Sentiment Analysis of Indonesian Government Policies Using the LSTM Model for Public Opinion Mapping

Abstract

Social media has evolved into a primary arena for citizens to express and negotiate opinions regarding government policies, creating vast opportunities for data-driven policy evaluation. This study aims to map public sentiment toward Indonesian government policies by integrating deep learning–based sentiment classification with linguistic and governance analysis. A dataset of approximately 50,000 Indonesian-language posts was collected from Twitter (X) and Facebook between January and June 2024. The data were processed through text cleaning, tokenization, stopword removal, and word embedding using Word2Vec and FastText, and subsequently classified into positive, negative, and neutral sentiments using a Long Short-Term Memory (LSTM) model. The results indicate that public opinion is predominantly negative (45%), particularly in relation to economic and taxation policies, while positive sentiment (34%) is mainly associated with education and health sectors. The LSTM model achieved an accuracy of 86.9%, outperforming Support Vector Machine (SVM) and Naïve Bayes models. Furthermore, linguistic analysis reveals that emotive and sarcastic expressions play a significant role in shaping critical public discourse, whereas colloquial language enhances engagement, especially among younger users. This study contributes by bridging computational sentiment analysis with linguistic interpretation and public policy evaluation within a unified framework. The findings provide practical implications for evidence-based policymaking by enabling governments to monitor public sentiment in real time, improve policy communication strategies, and foster more participatory and responsive governance.