The ability to programmatically and reliably infer the social attributes of a person from their name can be useful for a broad set of tasks, from estimating bias in coverage of women in the media to estimating bias in lending against certain social groups. But unlike the American Census Bureau, which produces a list of last names and first names, which can (and are) used to infer the gender, race, ethnicity, etc., from names, the Indian government produces no such commensurate datasets. Hence inferring the relationship between gender, ethnicity, language group, etc., and names has generally been done with small datasets constructed in an ad-hoc manner.
We fill this yawning gap. Using data from the Indian Electoral Rolls (parsed data here), we estimate the proportion female, male, and third sex (see here) for a particular first name, year, and state.
Please also check out pranaam that uses land record data from Bihar to infer religion based on the name. The package uses indicate to transliterate Hindi to English.
In all, we capitalize on information in the parsed electoral rolls from the following 31 states and union territories:
Andaman | Delhi | Kerala | Puducherry |
Andhra Pradesh | Goa | Madhya Pradesh | Punjab |
Arunachal Pradesh | Gujarat | Maharashtra | Rajasthan |
Assam | Haryana | Manipur | Sikkim |
Bihar | Himachal Pradesh | Meghalaya | Tripura |
Chandigarh | Jammu and Kashmir | Mizoram | Uttar Pradesh |
Dadra | Jharkhand | Nagaland | Uttarakhand |
Daman | Karnataka | Odisha |
We split the name into first name and last name (see the Python notebook for how we do this) and then aggregate per state and first_name, and tabulate prop_male, prop_female, prop_third_gender, n_female, n_male, n_third_gender. We produce native language rolls and English transliterations. (We use indicate to produce transliterations for Hindi rolls.)
This is used to provide the base prediction.
Given the association between prop_female and first_name may change over time, we exploited the age. Given the data were collected in 2017, we calculated the year each person was born and then did a group by year to create prop_male, prop_female, prop_third_gender, n_female, n_male, n_third_gender
Concerns:
- Voting registration lists may not be accurate, systematically underrepresenting poor people, minorities, and similar groups.
- Voting registration lists are, at best, a census of adult citizens. But to the extent there is prejudice against women, etc., that prevents them from reaching adulthood, the data bakes those biases in.
- Indian names are complicated. We do not have good parsers for them yet. We have gone for the default arrangement. Please go through the notebook to look at the judgments we make. We plan to improve the underlying data over time.
- For state electoral rolls that are neither in English nor Hindi, we use libindic. The quality of transliterations is consistently bad.
We start by providing a base model for first_name that gives the Bayes optimal solution---the proportion of people with that name who are women. We also provide a series of base models where the state of residence and year of birth is known.
If the name does not exist in the database, we use ML model that uses the relationship between sequences of characters in the first name and gender to predict gender from the name.
The model was trained as a regression problem instead of a classification problem because men and women share names. (See the histogram below for the female proportion for the dataset.) The model predicts the female proportion of the name. If it is less than 0.5, we classify it as male; otherwise, we classify it as female.
Test data
MSE no weights - loss: .05, metric: 0.05
RMSE no weights - loss: 0.22, metric: 0.22
Test data with weights
MSE with weights - loss: 0.05, metric: 0.04
RMSE with weights - loss: 0.22, metric: 0.22
Below are the inference results using different models.
