Pengaruh Augmentasi Data pada Fine-Tuning Model IndoBERT dalam Analisis Sentimen

Abstract

Class imbalance is one of the major challenges in sentiment analysis because classification models tend to favor the majority class, resulting in lower performance on minority classes. This study aims to improve Indonesian sentiment analysis performance on the SmSA dataset by applying three data augmentation techniques, namely Back-Translation (BT), Synonym Replacement (SR), and Antonym Replacement (AR), using the IndoBERT model. Data augmentation was applied only to the negative sentiment class until its size was balanced with the positive class. The Back-Translation approach employed a multi-hop translation scheme of Indonesian → English → Chinese → Indonesian. Synonym Replacement utilized the Indonesian Thesaurus with contextual embedding validation using IndoBERT, while Antonym Replacement incorporated clause-level segmentation, contextual antonym substitution, Negation Handling, and Negation Fallback. Before fine-tuning, several learning rate configurations were evaluated, and a learning rate of 1e-5 was selected as the optimal setting. Model performance was evaluated using Accuracy, Precision, Recall, and F1-Score across five experimental scenarios. The results showed that Back-Translation achieved the highest Accuracy of 91.80% with an F1-Score of 89.59%. The combination of all three augmentation methods achieved the highest F1-Score (89.65%) and Recall (89.39%), while Synonym Replacement and Antonym Replacement obtained F1-Scores of 88.80% and 88.90%, respectively. Compared with the baseline model, which achieved an Accuracy of 91.40% and an F1-Score of 88.64%, all augmentation methods improved the F1-Score, indicating better-balanced classification performance. These findings demonstrate that Back-Translation, Synonym Replacement, and Antonym Replacement effectively improve training data quality and enhance IndoBERT performance for Indonesian sentiment analysis.