SENTIMENT ANALYSIS OF THE MERDEKA CURRICULUM USING VADER AND LSTM

A Twitter-Based Study

Abstract

This study explores public sentiment toward the Merdeka Curriculum, a policy reform introduced to address educational challenges in Indonesia following the COVID-19 pandemic. Utilizing Twitter (now X) as a data source, this research collected 1,900 Indonesian-language tweets to assess how society perceives the curriculum’s flexibility and relevance. The analysis begins with data preprocessing, including cleaning, tokenization, stopword removal, and stemming. VADER (Valence Aware Dictionary and Sentiment Reasoner) is employed for initial sentiment labeling, classifying tweets into positive, neutral, and negative categories. The results reveal a significant class imbalance: 1,692 neutral, 117 positive, and 91 negative tweets. These labeled tweets are further analyzed using the Long Short-Term Memory (LSTM) algorithm for deep sentiment classification. The LSTM model demonstrates training accuracy improvements from 82% to 90.29% across five epochs, while validation accuracy remains steady at 88%. However, the model fails to accurately classify minority classes, with precision, recall, and F1-scores for positive and negative sentiments scoring zero. This indicates that while LSTM can effectively recognize the dominant neutral sentiment, it struggles with minority class identification due to data imbalance. The findings highlight the need for improved model training strategies and data augmentation to enhance classification performance across all sentiment categories. Overall, this study contributes to the integration of lexicon-based and deep learning approaches for sentiment analysis and offers valuable insights for educators and policymakers in optimizing the Merdeka Curriculum through data-informed decisions.