Rephrasing Toxic Comments using Unsupervised Text Style Transfer

This is the repository for the project I completed as part of the Spring 21 refresh of the online course Full Stack Deep Learning.

As part of the submission we had to write a project report:

Full Project Report.

We were encouraged to share online proposal and report videos as below:

• Project Proposal Loom Video

• Project Report Loom Video

Project Outline

• Investigate Unsupervised NLP Style Transfer Techniques & explore their application to toxic comment moderation

• Review and repurpose code from papers on Style Transfer including:

  • Style Transfer from Non-Parrallel Text by Cross-Alignment
  • Reformulating Unsupervised Style Transfer as Paraphrase Generation

• Fine tune models using the Civil Comments dataset on Hugging Face.

• GPT2 fine tuning for parahprase generation was done on Colab Pro using an Nvidia V100 GPU, base GPT2 and Bert Large models stored on google drive. Training script is available on Google Colab