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

Abstract: People tend to use language to mention surprising properties of events: for example, when a banana is blue, we are more likely to mention color than when it is yellow. This fact is taken to suggest that yellowness is somehow a typical feature of bananas, and blueness is exceptional. Similar to how a yellow color is typical of bananas, there may also be genders that are typical of occupations. In this work, we explore this question using information theoretic techniques coupled with corpus statistic analysis. In two distinct large corpora, we do not find strong evidence that occupations and gender display the same patterns of mentioning as do bananas and color. Instead, we find that gender mentioning is correlated with femaleness of occupation in particular, suggesting perhaps that woman-dominated occupations are seen as somehow ``more gendered'' than male-dominated ones, and thereby they encourage more gender mentioning overall.

• The paper finds that gender mentioning correlates significantly with occupational gender distributions, particularly in roles dominated by women.
• It employs NLP-based corpus analysis across datasets like US Labor Statistics, Reddit, and Wikipedia to calculate mutual information between gender and occupation mentions.
• The study debunks the 'blue banana' hypothesis, showing that atypical gender associations are less influential than established contextual gender norms.

Corpus Investigation of Occupation Gender Typicality in Language: An Analysis

The paper "Are Female Carpenters like Blue Bananas? A Corpus Investigation of Occupation Gender Typicality" explores the extent to which gender mentioning aligns with occupation gender typicality. The authors employ information-theoretic techniques and corpus statistical analysis to investigate how often and in what contexts gender is mentioned alongside occupations in large text corpora. They focus on two main hypotheses: (1) gender mentioning occurs more frequently with less typical gender-occupation pairings, akin to how mentioning "blue" with "banana" signals atypicality, and (2) gender mentioning is more frequent for women-dominated occupations where gender is more salient.

Research Approach and Findings

Data Sources and Methodology:

The study leverages four major datasets:

• US Labor Statistics: Provides empirical estimates of gender breakdowns across occupations.
• Pushshift.io Reddit: A large corpus of Reddit comments.
• Wikipedia: A structured and encyclopedic text corpus.
• Llama 2 Wikipedia: A synthesized text corpus generated by an open LLM to examine reproduction of gender biases.

The authors used a pipeline that integrated natural language processing tools for tokenization, part-of-speech tagging, and dependency parsing to extract and analyze gender-occupation pairs. They calculated mutual information to gauge the strength of the relationship between gender mentions and occupation gender typicality.

Key Findings:

1. Rejection of the Null Hypothesis:
   • The mutual information calculations indicated a correlation between gender mentioning and occupation genderedness, refuting the null hypothesis that there is no such relation.
2. Gender Mentioning and Femaleness of Occupation:
   • The analysis found that gender mentioning is generally more frequent in women-dominated occupations, particularly in the Pushshift.io Reddit dataset. Correlations observed were medium-sized ($r \approx 0.49$ for overall gender mentioning, $r \approx 0.50$ for femaleness mentioning, and $r \approx 0.42$ for maleness mentioning).
   • Wikipedia, with its stricter editorial standards and policies, did not demonstrate as strong a correlation, highlighting potential effects of text domain and moderation.
3. Debunking the "Blue Banana" Hypothesis:
   • The study did not find strong evidence that gender is mentioned more often when it is atypical for the occupation (i.e., the "blue banana" hypothesis).
   • Occupation genderedness based on empirical gender breakdowns did not correlate significantly with gender mentioning, suggesting alternative motives behind such mentions.
4. Corpus-Specific Effects:
   • Stronger correlations were found in Pushshift.io Reddit compared to Wikipedia, possibly due to the informal and less moderated nature of Reddit content.
   • Llama 2 Wikipedia, a LLM-generated corpus, showed even weaker correlations, suggesting that the modeled text may dilute real-world biases.
5. Perceived Gender Typicality vs. Empirical Gender Data:
   • Using word embeddings, the researchers compared empirical gender statistics with perceived gender codings of occupations. They found instances where societal stereotypes (female-coded vs. male-coded occupations) diverged from actual labor statistics.
   • Female-coded occupations displayed higher mutual information, indicating more frequent gender mentions compared to male-coded occupations.

Annotation and Sentiment Analysis: - An exploratory analysis revealed that discussions around woman-dominated occupations often contained negative sentiment or gender-related conversations, especially addressing gender balance and stereotypes. - Offensive content correlated with discussions about female-coded occupations like "nurse," whereas non-typical gender role references often drove gender-related discussions.

Implications and Future Directions

This corpus investigation provides nuanced insights into how gender and occupation intersect in linguistic mentions. The findings contribute significantly to the understanding of gender salience in language use and its socio-cultural implications.

Practical Implications:

• NLP Applications: Understanding how gender is mentioned in relation to occupations can inform the development of models aimed at mitigating gender bias in language.
• Policy and Education: These insights could shape language use guidelines in formal contexts (e.g., Wikipedia editing rules) and inform educational content on gender and occupation narratives.

Theoretical Implications:

• The divergence between empirical gender distributions and perceived gender codings underscores the importance of integrating psychological and sociological perspectives in linguistic studies.
• Future research might explore more diverse datasets, including multilingual contexts, to generalize these findings across languages and cultures.

Future Developments:

• Building on this work, future studies could leverage more refined techniques, such as transformer-based models, to capture subtle patterns of gender mentioning.
• Further exploration into non-binary gender representation in occupations could expand the binary framework currently constrained by existing resources.

In conclusion, this investigation elucidates the complexity behind gender and occupation mentions in language, emphasizing gender salience in women-dominated fields rather than non-typicality. The findings pave the way for more informed approaches in natural language processing and broader socio-linguistic endeavors.