PaperlayerView exact revision on arXiv

Source-linked AI summary

Degendering Resumes for Fair Algorithmic Resume Screening

Prasanna Parasurama, João Sedoc

arXiv:2112.08910v3cs.CLcs.CY

TL;DR

The paper asks whether removing gendered information from resumes can mitigate potential bias in algorithmic screening. It tests gender measurement and successive resume-obfuscation methods alongside screening performance. It finds that lexicon-based removal can reduce gender information, but more complete obfuscation can compromise screening, while embedding gender-subspace removal is ineffective.

  • Problem

    Resume-screening models may learn gendered information from resumes even without an explicit gender input, against a backdrop of biased hiring data and potential model discrimination.

    Source: §1 · p. 1

  • Method

    The authors classify applicants’ self-reported gender from resumes, iteratively remove gender cues, and assess the trade-off with a BigBird resume-screening model.

    Source: §1 · p. 1, §3.3 · p. 3, §3.5 · p. 4

  • Results

    Names alone do not anonymize resumes; lexicon-based obfuscation reduces gender information only to a point before screening performance suffers, while removing gender subspaces from embeddings is ineffective.

    Source: §5 · p. 6

  • Takeaways & Limitations

    Because perfect gender obfuscation rendered the screening model useless, the authors call for active consideration of fairness constraints alongside debiasing methods.

    Source: §5 · p. 6

    The experiments found no way to completely hide gender without significant screening costs; perfect obfuscation rendered the screening model useless.

    Source: §5 · p. 6

Abstract

from arXiv · show

We investigate whether it is feasible to remove gendered information from resumes to mitigate potential bias in algorithmic resume screening. Using a corpus of 709k resumes from IT firms, we first train a series of models to classify the self-reported gender of the applicant, thereby measuring the extent and nature of gendered information encoded in resumes. We then conduct a series of gender obfuscation experiments, where we iteratively remove gendered information from resumes. Finally, we train a resume screening algorithm and investigate the trade-off between gender obfuscation and screening algorithm performance. Results show: (1) There is a significant amount of gendered information in resumes. (2) Lexicon-based gender obfuscation method (i.e. removing tokens that are predictive of gender) can reduce the amount of gendered information to a large extent. However, after a certain point, the performance of the resume screening algorithm starts suffering. (3) General-purpose gender debiasing methods for NLP models such as removing gender subspace from embeddings are not effective in obfuscating gender.

1 Introduction

The paper asks whether removing gendered information from resumes can mitigate potential bias in resume screening. It links this concern to models learning gender cues and to bias in both training data and models.

1 Introduction

  • Resume-screening models can learn gendered information even when gender is not explicitly included, as illustrated by Amazon’s reported system rating male resumes higher.The passage says historical hiring data taught the model that men were more likely to be hired; men were also more likely to use words such as “executed” and “captured”.Source: §1 · p. 1
  • System bias depends on both biased hiring data and a model’s ability to distinguish genders, and sophisticated models may use gendered cues such as hobbies when predicting applications.Source: §1 · p. 1
  • The paper investigates whether resumes can be degendered to mitigate potential bias, first measuring gendered information, then testing obfuscation methods and screening-performance trade-offs.The experiments iteratively remove gender identifiers, gender-indicating words, hobbies, gender-predictive features, or gender subspaces in embeddings while seeking to preserve job-relevant content.Source: §1 · p. 1

2 Related Work

Prior work addresses societal stereotypes in general-purpose NLP models, while resume screening also raises the distinct issue of hiring bias encoded in training data. The literature also includes fairness methods that constrain decisions or outcomes rather than fully debiasing models.

2 Related Work

  • General-purpose NLP debiasing targets societal stereotypes learned from corpora, including occupational associations such as “computer programmer” with “man” and “homemaker” with “woman” (Bolukbasi et al., 2016).Source: §2 · p. 2
  • Debiasing approaches alter training data or learning algorithms; this paper’s approach belongs to the data-based category (Sun et al., 2019).Examples in prior work include gender obfuscation or swapping, removing gender subspaces, constraints, and adversarial learning (Zhao et al., 2018a; Bolukbasi et al., 2016; Zhao et al., 2018b; Zhou and Bansal, 2020).Source: §2 · p. 2
  • General-purpose debiased embeddings may help with gendered job-description language, but are unlikely to address hiring bias that teaches models to use job-irrelevant proxies such as hobbies or writing style.Fairness methods can instead constrain decisions or outcomes, as in gender-distribution constraints for LinkedIn Recruiter search results (Geyik et al., 2019).Source: §2 · p. 2

3 Data and Methods

The methods use applicant-tracking data from U.S. IT firms to construct resume representations, matched samples, and gender classifiers. The authors then test successive obfuscation procedures and train a screening model using resume and job information.

