Why text is harder
Images tolerate pixel level noise, but swapping one word can flip a sentence's meaning. Text augmentation must preserve the label while adding useful variation, which demands more care than image transforms.
Common techniques
- Synonym replacement swaps words for near synonyms using a thesaurus or embeddings.
- Back translation translates to another language and back, producing fluent paraphrases.
- Random insertion, deletion, and swap make small edits known together as EDA.
- Token masking randomly hides words, mirroring masked language model pretraining.
A back translation flow
Guarding the label
- Negation and sentiment words are fragile; replacing not or good can flip the class.
- Back translation is the safest for fluency but is slower and needs a translation model.
- Keep augmentation mild for short texts where every token carries weight.
Practical notes
- Validate that augmented examples still read naturally on a sample.
- Combine with dropout and label smoothing for stronger regularization.
Key idea
Text augmentation adds variation through synonym swaps, back translation, and small edits while protecting meaning. Watch fragile words like negations that can silently flip the label.