Optimizing Indobert for Imbalanced Sentiment Analysis Using Data Augmentation on Tapera and Benchmark Datasets

BERT has become one of the most effective Natural Language Processing (NLP) models for sentiment analysis due to its ability to generate contextual representations. However, previous studies have reported that BERT often exhibits imbalanced precision, recall, and F1-score across sentiment classes when trained on imbalanced datasets. This study investigates the effectiveness of data augmentation in improving the performance of two pretrained language models, IndoBERT and IndoBERTweet, on a TAPERA policy dataset containing 3,364 preprocessed tweets and a benchmark dataset with 12,693 instances. Data augmentation was applied only to the training data using Synonym Replacement (15%, 25%, and 35%), Back Translation, Masked Language Modeling (15%, 25%, and 35%), and Masked Language Modeling combined with Cosine Similarity. Model performance was evaluated using Accuracy and Macro F1-score. The best performance on the TAPERA dataset was achieved by IndoBERTweet with Masked Language Modeling (25%), yielding a Macro F1-score of 0.8421 and an Accuracy of 0.8880. On the benchmark dataset, IndoBERTweet with Synonym Replacement (25%) achieved the highest Macro F1-score (0.9346) and Accuracy (0.9520). These findings provide a comparative evaluation of augmentation strategies for IndoBERT-based sentiment analysis and offer practical guidance for handling class imbalance across different datasets.