Analisis Sentimen dan Persepsi Publik terhadap Pelaksanaan Program Makan Bergizi Gratis pada Media Sosial Threads

Abstract

Here is the translation of the summarized text in English (around 250 words): Indonesia's Free Nutritious Meal Program (MBG) is a national strategic policy targeting 82.9 million beneficiaries by 2026, driving extensive public discussion on social media. This study characterizes public perception of the MBG implementation on Threads through relevance detection, sentiment classification, and emotion classification of Indonesian-language posts. Following the CRISP-DM framework, 7,016 Threads posts were collected on June 20, 2026, using eight MBG-related keywords. An LLM-assisted annotation pipeline under a fixed schema generated labels, which were verified on a stratified sample. After data cleaning, 6,850 posts entered relevance detection, and 4,968 relevant posts proceeded to sentiment and emotion tasks. Five classifiers—Logistic Regression, Linear SVC, Multinomial Naive Bayes, Complement Naive Bayes, and Random Forest—were compared using TF-IDF features and 5-fold cross-validation. Evaluations on the held-out test set showed Random Forest performed best for relevance detection (Macro F1 = 0.8904), while Logistic Regression led sentiment (Macro F1 = 0.4810) and emotion classification (Macro F1 = 0.2382). Within the 4,968 relevant posts, negative sentiment dominated at 54.3% (2,696 posts). Frustration (37.7%) and anger (11.4%) were the most frequent emotions, with four risk-indicating emotions (frustration, anger, disappointment, worry) collectively accounting for 57.1%. A strong statistical association was found between sentiment and emotion ($\chi^2$ = 8,313.14, $p < 0.001$, Cramér's V = 0.747). Terms most strongly associated with negative sentiment included "korupsi" (corruption), "proyek" (project), "keracunan" (poisoning), and "anggaran" (budget). In conclusion, criticism dominates the MBG discourse on Threads, with emotion labels adding discriminative context beyond standard sentiment analysis. These findings provide platform-specific insights to help inform policy communication strategies.