Implementation of WEAT (and WEFAT) (Caliskan et al., 2017).

To use:

First, visit https://nlp.stanford.edu/projects/glove/ to download the pre-trained word embeddings. The embeddings in weat_1.json and wefat_1.json use Common Crawl (840B tokens, 2.2M vocab, cased, 300d vectors, 2.03 GB). Download the file, unzip, and place the resulting txt file in the same directory. The file was ommitted in this repository due to Github space constraints.

Next install requirements.txt with pip install.

Then run with:

python main.py --data_file_name DATA_FILE_NAME --embedded_data_file_name EMBEDDED_DATA_FILE_NAME --glove_file_name GLOVE_FILE_NAME --wefat_association_file_name WEFAT_ASSOCIATION_FILE_NAME --test TEST --iterations N --distribution_type DISTRIBUTION_TYPE

where:
DATA_FILE_NAME: name of target/attribute data file. In this example, data/weat_1.json or data/wefat_1.json
EMBEDDED_DATA_FILE_NAME: name of target/attribute embeddings file; if it does not exist it will be created. In this example, data/weat_1_embedded.json or data/wefat_1_embedded.json
GLOVE_FILE_NAME: name of GloVe embeddings file; required if EMBEDDED_DATA_FILE_NAME file does not exist
WEFAT_ASSOCIATION_FILE_NAME: name of file that maps target to other statistic. In this example, data/wefat_1_percentage_women.json maps profession (target) to % women
TEST: WEAT or WEFAT
N: number of iterations to calculate p-value
DISTRIBUTION_TYPE: type of distribution to compute p-value

Displays the effect size and the p-value.

Displays a graph of test statistic vs other statistic (e.g. percentage of women in an occupation) and Pearson's correlation coefficient. if wefat_assocation_file_name is specified. Otherwise displays effect size and p-value for each target word.

Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183–186. https://doi.org/10.1126/science.aal4230

Pennington, J. (2014). GloVe: Global Vectors for Word Representation. Stanford.edu. https://nlp.stanford.edu/projects/glove/

Guide to Using Pre-trained Word Embeddings in NLP. (2021, June 1). Paperspace Blog. https://blog.paperspace.com/pre-trained-word-embeddings-natural-language-processing/#using-glove-word-embeddings

U.S. Bureau of Labor Statistics. (2019, January 18). Employed persons by detailed occupation, sex, race, and Hispanic or Latino ethnicity. Bls.gov. https://www.bls.gov/cps/cpsaat11.htm

Toney-Wails, A., & Caliskan, A. (2021). ValNorm Quantifies Semantics to Reveal Consistent Valence Biases Across Languages and Over Centuries. ArXiv:2006.03950 [Cs]. https://arxiv.org/abs/2006.03950

Caliskan, A. (2017). Replication Data for: WEFAT and WEAT(UNF:6:C9yaa+UeGfFmtuHz734iKw==, V2). [Data set]. Harvard Dataverse. https://doi.org/10.7910/DVN/DX4VWP