Fair Text Classification with Wasserstein Independence

This repository allows to run the code to reproduce the experiments of the paper Fair Text Classification with Wasserstein Independence. The original code comes from the Fairlib library (https://github.com/HanXudong/fairlib), a framework in which we have integrated our approach.

Data

To download the dataset used in our experiments, please use the following commands:

• Bias in Bios dataset : python3 download_data.py Bios
• Moji dataset : python3 download_data.py Moji
• Marked Personas dataset : python3 download_data.py Marked_personas

Note: for the EEC dataset, no download is required, the dataset will be temporarely downloaded when training the demonic model.

Training demonic model

To train the demonic model, please use the following commands:

• for the Bios dataset :
  • without transfer :
    python3 train_demonic/train_demonic_Bios.py Bios
  • demonic pretrained on the EEC dataset :
    python3 train_demonic/train_demonic_Bios.py EEC
  • demonic pretrained on the Marked Personas dataset :
    python3 train_demonic/train_demonic_Bios.py dv2_story
• for the Moji dataset :
  python3 train_demonic/train_demonic_Moji.py Moji

Running experiments

To launch our default models, please run the following commands:

• for the Bios dataset: python3 main.py Bios Bios
• for the Moji dataset: python3 main.py Moji Moji

To launch our models on the Bios dataset, using a demonic model pretrained on other domains, please use:

• for the demonic model pretrained on the EEC dataset:
  python3 main.py Bios EEC
• for the demonic model pretrained on the Marked Personas dataset:
  python3 main.py Bios marked_personas