What Does Debiasing Really Remove? A Geometric Study of PCA-Based Gender Debiasing in Word Embeddings

• URL: http://arxiv.org/abs/2606.07964v1
• Date: Sat, 06 Jun 2026 03:44:22 GMT
• Title: What Does Debiasing Really Remove? A Geometric Study of PCA-Based Gender Debiasing in Word Embeddings
• Abstract summary: We conduct a systematic analysis of PCA-based gender debiasing and investigate what is actually removed from the embedding space.<n>Direct gender bias is primarily concentrated in the first principal component, supporting the low-rank bias hypothesis.<n>Associative bias measured by WEAT does not align with these principal directions and is instead spread across multiple embedding dimensions.
• Score: 0.0
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Debiasing methods based on principal component analysis (PCA) are broadly used to reduce gender bias in word embeddings used in LLMs, yet it remains unclear what aspects of bias they actually remove and how destructive this process is. These methods are based on the understanding that bias resides in a low-dimensional subspace, with the assumption that most of it can be captured by a few principal components. In this work, we conduct a systematic geometric analysis of PCA-based gender debiasing and investigate what is actually removed from the embedding space. Our experiments across multiple embeddings show that direct gender bias is primarily concentrated in the first principal component, supporting the low-rank bias hypothesis. However, associative bias measured by WEAT does not align with these principal directions and is instead spread across multiple embedding dimensions. Furthermore, as expected, we demonstrate that removing an increasing number of principal components leads to a consistent degradation of the embedding geometry, affecting semantic structure and vector relationships. These results reveal that PCA-based debiasing operates as a trade-off: while it effectively reduces certain forms of direct bias, it fails to eliminate distributed associations and introduces geometric distortion. Moreover, there is no universal optimal level of debiasing, as the balance between bias reduction and semantic preservation depends on the chosen metric and embedding. Overall, our findings suggest that bias in word embeddings is not purely low-rank and that simple subspace removal methods may be insufficient for comprehensive debiasing.

Related papers

• How Does Alignment Tuning Shape Representations of Sycophancy and Related Cue-Induced Biases in LLMs? [53.58941416780391]
  We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model.<n>We extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out transfer, and causal intervention.
  arXiv  Detail & Related papers  (2026-07-20T16:10:06Z)
• Bias Is a Subspace, Not a Coordinate: A Geometric Rethinking of Post-hoc Debiasing in Vision-Language Models [13.46898179372249]
  Vision-Language Models (VLMs) have become indispensable for multimodal reasoning, yet their representations often encode and amplify demographic biases.<n>We propose a geometrically principled framework that identifies and removes the entire subspace of linearly decodable bias.<n>Our method achieves more robust debiasing with an average improvement of $18.5%$ across four fairness metrics.
  arXiv  Detail & Related papers  (2025-11-22T17:04:30Z)
• Going Beyond Popularity and Positivity Bias: Correcting for Multifactorial Bias in Recommender Systems [74.47680026838128]
  Two typical forms of bias in user interaction data with recommender systems (RSs) are popularity bias and positivity bias. We consider multifactorial selection bias affected by both item and rating value factors. We propose smoothing and alternating gradient descent techniques to reduce variance and improve the robustness of its optimization.
  arXiv  Detail & Related papers  (2024-04-29T12:18:21Z)
• Revisiting the Dataset Bias Problem from a Statistical Perspective [72.94990819287551]
  We study the "dataset bias" problem from a statistical standpoint. We identify the main cause of the problem as the strong correlation between a class attribute u and a non-class attribute b. We propose to mitigate dataset bias via either weighting the objective of each sample n by frac1p(u_n|b_n) or sampling that sample with a weight proportional to frac1p(u_n|b_n).
  arXiv  Detail & Related papers  (2024-02-05T22:58:06Z)
• Partition-and-Debias: Agnostic Biases Mitigation via A Mixture of Biases-Specific Experts [24.055919128977195]
  We present the Partition-and-Debias (PnD) method that uses a mixture of biases-specific experts to implicitly divide the bias space into multiple subspaces. Experiments on both public and constructed benchmarks demonstrated the efficacy of the PnD.
  arXiv  Detail & Related papers  (2023-08-19T13:11:40Z)
• Exploring the Linear Subspace Hypothesis in Gender Bias Mitigation [57.292988892028134]
  Bolukbasi et al. present one of the first gender bias mitigation techniques for word representations. We generalize their method to a kernelized, nonlinear version. We analyze empirically whether the bias subspace is actually linear.
  arXiv  Detail & Related papers  (2020-09-20T14:13:45Z)
• MDR Cluster-Debias: A Nonlinear WordEmbedding Debiasing Pipeline [3.180013942295509]
  Existing methods for debiasing word embeddings often do so only superficially, in that words that are stereotypically associated with a particular gender can still be clustered together in the debiased space. This paper explores why this residual clustering exists, and how it might be addressed. We identify two potential reasons for which residual bias exists and develop a new pipeline, MDR Cluster-Debias, to mitigate this bias.
  arXiv  Detail & Related papers  (2020-06-20T20:03:07Z)
• Nurse is Closer to Woman than Surgeon? Mitigating Gender-Biased Proximities in Word Embeddings [37.65897382453336]
  Existing post-processing methods for debiasing word embeddings are unable to mitigate gender bias hidden in the spatial arrangement of word vectors. We propose RAN-Debias, a novel gender debiasing methodology which eliminates the bias present in a word vector but also alters the spatial distribution of its neighbouring vectors. We also propose a new bias evaluation metric - Gender-based Illicit Proximity Estimate (GIPE)
  arXiv  Detail & Related papers  (2020-06-02T20:50:43Z)
• Mitigating Gender Bias Amplification in Distribution by Posterior Regularization [75.3529537096899]
  We investigate the gender bias amplification issue from the distribution perspective. We propose a bias mitigation approach based on posterior regularization. Our study sheds the light on understanding the bias amplification.
  arXiv  Detail & Related papers  (2020-05-13T11:07:10Z)
• Double-Hard Debias: Tailoring Word Embeddings for Gender Bias Mitigation [94.98656228690233]
  We propose a technique that purifies the word embeddings against corpus regularities prior to inferring and removing the gender subspace. Our approach preserves the distributional semantics of the pre-trained word embeddings while reducing gender bias to a significantly larger degree than prior approaches.
  arXiv  Detail & Related papers  (2020-05-03T02:33:20Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.