Combining behavioral research & NLP to combat gender biases
summary
A prototype called Bias Eliminator app highlights masculine/feminine-coded words (using a 2011 gold‑standard 150-word list) and suggests context‑aware replacements via a RoBERTa masked LM with a mask‑query trick (inject biased word twice). Users can override suggestions; overrides pass lemma‑level checks and seed a one‑shot similarity hash, a shallow minibatch learner and periodic fine‑tuning to avoid catastrophic forgetting. Claims include a 13% drop in women’s application rate for strongly masculine ads; system is productionized with an MLOps pipeline and Kubernetes deployment.
AI-written summary · about
chapters (10)
- 0:00Start
- 11:54Using the Transformer model to find the best word replacement
- 12:14MASK
- 13:03Provide the information on the original word to get better synonyms
- 16:24Keeping the human in the loop Model fine-tuning
- 21:12Hash table of similarity
- 22:53Shallow model (REST)
- 24:14Framework of augmented intelligence: Decision flow
- 26:19Next step for Bias Eliminator Bias in performance reviews
- 29:02CHANGE THE WORLD THROUGH DATA
related talks
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
30mBiases in Language Models
22mGender Bias in Artificial Intelligence
20mDebiasing Contextualized Word Embeddings in NLP
28mBias in NLP
52mKeynote address
35mBias in NLP 101
38mSocial Implications of Bias in Machine Learning