Text augmentation creates modified input examples—such as deleting, swapping, or replacing words—to improve model robustness. It is less reliable than image augmentation because tiny edits can change meaning. section 14.5.3
A simple word-edit example
Random deletion or swapping can turn “The rain in Spain falls mainly on the plain” into “The rain Spain falls plain on the mainly.” Such examples may improve resistance to typos and grammar mistakes, but word order changes can unintentionally alter the label or meaning. section 14.5.3
Use these practical rules:
- Prefer edits that preserve meaning; synonym replacement can fail when a word has multiple meanings. section 14.5.3
- Seek additional text examples before investing heavily in augmentation techniques. section 14.5.3
