Kai-Wei Chang UCLA. What It Takes to Control Societal Bias in Natural Language Processing. References:

Transcription

1 What It Takes to Control Societal Bias in Natural Language Processing Kai-Wei Chang UCLA References: Kai-Wei Chang (kwchang.net/talks/sp.html) 1

2 A father and son get in a car crash and are rushed to the hospital. The father dies. The boy is taken to the operating room and the surgeon says, I can t operate on this boy, because he s my son. Can you explain why? Kai-Wei Chang (kwchang.net/talks/sp.html) 2

3 Implicit association test (IAT) Kai-Wei Chang (kwchang.net/talks/sp.html) 3

4 Concepts in semantic memory are assumed to be linked together with associated concepts having stronger links than unrelated concepts (Collins and Loftus, 1975). - Kai-Wei Chang (kwchang.net/talks/sp.html) 4

5 Concepts in semantic memory are assumed to be linked together with associated concepts having stronger links than unrelated concepts (Collins and Loftus, 1975). Kai-Wei Chang (kwchang.net/talks/sp.html) 5

6 So does computer Data with Societal Bias Model with Societal Bias Kai-Wei Chang (kwchang.net/talks/sp.html) 6

7 Word Embeddings can be Dreadfully Sexist [nips16] w/ Tolga Bolukbasi, James Zou, Venkatesh Saligrama, Adam Kalai v v "#$ v &'"#$ + v )$*+, v #)$/ he: brother beer physician programmer professor she: sister cocktail registered_nurse homemaker associate professor We use Google w2v embedding trained from the news Kai-Wei Chang (kwchang.net/talks/sp.html) 7

8

9 Related works Aylin, Joanna, and Arvind (2017) measure the biases in embedding using Implicit Association Test (IAT) and demonstrate it contain human-like biases Garg, Schiebinger, Jurafsky, Zou (2017) Word embeddings quantify 100 years of gender and ethnic stereotypes: 9

10 Gender Bias in Coref [NAACL 2018] Concurrent work (Rudinger et al., also studied gender bias in Coref. 10

11 Africa-American English Non Africa-American English Kai-Wei Chang (kwchang.net/talks/sp.html) 11

12 errors Majority Majority Minority Minority error rate: 6/30 = 80% error rate: 6/30 = 80% Kai-Wei Chang (kwchang.net/talks/sp.html) 12

13 Human Bias in Structured Prediction Models [EMNLP 17*] w/ Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez What s the agent for this image? Cooking Role Object agent food container tool place woman? vegetable bowl knife kitchen An example from a vsrl (visual Semantic Role Labeling) system *Best Long Paper Award at EMNLP 17 Kai-Wei Chang (kwchang.net/talks/sp.html) 13/9

14

15

16 Algorithmic Bias in Applications cooking dusting faucet fork World Dataset Model Credit: Mark Yatskar 16 16

17 Algorithmic Bias in Applications woman cooking cooking dusting faucet fork World Dataset Model Credit: Mark Yatskar 17 17

18 Algorithmic Bias in Applications woman cooking man fixing faucet cooking dusting faucet fork World Dataset Model Credit: Mark Yatskar 18 18

19 When AI products exhibit Societal Bias Kai-Wei Chang (kwchang.net/talks/sp.html) 19

20 Kai-Wei Chang (kwchang.net/talks/sp.html) 20

21 What It Takes to Control Societal Bias in NLP? Gender discrimination is illegal in the United States. Stanford Parser Kai-Wei Chang (kwchang.net/talks/sp.html) 21

22 A carton of ML (NLP) pipeline Prediction (Structured) Inference Representation E.g., word embedding, Knowledge bases. Etc. Auxiliary Corpus Data Kai-Wei Chang (kwchang.net/talks/sp.html) 22

23 Outline v Controlling Gender Bias in Representation Level A study of removing bias in Word Embedding v Reducing Gender Bias in Data Level A case study on co-reference resolution v Reducing Gender Bias in Inference Level Guiding predictions by corpus-wise constraints Kai-Wei Chang (kwchang.net/talks/sp.html) 23

