Welcome to the WEFE documentation!

WEFE: The Word Embeddings Fairness Evaluation Framework is an open source library for measuring and mitigating bias in word embedding models.

The following pages contain the documentation about WEFE: how to install the package, how to use it and how to contribute, as well as the detailed API documentation and extensive examples.

Getting Started

• About
  • Motivation and objectives
  • Similar Packages
  • Citation
  • Roadmap
  • License
  • Team
  • Acknowledgments
• Quick Start
  • Download and setup
  • Run your first Query

User Guide

• Bias Measurement
  • Run a Query
  • Run Query Arguments
  • Out of Vocabulary Words and Word Preprocessors
  • Running Multiple Queries
  • Aggregating Results
  • Model Ranking
• Bias Mitigation (Debias)
  • Hard Debias
  • Multiclass Hard Debias
• Loading Embeddings From Different Sources
  • Create a example query
  • Load from Gensim API
  • Using Gensim Load
  • Flair

API

• WEFE API
  • WordEmbeddingModel
  • Query
  • Metrics
  • Debias
  • Datasets
  • Preprocessing
  • Utils

Conceptual Guides

• Measurement Framework
  • Target set
  • Attribute set
  • Query
  • Query Template
  • Fairness Measure
  • Standard usage pattern of WEFE
  • Metrics Implemented So Far
• Mitigation Framework
  • Fit method
  • Transform method
  • Mitigation Methods Implemented So Far
• References
  • Measurements and Case Studies
  • Bias Mitigation
  • Surveys and other resources

Examples

• WEFE Case Study Replication
• Previous Studies Replication
  • Semantics derived automatically from language corpora contain human-like biases (WEAT)
  • A transparent framework for evaluating unintended demographic bias in word embeddings (RNSB)
• Multilingual Gender Bias Measurement Examples

Contributing

• Contributing
  • Get the repository
  • Testing
  • Build the documentation
  • How to implement your own metric
  • Mitigation Method Implementation Guide

Repository

You can find the project repository at the following link: WEFE repository on Github.