Man is to Computer Programmer as Woman is to Homemaker? Debiasing - PowerPoint PPT Presentation

Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings A paper presentation by Giacomo Alliata Computational Social Media 22/05/2020

Presentation of the paper ● Authors: Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama and Adam Kalai Presented at NIPS 2016 ● ● Outline of the paper: ○ Introduction and related works Geometry of Gender and Bias in Word Embeddings ○ Debiasing algorithms ○ ○ Determining gender neutral words ○ Debiasing results Discussion ○

What are word embeddings? A word embedding, trained on word co-occurrence in text corpora, represents each word (or common phrase) as a d-dimensional word vector. Main properties: words that have a similar semantic meaning tend to have vectors that ● are close together (measured with cosine similarity) vectors differences between word embeddings represent relationships ● between words Analogy: “ man is to king as woman is to x ” should return x = queen In this paper, the authors primarily use word2vec 300-dimensional embedding trained on a corpus of Google News texts consisting of 3 million English words.

Sexism in word embeddings “ man is to computer programmer as woman is to x ” returns x = housekeeper ● “ father is to a doctor as mother is to a x ” returns x = nurse ●

Quantifying bias To quantify bias, the authors compute the similarity between the target word and a pair of gender-specific words. Ex: nurse is both close to woman and man (since all are humans) but the distance ( nurse , woman ) is smaller than the distance ( nurse , man ), thus suggesting bias. Important distinction: gender-specific words: brother , king (associated with a gender by ● definition) gender-neutral words: nurse , babysit , shoes , flight attendant ● Gender pairs: set of words that differ mainly in gender (“ she-he ”, “ mother-father ”)

Geometry of Gender and Bias in Word Embeddings Occupational stereotypes: crowdworkers are asked to evaluate whether an occupation is considered female-stereotypic, male-stereotypic or neutral. Results: strong correlation (Spearman coefficient: 0.51) between the projection of the occupation words onto the she-he axis and the stereotypicality estimates (similar results both on w2cNEWS and GloVe trained on web-crawl corpus)

Geometry of Gender and Bias in Word Embeddings Analogies: take a seed pair (a, b) determining a direction and score all pairs of words (x, y) using the following matrix (with threshold δ ) Evaluation with US-based crowd-workers, with 2 questions: ● whether the pairing makes sense as an analogy ● whether it reflects a gender stereotype Results: (rated by 5+ people out of 10) 72/150 analogies are gender-appropriate ● 29/150 analogies exhibit gender stereotypes ●

Identifying the Gender Subspace Identify 10 gender-pairs directions to more robustly estimate bias different biases associated with different gender pairs ● ● polysemy (ex: “ to grandfather a regulation ”) PCA on these 10 vector differences identifies the gender subspace in a single direction 10 selected gender-pairs Percentage of variance explained by the PCA (on gender pairs and on 10 random vectors)

Debiasing Algorithms First step: Identify the Gender Subspace that captures the bias Second step: Neutralize and Equalize or Soften ● Neutralize ensures that gender neutral words are 0 in the gender subspace Equalize equalizes sets of words outside the subspace (thus any neutral ● word is equidistant to all words in each equality set) Ex: if { grandmother , grandfather } and { guy , gal } were two equality sets, then after ○ equalization babysit would be equidistant to grandmother and grandfather and also equidistant to gal and guy Soften reduces the difference between these sets while maintaining as ● much similarity as possible (with a parameter to control the trade-off) Ex: grandfather would preserve its other meaning in to grandfather a situation ○

Debiasing Algorithms

Determining Gender Neutral Words In practice, it is easier to identify the gender-specific words and take the complementary set of words as the gender-neutral words. Classification task: produce a training set (using ● dictionary definitions) of gender-specific words train the classification model ● (gender-specific or gender-neutral) linear SVM in the paper ○ F-score: 0.627 ± 0.102 ○ Results of the classificer

Debiasing Results Evaluate using analogy task as before (pair ( she , he ) as input and pair (x, y) as output, rated as gender stereotypical or not by crowd-workers). Results are also confirmed by testing on standard benchmarks the quality of the analogies

Discussion Main results: a single direction largely captures gender bias and projecting ● gender-neutral words on this direction helps to quantify gender bias the hard-debiasing algorithm significantly reduces the gender bias ● while preserving appropriate analogies ● the soft-debiasing algorithm could be useful in certain applications Overall, bias in word embeddings simply reflects bias in society but, if we are not able to directly reduce it, we should at least ensure machine learning does not increase it.

Thanks for your attention! Giacomo Alliata Computational Social Media 22/05/2020