=Paper=
{{Paper
|id=Vol-2812/RDAI-2021_paper_4
|storemode=property
|title=Connecting Underrepresented Minorities and Qualified Job Positions Using Online Data
|pdfUrl=https://ceur-ws.org/Vol-2812/RDAI-2021_paper_4.pdf
|volume=Vol-2812
|authors=Maysa M G Macedo,Marisa Affonso Vasconcelos,Andrea Britto Mattos,Rogerio Abreu de Paula
}}
==Connecting Underrepresented Minorities and Qualified Job Positions Using Online Data==
https://ceur-ws.org/Vol-2812/RDAI-2021_paper_4.pdf
Connecting Underrepresented Minorities and Qualified Job Positions Using
Online Data
Maysa M G Macedo, Marisa Affonso Vasconcelos, Andrea Britto Mattos, Rogerio Abreu de Paula
IBM Research
Rua Tutoia, 1157
Sao Paulo, SP, Brazil, 04007-900
{mmacedo, marisaav, abritto, ropaula}@br.ibm.com,
Abstract and race aptitudes, which are learnt by algorithms that in-
Several studies previously demonstrated that underrepre-
gest hiring historical data and determines who should see
sented minority (URM) groups often struggle to access high- hiring openings. In this context, (Hardt, Price, and Srebro
qualified jobs. At the same time, a wide range of researches 2016) and (Peña et al. 2020) proposed solutions to miti-gate
also indicates that diversifying the work environment can bias for a supervised learning. However, without any correc-
bring a very positive impact for the company, in terms of tion, job postings, for example, may not be reaching certain
productivity and revenue. However, many companies still fail groups of people, in particular the underrepresented ones.
in filing their positions with diverse candidates. In this re-
search, we aim to investigate the gap between companies of- Our Proposal
fering qualified job opportunities and underrepresented mi-
nority groups and attempt to increase the digital connection
In this research, we postulate that while the challenge of hir-
between them by making the job posting process more attrac- ing underrepresented candidates for qualified jobs is many-
tive and reachable for URMs. fold, two aspects are particularly critical and have been
greatly affected by emerging AI and social-media technolo-
gies in the past years: namely, AI for candidate-job matching
Introduction and the use of social media for reaching out to target candi-
Twenty years ago, an analysis by (Richard 2000) concluded dates. On the one hand, discriminatory hiring practices as
that racial diversity affected business strategy by means of well as implicit biases negatively affect the ability of un-
increasing productivity, return on equity, and market perfor- derrepresented candidate applications to be identified and
mance. Since then, several articles and reports have pointed thus vetted. On the other hand, companies might not even
to the social and financial benefits of a more diverse work be able to reach out to the most qualified underrepresented
environment. To name a few, the study by (Boston Consult- candidates or might not be perceived as creating equal and
ing Group 2018) found that diverse companies generate 19% just opportunities for all, thus reducing their attractiveness
more revenue and the report by (McKinsey 2018) concluded to URM candidates with the required skill.
that gender diversity in management positions actually in- Our research goals are to address these two complemen-
creases profitability more than previously thought. Based on tary challenges that together undermine the hiring oppor-
these findings, companies started creating efforts to hire in tunities for underrepresented candidates as well as a com-
more inclusive ways. pany’s ability to reach out to them. We aim at taking the first
In parallel, access to quality work opportunity becomes concrete steps toward this vision by exploring both (i) at-
a life-changing opportunity for underrepresented minority tractiveness and (ii) reach of job postings for URM groups.
(URM) groups (be they, Blacks, Latinxs, Native-Americans, To this end, this work proposes to investigate and devise an
LGBTQIA+, low-income individuals, or others). Several are AI-based approach for identifying biased and inhibiting lan-
the barriers and hurdles that hinder or even prevent them guage in job postings and investigating the extent to which
from accessing as well as reaching higher quality work op- such job-postings reach out and eventually influence those
portunities. They face hiring biases inherent in the hiring se- URM groups. More specifically, we will investigate and ad-
lection processes and data as documented by the HR com- dress two main research questions described as follows.
munity elsewhere.
In this context, emerging technologies, in particular AI, How can technology help bridge the social distance
can help address hiring URMs (e.g., via algorithms for between underrepresented candidates and
people-opportunity matching), but they may also exacerbate job-offering companies?
the existing gap by carrying over historical and social biases Are URMs being reached by job postings? A certain social
inherent in the training data. For instance, referral and selec- group may be involved in local social networks, as described
tion practices tend to reinforce existing stereotypical gender by (Hofstra et al. 2017), that may be cut off from major job
Copyright ©c2021
Copyright 2021,
for Association
this paper by for the Advancement
its authors. of Artificial
Use permitted under advertisement clusters, making some job opportunities un-
Creative Commons
Intelligence License Attribution
(www.aaai.org). All rights 4.0reserved.
