📄 Professional Presentation and Projected Power: A Case Study of Implicit Gender Information in English CVs

Summary

The paper asks whether self-identified men and women write their CVs differently, and whether that difference is visible to classifiers once obvious gender cues are removed. The authors recruited US full-time workers on Prolific. Each participant built a CV for a hypothetical promotion, mostly by copying from their real CV, using five fixed free-text sections. The resulting corpus has about 1.8K English CVs across 16 occupations, which lets the analysis partly control for the fact that occupations differ in gender balance.

Three analyses follow:

• Word choice: a TFIDF comparison of words unique to each gender.
• Power and agency: Sap et al. (2017) connotation frames applied to transitive verbs, assuming the CV author is the agent.
• Classification: gender classifiers trained after balancing the data and redacting pronouns and named entities.

Women used equal-power and low-power (theme-power) verbs significantly more often than men. Gender also stayed predictable above chance in every setting, for both linear and transformer models. The practical lesson is that removing names and entities does not make CVs gender-neutral, which matters for both human and automated hiring.

Key points

• Data: 2,000 Prolific participants (half women) were hired, and 1,789 CVs remained after quality filtering (50.5% female). Each CV has five segments: an optional summary, experience, education, skills, and optional certifications. All preprocessing used SpaCy's default English models.
• Verb lexicon: the frame-labeled set of about 2K verbs from Sap et al. (2017) covers 48% of verb types in the CVs. In the other direction, 57% of the labeled verbs appear in the corpus.
• Regression: a logistic regression on 1,503 CVs (those with 10–100 verbs) controlled for CV length and occupation. Two effects survived Holm–Bonferroni correction:
  • equal-power verbs (β=10.78, p=0.006), e.g. perform, assess;
  • theme-power verbs (β=15.25, p<0.001), e.g. assist, serve.
• Effects that did not hold up: men used more high- and low-agency verbs, but this did not survive correction. Agent-power verbs showed no significant gender difference.
• Education sections: women's Education sections were about 15% longer on average (247 vs 214 words).
• Classifier features: even under occupation-balanced training, stereotyped features stayed predictive, such as child and assist for women and supervise and technology for men.

Interesting ideas

• I1. Genre-based default agent. Because CVs are written in fragments, the authors simply assign every verb's agent to the writer. This genre-based default makes role analysis possible without explicit subjects.
• I2. Power and agency as verb attributes. Connotation frames give verbs reader-facing attributes: relative power between agent and theme, and the agent's degree of agency.
• I3. Coverage measured both ways. The authors report how much of the corpus the lexicon covers and how much of the lexicon the corpus uses. This shows quickly whether a resource fits a domain.
• I4. Redaction ablations. Retraining after progressively removing pronouns and named entities shows that a signal survives the removal of obvious cues. This is a clean way to expose hidden shortcuts.
• I5. Confounds and multiple comparisons. The authors control for length and occupation, then apply Holm–Bonferroni correction. This separates robust effects from ones that disappear under multiple testing.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): telegraphic text such as recipes, instructions and résumé-style lists could get a genre rule that fills a missing agent with a known default participant, recorded as an explainable assumption.
• I2 (world): these labels could sit alongside VerbNet classes in the compiled knowledge, so the world model can track who gains or loses standing in an event.
• I3 (world): any newly compiled lexicon (WordNet- or FrameNet-derived) could get this two-way coverage check per test corpus before its outputs are trusted.
• I4 (learn): classifiers or repair rules could be evaluated on entity-redacted copies as a negative control, revealing whether a decision rests on surface names rather than structure.
• I5 (learn): when many candidate rules or features are compared at once, the same correction would stop lucky improvements from being promoted.