Source codes for EMNLP 2022 paper "Balancing out Bias: Achieving Fairness Through Balanced Training"
If you use the code, please cite the following paper:
@inproceedings{han-etal-2022-balancing,
title = "Balancing out Bias: Achieving Fairness Through Balanced Training",
author = "Han, Xudong and
Baldwin, Timothy and
Cohn, Trevor",
booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
month = dec,
year = "2022",
address = "Abu Dhabi, United Arab Emirates",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.emnlp-main.779",
pages = "11335--11350",
abstract = "Group bias in natural language processing tasks manifests as disparities in system error rates across texts authorized by different demographic groups, typically disadvantaging minority groups. Dataset balancing has been shown to be effective at mitigating bias, however existing approaches do not directly account for correlations between author demographics and linguistic variables, limiting their effectiveness. To achieve Equal Opportunity fairness, such as equal job opportunity without regard to demographics, this paper introduces a simple, but highly effective, objective for countering bias using balanced training.We extend the method in the form of a gated model, which incorporates protected attributes as input, and show that it is effective at reducing bias in predictions through demographic input perturbation, outperforming all other bias mitigation techniques when combined with balanced training.",
}
In this work, we first propose a framework to balance demographic groups within each target class in training, and then present a demographic-augmented model, which further improves fairness.
The model is implemented using PyTorch and FairLib.
tqdm==4.62.3
numpy==1.22
docopt==0.6.2
pandas==1.3.4
scikit-learn==1.0
torch==1.10.0
PyYAML==6.0
seaborn==0.11.2
matplotlib==3.5.0
pickle5==0.0.12
transformers==4.11.3
sacremoses==0.0.53
Alternatively, you can install the fairlib directly:
pip install fairlib
from fairlib import datasets
datasets.prepare_dataset("moji", "data/deepmoji")
datasets.prepare_dataset("bios", "data/bios")
For a particular balanced training objective, e.g., joint balance, both adjusting instances weights and sampling can be employed.
Our implementations are presented in src\balanced_training.py
, where the get_weights
function and get_sampled_indices
correspond to instances reweighting and resampling respectively.
Specifically,
get_weights
: Given the balanced training objective, target labels, and protected labels, pre-calculate weights for each instance.get_sample_indices
: Given the balanced training objective, target labels, and protected labels, sampling instances for each group.
The implemented objectives includes:
- y: balance the target label distribution, which is typically used for long-tail learning
- g: balance the protected label distributions
- joint jointly balance protected attributes and target labels
- stratified_y balance protected labels within each target class
- stratified_g balance target labels within in each protected group
- EO maintain as much instances as possible for downsampling but degrade to stratified_y for reweighting and resampling.
The gated model learns a shared encoder and one extra encoder for each demographic groups.
Our implementation is presented in augmentation_layer.py
, which augment the naive encoder with demographic-specific encoders.
When using the augmentation layer, it takes the same inputs as the shared encoder along with groups labels, its outputs then will be added to the outputs of the main model to augment demographic information.
# Augmentation
if self.args.gated and self.args.n_hidden > 0:
assert group_label is not None, "Group labels are needed for augmentation"
specific_output = self.augmentation_components(input_data, group_label)
main_output = main_output + specific_output
All experiments can be reproduced by using fairlib.
-
python fairlib --BT Reweighting --BTObj joint
-
python fairlib --BT --BTObj joint --gated