Calibrating WEAT Against Anisotropy: ZCA Whitening as a Geometric Pre-Processing Step for Embedding Association Tests

WEAT bias measurements may be unreliable due to embedding space geometry: The widely-used Word Embedding Association Test (WEAT) assumes embedding spaces are "isotropic" (uniform in all directions), but many language models violate this assumption, potentially leading to incorrect bias measurements in AI fairness research.

ZCA whitening can fix the geometric problem: The researchers propose a mathematical pre-processing technique called Zero-phase Component Analysis (ZCA) whitening that transforms embedding spaces to restore the isotropy condition WEAT depends on, without substantially distorting the original data.

Calibration changes bias findings significantly: When ZCA whitening was applied to 70 different model-task combinations, over 30% of WEAT bias test results changed whether they were statistically significant, with effect sizes shifting in both directions—meaning some biases were previously overestimated while others were underestimated.

Current bias audits may be giving misleading results: The findings suggest that organizations currently using uncalibrated WEAT to measure bias in their language models may be getting inaccurate measurements, potentially missing real biases or flagging false positives.

Better calibration improves semantic understanding: After applying ZCA whitening, the embedding spaces performed better on standard semantic similarity benchmarks, indicating the calibrated spaces more accurately capture meaningful word associations.