Analisis Performa Teks Augmentasi Pada Sentimen Kebocoran Data Bank Syariah Indonesia Menggunakan IndoBERT

The data leak that occurred at Bank Syariah Indonesia (BSI) in May 2023 triggered a variety of public opinions spread on X social media. This study aims to classify public sentiment towards the incident and analyze the effect of various data augmentation techniques on the performance of the IndoBERT model. The dataset used consisted of 24,000 tweets and was preprocessed and sentiment labeled automatically. The data imbalance problem is solved by four augmentation techniques, namely SMOTE, synonym replacement, back translation, and a combination of synonym replacement and back translation. Each augmented dataset is trained using the IndoBERT model and evaluated based on accuracy, precision, recall, and f1-score. The evaluation results show that the SMOTE technique provides the best performance with an accuracy of 87%, outperforming the other augmentation techniques. This is because SMOTE works on numerical representations without causing meaning distortion, making it more stable in supporting the performance of the IndoBERT model. This research contributes to the selection of effective augmentation methods for sentiment analysis in the context of unbalanced data in Indonesian.