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

Research from arXiv proposes Zero-phase Component Analysis (ZCA) whitening as a pre-processing step to improve the reliability of Word Embedding Association Test (WEAT) measurements. WEAT, a widely used method in AI fairness research and computational social science, assesses bias using cosine similarity, which assumes an isotropic embedding space. However, many current language models do not meet this assumption, potentially compromising bias measurement accuracy. ZCA whitening aims to correct this by transforming the embedding space to restore isotropy with minimal vector perturbation, thereby enhancing the validity of bias assessments.

Why it matters

Accurate measurement of bias in AI systems is crucial for maintaining public trust and ensuring equitable outcomes across various applications. This research addresses a fundamental methodological challenge in bias detection, which directly impacts the integrity and fairness of AI technologies that are increasingly integrated into critical operational and decision-making processes.

Key insights

• The Word Embedding Association Test (WEAT) is a standard method for measuring bias in computational social science and AI fairness research.
• WEAT's reliability depends on the assumption that the underlying embedding space is isotropic.
• Many contemporary language models do not satisfy the isotropy assumption, leading to potential inaccuracies in bias measurement.
• Zero-phase Component Analysis (ZCA) whitening is proposed as a pre-processing technique.
• ZCA whitening transforms the covariance of the embedding space to achieve isotropy while minimizing changes to original vectors.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.06908