Perbandingan Analisis Sentimen dan Pemodelan Topik pada Universal Health Coverage di Indonesia Menggunakan BERT, IndoBERT, dan BERTopic

Abstract

The implementation of Universal Health Coverage (UHC) through the National Health Insurance (JKN) program in Indonesia has triggered diverse public opinions on social media platforms Twitter/X. This study aims to analyze public sentiment using two Transformer-based Deep Learning models, namely IndoBERT and Multilingual BERT, as well as to map the primary complaint topics using BERTopic. Data were collected from the Twitter/X platform covering the period from January 2014 to November 2025, resulting in 9,114 Indonesian-language tweet documents. After preprocessing and labeling into three sentiment categories (Negative, Neutral, and Positive), both models were trained using a fine-tuning approach with an 80:20 dataset split. Evaluation results show that IndoBERT outperforms BERT across all evaluation metrics, achieving an accuracy of 78.22%, Precision of 0.776, Recall of 0.782, and weighted F1-Score of 0.777. IndoBERT's best performance was achieved on the Negative class with an F1-Score of 0.857, surpassing BERT which obtained an accuracy of 71.64% and a Negative class F1-Score of 0.800. Topic analysis using BERTopic on 5,067 documents predicted as negative sentiment successfully identified 10 main complaint topics, with five dominant topics being bureaucratic complexity and premium burden, drug shortages at healthcare facilities, quality comparison with private insurance, demands for fund management transparency, and insufficient program socialization to the public. This study concludes that IndoBERT is the more optimal model for Indonesian sentiment classification, while the combination of IndoBERT and BERTopic is recommended as a real-time public sentiment monitoring instrument to support more responsive and data-driven evaluation of JKN/UHC policies.