NLPAug: Natural Language Processing Augmentation Library

1.1.11 · active · verified Mon Apr 13

NLPAug is a Python library designed for natural language processing data augmentation. It helps improve deep learning model performance by generating synthetic textual data, making models more robust and less prone to overfitting on small datasets. The library supports various augmentation techniques across character, word, and sentence levels. Currently at version 1.1.11, it maintains an active release cadence with several minor updates throughout the year.

Warnings

gotcha Many augmenters require external model or data downloads (e.g., NLTK data for SynonymAug, pre-trained word embeddings for WordEmbsAug, or transformer models). These are not automatically installed with the base package.
Fix: Use `nlpaug.util.file.download.DownloadUtil.download_xxx()` for models (e.g., word2vec, GloVe) and `nltk.download('wordnet')`, `nltk.download('omw-1.4')` for NLTK data before initializing the respective augmenters.
breaking The `augment()` method's output format changed from a single string to a list of strings when `n > 1` (default `n=1`) in version 0.0.9. Code expecting a direct string might break.
Fix: Always treat the output of `augment()` as a list. If only one augmented output is expected, access it via `augmented_text[0]`.
deprecated Several augmenter classes and parameters have been deprecated or replaced. For example, `WordNetAug` was replaced by `SynonymAug`, `QwertyAug` by `KeyboardAug`, and the `aug_n` parameter by `top_k`.
Fix: Consult the official documentation for the latest API and use the recommended replacement classes and parameters.
gotcha Performance of transformer-based augmenters (e.g., `ContextualWordEmbsAug`, `ContextualWordEmbsForSentenceAug`) can be slower than expected, especially with older `transformers` library versions or large `n` values.
Fix: Ensure `nlpaug` and `transformers` are updated to their latest versions for performance optimizations. Consider reducing `n` (number of augmented samples) or using batching for large datasets if performance is critical.
gotcha Compatibility issues with underlying libraries, particularly `transformers` and `torch`, have been observed. Specific versions may be required for certain augmenters to function correctly.
Fix: Pin `transformers` and `torch` versions as recommended in `nlpaug`'s GitHub README or installation guides. Updating `nlpaug` to the latest version often includes compatibility fixes.

Install

pip install nlpaug Basic Install
pip install nlpaug[transformers] nlpaug[nltk] nlpaug[gensim] nlpaug[audio] Install with all optional dependencies

Imports

KeyboardAug

from nlpaug.augmenter.char import KeyboardAug

SynonymAug

from nlpaug.augmenter.word import SynonymAug

ContextualWordEmbsAug

from nlpaug.augmenter.word import ContextualWordEmbsAug

RandomSentAug

from nlpaug.augmenter.sentence import RandomSentAug

Sequential
```
from nlpaug.flow import Sequential
```

DownloadUtil

from nlpaug.util.file.download import DownloadUtil

Quickstart

This quickstart demonstrates character-level augmentation using KeyboardAug. It initializes an augmenter to simulate typos by replacing characters with nearby keys on the keyboard. The `augment` method returns a list of augmented texts, even if `n=1` (default).

import nlpaug.augmenter.char as nac

text = "The quick brown fox jumps over the lazy dog."

# Initialize a Keyboard Augmenter
# Simulates typos based on keyboard proximity
aug = nac.KeyboardAug(aug_char_p=0.1, aug_word_p=0.1, aug_char_min=1)

# Augment the text
augmented_text = aug.augment(text)

print(f"Original: {text}")
print(f"Augmented: {augmented_text[0]}")

view raw JSON →