Quick Question • A doctor is walking down the street with a boy. The boy is the doctor’s son, but the doctor is not the boy’s father. How is that possible? GENDER BIAS IN WORD EMBEDDINGS 1
Simple Answer • The doctor is the boy’s mother… GENDER BIAS IN WORD EMBEDDINGS 2
Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them HILA GONEN AND YOAV GOLDBERG BAR ILAN UNIVERSITY INRIA PARIS 12/3/19 ACCEPTED TO NAACL 2019
Outline • Background • Gender Bias in word embeddings • Current debiasing methods • Post-processing (Bolukbasi et al.) • During training (Zhao et al.) • Experiments that reveal the remaining bias • Conclusion GENDER BIAS IN WORD EMBEDDINGS 4
Gender Bias in Applications 5
What do we mean by gender bias? GENDER BIAS IN WORD EMBEDDINGS 6
What do we mean by gender bias? GENDER BIAS IN WORD EMBEDDINGS 7
What do we mean by gender bias? (Zhao et al.,NAACL, 2018) GENDER BIAS IN WORD EMBEDDINGS 8
What do we mean by gender bias? (Hendricks et al., 2018) GENDER BIAS IN WORD EMBEDDINGS 9
Word Embeddings • TopK lists: nurse (Mikolov et al. 2013) GENDER BIAS IN WORD EMBEDDINGS 10
Word Embeddings • We will focus on gender bias in word embeddings GENDER BIAS IN WORD EMBEDDINGS 11
Bias in word embeddings 12
Bias in Word Embeddings GENDER BIAS IN WORD EMBEDDINGS 13
Bias in Word Embeddings • Caliskan et al. replicate a spectrum of known biases from the literature using word embeddings • Show that text corpora contain several types of biases: • morally neutral as toward insects or flowers • problematic as toward race or gender • veridical, reflecting the distribution of gender with respect to careers or first names • Introduce methods for identifying these biases GENDER BIAS IN WORD EMBEDDINGS 14
Bias in Word Embeddings GENDER BIAS IN WORD EMBEDDINGS 15
Bias in Word Embeddings • Permutation test: • X, Y: sets of target words (e.g. male names vs. female names) • A, B: sets of attribute words (e.g. career terms vs. family terms) • P-value is: GENDER BIAS IN WORD EMBEDDINGS 16
Bias in Word Embeddings Concepts 1 Concepts 2 Attributes 1 Attributes 2 Flowers : Insects : Pleasant : Unpleasant : buttercup, daisy, lily ant, caterpillar, flea freedom, health, love abuse, crash, filth European American names : African American names : Pleasant : Unpleasant : Brad, Brendan Darnell, Lakisha joy, love, peace agony, terrible Male attributes : Female attributes : Math words : Arts Words : male, man, boy female, woman, girl math, algebra, geometry poetry, art, dance GENDER BIAS IN WORD EMBEDDINGS 17
Definitions of Gender Bias in Word Embeddings 18
Definition of Gender Bias in Word Embeddings (NIPS, 2016) GENDER BIAS IN WORD EMBEDDINGS 19
Definition of Gender Bias in Word Embeddings • We check how similar a word is to “he” and “she” (cosine similarity) • Note that we care about the difference between the two • This is the projection on the direction of “he – she”* * This is the gender direction, can be computed using several pairs together (e.g. man-woman, brother-sister) GENDER BIAS IN WORD EMBEDDINGS 20
Definition of Gender Bias in Word Embeddings • bias(consultant) = -0.0023 • bias(nurse) = -0.2471 • bias(captain) = 0.1521 • bias(table) = -0.0003 Zhao et al. GENDER BIAS IN WORD EMBEDDINGS 21
Reducing of Gender Bias in Word Embeddings 22
Reduce Bias after Training • Bolukbasi et al. suggest to remove bias after training by removing the projection of every neutral word on the gender direction GENDER BIAS IN WORD EMBEDDINGS 23
Reduce Bias after Training • 1. Define a gender direction: • 10 gender pair difference vectors: • woman, man | girl, boy | she, he | mother, father daughter, son | gal, guy | female, male | her, his herself, himself | Mary, John • Compute and use their principal component GENDER BIAS IN WORD EMBEDDINGS 24
Reduce Bias after Training • 2. Define inherently neutral words: • Identify the set of gender specific words • The authors derive a list of 218 words from dictionary definitions: • mother, aunt, chairman, girlfriend, prince • The complementary set are the gender neutral words • The authors generalize the list to a broader vocabulary using SVM (~6500 words) GENDER BIAS IN WORD EMBEDDINGS 25
Reduce Bias after Training • 3. Zero the projection of all neutral words on the gender direction: 𝑥 Projection of on gender direction • The bias of all neutral words is now zero by definition GENDER BIAS IN WORD EMBEDDINGS 26
