NamSor AI Bias Estimator, Gender, Racial and Ethnic Fairness Toolkit
0 likes99 views
This document discusses NamSor, an AI tool that classifies names according to gender, ethnicity, country of origin, and other attributes. It can be used to audit algorithms for bias, but the classifications may oversimplify diversity. Limitations include how classifications may not fit all markets and regions due to complex, fractal human diversity between and within groups. The document advocates allowing people to self-identify beyond automated classifications.
NamSor AI Bias Estimator, Gender, Racial and Ethnic Fairness Toolkit
- 1. Expanding on Gender Diversity Report : NamSor algorithms for classification of names by Race/ethnicity or cultural origin/diasporas NamSor 1 2018-01
- 2. Gender, race/ethnicity or origin bias in AI ? Algorithms are used to assist human decision in funnel-based processes, ex. - recruitment, - credit allocation, AI especially used in the early stage of the selection process (ex resume sourcing or screening) : search, scoring, tagging Is the algorithm FAIR? 2
- 3. Estimating gender, racial/ethnic bias in algorithms ex. recruitment Two approaches : 1) Use Aequitas, an open source bias audit toolkit developed by the Center for Data Science and Public Policy at University of Chicago 2) Measure changes in diversity index (Shannon or Simpson) at each selective step What taxonomy for diversity analytics? What is race/ethnicity ? 3
- 4. NamSor sorts Names 4 Names reflect cultural Identity Since 2012, NamSor data mining software recognizes the linguistic or cultural origin of names in any alphabet / language, using both supervised and unsupervised machine learning (ie. clustering). 2014 : launch Gender API v1 2018 : software is re-written from scratch with standard ML frameworks : 1/ name embedding + neural networks 2/ na誰ve bayes classifier 2019 : launch NamSor API v2 with Gender, US Race/Ethnicity, Country/Origin/Diaspora classifiers
- 5. Our proud contribution to Gender Reports NamSor Gender API (v1) was used independently by both by Science-Metrix and Elsevier in 2015 and 2017 NamSor Gender API V2 was used for The Researcher Journey Through a Gender Lens and weve made specific improvements : Enhanced probability estimates for gender inference Improved support for East-Asian names (Chinese, Korean, Japanese) 5
- 6. Gender diversity is just one dimension, there are many other 6
- 7. An artistic illustration of ethnic diversity / diversity of origin among COVID-19 scientists Chinese sea at Ars Electronica 2020 by Dario Rodighiero (Harvard Metalab, https://github.com/rodighiero/COVID-19), Eveline Wandl-Vogt (Austrian Academy of Science) and Elian Carsenat (NamSor)
- 8. NamSor CORE taxonomies NamSor API* is available and already supports robust, fine-grained taxonomies for Gender US Race/Ethnicity Country/Origin Diaspora India Subclassification (States and Union Territories ISO 3166-2:IN) 8 * NamSor v2.0.16, 2021-10
- 9. 9 Classes Taxonomy Gender Male Female Field Example Description id ref12315 The input identifier firstName John The input given name / firstName lastName Smith The input family name / surname / lastName likelyGender male The likely gender : male or female probabilityCalibrated 0.99 The calibrated probability : 0.5 is Unknown, +1 is sure genderScale -0.99 The scale is -1..0..+1 and is based on the probability (Probability = 0.5 -> Scale = 0; Gender = Male & Probabilty = 1 -> Scale = -1; Gender = Female & Probability = 1 -> Scale = +1) score 41 A non calibrated Score (use Probability instead) : score = Math.log(getProbaFirst() / getProbaNotFirst()) maxed to 100 Gender classification model infers the likely gender, with probability :
- 10. 10 4 Classes or 6 classes* Taxonomy US Census Race/Ethnicity W_NL (White) B_NL (Black) HL (Hispano- Latino) A (Asian) Field Example Description id ref12315 The input identifier firstName Mary The input first name / given name lastName Cao The input last name / surname countryIso2 US The country of residence, the host country (ex. US, CA, NZ, GB) raceEthnicity A The likely 'race'/ethnicity : W_NL (white, non latino), HL (hispano latino), A (asian, non latino), B_NL (black, non latino) raceEthnicityAlt W_NL The best alternative 'race'/ethnicity raceEthnicitiesTop A, W_NL, ... The likely 'race'/ethnicities probabilityCalibrated 0.91 The calibrated probability of having guessed right the 'race'/ethnicity as A (Asian) probabilityCalibratedAlt 0.95 The calibrated probability of having guessed right the 'race'/ethnicity as either A or W_NL (White Non Latino) US Race/Ethnicity classifies names by race/ethnicity according to US Census taxonomy, along with probabilities. *add header X-OPTION-USRACEETHNICITY-TAXONOMY: USRACEETHNICITY-6CLASSES for two additional classes, AI_AN (American Indian or Alaskan Native) and PI (Pacific Islander)
- 11. 11 Classes Taxonomy Country IE DE ES MX id ref12315 The input identifier name Jing Cao The input full name country CN The likely residence country ISO2 code, which CAN include melting-pot countries countryAlt TW The best alternative residence country region Asia An arbitrary grouping of countries by topRegion/Region/subRegion topRegion Asia An arbitrary grouping of countries by topRegion/Region/subRegion subRegion Eastern Asia An arbitrary grouping of countries by topRegion/Region/subRegion countriesTop CN, TW, HK... The top 10 likely residence country ISO2 codes probabilityCalibrated .89 The calibrated probability of having guessed right the country of residence (CN) probabilityCalibratedAlt 0.92 The calibrated probability of having guessed right the country of residence as either CN or TW. Country classifies names to ~250 countries with valid ISO2 codes, from Ireland (IE) to Spain (ES) or Mexico (MX) including all African and Asian countries.
