Abstract
Word embeddings are often criticized for capturing undesirable word associations such as gender stereotypes. However, methods for measuring and removing such biases remain poorly understood. We show that for any embedding model that implicitly does matrix factorization, debiasing vectors post hoc using subspace projection (Bolukbasi et al., 2016) is, under certain conditions, equivalent to training on an unbiased corpus. We also prove that WEAT, the most common association test for word embeddings, systematically overestimates bias. Given that the subspace projection method is provably effective, we use it to derive a new measure of association called the relational inner product association (RIPA). Experiments with RIPA reveal that, on average, skipgram with negative sampling (SGNS) does not make most words any more gendered than they are in the training corpus. However, for gender-stereotyped words, SGNS actually amplifies the gender association in the corpus.
Introduction
Word embeddings like Word2Vec and GloVe are essential tools in NLP but are notorious for reflecting societal biases. Common examples include stereotypical analogies such as doctor:nurse :: man:woman. While researchers have developed metrics like the Word Embedding Association Test (WEAT) and debiasing techniques like subspace projection, the theoretical foundations of these tools have remained opaque. This paper investigates whether subspace projection is a mathematically sound correction and evaluates the reliability of standard bias metrics.
The Flaws of WEAT
The Word Embedding Association Test (WEAT) is the current gold standard for measuring bias. However, this research proves that WEAT is fundamentally flawed because it systematically overestimates bias. WEAT implicitly assumes that attribute words (e.g., "male" vs. "female") occur with equal frequency in the training corpus. When this assumption is violated, the metric breaks.
As shown in the table above, one can manipulate WEAT's attribute sets to "prove" with high statistical significance that completely neutral words like "door" or "dog" are highly gender-biased. Because of these flaws, the authors propose the Relational Inner Product Association (RIPA), a more robust metric defined as the inner product of a word vector and a one-dimensional relation vector.
Bias Amplification in Embeddings
A key finding is that embedding models do not necessarily invent bias in neutral words; rather, they amplify existing biases present in the training data. By comparing the genderedness of words in the raw corpus versus the resulting embedding space using RIPA, the authors found that Skip-Gram with Negative Sampling (SGNS) significantly increases the gender association for words that are already gender-stereotyped or gender-appropriate.
Quantitatively, SGNS amplified the absolute genderedness of gender-appropriate words by an average of 0.291 and gender-biased words by 0.197. In contrast, truly neutral words saw a negligible change (-0.006), suggesting that the model preserves the neutrality of words that lack a strong prior association in the corpus.
Theoretical Soundness of Subspace Projection
The paper provides a rigorous proof that debiasing via subspace projection is not merely a heuristic. Under the condition of perfect reconstruction in matrix factorization models, post-hoc projection is mathematically equivalent to training the model from scratch on a perfectly unbiased corpus. However, the authors caution that vectors should not be normalized before debiasing, as length encodes critical frequency information.
Unsupervised Debiasing and Results
To improve debiasing without losing valid gender-specific information (e.g., keeping the distinction between king and queen), the authors introduced an unsupervised heuristic. This method automatically decides whether a word should be debiased by comparing its projection on "gender-defining" versus "bias-defining" vectors.
This approach significantly outperformed previous supervised baselines. It preserved 94.9% of valid gender-appropriate analogies (compared to 16.5% in prior work) while successfully removing 63.3% of gender-biased analogies.
Conclusion
This work clarifies the mathematical nature of bias in word embeddings. While standard metrics like WEAT can be misleading, RIPA provides a more reliable measure that reveals how models amplify corpus-level stereotypes. Furthermore, the proof of equivalence for subspace projection validates it as a theoretically sound method for creating fairer embeddings. Future work will extend these findings to other forms of bias, such as race and religion.
Summary updated