<![CDATA[TikTok has become an important platform for public discussion of government programs, including the Free Nutritious Meals (MBG) Program. The large volume of user comments makes manual analysis inefficient and motivates the use of automated sentiment classification. This study investigates public sentiment toward the MBG Program through comments collected from the official TikTok account of the National Nutrition Agency (@badangizinasional.ri). The study applies a pipeline consisting of text preprocessing, automatic sentiment labeling using the IndoRoBERTa sentiment-classification model, undersampling, TF-IDF feature extraction, and Multinomial Naïve Bayes classification. A total of 1,048 comments were collected, of which 1,028 remained after preprocessing. The IndoRoBERTa-based pseudo-labeling process assigned 333 comments (32.39%) to the positive class and 695 comments (67.61%) to the negative class. Because the class distribution was imbalanced, undersampling was applied, resulting in 666 comments consisting of 333 comments per class. Hyperparameter optimization using GridSearchCV with 5-fold cross-validation selected alpha = 0.5, max_df = 0.8, min_df = 3, ngram_range = (1,3), and sublinear_tf = False. On the 20% stratified test set, Multinomial Naïve Bayes achieved 80.60% accuracy, 81.04% precision, 80.60% recall, and 80.53% F1-score when evaluated against the IndoRoBERTa-generated labels. These results should be interpreted as agreement with automatically generated pseudo-labels rather than direct evidence of performance against human-verified ground truth. Therefore, the observed negative sentiment distribution reflects the original post-preprocessing dataset before undersampling, while the classifier was trained and tested on a balanced subset. Manual validation of the pseudo-labels and comparison with alternative classifiers are recommended for future work]]>
