← All Cracked Math projects
PM-B03NOT STARTEDBEGINNERMATHPYTHONIMPACT ★★☆☆☆

Project the bias out of GloVe, then prove it is still there

Dot products, norms and one orthogonal projection are enough to 'debias' word vectors. They are also enough to show why that fails.

EST.5 h
REWARD+200 XP
AFTERBolukbasi et al. 2016 + 3

ML-ARENA BENCHMARKS

PROOF: FINDING
METRICINDUSTRYMINE
WEAT effect size (career/family x male/female word sets), analogy accuracy, and k-means bias-recoverability accuracy (how well k-means clustering on the debiased vectors can still sort words by their original gender bias)
GOAL · Cut the direct WEAT effect size by >= 80% while analogy accuracy drops by <= 1 pt. Also report a measured recoverability score that is clearly above chance, rather than hiding it.
Caliskan et al. 2017 WEAT protocol; Gonen & Goldberg 2019 recoverability test · numbers measured locally—
no results yet
YOUR RESULTS

Your recorded results appear here.

Brief

Word vectors such as GloVe place each word as a point in a 300-dimensional space, and some directions in that space encode gender. Implement Bolukbasi-style hard debiasing (the method from Bolukbasi et al. that removes a gender direction from word vectors) on GloVe 6B-300d, using only dot products, norms and orthogonal projections. First, estimate a gender direction from definitional pairs (word pairs such as he/she that differ only in gender). Then neutralise: project that direction out of words that should be gender-neutral. Then equalise: make each definitional pair equally distant from those neutral words. Implement two tests yourself: the WEAT effect size (a score of how much more strongly one set of words associates with one group than with another) and the Google analogy test (questions like 'man is to king as woman is to ?'). Then replicate Gonen & Goldberg's finding that clustering still recovers the removed bias, and report it honestly.

VICTORY CONDITIONS

  1. Cut the direct WEAT effect size by >= 80% while analogy accuracy drops by <= 1 pt. Also report a measured recoverability score that is clearly above chance, rather than hiding it. (WEAT effect size (career/family x male/female word sets), analogy accuracy, and k-means bias-recoverability accuracy (how well k-means clustering on the debiased vectors can still sort words by their original gender bias)).
  2. A public repository with a README that shows your result and how to reproduce it.

Terms in plain words

Orthogonal projection
Removing the part of a vector that points along a given direction, leaving only the part perpendicular to it.
Hard debiasing
Projecting a bias direction out of word vectors, then adjusting word pairs so they sit symmetrically around the neutral words.
WEAT effect size
A standardised score of how much closer one set of target words (for example career words) sits to one attribute group (for example male words) than to another.
Bias recoverability
Whether a simple method such as k-means clustering can still separate words by their original bias after debiasing, which shows the bias was hidden rather than removed.

Why it rates ★★☆☆☆

This replicates known papers. The honest negative result, showing the bias is still there, is the only part a reviewer would remember.