Anyone familiar with HR practices will probably know this. decades of research Resumes with black and/or female names at the beginning have been shown to receive fewer callbacks and interviews than resumes with white and/or male names, even if the rest of the resume is the same. has been. New research shows that similar biases emerge when large-scale language models are used to evaluate resumes instead of humans.
in new paper published within the last month AAAI/ACM Conference on AI, Ethics, and Societytwo researchers at the University of Washington ran hundreds of publicly available resumes and CVs using three different large-scale text embedding (MTE) models. According to the researchers, these models, based on the Mistal-7B LLM, were each tested on slightly different datasets to improve the base LLM’s capabilities in “representational tasks such as document retrieval, classification, and clustering.” It has been fine-tuned. Achieved “state-of-the-art performance” with MTEB Benchmark.
Rather than seeking exact matches of terms from job descriptions or assessing them via prompts (e.g., “Does this resume fit the job description?”), researchers was used to generate a relevance score embedded in each resume-job description pair. To measure potential bias, resumes are first run through the MTE without naming (to check reliability), and then based on actual use among groups in the general population, high It was run again using different names that achieved “distinctiveness scores” for race and gender. . We then analyze the top 10% of resumes that MTE determines are most similar to each job description to see if names from racial or gender groups are selected at higher or lower rates than expected. I did.
consistent pattern
After comparing over 3 million resumes and job descriptions, we found some pretty obvious biases. Across all three MTE models, white names were preferred in an overall 85.1 percent of the tests conducted, while black names were preferred in only 8.6 percent (the rest were too small to be judged as insignificant). showed a score difference close to zero). When it comes to gender-specific names, 51.9 percent of the tests preferred male names, compared to 11.1 percent that preferred female names. “Cross-sectional” comparisons that include both race and gender may further clarify the results. In “0% of bias tests,” black men’s names were preferred over white men’s names, the researchers wrote.
