Part of my series of notes from NAACL-HLT 2019 in Minneapolis.
Introduction
- Amazon’s biased recruiting tool
- the data was biased because the recruiting practices were biased
- the model just learned to faithfully reflect that
- complementary research agendas
- studying people can help us mitigate issues in ML
- ML models can help us understand people
- some examples: toxic language detection, face recognition, predictive policing
- ML is pattern recognition in existing data, we shouldn’t be surprised about any of this
- real phase shift in interest in fairness in ML around 2016/2017
Word Embeddings
- Bolukbasi et al., NIPS 2016 – analogies paper
- Caliskan et al., Science 2017 – speaker’s paper, dive into this one
- implicit association test
- differential reaction time is a measure of bias
- create word embedding association test
- differential word similarity is a measure of bias
- not symmetric, interestingly
- analysis applied to pretrained GloVe (web data) and word2vec (news data)
- actually not many differences even though nature of data quite different
- warmup: universal associations
- e.g. between flowers & pleasantness vs. insects & unpleasantness
- validates that this approach seems to work
- also show effect with gender, race, age, etc.
Understanding culture helps discover bias in ML
- don’t need to be a psychology or sociology expert, some commonsense knowledge helps
- NB. this is why we need diversity in ML! the speaker didn’t say this, but I am definitely saying it!
- Rudinger et al., NAACL 2018 – gender bias in coreference resolution for occupations
- Blodget & O’Connor 2017 – racial bias in language identification for different dialects
- use census data to get probability distribution of demographics according to location
- topic modeling with demographic groups as topics
- validate by reproducing linguistic phenomena
- proposed mitigation: ensemble models
- raises further ethical qustions…
- what counts as “African American English”?
- when is it appropriate to treat different groups differently?
- this is very tricky, legally and ethically
Interlude: Debiasing
- debiasing word embeddings – “is vs. ought”
- maybe be a bit cautious – doing surgery on things we don’t understand so well
- word embeddings are quite brittle
- bias detection is also brittle
- debiasing is not always the answer!
- does bias in representation translate to bias in downstream tasks?
- mitigation should be application specific
ML as a magnifying lens into human culture
- doing this with corpus linguistics for a while
- word embeddings can extend this by using real-world statistics accurately
- Garg et al., PNAS 2018 – word embeddings quantify 100 years of gender and ethnic stereotypes
- can study these to understand e.g. historical stereotypes
- Lewis & Lupyan, PNAS 2019 – does language merely reflect or also cause stereotypes?
- study 25 different languages and correlate to implicit bias in the countries
- also structural encoding of gender in different languages – maybe implies causality
- not just distributional semantics, but also linguistic structure
- many under-explored research directions here