- 12. 12 Classes Taxonomy Origin IE DE ES PT id ref12315 The input identifier name Jing Cao The input full name country CN The likely residence country ISO2 code, which CAN include melting-pot countries countryAlt TW The best alternative residence country region Asia An arbitrary grouping of countries by topRegion/Region/subRegion topRegion Asia An arbitrary grouping of countries by topRegion/Region/subRegion subRegion Eastern Asia An arbitrary grouping of countries by topRegion/Region/subRegion countriesTop CN, TW, HK... The top 10 likely residence country ISO2 codes probabilityCalibrated .89 The calibrated probability of having guessed right the country of residence (CN) probabilityCalibratedAlt 0.92 The calibrated probability of having guessed right the country of residence as either CN or TW. Origin infers the likely country of origin from a name, based on naming patterns among ~130 countries with strong name identity (IE, DE, ES, PT etc.)
- 13. 13 Classes Taxonomy Diaspora Irish German Hispanic Chinese Field Example Description id ref12315 The input identifier firstName Mary The input first name / given name lastName Cao The input last name / surname countryIso2 US The country of residence, the host country (ex. US, CA, NZ, GB) ethnicity Chinese The likely ethnicity ethnicityAlt Vietnamese The best alternative ethnicity ethnicitiesTop Chinese, Vietnamese , Korean ... The top 10 likely ethnicities probabilityCalibrated 0.84 The calibrated probability of having guessed right the ethnicity as Chinese probabilityCalibratedAlt 0.85 The calibrated probability of having guessed right the country of residence as either Chinese or Vietnamese. Diaspora infers the likely ethnicity, diaspora or country of origin from a name, given a geographic context (ex. US, CA, ...) with ~130 ethnicities (Irish, Chinese, etc)
- 14. 14 Classes Taxonomy Subclassification (India) IN-AP Andhra Pradesh IN-AR Arunchal Pradesh IN-AS Assam Field Example Description id ref12315 The input identifier firstName Bhupen The input first name / given name lastName Borah The input last name / surname countryIso2 IN The country (initially only IN : India is supported) subClassification IN-AR The likely state/region subClassificationAlt IN-ML The best alternative state/region subClassificationTop IN-AR, IN- ML... The top 10 likely states/regions probabilityCalibrated 0.84 The calibrated probability of having guessed right the likely state/region as IN- AR probabilityCalibratedAlt 0.85 The calibrated probability of having guessed right the likely state/region as IN- AR or as IN-ML Subclassification infers the likely state/region (a sub-level of country). Initially this model is calibrated only for India (IN) States or Union Territories (ISO 3166-2:IN). We can expand this model to other countries, let us know.
- 15. Limitations to such taxonomies Human societies are fractal in their diversity : A coarse-grained classification model may not fit all markets (ex. African- American/Black vs. White vs. African / Black : how does North-African fit?) A fine-grained classification model may be too fine-grained or controversial in specific regions For example, IN/Indian is one class among 130 classes in our Origin/Diaspora taxonomy, but there are ~30 states in India with many ethnic/clan/caste system sub-groups 15 Liberia - a regional onomastics 'mille-feuille' Example of complex regional or ethnic identities in Africa : Liberia. This visualization utilizes unsupervised name classification algorithm, to recognize subgroups in different regions of Liberia. Privacy and self-identification : how can people override the classification ?
- 16. Thank you ! Elian CARSENAT, elian.carsenat@namsor.com Phone : +33 6 52 77 99 07 Try NamSor for yourself at, https://namsor.app/ 16