3 Data and Methods

  • The dataset contains applicant resumes and application outcomes from eight U.S. IT firms, alongside applicant attributes and the jobs to which applicants applied.The aggregated ATS data include names, gender, experience, degree, field of study, and callback outcomes.Source: §3 · p. 2

3.1 Vector Representations of Resumes

  • Each resume’s skills-space representation averages Word2Vec vectors for dictionary-filtered skills, competencies, job titles, and job-relevant keywords.The dictionary was assembled from skills and job titles in a secondary LinkedIn dataset; Figure 1 illustrates the resume vector representation.Source: §3.1 · p. 2, §3.2 · p. 2, Figure 1 · p. 3

3.2 Matched Sample of Resumes

  • To control observable occupational characteristics, the authors matched male and female resumes on experience, degree, field of study, and resume similarity, yielding 348,000 resumes.Matching used 1-to-1 pairing without replacement, required resumes to be within two years of experience and have cosine similarity of at least 0.7, and selected the highest-similarity eligible match.Source: §3.2 · pp. 2–3

3.3 Measuring Gendered Information in Resumes – Gender Classifier

  • Gendered information is operationalized as anything predictive of applicant gender, measured by training resume classifiers and testing predictive power on a holdout set.The models include Tf-Idf with logistic regression, BigBird, and word embeddings with logistic regression; the data use an 80/10/10 train/evaluation/test split.Source: §3.3 · p. 3

3.4 Gender-Predictive Features – SHAP Values

  • SHAP estimates feature importance by measuring how perturbing a resume feature changes predictions, with instance-level values aggregated across resumes for model-level importance.Source: §3.4 · pp. 3–4

3.5 Obfuscating Gendered information

  • Obfuscation removes names and contact identifiers, gender-indicating words, hobbies, and iteratively selected gender-predictive features, then compares debiased with off-the-shelf embeddings.The sequence is designed to investigate obfuscation while preserving job-relevant content.Source: §3.5 · p. 4

3.6 Trade-off Between Gender Obfuscation and Screening Algorithm Performance

  • The screening model uses BigBird to predict callbacks from resume text plus selected job details, including company, role, location, employment type, skills, and keywords.The authors omit full job descriptions to avoid lengthy boilerplate, instead concatenating ATS job characteristics with resume text as model input.Source: §3.6 · p. 4

4 Results

Gender-classification AUROC declines as resume cues are removed, but BigBird recovers some signal and embedding debiasing changes little. The paper also evaluates screening performance across progressively larger feature removals.

Gender Obfuscation Experiments

  • AUROC falls from 0.88 without obfuscation to 0.75 after removing names, gender-indicating words, and hobbies; BigBird raises it from 0.75 to 0.8.The matched-sample baseline is 0.85; the experiments use a bag-of-words Tf-Idf model except where BigBird is substituted.Source: Gender Obfuscation Experiments · p. 5
  • Removing 100 to 40,000 SHAP-ranked gender features drives classification toward AUROC=0.5, but each hold-out set is scored with the fixed Model 6 classifier.The authors do not retrain the classifier at each removal step.Source: Gender Obfuscation Experiments · p. 5
  • Debiased embeddings reduce AUROC only slightly, from 0.68 with the base embeddings to 0.67 with debiased embeddings.The debiased embeddings are compared with the off-the-shelf model on which they are based (Bolukbasi et al., 2016).Source: Gender Obfuscation Experiments · p. 5

4.1 Trade-off Between Gender Obfuscation and Screening Performance

  • Figure 2 evaluates screening and gender classification on hold-out sets spanning removal of 100 to 40,000 predictive features.The leftmost point uses 100 removed features and the rightmost uses 40,000.Source: §4.1 · p. 5, Figure 2 · p. 5

5 Conclusion

The conclusion reports that resumes retain gender signal after basic anonymization, while stronger obfuscation trades off against screening performance. It argues that fairness constraints remain relevant alongside feature obfuscation.

5 Conclusion

  • Gendered information remains substantial: after removing names, gender-indicating words, and hobbies, gender classification reaches AUC=0.75 with Tf-Idf and AUC=0.82 with BigBird.Lexicon-based obfuscation can reduce gender information up to AUC=0.7 without screening-performance loss, but experiment 11 achieves perfect obfuscation while rendering screening useless.Source: §5 · p. 6
  • Removing names alone does not anonymize resumes from algorithms, which can use gender proxies and propagate bias in hiring data.The conclusion notes that commercial hiring algorithms provide little information about how they handle training-data bias.Source: §5 · p. 6
  • The authors call for explicit fairness considerations, such as fairness constraints, alongside rather than instead of debiasing methods.They describe the two approaches as compatible.Source: §5 · p. 6

A Model Specifications

The appendix specifies classifiers and training choices used in the paper's model comparisons.

A Model Specifications

B List of Gender Indicating Words

The appendix lists gender-indicating terms spanning gender labels, pronouns, kinship terms, and gendered occupational words.

B List of Gender Indicating Words

C Sample Input Instance for the Screening Model

Figure 3 presents a sample input data instance for the screening model.

C Sample Input Instance for the Screening Model

  • Figure 3 presents a sample input data instance for the screening model.Source: Figure 3 · p. 9
Loading 2112.08910v3…