24 Word Embeddings can be Dreadfully Sexist [nips16] w/ Tolga Bolukbasi, James Zou, Venkatesh Saligrama, Adam Kalai v v "#$ v &'"#$ + v )$*+, v #)$/ he: brother barbershop beer physician programmer professor she: sister salon cocktail registered_nurse homemaker associate professor We use Google w2v embedding trained from the news Kai-Wei Chang (kwchang.net/talks/sp.html) 24

25 Geometry of Gender and Bias v Identifying the gender subspace Top 10 Eigenvalue Top 10 Eigenvalue PCA ( he - she, father - mother, ) Gender Pair PCA ( dog - cat, house - building, ) Random Pair Kai-Wei Chang (kwchang.net/talks/sp.html) 25

26 Reducing bias SEXIST FEMALE MALE DEFINITIONAL

27 SEXIST FEMALE tote browsing tanning scrimmage dress sewing brilliant nurse cocky genius homemaker MALE DEFINITIONAL (related [Schmidt 15])

28 Approach 1: Project out gender dimension (hard version) v Step 1: Remove gender dimension from gender-neutral words

29 Approach 1: Post-processing (Hard) Project Out Gender Dimension v Step 2: re-center gender-definitional pairs

30 Approach 2: Post-processing (Soft) Find a linear transformation T of the gender-neutral words to reduce the gender component while not moving the words too much. W = matrix of all word vectors. N = matrix of neutral word vectors. min T (TW) T (TW) W T W 2 F + (TN) T (TB) 2 F don t move too much minimize gender component

31 Approach 3: Learning Gender-Neutral Word Embedding [Jieyu+EMNLP18] v How can we not to encode gender information in word vectors? corpus man, woman father, mother he, she Kai-Wei Chang (kwchang.net/talks/sp.html) 31

32 Approach 3: Learning Gender-Neutral Word Embeddings [Jieyu+ EMNLP18] dimensions for other latent aspects w 1 1-1? mother father doctor dimensions reserve for gender information Kai-Wei Chang (kwchang.net/talks/sp.html) 32

33 w 2 w 1 Kai-Wei Chang (kwchang.net/talks/sp.html) 33

34 Are these debiased vectors actually useful? Kai-Wei Chang (kwchang.net/talks/sp.html) 34

35 Outline v Controlling Gender Bias in Representation Level A study of removing bias in Word Embedding v Reducing Gender Bias in Data Level A case study on co-reference resolution v Reducing Gender Bias in Inference Level Guiding predictions by corpus-wise constraints Kai-Wei Chang (kwchang.net/talks/sp.html) 35

36 Gender Bias in Coref [NAACL 2018] Concurrent work (Rudinger et al., also studied gender bias in Coref. 36

37 Wino-bias data v Stereotypical dataset v Anti-stereotypical dataset Kai-Wei Chang (kwchang.net/talks/sp.html) 37

38 Gender bias in Coref System E2E E2E (Debiased WE) E2E (Full model) Steoetype Anti-Steoretype Avg Kai-Wei Chang (kwchang.net/talks/sp.html) 38

39 Source of gender bias Co-reference Prediction Use gender-neutral embedding (Structured) Inference Representation Word Embedding Data 80% entities in OntoNotes are male Kai-Wei Chang (kwchang.net/talks/sp.html) 39

40 Gender bias in Coref System E2E E2E (Debiased WE) E2E (Full model) Steoetype Anti-Steoretype Avg Kai-Wei Chang (kwchang.net/talks/sp.html) 40

41 How to deal with bias in data v Idea: simulate sentence in opposite gender John went to his house F2 went to her house Named Entity are anonymized Gender words are swapped Kai-Wei Chang (kwchang.net/talks/sp.html) 41

42 Gender bias in Coref System E2E E2E (Debiased WE) E2E (Full model) Steoetype Anti-Steoretype Avg Kai-Wei Chang (kwchang.net/talks/sp.html) 42