International (CC BY 4.0). reachable. By analyzing the job posting (social) graph in a
social network, we will be able to devise ways to reach dif- Data from Kaggle
ferent social groups. We will also make use of the social- NLP Burning Glass,
Company T Tisis Company’s HR,
graph as means to identify and determine specific social Company
Company
searching
searchingforT is
searchingfor for:
and dictionary of
group languages and determine the semantic social distance HR ___________ biased terms
between the social groups of which underrepresented can- Neural network
HR staff Job posting
didates are members and the companies offering qualified
Bias detection
jobs. Figure 1 depicts all these aspects of the investigation.
HR
Reach out analysis
URM
Bias report Spread post
Homophily
analysis
Figure 2: Scheme for bias mitigation in job postings. It starts
Company T is
Company
CompanyTTisis
searching
searching
with an HR member preparing a job post description which
searchingfor
for:________
for:______
__________ Keywords analysis will be the input data. The raw text is analyzed by an NLP
Social networks algorithm that identifies potentially problematic terms or ex-
HR staff Job posting
Best keyword set (based on homophily score) pressions. As training data, a dictionary created from the
URM groups reach phase can be used, as well as data from
Burning Glass, Kaggle, and the company’s HR. The analysis
outputs a report of such expressions so that the job posting
Figure 1: Scheme for reach out URM candidates. The input may be revised by an HR member. The revised job posting
for this solution would be job posting texts along with the may undergo the bias detection until the bias report outputs
information of the URM group being sought on a social net- that the text is OK. Finally, the revised job posting may be
work. The methodology comprises the choice of keywords spread on social media and other webpages.
that assists to define the target audience for the job post, and
this audience should be as diverse as possible. This choice
will be based on the calculation of the homophily score, References
which is described in (Karimi et al. 2018). Boston Consulting Group. 2018. How diverse leadership
teams boost innovation. https://www.bcg.com/publications/
2018/how-diverse-leadership-teams-boost-innovation.
How do job descriptions drive away Hardt, M.; Price, E.; and Srebro, N. 2016. Equality of Op-
underrepresented candidates? portunity in Supervised Learning. In Advances in Neural In-
It is well-recognized that particular languages convey spe- formation Processing Systems, volume 29, 3315–3323. Cur-
cific sets of social values that directly affect how a message ran Associates, Inc.
might be differently interpreted by distinct social groups. Hofstra, B.; Corten, R.; van Tubergen, F.; and Ellison, N.
For example, in seeking for a “ninja programmer”, which 2017. Sources of Segregation in Social Networks: A Novel
is widely perceived as a male-oriented attribute, a job post- Approach Using Facebook. American Sociological Review
ing conveys the idea of a male-oriented or male-preferred 82(3): 625–656.
work environment, thus reducing the likelihood of female
programmers to apply for that particular job offering. To Karimi, F.; Génois, M.; Wagner, C.; Singer, P.; and
what extent does a job posting carry, at times inconspic- Strohmaier, M. 2018. Homophily influences ranking of mi-
uously, implicit bias, or structural forms of discriminatory norities in social networks. Scientific Reports 8(1): 11077.
practices? We aim to evaluate AI-based technologies of NLP McKinsey. 2018. Delivering Through Diversity. https:
for automatically flagging biased or discriminatory language //www.mckinsey.com/business-functions/organization/our-
in job postings. In creating AI tools that can detect language insights/delivering-through-diversity#.
biases and prejudices, we will be able to devise an overar-
Peña, A.; Serna, I.; Morales, A.; and Fierrez, J. 2020. Fair-
ching solution for supporting more equitable and just hiring
CVtest Demo: Understanding Bias in Multimodal Learning
practices by recommending more appropriated languages as
with a Testbed in Fair Automatic Recruitment. In Proceed-
well as identifying ‘hot-spots’ of inappropriate job postings.
ings of the 2020 International Conference on Multimodal
Figure 2 shows in details our proposal for bias detection.
Interaction, ICMI ’20, 760–761. Association for Computing
Machinery.
Conclusion
Richard, O. C. 2000. Racial Diversity, Business Strat-
We believe that we have still a lot to advance in science and
egy, and Firm Performance: A Resource-Based View. The
technology to achieve equitable and just hiring practices. In
Academy of Management Journal 43(2): 164–177.
particular, we think that an approach that assesses and im-
proves the reaching out to underrepresented candidates has
potential to improve hiring processes and therefore increase
the diversity of the companies’ workforce.