We strongly recommend installing naampy inside a Python virtual environment (see venv documentation)
pip install naampy
usage: in_rolls_fn_gender [-h] -f FIRST_NAME [-s {andaman,andhra,arunachal,assam,bihar,chandigarh,dadra,daman,delhi,goa,gujarat,haryana,himachal,jharkhand,jk,karnataka,kerala,maharashtra,manipur,meghalaya,mizoram,mp,nagaland,odisha,puducherry,punjab,rajasthan,sikkim,tripura,up,uttarakhand}] [-y YEAR] [-o OUTPUT] input Appends Electoral roll columns for prop_female, n_female, n_male n_third_gender by first name positional arguments: input Input file optional arguments: -h, --help show this help message and exit -f FIRST_NAME, --first-name FIRST_NAME Name or index location of column contains the first name -s {andaman,andhra,arunachal,assam,bihar,chandigarh,dadra,daman,delhi,goa,gujarat,haryana,himachal,jharkhand,jk,karnataka,kerala,maharashtra,manipur,meghalaya,mizoram,mp,nagaland,odisha,puducherry,punjab,rajasthan,sikkim,tripura,up,uttarakhand}, --state {andaman,andhra,arunachal,assam,bihar,chandigarh,dadra,daman,delhi,goa,gujarat,haryana,himachal,jharkhand,jk,karnataka,kerala,maharashtra,manipur,meghalaya,mizoram,mp,nagaland,odisha,puducherry,punjab,rajasthan,sikkim,tripura,up,uttarakhand} State name of Indian electoral rolls data (default=all) -y YEAR, --year YEAR Birth year in Indian electoral rolls data (default=all) -o OUTPUT, --output OUTPUT Output file with Indian electoral rolls data columns choices=["v1", "v2", "v2_1k", "v2_native", "v2_en"],
>>> import pandas as pd >>> from naampy import in_rolls_fn_gender >>> names = [{'name': 'gaurav'}, {'name': 'nabha'}, {'name': 'yasmin'}, {'name': 'deepti'}, {'name': 'hrithik'}, {'name': 'vivek'}] >>> df = pd.DataFrame(names) >>> in_rolls_fn_gender(df, 'name') name n_male n_female n_third_gender prop_female prop_male prop_third_gender pred_gender pred_prob 0 gaurav 25625.0 47.0 0.0 0.001831 0.998169 0.0 NaN NaN 1 nabha NaN NaN NaN NaN NaN NaN female 0.755028 2 yasmin 58.0 6079.0 0.0 0.990549 0.009451 0.0 NaN NaN 3 deepti 35.0 5784.0 0.0 0.993985 0.006015 0.0 NaN NaN 4 hrithik NaN NaN NaN NaN NaN NaN male 0.922181 5 vivek 233622.0 1655.0 0.0 0.007034 0.992966 0.0 NaN NaN >>> help(in_rolls_fn_gender) Help on method in_rolls_fn_gender in module naampy.in_rolls_fn: in_rolls_fn_gender(df, namecol, state=None, year=None) method of builtins.type instance Appends additional columns from Female ratio data to the input DataFrame based on the first name. Removes extra space. Checks if the name is the Indian electoral rolls data. If it is, outputs data from that row. Args: df (:obj:`DataFrame`): Pandas DataFrame containing the first name column. namecol (str or int): Column's name or location of the name in DataFrame. state (str): The state name of Indian electoral rolls data to be used. (default is None for all states) year (int): The year of Indian electoral rolls to be used. (default is None for all years) Returns: DataFrame: Pandas DataFrame with additional columns:- 'n_female', 'n_male', 'n_third_gender', 'prop_female', 'prop_male', 'prop_third_gender' by first name # If you want to use model prediction use `predict_fn_gender` like below from naampy import predict_fn_gender input = [ "rajinikanth", "harvin", "Shyamsingha", "srihan", "thammam", "bahubali", "rajarajeshwari", "shobby", "tamannaah bhatia", "mehreen", "kiara", "shivathmika", "komalee", "nazriya", "nabha", "taapsee", "parineeti", "katrina", "ileana", "vishwaksen", "sampoornesh", "hrithik", "emraan", "rajkummar", "sharman", "ayushmann", "irrfan", "riteish" ] print(predict_fn_gender(input)) name pred_gender pred_prob 0 rajinikanth male 0.994747 1 harvin male 0.840713 2 shyamsingha male 0.956903 3 srihan male 0.825542 4 thammam female 0.564286 5 bahubali male 0.901159 6 rajarajeshwari female 0.942478 7 shobby male 0.788314 8 tamannaah bhatia female 0.971478 9 mehreen female 0.659633 10 kiara female 0.614125 11 shivathmika female 0.743240 12 komalee female 0.901051 13 nazriya female 0.854167 14 nabha female 0.755028 15 taapsee female 0.665176 16 parineeti female 0.813237 17 katrina female 0.630126 18 ileana female 0.640331 19 vishwaksen male 0.992237 20 sampoornesh male 0.940307 21 hrithik male 0.922181 22 emraan male 0.795963 23 rajkummar male 0.845139 24 sharman male 0.858538 25 ayushmann male 0.964895 26 irrfan male 0.837053 27 riteish male 0.950755
When you first run in_rolls_fn_gender, it downloads data from Harvard Dataverse to the local folder. Next time you run the function, it searches for local data and if it finds it, it uses it. Use predict_fn_gender to get gender predictions based on first name.
Suriyan Laohaprapanon, Gaurav Sood, and Rajashekar Chintalapati
The package is released under the MIT License.