43 Outline v Controlling Gender Bias in Representation Level A study of removing bias in Word Embedding v Reducing Gender Bias in Data Level A case study on co-reference resolution v Reducing Gender Bias in Inference Level Guiding predictions by corpus-wise constraints Kai-Wei Chang (kwchang.net/talks/sp.html) 43

44 Human Bias in Structured Prediction Models [EMNLP 17*] w/ Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez What s the agent for this image? Cooking Role Object agent food container tool place woman? vegetable bowl knife kitchen An example from a vsrl (visual Semantic Role Labeling) system *Best Long Paper Award at EMNLP 17 Kai-Wei Chang (kwchang.net/talks/sp.html) 44/9

45 imsitu Visual Semantic Role Labeling (vsrl) Convolutional Neural Network COOKING ROLES AGENT FOOD CONTAINER TOOL (events) NOUNS woman vegetable pot spatula Kai-Wei Chang (kwchang.net/talks/sp.html) Yatskar et al. CVPR 16, Yang et al. NAACL 16, Gupta and Malik arxiv 16 Regression Conditional Random Field 45

46 COCO Multi-Label Classification (MLC) Convolutional Neural Network WOMAN PIZZA yes ZEBRA no FRIDGE yes CAR no Regression Conditional Random Field Kai-Wei Chang (kwchang.net/talks/sp.html) 46

47 Defining Dataset Bias (events) 47

48 Defining Dataset Bias (objects) Training Gender Ratio ( noun) 48

49 Gender Dataset Bias Kai-Wei Chang (kwchang.net/talks/sp.html) 49

50 Model Bias Amplification Kai-Wei Chang (kwchang.net/talks/sp.html) 50

51 Model Bias Amplification Kai-Wei Chang (kwchang.net/talks/sp.html) 51

52 Reducing Bias Amplification (RBA) Dataset Model RBA v Corpus level constraints on model output (ILP) v Doesn t require model retraining v Reuse model inference through Lagrangian relaxation v Can be applied to any structured model Kai-Wei Chang (kwchang.net/talks/sp.html) 52

53 Reducing Bias Amplification (RBA) 53

54 Reducing Bias Amplification (RBA) Sontag et al., 2011; Rush and Collins, 2012; Chang and Collins, 2011; Peng et al., 2015, Chang et al., 2013; Dalvi,

55 Reducing Bias Amplification (RBA) Lagrangian : λ 4 0 Sontag et al., 2011; Rush and Collins, 2012; Chang and Collins, 2011; Peng et al., 2015, Chang et al., 2013; Dalvi,

56 Lagrangian Relaxation Sontag et al., 2011; Rush and Collins, 2012; Chang and Collins, 2011; Peng et al., 2015, Chang et al., 2013; Dalvi,

57 Lagrangian Relaxation Sontag et al., 2011; Rush and Collins, 2012; Chang and Collins, 2011; Peng et al., 2015, Chang et al., 2013; Dalvi,

58 Lagrangian Relaxation Sontag et al., 2011; Rush and Collins, 2012; Chang and Collins, 2011; Peng et al., 2015, Chang et al., 2013; Dalvi,

59 Lagrangian Relaxation 59 Sontag et al., 2011; Rush and Collins, 2012; Chang and Collins, 2011; Peng et al., 2015, Chang et al., 2013; Dalvi, 2015

60 Gender Bias De-amplification in imsitu Male bias 60

61 Gender Bias De-amplification in imsitu Male bias 61

62 UCLA NLP NLP Applications Efficient Algorithms activity agent food cooking woman vegetable Learning from weak signals Fairness (data biases) 62

63 Conclusions v Like other AI systems, NLP systems affect by societal bias present in data v The issues cause unfair predictions are not new vdomain adaptation / Data collection bias v Ultimate goal: robust NLP systems for social good v References: Kai-Wei Chang (kwchang.net/talks/sp.html) 63