Performance Evaluation of Text Augmentation Methods with BERT on Small-sized, Imbalanced Datasets

Recently deep learning methods have achieved great success in understanding and analyzing text messages. In real-world applications, however, labeled text data are often small-sized and imbalanced in classes due to the high cost of data collection and human annotation, limiting the performance of deep learning classifiers. Therefore, this study explores an understudied area—how sample sizes and imbalance ratios influence the performance of deep learning models and augmentation methods—and provides a solution to this problem. Specifically, this study examines the performance of BERT, Word2Vec, and WordNet augmentation methods with BERT fine-tuning on datasets of sizes 500, 1,000, and 2,000 and imbalance ratios of 4:1 and 9:1. Experimental results show that BERT augmentation improves the performance of BERT in detecting the minority class, and the improvement is most significantly (15.6–40.4% F1 increase compared to the base model and 2.8%–10.4% F1 increase compared to the model with the oversampling method) when the data size is small (e.g., 500 training documents) and highly imbalanced (e.g., 9:1). When the data size increases or the imbalance ratio decreases, the improvement generated by the BERT augmentation becomes smaller or insignificant. Moreover, BERT augmentation plus BERT fine-tuning achieves the best performance compared to other models and methods, demonstrating a promising solution for small-sized, highly imbalanced text classification tasks.