Biases in Language Models
summary
Claims measurement of gender bias across Flair BERT and GloVe using Winograd-schema sentences about occupations and applies projection and fine-tuning to reduce it. Non‑contextual projection moves embeddings off a gender axis so associations like “teacher” fade while “king/queen” stay, and contextual fine‑tuning cuts a pronoun prediction weight (example: "him") from ~0.4 to ~0.01. Uses visualization (t‑SNE), fairness tooling (AI Fairness 360), resampling/SMOTE for balancing and null‑space projection for residual signals, reports comparative metrics rather than vague examples.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:55Topics
- 1:53What is Language Model
- 9:09Noncontextual Language Models
- 12:21Past Work
- 14:34Contextual Models
- 20:56Biases and Data
- 25:07Questions
related talks
28mBias in NLP
31mCombining behavioral research & NLP to combat gender biases
31mCombining behavioral research and NLP to combat gender biases
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
02mThe real project of AI ethics
31mApproaches to Fairness and Bias Mitigation in Natural Language Processing
35mBias in NLP 101
28mKeynote, Why and how to care about ethics in ML