Removing the "gender direction" from word embeddings hides bias without removing it. Words like nurse and receptionist still sit together, and surgeon and engineer still sit together, so a downstream model can recover the original stereotype from the neighborhoods alone. The obvious response is to break those neighborhoods up. When methods do that, they often damage meaning, because part of every neighborhood is real semantics: nurses and receptionists really do share contexts about patients and appointments. What the field lacks is a measurement that says, for each pairwise relation between two words, whether it is carried by gendered co-occurrence or whether it is ordinary meaning that happens to line up with a stereotype.
Evolutionary biology faced the same problem and solved it with a calibrated neutral baseline. To detect selection on a gene, you compare protein-changing substitutions ($d_N$) against silent substitutions ($d_S$), which accumulate at the neutral rate. The ratio $\omega = d_N/d_S$ only means something because $d_S$ measures how much change happens without selection. We propose the same logic for embedding geometry. For each word-pair relation, we compare how much it moves when gendered signal is removed from the training corpus against how much it moves when the corpus is merely resampled. Relations that shift far beyond resampling noise are carried by gender. Relations that do not are left alone. The resulting score gives a principled answer to which edges a structural debiasing method should cut, and it lets us test whether that answer resolves the observed trade-off between removing residual clusters and keeping downstream utility.
Core Claim
Residual cluster bias in static embeddings is a mixture. Some relations hitchhike on gendered co-occurrence; others are conserved semantics. Only an intervention-versus-resampling ratio can separate the two reliably, because the raw size of a relation's shift under any intervention is dominated by estimation noise for rare words. This mirrors how masked protein models confuse rare mutations with harmful ones, a confusion that Separating selection from mutation in antibody language models fixed by fitting a separate neutral mutation model before estimating selection.
We make three claims. First, editing only the high-ratio relations will match the strongest neighborhood-aware debiasing on bias benchmarks while costing less semantic and downstream utility. Second, editing low-ratio relations of the same number will cost utility and buy little bias reduction. Third, a sizable share of the clustering that Lipstick on a Pig attributes to bias consists of low-ratio, semantically conserved relations. If that third claim holds, cluster-purity audits overstate how much bias survives.
The Measurement
We train GloVe on the English Wikipedia dump used throughout this literature, under two conditions. The original corpus is one condition. The other is a gender-neutralized corpus, in which pronouns, gendered nouns and given names are replaced by a single neutral person token. Each condition gets $B$ bootstrap replicates, which resample documents with replacement, retrain, and are aligned by orthogonal Procrustes rotation on a stable anchor vocabulary.
For words $i$ and $j$, let $s^{(b)}_{ij}$ be their cosine similarity in replicate $b$. Define the neutral divergence $d^0_{ij}$ as the mean squared difference in $s_{ij}$ between pairs of replicates from the same corpus. Define the intervention divergence $d^I_{ij}$ as the mean squared difference between replicates of the original corpus and replicates of the neutralized corpus. The relational ratio is
$$\omega_{ij} = \frac{d^I_{ij}}{d^0_{ij}}.$$
Values near 1 mean that removing explicit gender moves the relation no more than resampling does. Large values mean the relation depends on gendered co-occurrence.
The direction of interpretation is inverted relative to biology. Here a high ratio flags the dispensable part, and a ratio near 1 flags structure that is stable with respect to the demographic signal. The ratio is computed over pairs among the union of each target word's 100 nearest neighbors within the top 50,000 words, which keeps the number of relations tractable.
A Bayesian approach to uncertainty in word embedding bias estimation motivates the denominator: it argues that embedding bias estimates carry real uncertainty. We turn that observation into a per-relation noise floor rather than a confidence interval on an aggregate score.
Repair Step
Given a threshold $\tau$ on $\omega$, we post-process the original embedding vectors $u$. Flagged relations are pulled toward their neutral-corpus similarity $\bar s^{\text{neut}}_{ij}$. Unflagged relations are held at their original similarity $\bar s^{\text{orig}}_{ij}$:
$$\min_{u} \sum_{\omega_{ij} > \tau} \big(\cos(u_i,u_j) - \bar s^{\text{neut}}_{ij}\big)^2 + \lambda \sum_{\omega_{ij} \le \tau} \big(\cos(u_i,u_j) - \bar s^{\text{orig}}_{ij}\big)^2,$$
where $\lambda$ trades bias removal against preservation. Gendered words keep their original vectors, so definitional information survives.
