Bias in NLP
summary
Shows that word2vec trained on English Wikipedia encodes measurable gender skew: embeddings are pulled toward gendered contexts via negative sampling, causing machine translation to pick masculine or feminine pronouns incorrectly (example with Farsi). Builds a bias score using female/male bias factors and applies it to 350 occupations, finding extreme cases (housekeeper strongly female-skewed) and systematic clustering by gender. Publishing the quantification succeeds; mitigation is left as an open, human-in-the-loop decision.
AI-written summary · about
chapters (8)
- 0:00Start
- 2:23What Is Deep Learning
- 3:18Deep Learning
- 4:58Safety and Security Uncertainty
- 7:12Basic Neural Machine Translation Architecture
- 16:07Context Vectors
- 20:59What Happens When the Data Has Bias
- 24:22Conclusion
related talks
35mBias in NLP 101
30mBiases in Language Models
31mCombining behavioral research & NLP to combat gender biases
27mDesigning More Inclusive AI with More Inclusive Workflows
31mCombining behavioral research and NLP to combat gender biases
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
01h 31mHow To Measure And Mitigate Unfair Bias in Machine Learning Models
02mThe real project of AI ethics