TL;DR
This paper reveals significant gender-based differences in how job applicants write their resumes, with women using more diverse language and men tending to be more direct. AI models can predict gender from resumes with 80% accuracy, but this accuracy decreases with higher education and more standardized text. Crucially, while better language skills generally correlate with higher salary expectations, this "language premium" is amplified for men in male-dominated fields and diminished for women in female-dominated fields.
What Problem Does This Paper Address?
Gender disparities in the labor market are well-documented, extending from hiring to promotions and salary. While many studies attribute these gaps to stereotypes and discrimination, this paper investigates a less-explored area: how the language used in job application materials, specifically resumes, might contribute to these disparities. The authors aim to understand if and how men and women differ in their resume writing styles and how these differences, in turn, relate to salary expectations, particularly in the context of an increasingly AI-driven hiring landscape. They also explore how these language differences might be influenced by education and text standardization.
Methods
The researchers employed a multi-faceted approach:
- Linguistic Analysis: They analyzed resume text for various linguistic features, including sentence length, lexical diversity (number of unique words), and the prevalence of different word categories (nouns, verbs, adjectives) and syntactic structures.
- Word Embeddings: To capture higher-order semantic differences, they used a generative pre-trained transformer (GPT)-based model to create word embeddings. These embeddings represent words in a high-dimensional space, allowing for the visualization and analysis of semantic associations. Principal Component Analysis (PCA) was used to reduce the dimensionality for visualization.
- AI-based Gender Prediction: Two types of neural network models, GPT-based and LSTM-based, were trained on resume text to predict the gender of the applicant. This was done to quantify how much gender information is encoded in resume language.
- Regression Analysis: Ordinary Least Squares (OLS) regression was used to examine the relationship between resume language diversity (LD) and expected salary, controlling for demographic and job-related factors. They also investigated how the share of women in an occupation (SW) moderates this relationship.
Datasets
The primary dataset comprised 6.9 million resumes of Chinese job applicants collected from recruitment websites between 2014-2015. This dataset included information on gender, age, academic degree, occupation, expected salary, and a summary of work experience.
To supplement their analysis, they also used:
- Sina Weibo microblogs: Data from 900,000 users to compare text standardization requirements.
- China National Knowledge Infrastructure (CNKI) publications: Abstracts from 57,184 publications to further assess the impact of text standardization.
Key Insights and Ideas
- Gendered Language in Resumes: Men and women exhibit distinct patterns in their resume language, not just in simple metrics but also in high-level semantic structures.
- AI Can Detect Gender: Neural networks can predict an applicant's gender from their resume with a surprisingly high accuracy (80%), suggesting that gender is subtly encoded in written language.
- Education and Standardization Matter: The ability of AI to predict gender decreases with higher education levels and increased text standardization. This implies that more formal or educated writing styles might blur gender-specific linguistic cues.
- The "Language Premium" is Gendered: While better language skills generally lead to higher salary expectations, this benefit is not uniform. It's amplified for men in male-dominated fields and significantly weakened for women in female-dominated fields.
Key Results
- Resume Differences: Women tend to use longer sentences and a more diverse set of unique words, while men's resumes are generally shorter.
- Gender Prediction Accuracy: AI models achieved 80% accuracy in predicting gender from resumes. This accuracy dropped to 72% when datasets were balanced for demographic factors, indicating that even after controlling for these, language differences remain predictive.
- Education Impact: Gender prediction accuracy decreased with higher academic degrees. For instance, applicants with a Master's degree showed smaller gender differences in language metrics compared to those with secondary education.
- Text Standardization Impact: Gender prediction accuracy was highest for microblogs (least standardized), lower for resumes, and lowest for scientific papers (most standardized).
- Salary Expectations: Resume language diversity (LD) had a positive effect on expected salary. However, this effect was moderated by the share of women (SW) in an occupation. In occupations with a high SW, the positive relationship between LD and expected salary was diminished. For women, this effect was even more pronounced, while for men, it was reversed, suggesting a "language premium" that benefits men more in male-dominated fields.
Limitations and Future Work
The authors acknowledge several limitations:
- Representativeness of Data: The resume data, while large-scale, might not fully represent the formal labor market due to the demographics of users on recruitment platforms.
- Cultural Specificity: The study focuses on the Chinese labor market and Chinese language. Generalizability to other languages and cultural contexts requires further investigation.
- Expected vs. Actual Salary: The study uses expected salary, which captures self-assessment but not actual received salary.
- Causality: The study identifies correlations but cannot definitively establish causality between resume language differences and gender pay gaps.
Future work could involve:
- Field Experiments: Designing experiments to explore the causal mechanisms underlying these observations.
- Cross-Cultural Studies: Investigating these phenomena in different linguistic and cultural settings.
- Exploring Other Language Modalities: Examining gender differences in spoken language or other forms of written communication.
- Developing AI for Fairness: Using these findings to inform the development of AI tools that mitigate gender bias in hiring.
Summary updated