LiDA: Language-Independent Data Augmentation for Text Classification

Developing a high-performance text classification model in low-resource language is challenging due to the lack of labeled data. Meanwhile, collecting large amounts data cost-inefficient. One approach increase amount create synthetic using augmentation techniques. However, most available techniques work on English and are highly language-dependent as they perform at word sentence level, such replacing some words or paraphrasing sentence. We present Language-independent Data Augmentation (LiDA), technique that utilizes multilingual from training dataset. Unlike other methods, our worked embedding level independent any particular language. evaluated LiDA three languages various fractions dataset, result showed improved performance both LSTM BERT models. Furthermore, we conducted an ablation study determine impact components method overall performance. The source code https://github.com/yest/LiDA .