Comparative performance of IndoBERT-based deep learning models with SMOTE for trade tariff sentiment analysis

Abstract

The rapid growth of social media discussions on trade tariff policies has produced large volumes of Indonesian-language opinion data, making sentiment analysis an important tool for understanding public responses. However, studies in this domain remained limited, particularly those addressing class imbalance and comparing multiple hybrid architectures. This study aimed to compare four IndoBERT-based model configurations, with and without the Synthetic Minority Over-sampling Technique (SMOTE), for classifying sentiment in Indonesian trade tariff discussions under class imbalance. The dataset consisted of Indonesian tweets related to trade tariffs collected from X (Twitter) between January and July 2025. The configurations were Logistic Regression as a baseline, IndoBERT-BiLSTM, IndoBERT-CNN, and IndoBERT-BiLSTM-CNN. The tweets were passed through IndoBERT to generate contextual embeddings without fine-tuning. Performance was evaluated using accuracy, precision, recall, and macro F1-score, with macro F1-score as the primary metric. The results showed that IndoBERT-CNN with SMOTE achieved the best macro F1-score of 0.7657 and an accuracy of 0.8654. SMOTE consistently improved recall across all deep learning architectures, with IndoBERT-CNN gaining the most, from 0.7304 to 0.7875. These findings showed that, among the four compared configurations, IndoBERT-CNN with SMOTE provided the most balanced performance for imbalanced Indonesian trade tariff sentiment classification.