nlpaug

Augmentation techniques span word-level substitution (synonym replacement via WordNet, TF-IDF weighted substitution, contextual word substitution via BERT and XLNet), character-level perturbations (keyboard error simulation, OCR error simulation, random insertion/deletion), sentence-level transformations (back-translation, abstractive summarization), and audio augmentation for speech processing pipelines.

The library is structured around an augmentation pipeline model where multiple augmenters can be composed in sequence or applied randomly, with configurable augmentation rates that balance between data diversity and label-preserving fidelity.

NLPAug supports both synchronous and asynchronous augmentation for large dataset processing, and integrates with standard NLP frameworks including Hugging Face Transformers for contextual augmentation methods that produce semantically coherent substitutions rather than random word swaps.

NLPAug is open-source under the MIT license and is used across NLP applications where training data scarcity limits model generalization medical NLP where annotated clinical text is expensive to produce, legal NLP where specialized corpora are proprietary, and low-resource language modeling.

By artificially expanding training datasets with semantically consistent variations, NLPAug helps models learn more robust representations that generalize better to the surface form variation found in real-world text. It is installable via pip and provides tutorial notebooks demonstrating each augmentation technique.