Share
Embed
Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
fatml
November 18, 2016
Transcript
33 slides
-
computer programmer (kəәmˈpjuːtəә ˈprəәʊɡræməә) n., pl., -s 1. A man
who writes programs for the opera4on of computers, especially as an occupa4on. nurse ('nəәrs) n., pl., -s 1. A woman trained to care for the sick or infirm, especially in a hospital. -
programmer javascript python engineer computer_science Mary
John she he soAball football Mary Jones She was valedictorian So-ball team captain Javascript, python… John Jones He was valedictorian Football team captain Javascript, python… 1) 2) -
MALE FEMALE SEXIST DEFINITIONAL (related [Schmidt
‘15]) The embedding captures gender stereotypes and sexism. -
MALE FEMALE SEXIST DEFINITIONAL tote
browsing tanning scrimmage dress sewing brilliant nurse cocky genius homemaker Easier to debias an embedding than to debias a human (related [Schmidt ‘15]) -
Based on word2vec trained on Google News corpus
Analogies Parallelograms capture semantics: [MikolovYZ 13] • Man:King :: Woman:Queen • Paris:France :: Tokyo:Japan • He:Blue :: She:Pink • He:Brother :: She:Sister • He:Doctor :: She:Nurse • He:Sausage :: She:Buns • He:Realist :: She:Feminist • She:Pregnancy :: He:Kidney stone • He:Computer programmer :: She:Homemaker -
Metric1: occupations. 327 gender neutral occupations. Project on to she—he
direction. she he homemaker nurse recep4onist maestro boss philosopher -
Metric1: occupations. 327 gender neutral occupations. Project on to she—he
direction. 'SVV(TVSNIGXMSR WLI LI , GVS[HVEXMRK) = . Crowdworkers rate each occup. for gender stereotype she he homemaker nurse recep4onist maestro boss philosopher -
Consistency of embedding stereotype word2vec trained on Google news GloVe
trained on web crawl Each dot is an occupation; Spearman = 0.8 -
Metric 2: analogies. Automatically generate he : x :: she
: y analogies. she he sister brother x = y = -
Metric 2: analogies. Automatically generate he : x :: she
: y analogies. she he sister brother min cos(LI WLI, \ ]) WYGLXLEX ||\ ]|| < small angle x = y = -
Metric 2: analogies. Automatically generate he : x :: she
: y analogies. she he sister brother min cos(LI WLI, \ ]) WYGLXLEX ||\ ]|| < homemaker programmer cupcake pizza -
Metric 2: analogies. Automatically generate he : x :: she
: y analogies. she he sister brother min cos(LI WLI, \ ]) WYGLXLEX ||\ ]|| < cupcake pizza 29/150 analogies rated as gender stereotypic by majority of crowdworkers -
Metric 3: indirect bias. so-ball football pitcher
recep4onist footballer maestro • Gender stereotype could affect the geometry between words that should be gender-neutral. • Project occupations onto softball—football axis. -
The geometry of gender she he
her his woman man female male Mary John Select pairs of words that reflect gender opposites. -
The geometry of gender she he
her his woman man female male Mary John Principal components Select pairs of words that reflect gender opposites. -
Geometry of gender 1 2 3 4 5 6 7
8 9 10 % of variance explained Principal components The top PC seems to capture the gender subspace B. -
Debiasing algorithm (ver.1) 1. Identify words that are gender-neutral N
and gender- definitional S. 2. Project away the gender subspace from the gender- neutral words. 3. Normalize vectors. [ := [ [ · & JSV [ 2 & MWXLIKIRHIVWYFWTEGI -
Identify gender-definitional words he she king
queen programmer homemaker smart cute blue pink Linear SVM 218 gender-definitional words -
Projecting away gender component he she
king queen programmer homemaker smart cute blue pink & & -
Projecting away gender component he she
king queen homemaker smart cute blue pink & programmer & 299 dimensions “hard debiasing” -
Advanced debiasing Find a linear transformation T of the gender-neutral
words to reduce the gender component while not moving the words too much. min 8 ||(8;)8(8;) ;8;|| * + ||(82)8(8&)|| * ; = QEXVM\SJEPP[SVHZIGXSVW 2 = QEXVM\SJRIYXVEP[SVHZIGXSVW don’t move too much minimize gender component -
Debiasing results: analogies # analogies generated # stereotypic analogies #
analogies generated # appropriate analogies -
Debiasing results: analogies Debiasing reduced stereotypic analogies while preserving the
utilities of the embedding. # analogies generated # stereotypic analogies # analogies generated # appropriate analogies -
Debiasing results: indirect bias so-ball football pitcher
recep4onist footballer maestro Original embedding -
Debiasing results: indirect bias so-ball football pitcher
recep4onist footballer maestro Original embedding so-ball football pitcher major leaguer footballer midfielder Debiased embedding -
Other types of embedding biases Embedding captures other types of
biases: cultural, religious, ethnic, etc. Example: • Occupation words closest to minorities: butler, footballer. • Occupation words closest to whites: legislator, lawyer. Talk to Max Leiserson! -
computer programmer (kəәmˈpjuːtəә ˈprəәʊɡræməә) n., pl., -s 1. A man
who writes programs for the opera4on of computers, especially as an occupa4on. nurse ('nəәrs) n., pl., -s 1. A woman trained to care for the sick or infirm, especially in a hospital. Debias word-embedding à debias machine learning’s dictionary -
Discussion points • Who’s responsible: data, algorithm or user? •
Geometry captures bias. • Fairness and transparency of embeddings (interpretable). • Using debiased embedding for downstream applications. Paper: Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. NIPS’16. -
Thanks! • Who’s responsible: data, algorithm or user? • Geometry
captures bias. • Fairness and transparency of embeddings (interpretable). • Using debiased embedding for downstream applications. Paper: Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. NIPS’16.