Sentiment analysis on public opinion trends on #MakanBergiziGratis programs on platform X using Long Short-Term Memory (LSTM) networks

Abstract

The Free Nutritious Meal Program (Makan Bergizi Gratis/MBG) has sparked diverse public reactions on X (formerly Twitter), making it important to understand sentiment dynamics at scale. This study analyzes sentiment in 5,516 Indonesian-language tweets collected between January to November 2025 using a hybrid embedding of Bidirectional Encoder Representations from Transformers (BERT) with a Long Short-Term Memory (LSTM) model. Initial labeling using a lexicon-based approach reveals an imbalanced distribution (positive 34.34%, neutral 60.80%, negative 4.86%). To address this issue, the Synthetic Minority Over-sampling Technique (SMOTE) is used to build a synthetic representation for minority classes in the embedding space. This approach helps improve class balance while maintaining the overall data representation. The proposed model achieves good performance, with 82.98% accuracy, 83.27% precision, 82.98% recall, and an F1-score of 83.01%. Cross-validation indicates relatively consistent performance, with accuracy ranging from 80.51% to 84.41%. Despite these results, the negative class remains challenging (F1-score 0.57), highlighting the impact of linguistic complexity, including implicit and nuanced expressions in social media text. Overall, the findings suggest that integrating BERT embeddings with LSTM and feature-level SMOTE can be a suitable approach for handling imbalanced sentiment classification tasks. However, further improvements, particularly in advanced transformer fine-tuning and deeper linguistic modeling, are needed to better capture subtle sentiment patterns.