IMPACT OF TEXT AUGMENTATION ON INDOBERT PERFORMANCE FOR HOSPITAL REVIEW SENTIMENT ANALYSIS

Abstract

Sentiment analysis of hospital patient reviews plays a critical role in evaluating healthcare service quality. However, limited labeled data and class imbalance often affect classification performance and reduce minority-class detection. This study empirically investigates the impact of text augmentation techniques on improving IndoBERT performance for sentiment classification of patient reviews at Dr. Mohammad Hoesin Palembang General Hospital. A total of 1,464 reviews were collected, preprocessed, and weakly labeled using a VADER-based approach, resulting in 1,168 positive and 296 negative instances. To address class imbalance, augmentation was applied exclusively to the training set using back-translation and Easy Data Augmentation (EDA), including synonym replacement, random insertion, random swap, and random deletion. IndoBERT was fine-tuned under consistent hyperparameter settings and evaluated using accuracy, precision, recall, F1-macro, and AUC. The baseline model achieved an F1-macro of 70.2%, indicating limited minority-class sensitivity. After augmentation, the random swap technique achieved the highest observed performance within the experimental setup, reaching 96.8% accuracy, 95.3% F1-macro, and 98.9% AUC. These results suggest improved performance within the experimental setting, particularly in minority-class detection. However, it should be noted that the labels were generated through a translation-based weak labeling approach, which may introduce noise and affect the accuracy of the labels. Therefore, the findings should be interpreted within the scope of this experimental setting.