FairLangProc 0.1.3

A comprehensive set of fairness metrics, datasets and algorithms to address and mitigate bias in Natural Language Processing.

Fair Language Processing (FairLangProc)

The Fair Language Processing package is a extensible open-source Python library containing techniques developed by the research community to help detect and mitigate bias in Natural Language Processing throughout the AI application lifecycle.

The FairLangProc package includes:

1. Data sets to test for biases in NLP models.
2. Metrics based on different philosophies to quantified said biases.
3. Algorithms to mitigate biases. It has been created with the intention of encouraging the use of bias mitigation strategies in the NLP community, and with the hope of democratizing these tools for the ever-increasing set of NLP practitioners. We invite you to use it and improve it.

The companion paper provides a comprehensive introduction to the concepts and capabilities, with all code available in notebooks. Furthermore, we are working torwards a comprehensive documentation.

We have developed the package with extensibility in mind. This library is still in development. We encourage your contributions.

Supported fairness datasets

Data Set	Size	Reference
BBQ	58,492	Parrish et al., 2021
BEC-Pro	5,400	Bartl et al., 2020
BOLD	23,679	Dhamala et al., 2021
BUG	108,419	Levy et al., 2021
Crow-SPairs	1,508	Nangia et al., 2020
GAP	8,908	Webster et al., 2018
HolisticBias	460,000	Smith et al., 2022
HONEST	420	Nozza et al., 2021
StereoSet	16,995	Nadeem et al., 2020
UnQover	30	Li et al., 2020
WinoBias+	1,367	Vanmassenhove et al., 2021
WinoBias	3,160	Zhao et al., 2018
WinoGender	720	Rudinger et al., 2018

Supported fairness metrics

• Generalized association tests (WEAT) (Caliskan et al., 2016)
• Log Probability Bias Score (LPBS) (Kurita et al., 2019)
• Categorical Bias Score (CBS) (Ahn et al., 2021)
• CrowS-Pairs Score (CPS) (Nangia et al., 2020)
• All Unmasked Score (AUL) (Kaneko et al., 2021)
• Demographic Representation (DR) (Liang et al., 2022)
• Stereotypical Association (SA) (Liang et al., 2022)
• HONEST (Nozza et al., 2021)

Supported bias mitigation algorithms

• Counterfactual Data Augmentation (CDA) (Webster et al. 2020)
• Projection based debiasing (Bolukbasi et al., 2023)
• Bias removaL wIth No Demographics (BLIND) (Orgad et al., 2023)
• Adapter-based DEbiasing of LanguagE models (Lauscher et al., 2021)
• Modular Debiasing with Diff Subnetworks (Hauzenberger et al., 2023)
• Entropy Attention Temperature (EAT) scaling (Zayed et al., 2023)
• Entropy Attention Regularizer (EAR) (Attanasio et al., 2022)
• Embedding based regularizer (Liu et al., 2020)
• Selective unfreezing (Gira et al., 2024)

Setup

Python

Has been tested and ran with Python 3.13. Compatibility with older versions is possible and expected, although no tests have been run to check the possible configurations.

To install the latest stable version from PyPI, run:

pip install FairLangProc

Manual installation

Clone the latest version of this repository:

git clone https://github.com/arturo-perez-peralta/FairLangProc

Using FairLangProc

The notebooks directory contains a diverse collection of jupyter notebooks that showcase how to use the different processors, metrics and data sets. If you'd like to run the examples requiring , download the data sets now and place them in a folder named Fair-LLM-Benchmarks inside the 'FairLangProc/datsets' path or simply clone the repository from Gallegos et al.