Generating Natural Language Adversarial Examples on a Large Scale with Generative Models

An open-source Pytorch implementation Generating Natural Language Adversarial Examples on a Large Scale with Generative Models paper

Links and References:

• https://github.com/leohsuofnthu/Pytorch-TextCNN
• https://github.com/wiseodd/controlled-text-generation

Project Structure

• TextCNN - scripts to train TextCNN (required to run pipeline)
• VAE - scripts to train VAE (not needed to run the pipeline, it is included separately)
• notebooks - an end-to-end pipeline for training proposed model
• datasets - torchtext based dataloaders
• scripts - some helpful scripts

Requirements:

• Torch >= 1.0.
• Torchtext
• Spacy
• default data science libraries NumPy, SciPy, Pandas, Matplotlib, etc.

All libs should be available via pip or conda,

For Spacy don't forget to download a language model, python -m spacy download en_core_web_sm

In case of any issues, please contact us by:

• email: Konstantin.Sozykin (dot) skoltech.ru, Nikita.Stroev (dot) skolkovotech.ru
• telegram @Konstantin.Sozykin

Also dataset and trained weights are available here:

https://yadi.sk/d/420Wdc6M7Y3cgg

you can extract it into ./notebooks folder