This is deliberately close in form to RAN-Debias from Nurse is Closer to Woman than Surgeon?. That method repels words from neighbors flagged by a bias-based network, attracts them to their original positions, and neutralizes the gender axis, using WordNet and dictionary lookups to decide which words are gender-specific. Our contribution is the criterion for which relations to edit, not a new optimizer. The criterion needs no lexicon beyond a list of gendered tokens, and it is calibrated against noise.
Relation to Prior Work
The motivating evidence comes from three papers. Lipstick on a Pig shows that projection debiasing leaves clusters recoverable at high accuracy. MDR Cluster-Debias unfolds the embedding manifold nonlinearly and clusters around definitional pairs. It passes most WEAT tests but gains little on OntoNotes coreference and drops notably on IMDB sentiment. We read that as evidence that it cut conserved relations. Investigating the Effects of Fairness Interventions Using Pointwise Representational Similarity audits interventions point by point with kernel alignment (PNKA, which measures how much each point's position relative to all others changes). It finds that interventions move gender-defining anchor words more than stereotypical words. That is the opposite of what one wants, and the low-ratio constraint above is designed to prevent it.
The method is supplied by the antibody work. Separating selection from mutation in antibody language models fits a fixed neutral mutation model on unselected DNA and learns selection factors on top of it. A Sitewise Model of Natural Selection on Individual Antibodies via a Transformer–Encoder moves from whole-gene to per-site selection and finds that the textbook partition of antibody regions into conserved scaffold versus variable loops is wrong at the level of individual sites. We borrow two ideas from these papers: a separately estimated neutral process, and per-unit rather than per-region scores. We do not borrow their discrete phylogenetic machinery. Our neutral process comes from corpus resampling, and the counterfactual corpus plays the role that the unmutated germline template plays in Explicit representation of germline and non-germline residues improves antibody language modeling.
The closest existing NLP mechanism is Diffusion Theory as a Scalpel. It also flags units whose drift exceeds a calibrated expectation, using squared parameter drift relative to curvature, and resets them to pretrained weights. It explicitly cannot address bias already present in the pretrained reference. Our score works on relations learned during pretraining and uses a counterfactual corpus, not a clean checkpoint, as the reference.
Research Program
The work has three stages. The first is measurement. We compute $\omega$ for all candidate relations with $B=10$ replicates per corpus. We report how many relations exceed the noise floor, how $\omega$ depends on word frequency, and how high- and low-ratio relations are distributed inside the gender clusters found on the Lipstick top-1,000 list.
The second is a causal edit study. Editing a set of relations should change downstream bias and utility in a way that the ratio predicts. The third is the full repair, compared against existing debiasing methods. A secondary robustness check recomputes $\omega$ with counterfactual swapping (exchanging he/she, man/woman, and gendered names) as the intervention, instead of neutralization, to test whether the flagged set depends on how gender is removed.
One limitation is acknowledged up front. Neutralization removes explicit gender tokens, but second-order paths through gender-correlated words (pregnant, beard) survive. Our score therefore detects bias mediated by explicit gendered context, and we will characterize what it misses using the nonlinear SVM audit.
Baselines
The do-nothing reference is the original GloVe trained by us on the same corpus with the same hyperparameters. It is the starting point for every post-processing method and fixes the bias numbers that everything else must reduce. The original-GloVe WinoBias pro- versus anti-stereotype gap reported in Nurse is Closer to Woman than Surgeon? is 30.2. We will check that our retrained baseline lands near it.
Hard-Debias projection and GN-GloVe are the linear and training-loss baselines. They are run from the same base vectors and the same definitional pairs. On Hard-Debias, Lipstick on a Pig reports 92.5% K-means cluster accuracy, 88.88% RBF-SVM recovery, and a neighbor-bias correlation of 0.6 to 0.8. These anchor the structural audits.
Two retraining baselines test whether the reference corpus alone suffices. One is GloVe trained on the counterfactually swapped corpus. The other is GloVe trained on the gender-neutralized corpus itself and used directly as the debiased space.
MDR Cluster-Debias represents aggressive nonlinear cluster breaking. Its paper gives qualitative downstream outcomes: similar MR accuracy, a notable IMDB drop, and minimal OntoNotes differences. The file contains no reported numbers for these, so we will rerun it.
The strongest competitor is RAN-GloVe. It is the only method in the file that reports strong bias reductions while keeping SimLex-999 stable: a GIPE proximity-bias reduction of over 42% relative to baselines, a 21.4% improvement over GN-GloVe on SemBias, and a WinoBias gap of 0.4. No reported result was found on Lacuna for RAN-GloVe on the Lipstick K-means or SVM audits, and no numeric SimLex-999 value was found for it, so we will measure both.
