Liza Wikarsa, Lanny Sitanayah, Vicken Evan Manginsela, Minsoo Kim
Indonesia’s Makan Bergizi Gratis (MBG) Program is a nationwide free nutritious meal program to reduce child malnutrition and stunting, yet its implementation has raised public concern over food safety and management. This study analyzes public emotions toward the MBG Program using social media data classified into seven emotions – happiness, trust, surprise, neutral, fear, sadness, and anger. A hybrid BERTKNN model was applied, where BERT extracted contextual features and KNN performed interpretable emotion classification. The hybrid model showed 51.8% accuracy (60: 40 split), while the standard BERT reached 88.0% accuracy highlighting the superiority of transformer-based contextual learning. Though this study is a rare trial detecting seven Indonesian emotional polarities with a hybrid BERT-KNN model, the significant performance gap underlines the pitfall of this naive integration. It specifically highlights the challenge of the structural mismatch between BERT’s high-dimensional embeddings and KNN’s simple distance metrics, which fails to delineate the subtle boundaries between numerous classes. Future work should optimize BERT embeddings for distance-based classification through dimensionality reduction techniques such as PCA, and explore adaptive, non-linear mapping methods. © 2026.
Faculty of Engineering, Universitas Katolik De La Salle Manado, Kairagi I Kombos Manado, Belakang Wenang Permai II, Manado, 95000, Indonesia; Department of Industrial and Data Engineering, College of Engineering, Pukyong National University, 45 Yongso-ro, Nam-gu, Busan, 48513, South Korea
