Saliency-Based Token Swap - A Language-Agnostic Data Augmentation Method for Text Classification

Data scarcity remains a significant challenge in text classification, often resulting in suboptimal performance of machine learning models. To address this issue, this paper introduces Saliency-based token Swap (SSwap), an innovative data augmentation technique designed to enhance classification performance by operating through strategies that utilize saliency values to swap tokens between sentences. SSwap was assessed across varying levels of data availability, evaluated on low-resource languages (Sinhala and Tamil), and compared with existing data augmentation methods. Experiments demonstrated consistent improvement in text classification performance, particularly in low-resource settings. Results underscored SSwap's potential as a valuable tool for sustaining robust model performance in scenarios with limited data and low-resource languages, with implications for a wide range of text classification tasks.