BDA: Bangla Text Data Augmentation Framework

Abstract

Data augmentation can be a valuable technique, particularly in resource-scarce linguistic domains for improving the performance of natural language processing tasks by creat ing new synthetic data instances. This paper introduces a Bangla text Data Augmen tation Framework (BDA) using pre-trained model-based and rule-based approaches, along with a filtering pipeline to ensure semantic similarity and lexical variance be tween augmented and original text. We provide a comprehensive pipeline for the pro posed framework and perform an in-depth analysis of how well it performs in the Bangla text classification tasks. Our framework improved the F1 score of classification tasks by up to 13.92%, 8.58%, and 10.55% among 15%, 50%, and 100% clipping ranges re spectively, across five different datasets. Training with BDA while using only 50% of the available training set achieved the comparable F1 score as normal training with all available data. We provide an extensive study of the performance of each augmentation approach at the clipping ranges of datasets using BanglaBERT and variants of SVM. Furthermore, we discuss the indicators for optimal performance of the BDA framework and its shortcomings with in-depth analysis