RaceBERT -- A transformer based model to predict race and ethnicty from names

Installation

pip install racebert

Using a virtual environment is highly recommended! You may need to install pytorch as instructed here: https://pytorch.org/get-started/locally/

raceBERT predicts race (U.S census race) and ethnicity from names.

from racebert import RaceBERT

model = RaceBERT()

# To predict race
model.predict_race("Barack Obama")

>>> {"label": "nh_black", "score": 0.5196923613548279}

The race categories are:

Race	Label
Non-hispanic White	nh_white
Hispanic	hispanic
Non-hispanic Black	nh_black
Asian & Pacific Islander	api
American Indian & Alaskan Native	aian

# Predict ethnicity
model.predict_ethnicty("Arjun Gupta")

>>> {"label": "Asian,IndianSubContinent", "score": 0.9612812399864197}

The ethnicity categories are:

Ethnicity
GreaterEuropean,British
GreaterEuropean,WestEuropean,French
GreaterEuropean,WestEuropean,Italian
GreaterEuropean,WestEuropean,Hispanic
GreaterEuropean,Jewish
GreaterEuropean,EastEuropean
Asian,IndianSubContinent
Asian,GreaterEastAsian,Japanese
GreaterAfrican,Muslim
Asian,GreaterEastAsian,EastAsian
GreaterEuropean,WestEuropean,Nordic
GreaterEuropean,WestEuropean,Germanic
GreaterAfrican,Africans

If you have a GPU, you can speed up the computation by specifying the CUDA device when you instantiate the model.

from racebert import RaceBERT

model = RaceBERT(device=0)

# predict race in batch
model.predict_race(["Barack Obama", "George Bush"])

>>>
[
        {"label": "nh_black", "score": 0.5196923613548279},
        {"label": "nh_white", "score": 0.8365859389305115}
]

# predict ethnicity in batch
model.predict_ethnicity(["Barack Obama", "George Bush"])

Alternatively, you can work with the transformers models hosted on the huggingface hub directly.

Please refer to the transformers documentation.