Abstract
Predicting organizational hierarchy from communication data is a foundational task in social NLP. While lexical and structural features are commonly used, the role of gender is often overlooked. This paper investigates how gender and the "gender environment"—the gender composition of a conversation—influence manifestations of power in the Enron email corpus. By introducing a gender-identified version of the corpus, the authors demonstrate that gender-aware features significantly improve hierarchy prediction models and reveal distinct sociolinguistic patterns, such as the tendency for male superiors to use more overt displays of power compared to female superiors.
Introduction
Sociolinguistic research suggests that men and women employ different strategies to navigate power dynamics and social maintenance. However, computational models for power prediction frequently ignore these nuances due to a lack of large-scale, gender-labeled datasets. This study addresses this gap by enriching the Enron email corpus with gender identifiers and examining how a speaker's behavior shifts based on their own gender and the gender makeup of their audience.
Data and Methodology
The authors developed a gender-identified version of the Enron corpus using a semi-automatic approach. For 145 core employees, gender was manually verified with 100% accuracy. For the broader participant pool, a conservative thresholding method was applied using the U.S. Social Security Administration (SSA) baby names database.
The researchers extracted three types of features:
- Structural (
THR_STR): Positional data (e.g., thread starter) and verbosity. - Metadata (
THR_META): Average number of recipients, reply rates, and "To" vs. "CC" status. - Dialogic (
DIA): Automated tagging of Dialog Acts (Requests, Informing) and Overt Displays of Power (ODP)—language that imposes constraints on the addressee.
A key innovation is the "Gender Environment" metric, defined as the percentage of women in a thread (excluding the writer). Environments were categorized as Male (<33.3% women), Female (>66.7% women), or Mixed.
Sociolinguistic Analysis: The "Face-Saving" Superior
The study tested whether female managers utilize more "face-saving" strategies than their male counterparts. The analysis of Overt Displays of Power (ODP) revealed a striking disparity. While male superiors exhibited the highest frequency of ODP (0.135), female superiors exhibited the lowest (0.072), even lower than female subordinates.
This suggests that the "power gap" in aggressive or commanding language is primarily driven by male behavior. Furthermore, the "gender environment" proved to be a significant factor: subordinates in predominantly female environments were found to use more "conventional" language (e.g., formal greetings and social maintenance) than any other group, highlighting how audience composition shapes linguistic choices.
Machine Learning Results
The inclusion of gender and environment features provided a measurable boost to automated hierarchy prediction. A baseline model using lexical and metadata features achieved 68.9% accuracy in determining the direction of power between pairs. Adding gender (GEN) and environment (ENV) features increased this to 70.2% on a blind test set.
More importantly, gender features acted as a contextual "multiplier." In models that excluded raw lexical content to focus on structural and dialogic patterns, the removal of gender features led to a significant 5.4% decrease in accuracy, proving that gender is essential for correctly interpreting social signals.
Conclusion
This research underscores that power is not a gender-neutral construct in corporate communication. By accounting for both the speaker's gender and the gender environment of the discourse, NLP models can achieve a more nuanced and accurate understanding of organizational hierarchy. Future work may explore these patterns in different cultural and digital contexts to determine the universality of these findings.
Summary updated