TL;DRWe use word embeddings (word2Vec + transformer-based) on an 11-year corpus of UW Plastic Surgery residency letters of recommendation to quantify, interpret, and visualize implicit gender and ethnic biases/trends.
Abstract
The integrity of medical residency selection is paramount in cultivating competent physicians. With recent changes to the selection process, the weight of the letter of recommendation has increased, bringing potential gender and ethnic biases to the forefront. This study introduced a novel computational approach in applicant assessment research to quantify, interpret, and visualize the root causes of implicit gender and racial biases. The study used natural language processing techniques, including word embeddings through Word2Vec, other custom transformer-based models, and a suite of bias-detection algorithms.