Neutralising bias on word embeddings

1. NEUTRALISING BIAS ON WORD EMBEDDINGS –Wilder Rodrigues

2. Wilder Rodrigues • Machine Learning Engineer at Quby; • Coursera Mentor; • City.AI Ambassador; • School of AI Dean [Utrecht] • IBM Watson AI XPRIZE contestant; • Kaggler; • Public speaker; • Family man and father of 3. @wilderrodrigues https://medium.com/@wilder.rodrigues

3. How do you see racism? • Before you proceed, please watch this video: https://www.youtube.com/watch?v=5F_atkP3pqs • The audio is in Portuguese, but in the next slide you will ﬁnd translations for what people said in the interviews. Source: Canal deTV da FAP (Astrojildo Pereira Foundation)

4. Translations • Group 1 • He is late; • She is a fashion designer; • Holds an executive position in either the HR or Finance area; • Taking care of his garden. Doesn’t look like a gardener; • She is cleaning her own house; the countertop; • Grafﬁti artist; it’s an art, it’s not vandalism. • Group II • Vandalising the wall; she is a spitter; • She is a housekeeper; cleaning the house; • He is a gardener; • He looks like a security guard or a chauffeur; • Seamstress; saleswoman; • He is running away; he is a thief.

5. Unconscious bias • Blue is for boys, pink for girls. • Boys are better at maths and science. • Tall people make better leaders. • New mothers are more absent from work than new fathers. • People with tattoos are rebellious. • Younger people are better with technology than older people.

6. –Joanna Bryson, University of Bath and Princeton University "AI is just an extension of our existing culture.”

7. Racialized code & Unregulated algorithms Source: https://www.theguardian.com/technology/2017/dec/04/racist-facial-recognition-white-coders-black-people-police Joy Buolamwini, Code4Rights and MIT Media Lab Researcher.

8. How white engineers built racist code – and why it's dangerous for black people Source: https://www.theguardian.com/technology/2017/dec/04/racist-facial-recognition-white-coders-black-people-police

9. Implicit AssociationTest Both black and white Americans, for example, are faster at associating names like “Brad” and “Courtney” with words like “happy” and “sunrise,” and names like “Leroy” and “Latisha” with words like “hatred” and “vomit” than vice versa. Source: http://www.sciencemag.org/news/2017/04/even-artiﬁcial-intelligence-can-acquire-biases-against-race-and-gender

10. W.E.A.T Names like “Brett” and “Allison” were more similar to those for positive words including love and laughter, and those for names like “Alonzo” and “Shaniqua” were more similar to negative words like “cancer” and “failure.” Source: http://www.sciencemag.org/news/2017/04/even-artiﬁcial-intelligence-can-acquire-biases-against-race-and-gender

11. W.E.F.A.T How closely related the embeddings for words like “hygienist” and “librarian” were to those of words like “female” and “woman.” It then compared this computer-generated gender association measure to the actual percentage of women in that occupation. Source: http://www.sciencemag.org/news/2017/04/even-artiﬁcial-intelligence-can-acquire-biases-against-race-and-gender

12. Word Embeddings A ⋅ B ∥A∥∥B∥ = ∑ n i=1 AiBi ∑ n i=1 A2 i ∑ n i=1 B2 i Source: https://medium.com/cityai/deep-learning-for-natural-language-processing-part-i-8369895ffb98 Father (L2 norm): 5.31 Mother (L2 norm): 5.63 d: 26.67 p: 29.89 Similarity: d / p = 0.89 Car (L2 norm): 5.73 Bird (L2 norm): 4.83 d: 5.96 p: 27.67 Similarity: d / p = 0.21

13. Identifying gender [woman] - [man] = [female]

14. What about other words?

15. Neutralising bias from non-gender speciﬁc words ebias_comp = e ⋅ g ∥g∥2 2 g edebiased = e − ebias Source: Bolukbasi et al., 2016, https://arxiv.org/pdf/1607.06520.pdf

16. Does it work? • Cosine similarity between receptionist and gender, before neutralising: • 0.3307794175059373 • Cosine similarity between receptionist and gender, after neutralising: • 5.2021694209043796e-17

17. Equalising gender-speciﬁc words Tricky parts!

18. Equalising gender-speciﬁc words • Cosine similarity between actor and gender, before equalising: • -0.08387555382505694 • Cosine similarity between actress and gender, before equalising:: • 0.33422494897899785 • Cosine similarity between actor and gender, after equalising: • -0.8796563888581831 • Cosine similarity between actress and gender, after equalising: • 0.879656388858183

19. How far is actor from babysitter? • Cosine similarity between actor and babysitter, before neutralising: • 0.2766562472128601 • Cosine similarity between actress and babysitter, before neutralising:: • 0.3378475317457311 • Cosine similarity between actor and babysitter, after neutralising: • 0.1408988327631711 • Cosine similarity between actress and babysitter, after neutralising: • 0.14089883276317122

20. References • https://www.youtube.com/watch?v=5F_atkP3pqs • https://www.theguardian.com/technology/2017/dec/04/racist-facial-recognition-white-coders-black-people-police • http://www.sciencemag.org/news/2017/04/even-artiﬁcial-intelligence-can-acquire-biases-against-race-and-gender • https://medium.com/cityai/deep-learning-for-natural-language-processing-part-i-8369895ffb98 • Bolukbasi et al., 2016, https://arxiv.org/pdf/1607.06520.pdf • Jeffrey Pennington, Richard Socher, and Christopher D. Manning, https://nlp.stanford.edu/projects/glove/ • https://github.com/ekholabs/DLinK/blob/master/notebooks/nlp/neutralising-equalising-word-embeddings.ipynb