WEFE: The Word Embedding Fairness Evaluation Framework

Word Embedding Fairness Evaluation (WEFE) is an open source library for measuring an mitigating bias in word embedding models. It generalizes many existing fairness metrics into a unified framework and provides a standard interface for:

• Encapsulating existing fairness metrics from previous work and designing new ones.
• Encapsulating the test words used by fairness metrics into standard objects called queries.
• Computing a fairness metric on a given pre-trained word embedding model using user-given queries.

WEFE also standardizes the process of mitigating bias through an interface similar to the scikit-learn fit-transform. This standardization separates the mitigation process into two stages:

• The logic of calculating the transformation to be performed on the model (fit).
• The execution of the mitigation transformation on the model (transform).

The official documentation can be found at this link.

Installation

There are two different ways to install WEFE:

To install the package with pip

pip install wefe

• With conda:

To install the package with conda:

conda install -c pbadilla wefe

Requirements

These package will be installed along with the package, in case these have not already been installed:

1. numpy
2. scipy
3. scikit-learn
4. scipy
5. pandas
6. gensim
7. plotly
8. requests
9. tqdm
10. semantic_version

Contributing

You can download the code executing

git clone https://github.com/dccuchile/wefe

To contribute, visit the Contributing section in the documentation.

Development Requirements

To install the necessary dependencies for the development, testing and compilation of WEFE documentation, run

pip install -r requirements-dev.txt

Testing

All unit tests are in the wefe/tests folder. It uses pytest as a framework to run them.

To run the test, execute:

pytest tests

To check the coverage, run:

pytest tests --cov-report xml:cov.xml --cov wefe

And then:

coverage report -m

Build the documentation

The documentation is created using sphinx. It can be found in the docs folder at the root of the project. To compile the documentation, run:

cd docs
make html

Then, you can visit the documentation at docs/_build/html/index.html

Changelog

Version 0.4.1

• Fixed a bug where the last pair of target words in RIPA was not included.
• Added a benchmark that compares WEFE with another measurement and bias mitigation libraries in the documentation.
• Added a library changes since original paper release page in the documentation.

Version 0.4.0

• 3 new bias mitigation methods (debias) implemented: Double Hard Debias, Half Sibling Regression and Repulsion Attraction Neutralization.
• The library documentation of the library has been restructured. Now, the documentation is divided into user guide and theoretical framework The user guide does not contain theoretical information. Instead, theoretical documentation can be found in the conceptual guides.
• Improved API documentation and examples. Added multilingual examples contributed by the community.
• The user guides are fully executable because they are now on notebooks.
• There was also an important improvement in the API documentation and in metrics and debias examples.
• Improved library testing mechanisms for metrics and debias methods.
• Fixed wrong repr of query. Now the sets are in the correct order.
• Implemented repr for WordEmbeddingModel.
• Testing CI moved from CircleCI to GithubActions.
• License changed to MIT.

Version 0.3.2

• Fixed RNSB bug where the classification labels were interchanged and could produce erroneous results when the attributes are of different sizes.
• Fixed RNSB replication notebook
• Update of WEFE case study scores.
• Improved documentation examples for WEAT, RNSB, RIPA.
• Holdout parameter added to RNSB, which allows to indicate whether or not a holdout is performed when training the classifier.
• Improved printing of the RNSB evaluation.

Version 0.3.1

• Update WEFE original case study
• Hotfix: Several bug fixes for execute WEFE original Case Study.
• fetch_eds top_n_race_occupations argument set to 10.
• Preprocessing: get_embeddings_from_set now returns a list with the lost preprocessed words instead of the original ones.

Version 0.3.0

• Implemented Bolukbasi et al. 2016 Hard Debias.
• Implemented Thomas Manzini et al. 2019 Multiclass Hard Debias.
• Implemented a fetch function to retrieve gn-glove female-male word sets.
• Moved the transformation logic of words, sets and queries to embeddings to its own module: preprocessing
• Enhanced the preprocessor_args and secondary_preprocessor_args metric preprocessing parameters to an list of preprocessors preprocessors together with the parameter strategy indicating whether to consider all the transformed words ('all') or only the first one encountered ('first').
• Renamed WordEmbeddingModel attributes `model` and `model_name` to `wv` and `name` respectively.
• Renamed every run_query `word_embedding` argument to `model` in every metric.

Version 0.2.2

• Added RIPA metrics (thanks @stolenpyjak for your contribution!).
• Fixed Literal typing bug to make WEFE compatible with python 3.7.

Version 0.2.1

• Compatibility fixes.

Version 0.2.0

• Renamed optional `run_query` parameter `warn_filtered_words` to warn_not_found_words.
• Added `word_preprocessor_args` parameter to `run_query` that allow specifying transformations prior to searching for words in word embeddings.
• Added `secondary_preprocessor_args` parameter to `run_query` which allows specifying a second pre-processor transformation to words before searching them in word embeddings. It is not necessary to specify the first preprocessor to use this one.
• Implemented `__getitem__` function in `WordEmbeddingModel`. This method allows obtaining an embedding from a word from the model stored in the instance using indexers.
• Removed underscore from class and instance variable names.
• Improved type and verification exception messages when creating objects and executing methods.
• Fix an error that appeared when calculating rankings with two columns of aggregations with the same name.
• Ranking correlations are now calculated using pandas `corr` method.
• Changed metric template, name and short_names to class variables.
• Implemented `random_state` in RNSB to allow replication of the experiments.
• run_query now returns as a result the default metric requested in the parameters and all calculated values that may be useful in the other variables of the dictionary.
• Fixed problem with api documentation: now it shows methods of the classes.
• Implemented p-value for WEAT

Citation

Please cite the following paper if using this package in an academic publication:

P. Badilla, F. Bravo-Marquez, and J. Pérez WEFE: The Word Embeddings Fairness Evaluation Framework In Proceedings of the 29th International Joint Conference on Artificial Intelligence and the 17th Pacific Rim International Conference on Artificial Intelligence (IJCAI-PRICAI 2020), Yokohama, Japan.

Bibtex:

@InProceedings{wefe2020,
    title     = {WEFE: The Word Embeddings Fairness Evaluation Framework},
    author    = {Badilla, Pablo and Bravo-Marquez, Felipe and Pérez, Jorge},
    booktitle = {Proceedings of the Twenty-Ninth International Joint Conference on
               Artificial Intelligence, {IJCAI-20}},
    publisher = {International Joint Conferences on Artificial Intelligence Organization},
    pages     = {430--436},
    year      = {2020},
    month     = {7},
    doi       = {10.24963/ijcai.2020/60},
    url       = {https://doi.org/10.24963/ijcai.2020/60},
    }

Team

• Pablo Badilla.
• Felipe Bravo-Marquez.
• Jorge Pérez.
• María José Zambrano.

Contributors

We thank all our contributors who have allowed WEFE to grow, especially stolenpyjak and mspl13 for implementing new metrics.

We also thank alan-cueva for initiating the development of metrics for contextualized embedding models and harshvr15 for the examples of multi-language bias measurement.

Thank you very much 😊!