nlpaug

The nlpaug library provides tools for augmenting NLP data in machine learning projects. It is built around two core concepts:

1. Augmenter: The basic element used to perform a specific type of augmentation.
2. Flow: A pipeline used to orchestrate multiple augmenters together in a sequence.

As of version 2.0.0 (June 2026), the runtime baseline for NLPaug has been upgraded. Ensure your environment meets the following requirements:

• Python: 3.12 or higher
• Major Dependencies:
  • transformers 5.9+
  • gensim 4.4+
  • librosa 0.11+
  • NumPy 2.x+

Install the core nlpaug package using pip. The library targets Python 3.12+.

pip install nlpaug

The nlpaug library provides several specialized augmentation workflows. You can find detailed implementation examples in the following Jupyter notebooks:

• Audio augmenters: Techniques for augmenting audio data.
• Textual augmenters: Techniques for augmenting text data.
• Spectrogram augmenters: Techniques for augmenting spectrograms.
• Custom augmenter: How to implement and use your own augmentation logic.
• TF-IDF model training: Using TF-IDF for augmentation-related tasks.
• Flow: Managing augmentation flows.

If using WordEmbsAug with word2vec, GloVe, or fasttext, you must download the pretrained assets first using DownloadUtil.

from nlpaug.util.file.download import DownloadUtil
DownloadUtil.download_word2vec(dest_dir='.')
DownloadUtil.download_glove(model_name='glove.6B', dest_dir='.')
DownloadUtil.download_fasttext(model_name='wiki-news-300d-1M', dest_dir='.')

Install optional dependencies for specific augmentation types using bracket notation:

• Transformers: pip install "nlpaug[transformers]"
• NLTK: pip install "nlpaug[nltk]"
• Word Embeddings: pip install "nlpaug[word-embs]"
• Audio: pip install "nlpaug[audio]"
• Lambada: pip install "nlpaug[lambada]"

Character-level augmentation simulates errors like OCR mistakes or keyboard typos.

• OcrAug: Simulates optical character recognition errors (e.g., substituting 'o' with '0').
• KeyboardAug: Simulates typos based on keyboard distance.
• RandomCharAug: Performs random character operations using the action parameter.

Word-level augmentation substitutes or inserts words using various techniques:

• SpellingAug: Substitutes words with spelling mistakes from a dictionary.
• WordEmbsAug: Uses word embeddings (word2vec, glove, or fasttext) to find similar words.
• TfIdfAug: Uses TF-IDF similarity to insert or substitute words.
• ContextualWordEmbsAug: Uses language models (BERT, DistilBERT, RoBERTA, or XLNet) to predict target words.
• SynonymAug: Uses WordNet or PPDB to find synonyms.
• AntonymAug: Substitutes words with their antonyms.
• RandomWordAug: Performs random word operations (swap, delete, or crop).
• SplitAug: Splits a word into two tokens.
• BackTranslationAug: Translates text to another language and back to create variations.
• ReservedAug: Protects specific tokens from being augmented.

Before using LAMBADA augmentation, you must train a classifier on your input data. The input data should be a CSV file with text and label columns. Use the train_cls.py script to train the model.

python ../scripts/lambada/train_cls.py  \
    --train_data_path ../test/res/text/classification.csv \
    --val_data_path ../test/res/text/classification.csv \
    --output_dir ../model/lambada/cls \
    --device cuda \
    --num_epoch 2

Use AudioLoader.load_mel_spectrogram to load an audio file as a Mel spectrogram and AudioVisualizer.spectrogram to display it. This is useful for preparing data for spectrogram-based augmentation.

from nlpaug.util.audio.loader import AudioLoader
from nlpaug.util.audio.visualizer import AudioVisualizer

path = 'Yamaha-V50-Rock-Beat-120bpm.wav'

data = AudioLoader.load_mel_spectrogram(path, n_mels=128)
AudioVisualizer.spectrogram('Original', data)