KLASIFIKASI SENTIMEN MASYARAKAT TERHADAP AKSI 17+8 DI MEDIA SOSIAL MENGGUNAKAN LSTM DAN BERT

Abstract

This study aims to compare the performance of Long Short-Term Memory (LSTM) and Bidirectional Encoder Representations from Transformers (BERT) models in classifying public sentiment on platform X (Twitter) regarding the "17+8 Tuntutan Rakyat" movement. The dataset consists of 1,000 Indonesian tweets collected between August and September 2025, with a subset of 200 data evaluated using a 5-Fold Cross Validation scheme. The average evaluation results show that the LSTM model achieved an accuracy of 0.6900, whereas BERT achieved 0.5400. However, per-class metric analysis reveals that LSTM suffered from severe majority-class bias by predicting all instances as neutral (F1-score of 0.0000 for both positive and negative classes), whereas BERT demonstrated discrimination capability on minority classes (negative recall of 26.83% and positive recall of 14.29%). This research is limited by a small evaluation subset size and the absence of class imbalance handling techniques, which implies the crucial need for GPU acceleration and resampling methods in future social media text sentiment analysis studies.