Gender-based Language Differences in Letters of Recommendation.

Area of Science:

• Medical Education
• Natural Language Processing
• Gender Studies

Background:

• Letters of Recommendation (LOR) are crucial in biomedical education for candidate evaluation.
• The subjective nature of LORs can introduce gender-based language bias.
• Proactive identification and mitigation of gender stereotypes are essential for equality.

Purpose of the Study:

• To develop and evaluate Natural Language Processing (NLP) methods for detecting gender bias in physical therapy residency LORs.
• To extract psycholinguistic and thematic characteristics from LORs to identify biased language.
• To compare language use patterns between male and female writers and applicants.

Main Methods:

• Four NLP approaches were developed: sublanguage analysis, dictionary-based, rule-based, and deep learning.
• Models were evaluated on LORs from three Mayo Clinic physical therapy residency programs.
• MedTaggerIE and Bidirectional Encoder Representations from Transformers (BERT) models showed moderate-to-high performance in thematic category extraction.

Main Results:

• Male writers more frequently used terms like 'intelligence' and 'exceptional'.
• Male applicants were associated with terms such as 'strength' and 'interpersonal skills'.
• Significant differences were observed in how male and female writers expressed doubt, motivation, and recommendations.

Conclusions:

• NLP effectively identifies gender-biased language in LORs.
• Specific word choices and thematic expressions differ between genders in LORs.
• Findings require careful interpretation within the study's context, necessitating further research.