Application of IndoBERT for Granular Emotion Detection in Public Sentiment toward the Free Nutritious Meals Program (MBG)

Abstract

The Free Nutritious Meals Program (Program Makan Bergizi Gratis/MBG) is a government policy that has generated diverse public responses on social media. Previous studies have generally focused on sentiment analysis, which may not adequately capture the fine-grained emotional responses of the public, while the use of data augmentation and misclassification analysis remains limited. This study aims to classify public emotions toward the MBG Program using IndoBERT based on Robert Plutchik’s emotion theory, which consists of eight emotion categories: anger, happiness, sadness, trust, fear, disgust, interest, and surprise. Data were collected through web scraping from X (Twitter), TikTok, and Facebook and were subsequently annotated by the author and two expert annotators. The preprocessing steps included case folding, data cleaning, and normalization. Back-translation data augmentation was applied to minority classes in the training set to address class imbalance. The experiments compared baseline and augmented data scenarios using several learning-rate configurations. Model performance was evaluated using accuracy, precision, recall, and F1-score, along with misclassification analysis. The results show that Model M4, which employed data augmentation with a learning rate of 1 × 10⁻⁵, a batch size of 16, and 5 epochs, achieved an accuracy of 85.13% and a macro F1-score of 85.39%. Based on the 733 test records, anger was the dominant emotion (31.79%), followed by happiness (16.78%) and trust (15.96%). Anger was the dominant emotion on Facebook and TikTok, whereas trust was dominant on X (Twitter). The misclassification analysis revealed classification errors among several emotions with semantically similar characteristics. These findings demonstrate the applicability of IndoBERT combined with data augmentation for granular emotion classification of public responses to the MBG Program.