This repository contains the code, data, and supplementary material for the experiments included in the paper 'Explaining text classifiers with counterfactual representations', accepted at ECAI 2024 (to appear).

Create and start a new virtual environment:

conda create -n CFR python=3.10.9 anaconda
conda activate CFR

Download and pre-process the data before running the experiments.

• EEEC+: data and pre-processing in "./datasets/EEEC/EEEC_3race"
• BiasInBios: data and pre-processing in "./datasets/biasbios"
• CEBaB:
  • data: https://cebabing.github.io/CEBaB/
  • pre-processing in "./datasets/CEBaB-v1.1"
• GloVe :
  • data:
    • https://nlp.biu.ac.il/~ravfogs/rlace/glove/glove-gender-data.pickle (GloVe embeddings with gender-bias labels)
    • https://nlp.biu.ac.il/~ravfogs/rlace/glove/glove-top-50k.pickle (150k GloVe embeddings)

Run the notebooks:

• CFRs_EEECp_gender_balanced.ipynb
• CFRs_EEECp_gender_aggressive.ipynb
• CFRs_EEECp_race_balanced.ipynb
• CFRs_EEECp_race_aggressive.ipynb

Run the notebook:

• CFRs_biasbios.ipynb

Run:

• CFRs_CEBaB_compare_methods.ipynb

Run the notebook:

• CFRs_GloVe.ipynb