Modeling Public Issues and Attitudes Toward the Free Nutritious Meals Program Using a Hybrid BERTopic and IndoBERT Architecture

Abstract

The Free Nutritious Meal (MBG) program is a large-scale social policy that has generated extensive public discussion on YouTube. However, existing public opinion analyses primarily rely on sentiment analysis, which often fails to distinguish operational criticism from fundamental opposition to the policy. This study proposes a hybrid machine learning pipeline to analyze discussion topics and public stance using 5,509 preprocessed YouTube comments. Topic modeling was performed using BERTopic with the paraphrase-multilingual-MiniLM-L12-v2 embedding model, UMAP, and HDBSCAN, while stance classification was conducted by fine-tuning the IndoBERT-base-p1 model. The results identified 19 coherent topics with a C_v coherence score of 0.4918. The fine-tuned IndoBERT achieved an accuracy of 70.00% and a Macro F1-score of 0.7002. Oppositional stances dominated the discussions (50.0%), particularly on food safety concerns (91.1%) and allegations of project corruption (86.1%). In contrast, supportive opinions (22.2%) primarily focused on the program's nationwide equity and social welfare objectives. These findings suggest that strengthening kitchen hygiene standard operating procedures (SOPs), enhancing budget transparency, and improving public communication are critical to increasing public trust and supporting the effective implementation of the MBG program.