Debiasing Contextualized Word Embeddings in NLP
summary
Lexalytics and UMass rewrite a large corpus (replace Mister→prefix, sister→sibling, neutral pronouns) and retrain contextual models (ELMo) to reduce gender signals; initial classifiers drop from 97% detectability to about 14% residual gender prediction, but the gender cue re-emerges with continued training (noted by epoch 9). Downstream GLUE performance falls ~1.5%, names overlap with common words breaks anonymization, and baby‑name SSN lists are used as a resource.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:52ABSTRACT SUMMARY
- 1:40What Are Language Models? • Predicting hidden words in a sentence can be used to induce general knowledge in deep networks
- 2:58Why We're Working on This Lexalytics Experience: • 17 years of natural language processing • Bilions of documents processed across 20+ languages, per day • We maintain high accuracy NLP models and build custom solutions for our needs are crucial for us
- 3:46Language Models & Bias
- 6:08Bias can be subtle - Employee Reviews Male-Leaning Terms Female-Leaning Terms
- 7:29General Approaches
- 10:03Challenges • Simple geometric approaches hide but do not remove blas
- 11:29Some Numbers
- 13:59Hiding Gender Markers Helped
- 15:30Language Is Hard... Non obvious gendered phrases . "Salesperson" tends to be female
- 17:08Recommendations PLEASE THINK CAREFULLY WHEN USING MACHINE LEARNING • Data Cleaning and Adversarial Learning look most promising to me • Make sure to track correlations between outcomes and target
related talks
31mCombining behavioral research and NLP to combat gender biases
31mCombining behavioral research & NLP to combat gender biases
30mBiases in Language Models
52mKeynote address
28mBias in NLP
35mBias in NLP 101
28mBreaking the Chains of Oppressive Software
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models