The Geometry of Societal Bias in Word Embeddings

66 concepts68 papers1 proposal

When we train word embeddings, we map the statistical patterns of human language into a high-dimensional vector space. The core assumption of distributional semantics is that words appearing in similar contexts should have similar vector representations. However, because these models learn from uncurated web text, news articles, and historical corpora, they don't just learn linguistic definitions—they inherit the cultural prejudices, stereotypes, and systemic inequalities of the society that produced the text.

Understanding this phenomenon requires moving past the idea of bias as a vague ethical flaw and looking at it as a concrete geometric distortion within vector spaces.

The Geometric Anatomy of Bias

How does a social stereotype manifest mathematically? In a seminal paper, Tolga Bolukbasi et al. exposed this "ghost in the machine" by demonstrating that word embeddings capture gender stereotypes in the form of directional vectors in Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings.

If you take the vector difference between "man" and "woman" ($\vec{v}_{\text{man}} - \vec{v}_{\text{woman}}$), you isolate a direction in space that ideally should only represent biological gender. Yet, when you project gender-neutral occupations onto this axis, words like "programmer" or "doctor" align heavily with the masculine direction, while "homemaker" or "nurse" align with the feminine direction. This spatial distortion means that neutral concepts become clumped or entangled with specific demographic groups, creating a biased baseline for downstream systems.

This inheritance is not limited to binary gender. As Thomas Manzini et al. showed in Black is to Criminal as Caucasian is to Police: Detecting and Removing Multiclass Bias in Word Embeddings, these geometric distortions expand into multiclass biases encompassing race and religion.

The consequences are severe when these embeddings are used in real-world applications like resume screening. As explored by Kyra Wilson in Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval, embedding models systematically penalize marginalized groups (such as Black men) by ranking their names lower due to illicit proximities in the vector space. Furthermore, Fatma Elsafoury et al. demonstrated in SOS: Systematic Offensive Stereotyping Bias in Word Embeddings that embeddings exhibit systematic offensive stereotyping, mathematically linking profanity and slurs directly to marginalized social groups.

Confounding Signals: Grammatical Gender and Frequency

Disentangling pure social bias from other geometric properties is a major challenge. In languages with complex grammatical gender systems, the vector space is often dominated by syntactic rules rather than semantic stereotypes.

Shiva Omrani Sabbaghi argued in Measuring Gender Bias in Word Embeddings of Gendered Languages Requires Disentangling Grammatical Gender Signals that the arbitrary grammatical gender assigned to inanimate objects masks the true intensity of underlying social stereotypes. When these syntactic signals are stripped away, the measured social bias actually becomes significantly stronger, proving that standard metrics often underestimate bias in non-English languages.

Similarly, the geometric magnitude ($L_2$ norm) of an embedding often acts as a proxy for word frequency. As Yi Zhou et al. noted in On the Curious Case of l2 norm of Sense Embeddings, using frequency-based geometric features introduces a risk of inadvertently encoding or amplifying social biases derived from unbalanced training corpora.

Mitigation: Straightening the Vector Space

To fix these distortions, researchers have developed techniques that operate either after training (post-processing) or during training (train-time constraints).

Post-Processing Projections

Early methods relied on complex iterative projections to remove protected attributes. However, Pantea Haghighatkhah et al. showed in Better Hit the Nail on the Head than Beat around the Bush: Removing Protected Attributes with a Single Projection that a single, targeted Mean Projection can achieve the same linear guarding against bias while preserving the underlying semantic structure better.

Similarly, Sunipa Dev et al. in Attenuating Bias in Word vectors demonstrated that linear projections using common names as masked carriers can automatically scrub problematic directions. Going beyond simple projections, Tenzin Singhay Bhotia et al. proposed in Nurse is Closer to Woman than Surgeon? Mitigating Gender-Biased Proximities in Word Embeddings a neighborhood-rearrangement approach that actively pushes biased neighbors away while pulling semantic anchors closer.

Train-Time Disentanglement

Rather than scrubbing vectors post-hoc, some approaches isolate bias during training. Jieyu Zhao et al. developed GN-GloVe in Learning Gender-Neutral Word Embeddings, which forces gender information into specific, isolated dimensions while keeping neutral concepts unpolarized.

In a similar vein, Kyungwoo Song et al. utilized latent disentanglement and counterfactual generation in Neutralizing Gender Bias in Word Embedding with Latent Disentanglement and Counterfactual Generation to create gender-flipped variants of words, averaging them out to neutralize stereotypical associations during the learning process.

Dictionary Bootstrapping

A powerful alternative to uncurated corpora is using authoritative dictionary definitions as "neutral anchors." Byungkon Kang et al. demonstrated in Embedding Senses via Dictionary Bootstrapping that treating a dictionary as a closed graph allows models to learn unbiased sense-specific vectors.

This insight was expanded by Masahiro Kaneko et al. in Dictionary-based Debiasing of Pre-trained Word Embeddings and Haozhe An et al. in Learning Bias-reduced Word Embeddings Using Dictionary Definitions, showing that dictionary definitions can serve as a gold standard to filter out discriminatory noise while preserving or even enhancing semantic richness. By anchoring words to their formal definitions rather than their messy real-world co-occurrences, the geometry of the vector space remains aligned with objective meaning rather than historical prejudice.

Key Foundational Work

Embedding Senses via Dictionary Bootstrapping (2017) — contributed 1 concept elements to this direction.
Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings (2016) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_societal_bias_inheritance_ae83ca98,
  title = {societal bias inheritance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/societal-bias-inheritance/rdr_ae83ca980d574b7b81779a4740713118},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_societal_bias_inheritance_ae83ca98_v1,
  title = {societal bias inheritance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/societal-bias-inheritance/rdr_ae83ca980d574b7b81779a4740713118/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake