Data augmentation

Data augmentation encompasses techniques for expanding training datasets by generating synthetic examples or applying transformations to existing data. In NLP, data augmentation methods range from simple rule-based transformations (token swapping, deletion) to adversarial example generation designed to improve model robustness and generalization.

Key papers

• TextAttack: A Framework for Adversarial Attacks, Data Augmentation, and Adversarial Training in NLP — Provides automated toolkit for data augmentation via pre-implemented adversarial attack recipes, enabling generation of training data for adversarial training.

Related topics

• Adversarial training (using augmented data for training)
• Adversarial robustness (motivation for augmentation)
• Model Evaluation (evaluating augmentation effectiveness)