Sentiment Analysis and Topic Modeling Using BERT And LDA Methods (Case Study of Free Meal Program on Twitter)

Abstract

The Free Meal Program is one of the Indonesian government's strategic efforts to structurally address poverty and malnutrition (stunting). As a new policy with massive social and fiscal impacts, an in-depth evaluation is required to measure public acceptance. This study aims to categorize public sentiment into positive and negative categories and identify the dominant topics discussed on Twitter (X) regarding the program. The methodology involved crawling Twitter data, resulting in a total of 8,307 datasets. Sentiment labeling was performed automatically using the IndoBERT deep learning model, followed by topic modeling using the Latent Dirichlet Allocation (LDA) method for each sentiment category. The results of the topic modeling were validated through topic coherence tests using word instruction task and topic instruction task techniques. The results showed an imbalanced sentiment distribution, with 7,034 negative sentiments and 1,273 positive sentiments. LDA modeling successfully extracted 5 optimal topics for both sentiment categories. Positive sentiments included topics such as budget efficiency, the role of government institutions (National Police), technical implementation, and local economic empowerment. Meanwhile, negative sentiments encompassed concerns regarding state budget (APBN) priorities, health/poisoning issues, and the comparative urgency between the free meal program and the education and health sectors. The coherence test results showed an interpretation accuracy rate of 93% for keywords and 79% for topic relevance, indicating that the developed LDA model was optimal in extracting public opinion.