IDA: An Imbalanced Data Augmentation for Text Classification

With the increasing amount of textual data generated online, an automatic system for text classification is imperative. However, classification models face the challenge of limited and imbalanced data, resulting in poor performance on minority classes. This paper presents a data augmentation technique for imbalanced text classification called Imbalanced Data Augmentation (IDA). The proposed technique consists of three main components: word selection, synonym substitution, and stop word insertion. We evaluate IDA’s performance using an imbalanced dataset of user-generated feedback on Algerian higher education sourced from tweets. Our proposed technique significantly improves the detection of the minority class by achieving the highest F1-score compared to the other evaluated data augmentation methods. Overall, IDA is a useful tool for enhancing the performance of text classifiers on imbalanced datasets by preventing overfitting, improving model generalization, addressing class imbalances, and reducing the cost of collecting and labeling data.