Reduce Bias after Training • 4. Equalize: • A family of equality sets (pairs): • For each pair, compute: normalized, same bias for both words • Equalize the words in the pair: GENDER BIAS IN WORD EMBEDDINGS 27
Reduce Bias after Training • We will address these embeddings as HARD-DEBIASED GENDER BIAS IN WORD EMBEDDINGS 28
Reduce bias during training (EMNLP, 2018) GENDER BIAS IN WORD EMBEDDINGS 29
Reduce bias during training • Zhao et al. suggest to reduce bias during training: • Train word embeddings using GloVe (Pennington et al., 2014) • Alter the loss to encourage the gender information to concentrate in the last coordinate • To ignore gender information – simply remove the last coordinate GENDER BIAS IN WORD EMBEDDINGS 30
Reduce bias during training • Loss: GloVe component (captures word proximity) GENDER BIAS IN WORD EMBEDDINGS 31
Reduce bias during training • Use two groups of male/female seed words, and encourage words from different groups to differ in their last coordinate : Gender coordinate(s) • Seed words – according to WordNet • *The authors also experiment with another variant for this component GENDER BIAS IN WORD EMBEDDINGS 33
Reduce bias during training • Encourage the representation of neutral-gender words (excluding the last coordinate) to be orthogonal to the gender direction : Vector w/o gender coordinate(s) GENDER BIAS IN WORD EMBEDDINGS 34
Reduce bias during training • Gender direction is estimated on the fly • Averaging differences between pairs of male-female words GENDER BIAS IN WORD EMBEDDINGS 35
Reduce bias during training • We will address these embeddings as GN-GLOVE GENDER BIAS IN WORD EMBEDDINGS 36
Some Results • Bolukbasi et al.: • Bias of all inherently-neutral words is zero by definition • Generated analogies are less stereotyped GENDER BIAS IN WORD EMBEDDINGS 37
Some Results • Zhao et al.: • Decrease bias in co-reference resolution GENDER BIAS IN WORD EMBEDDINGS 38
Problem solved? • Not so fast… GENDER BIAS IN WORD EMBEDDINGS 39
Clustering male- and female- biased words • We take the most biased words in the vocabulary according to the original bias (500 male-biased and 500 female-biased) • We cluster them into two clusters using K-means • The clusters align with gender with accuracy of: • 92.5% compared to 99.99% (HARD-DEBIASED) • 85.6% compared to 100% (GN-GLOVE) GENDER BIAS IN WORD EMBEDDINGS 40
Clustering male- and female- biased words HARD-DEBIASED GN-GLOVE GENDER BIAS IN WORD EMBEDDINGS 41
Bias-by-neighbors • Bias is still manifested by the word being close to socially-marked feminine words • A new mechanism for measuring bias: • The percentage of male/female socially-biased words among the k nearest neighbors of the target word • Pearson correlation with bias-by-projection: • 0.686 compared to 0.741 (HARD-DEBIASED) • 0.736 compared to 0.773 (GN-GLOVE) GENDER BIAS IN WORD EMBEDDINGS 42
Professions • We take a predefined list of professions • We show correlation between the bias-by-projection and bias-by- neighbors, before and after debiasing GENDER BIAS IN WORD EMBEDDINGS 43
Professions HARD-DEBIASED GN-GLOVE GENDER BIAS IN WORD EMBEDDINGS 44
Association with stereotyped words • We evaluate the association between female/male names and female/male stereotyped words (experiments taken from Caliskan et al.) Female-associated Male-associated names Amy, Joan, Lisa John, Paul, Mike family vs. carrer home, parents, children executive, management, professional arts vs. math poetry, art, dance math, algebra, geometry arts vs. science dance, literature, novel science, technology, physics • All the associations have very small p-values GENDER BIAS IN WORD EMBEDDINGS 45
Classifying to gender • Can a classifier learn to generalize from some gendered words to others based only on their representations? train 1000 SVM 5000 most biased words 4000 test GENDER BIAS IN WORD EMBEDDINGS 46
Classifying to gender • Results: GENDER BIAS IN WORD EMBEDDINGS 47
Conclusion • Word embeddings exhibit gender bias • Debiasing is hard! • Social gender bias is picked up from the data by the models • A lot of the bias information is still recoverable (even when the bias is low/zero according to the definition usually used) • The way we define the bias is important, and needs to be reconsidered when trying to solve the problem GENDER BIAS IN WORD EMBEDDINGS 48
Questions? GENDER BIAS IN WORD EMBEDDINGS 49
Thank you! GENDER BIAS IN WORD EMBEDDINGS 50
Recommend
More recommend
