bolukbasi-2016-gender-analogies-word2vec-google-news
IN premise — summaries/2026/08/24/wiki-Word_embedding-chunk-1.md
Created 2026-08-24T17:11:28+00:00
Bolukbasi et al. (2016) demonstrated that word2vec vectors trained on Google News reproduce gender/racial stereotype analogies such as 'man : programmer :: woman : homemaker', and Zhao et al. (2017) showed embeddings can amplify societal biases beyond what is present in training data
Summary
Word-embedding models like word2vec don't just mirror the gender and racial stereotypes in their training text; they can actually amplify those biases, producing outputs more prejudiced than the source material. This matters because any downstream system that builds on these vectors silently inherits and can worsen societal biases unless it explicitly corrects for them.