YouTube Sentiment Analysis on Felt Earthquake News Using LSTM and IndoBERT

Abstract

This study analyzes public sentiment in YouTube comments on felt-earthquake news in Indonesia and compares a bidirectional Long Short-Term Memory implementation (LSTM) with IndoBERT. An experimental quantitative design was used. Comments were collected through the YouTube Data API v3 using official earthquake-event references from the Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG) and keyword-based video searches covering January 2021 to August 2025. The acquisition stage produced 51,870 comments from 1,626 unique videos. Data were processed through duplicate removal, text cleaning, case folding, slang normalization, tokenization, and stopword removal, resulting in 49,041 clean comments. Positive, negative, and neutral labels were assigned by aggregating word-polarity scores from the Indonesian Sentiment Lexicon (InSet), which served as weak supervision. Stratified sampling divided the dataset into 80% training data and 20% testing data. The LSTM model used a 100-dimensional embedding, a 64-unit bidirectional LSTM layer, global max pooling, and early stopping; IndoBERT was fine-tuned from indobenchmark/indobert-base-p2 for four epochs. Performance was assessed with accuracy, macro precision, macro recall, macro F1-score, and confusion matrices. Positive sentiment accounted for 41.2% of the corpus, negative sentiment for 35.5%, and neutral sentiment for 23.2%. IndoBERT achieved 91.50% accuracy and a 91.03% macro F1-score, outperforming LSTM at 90.91% accuracy and a 90.34% macro F1-score. IndoBERT provided the strongest contextual classification, while LSTM remained a competitive and substantially lighter option for resource-constrained monitoring