Every post-processing method starts from identical base vectors and receives the same tuning budget: twenty hyperparameter configurations, selected on held-out halves of SimLex-999 and the WinoBias development split. Beating RAN-GloVe means matching its WinoBias gap within noise, defined as a gap no larger than 1.0 across five seeds. It also means reaching equal or better GIPE and SemBias scores together with significantly higher SimLex-999 and sentiment accuracy, or significantly lower Lipstick-audit recovery at equal utility.
Two ablations isolate the key ingredient. In the first, relations are selected by raw shift $d^I_{ij}$ with no resampling denominator. In the second, an equal-size random set of relations is selected. Both are passed through the identical repair objective.
Why not just…?
Why not just keep projecting out the gender direction? This is exactly what Lipstick on a Pig tested. Clusters remained recoverable at 92.5% and a nonlinear classifier recovered gender at 88.88%, because projection leaves the distances among neutral words almost untouched. We include it only as a reference point.
Why not just use the embeddings trained on the gender-neutralized or swapped corpus? This is the most serious threat, because our reference already contains most of the signal. There are two reasons we expect it to fall short. Neutralization collapses all gendered words into one token, which destroys the definitional information that coreference needs. And retraining reshuffles every relation, including conserved ones, by at least resampling noise. Swapping-based retraining also has documented limits: It's All in the Name explores name-based substitution, and leakage persists through surrounding context. The direct comparison in the Baselines section settles the question. If the neutralized-corpus embeddings match our repair on WinoBias, OntoNotes and SimLex, the ratio adds nothing, and we will say so.
Why not just run RAN-Debias as published? RAN already improves neighborhoods and preserves SimLex, so on WinoBias, where its gap is 0.4, there is almost no room left. Our expected advantage lies elsewhere. RAN's choice of which neighbors to repel comes from a bias-based network built on projections, which is the very geometry that Lipstick shows to be unreliable, and its gender-word classifier needs WordNet and dictionaries. If RAN matches our repair on the cluster audits and sentiment, the contribution shrinks to the diagnostic in the first stage.
Why not just rank relations by how much they move under the intervention, without the resampling baseline? Rare-word similarities are noisy, so raw shift will preferentially flag them. This is the language analogue of masked protein models scoring rare mutations as harmful. The raw-shift ablation tests this directly. We predict it selects relations with lower average frequency and yields a worse bias-utility trade-off. If it does not, the biological analogy contributes nothing here.
Decisive Experiments
The decisive test is the edit study. We draw equal-size sets of relations from the top, middle and bottom $\omega$ deciles, apply only the edit term of the repair to each set, and measure the change in WinoBias gap, GIPE and SemBias against the change in SimLex-999, MR and IMDB accuracy, and OntoNotes F1. The proposal predicts a monotone pattern. Editing the top decile should reduce bias at little utility cost. Editing the bottom decile should cost utility with little bias reduction. If the pattern is flat or reversed, the ratio does not identify bias-carrying structure and the method fails.
The second check revisits the Lipstick audit. We report what fraction of within-cluster relations among the top-1,000 biased words are low-ratio. After repair, we rerun K-means, the RBF-SVM and neighbor-bias correlation, and we also report PNKA shifts on gender-defining words. The prediction is that definitional anchors move less than under Hard-Debias. If the remaining cluster purity is concentrated in low-ratio relations, that is evidence that some residual clustering is legitimate semantics rather than hidden bias, which would sharpen how the field reads Lipstick-style audits.
References
- Separating selection from mutation in antibody language models
- Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them
- A Bayesian Approach to Uncertainty in Word Embedding Bias Estimation
- Nurse is Closer to Woman than Surgeon? Mitigating Gender-Biased Proximities in Word Embeddings
- MDR Cluster-Debias: A Nonlinear Word Embedding Debiasing Pipeline
- Investigating the Effects of Fairness Interventions Using Pointwise Representational Similarity
- A Sitewise Model of Natural Selection on Individual Antibodies via a Transformer–Encoder
- Explicit representation of germline and non-germline residues improves antibody language modeling
- Diffusion Theory as a Scalpel: Detecting and Purifying Poisonous Dimensions in Pre-trained Language Models Caused by Backdoor or Bias
- It's All in the Name: Mitigating Gender Bias with Name-Based Counterfactual Data Substitution