Bias in NLP 101
summary
Uses an Indian movie corpus of about 400 films (1970–2017) with roughly 5,858 female vs ~9,380 male cast entries to measure gendered descriptions and occupations, applies IBM’s DeCogTeller to swap gendered tokens and re-evaluate plausibility, and reports bias reduction but not elimination — some role attributes (like “aspiring” or “deliverer”) remain gendered. Training shows best KNN performance at k=1, and the pipeline flags underlabelling of women (shorter, relational descriptions) as a primary source of model bias.
AI-written summary · about
chapters (8)
- 0:00Introduction to bias in NLP
- 1:10Real-world examples of bias
- 6:12Gender bias in media and books
- 9:12Detecting and analyzing stereotypes
- 13:26Bias in movie datasets
- 16:42Removing bias with Decockteller
- 25:37Resources and further reading
- 27:57Q&A and closing remarks
related talks
28mBias in NLP
52mKeynote address
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
22mGender Bias in Artificial Intelligence
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
30mBiases in Language Models