To classify litigant and judge names by gender and religion in our paper on In-group bias in the Indian judiciary, we developed and applied a machine classifier that reads name strings and assigns probabilities of names belonging to female/male or Muslim/non-Muslim individuals. Specifically, we trained a Character-level Bidirectional Long Short-Term Memory (LSTM) network model. The LSTM is a specific type of Recurrent Neural Network which reads over name characters and interprets them based on a “memory” of the history of characters.
We trained the LSTM classifier on two large underlying datasets containing millions of names mapped to identities. To classify gender, we use a dataset of 13.7 million names with labeled gender from the Delhi voter
rolls. To classify religion, we use a database of 1.4 million names with a religion label for individuals who sat for the National Railway Exam. The resulting model
to classify name strings by gender is saved in this zipped folder as
delhi_names_gender.hdf5. For those interested in using the underlying religion classifier, we request you to get in touch with us for access.
The gender classifier model and a sub-set of 10,000 random names have been provided in the classifier demonstration data packet to illustrate how the LSTM classifier built for our study works. These are called in by classifier/classify_names.do for name classification.
Additional details about the LSTM network, underlying training data and accuracy measures can be found in our paper here.