Are Female Carpenters like Blue Bananas? A Corpus Investigation of Occupation Gender Typicality

Da Ju, Karen Ullrich, Adina WilliamsACL 2024 · DOI

Abstract

In linguistics, the "typicality effect" suggests that attributes are specified only when they deviate from the norm—we say "blue banana" because bananas are usually yellow. This paper investigates whether this logic applies to gender in occupational discourse. By analyzing U.S. Bureau of Labor Statistics alongside massive text corpora from Reddit and Wikipedia, we demonstrate that the "blue banana" hypothesis fails. Instead of mentioning gender only when it is surprising (e.g., "female carpenter"), gender is mentioned significantly more frequently in woman-dominated occupations across the board. This suggests that female-dominated fields are subject to higher levels of social salience and policing.

Introduction

Does language use follow a rule of "surprisingness" regarding gender? If a "carpenter" is stereotypically male, do we only specify gender when the individual is female? Previous theories suggested that mentioning gender serves as a marker of atypicality. This study tests this "blue banana" hypothesis against an alternative: that certain occupations, specifically those dominated by women, are simply more "saliently gendered" in collective discourse, leading to higher rates of gender specification regardless of typicality.

Methodology

The researchers employed a multi-faceted approach combining information theory and corpus linguistics:

  • Information Theoretic Markers: They used Conditional Mutual Information (CMI) to quantify how much an occupation predicts a gender mention.
  • Correlation Analysis: Pearson correlations were calculated between the empirical "femaleness" of an occupation (from 2023 U.S. Labor Statistics) and the frequency of gendered adjectives in text.
  • Corpus Comparison: The study utilized the Pushshift.io Reddit corpus (118B tokens) for informal speech and Wikipedia (3B tokens) for formal discourse.
  • Synthetic Baseline: Llama 2 70B was used to regenerate Wikipedia articles to observe how LLMs handle these linguistic patterns.

Key Results

The central finding of the study is the failure of the "blue banana" hypothesis. If gender mentioning were driven by surprise, we would expect high "male" mentions in woman-dominated jobs and high "female" mentions in man-dominated jobs. However, the data reveals a strong positive correlation between an occupation's femaleness and the frequency of all gender mentions.

Figure 1: Correlation between occupation femaleness and the frequency of gender mentioning on Reddit.

As shown in Figure 1, occupations with higher proportions of women (like nurse, hairdresser, and teacher) show a much higher probability of gender being mentioned ($r = 0.51$). This correlation persists whether the mention is "female" or "male," indicating that these roles are viewed through a more gendered lens than male-dominated roles like "carpenter" or "constructor."

Reddit vs. Wikipedia

The intensity of this effect varies significantly by corpus. Informal spaces like Reddit exhibit much stronger gender-typicality biases than curated platforms like Wikipedia.

Figure 4: Mutual information between gender and its mention across Wikipedia and Reddit, highlighting specific occupational outliers.

Figure 4 illustrates that the Mutual Information (MI) between an occupation and gender mentions is an order of magnitude higher for "female-coded" occupations (e.g., hairdresser, housekeeper) on Reddit compared to Wikipedia. On Reddit, these occupations are extreme outliers, suggesting that informal discourse is where social policing and gender-based commentary are most prevalent.

Discussion: Asymmetric Gendering

The results support the concept of "asymmetric gendering," where "man" is treated as the default human category, while "woman" is viewed as an additional semantic attribute. Consequently, woman-dominated fields are seen as "more gendered," leading to more frequent specification of gender even when it matches the statistical norm.

A qualitative audit of Reddit threads further revealed that mentions of "male nurses" or "female teachers" were often embedded in derogatory, offensive, or sexually explicit contexts, or discussions focused on gender balance. This suggests that the high frequency of gender mentions in these fields is not merely for clarity but is often a byproduct of social friction and bias.

Conclusion

This investigation concludes that gender mentioning in occupations is driven by social salience rather than atypicality. While LLMs like Llama 2 show "cleaner" data with fewer explicit correlations—likely due to RLHF—the underlying human tendency to over-gender woman-dominated roles remains a significant feature of natural language. Future work should explore these dynamics in languages with grammatical gender to see if the "masculine default" further amplifies these effects.

Summary updated

BibTeX
@inproceedings{ju2024are,
  author = {Da Ju and Karen Ullrich and Adina Williams},
  title = {{Are Female Carpenters like Blue Bananas? A Corpus Investigation of Occupation Gender Typicality}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.findings-acl.253},
}
Figure