Data augmentation for Thai natural language processing using different tokenization

Abstract

Tokenization is one of the most important data pre-processing steps in the text classification task and also one of the main contributing factors in the model performance. However, getting good tokenizations is non-trivial when the input is noisy, and is especially problematic for languages without an explicit word delimiter such as Thai. Therefore, we proposed an alternative data augmentation method to improve the robustness of poor tokenization by using multiple tokenizations. We evaluated the performance of our algorithms on different Thai text classification datasets. The results suggested our augmentation scheme makes the model more robust to tokenization errors and can be combined well with other data augmentation schemes.