Signal Calibrating WEAT Against Anisotropy: ZCA Whitening for Embedding Association Tests

Summary

This arXiv preprint proposes Zero-phase Component Analysis (ZCA) whitening as a geometric pre-processing step for the Word Embedding Association Test (WEAT), a widely used method for measuring bias in computational social science and AI fairness research. WEAT relies on cosine similarity to measure semantic association, an approach that assumes the embedding space is approximately isotropic. The authors note that prior work has reported many widely used language models do not satisfy this isotropy assumption, raising concerns about the reliability of WEAT-based bias measurements. The paper presents ZCA whitening as a method to correct this geometric distortion before applying WEAT. It was posted to arXiv's cs.CY category on 2026-08-07.