously comparing the performance of two prominent Transformer-based models, namely IndoBERT-Large and NusaBERT-Large. This research adopts a quantitative approach employing supervised learning on 10,201 Indonesian-language posts (tweets) collected through web scraping from February 2024 to September 2025. A total of 2,000 samples were manually annotated as ground truth, achieving a high level of inter-annotator reliability (Cohen’s Kappa, κ = 0.81). The experimental results indicate that IndoBERT-Large outperforms NusaBERT-Large, achieving an accuracy of 83.00%, while NusaBERT-Large demonstrates competitive performance with an accuracy of 80.50%. Substantively, public discourse is dominated by negative sentiment, accounting for nearly 50% of the total data, reflecting public concerns regarding budgetary constraints and technical implementation issues. Positive sentiment ranges between 33% and 36%, indicating sustained and substantial public support for the program. These findings confirm the effectiveness of Transformer-based models in accurately capturing the dynamics of public opinion toward government policies using social media data.