Hundreds of papers missed blatant sexism in word embeddings

Evidence

We are only now coming to a full appreciation of the problem. “There have been hundreds of papers written about word embeddings and their applications, from Web search to parsing Curricula Vitae,” as Tolga Bolukbasi, Adam Kalai, and their collaborators write. “However, none of these papers have recognized how blatantly sexist the embeddings are and hence risk introducing biases of various types into real-world systems.” — Brian Christian, The Alignment Problem, 186

Categorization

• XREF: Speaks to the reader’s interest in humanity’s blind spots—a whole research community missed something obvious, paralleling Kodak’s own blindness to why they improved their film. — connects to: Word embeddings’ gender skew correlates with real occupational gender ratios; Word embeddings serve as novel objects for sociological analysis of bias; Common language carries more gender bias than objective metrics show — Brian Christian, The Alignment Problem, 221 (+4 more)
• QUESTION: What blind spots in our own work are we similarly unable to see?
• corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 186

Broader Connections

• Even massive, comprehensive text datasets reflect biases present in the real world #corroborates (0.80)
• Word embeddings’ gender skew correlates with real occupational gender ratios #corroborates (0.70)
• Common language carries more gender bias than objective metrics show #corroborates (0.70) — Brian Christian, The Alignment Problem, 221
• Machine-learning bias often stems from unrepresentative training data #expands (0.65)
• Word embeddings serve as novel objects for sociological analysis of bias #expands (0.60)
• Ideological premises in media encoding operate unconsciously even for the encoders #exemplifies (0.45) — Hall, Stuart, Essential Essays, Volume 1, 921