Source-linked AI summary
Learning Gender-Neutral Word Embeddings
Jieyu Zhao, Yichao Zhou, Zeyu Li, Wei Wang, Kai-Wei Chang
TL;DR
Word embeddings can inherit gender stereotypes that affect downstream applications, while existing projection-based debiasing can propagate classifier errors and remove useful gender information. The paper proposes GN-GloVe, which reserves gender information in selected dimensions and neutralizes the others during training. Experiments show that GN-GloVe reduces occupational-word gender bias, identifies gender-definition pairs accurately, and reduces downstream bias while retaining comparable coreference performance.
Problem
Word embeddings trained on human-generated corpora can encode gender stereotypes that affect downstream applications, while prior projection-based debiasing can propagate classifier errors and remove useful gender information.
Method
GN-GloVe jointly learns word vectors while reserving protected gender information in certain dimensions and neutralizing gender influence in the others.
Results
GN-GloVe reduces occupational-word gender bias by 35% relative to GloVe, achieves 97.7% accuracy on gender-definition word-pair analogies, and reduces WinoBias bias with comparable OntoNotes performance.
Takeaways & Limitations
The approach isolates protected attributes while preserving word proximity and functionality, and it can be applied to other embedding models, languages, and societal stereotypes.
Takeaways & Limitations
Projection-based alternatives require identifying gender-neutral words with a classifier and can remove gender information needed in domains such as medicine and social science.
Abstract
from arXiv · showhide
Word embedding models have become a fundamental component in a wide range of Natural Language Processing (NLP) applications. However, embeddings trained on human-generated corpora have been demonstrated to inherit strong gender stereotypes that reflect social constructs. To address this concern, in this paper, we propose a novel training procedure for learning gender-neutral word embeddings. Our approach aims to preserve gender information in certain dimensions of word vectors while compelling other dimensions to be free of gender influence. Based on the proposed method, we generate a Gender-Neutral variant of GloVe (GN-GloVe). Quantitative and qualitative experiments demonstrate that GN-GloVe successfully isolates gender information without sacrificing the functionality of the embedding model.
1 Introduction
Word embeddings can encode gender stereotypes that affect downstream systems, motivating a training method that separates protected gender information from gender-neutral content. GN-GloVe jointly learns this separation while preserving selected gender information and avoiding a separate gender-neutral-word classifier.
- Embedding models trained on human-generated corpora can associate gender-neutral words such as “programmer” more closely with “male” than “female.”
- Gender-biased embeddings can contribute to sexist coreference systems and discriminatory ranking in resume filtering, job recommendation, and search applications.
- Prior projection-based debiasing can propagate classifier errors and remove gender information needed in domains such as medicine and social science.
- GN-GloVe reserves protected gender information in certain vector dimensions while neutralizing other dimensions during training.
- By jointly identifying gender-neutral words during learning, GN-GloVe avoids a separate classifier and its associated error-propagation issue.
- Experiments report that GN-GloVe isolates protected attributes while preserving word proximity and improving interpretability of word representations.
2 Related Work
Related work establishes that social stereotypes appear across NLP systems, datasets, and embeddings. Prior studies document these biases and their temporal variation, but the cited work did not provide constructive solutions.
- Word Embeddings: Word embeddings are widely used as foundational representations across many NLP applications and training approaches.
- Stereotype Analysis: Gender stereotypes have been observed in online advertising, web search, online reviews, and coreference resolution systems.
- Stereotype Analysis: Studies have identified gender, racial, religious, and age-based stereotypes in corpora, along with changes in gender and ethnic stereotypes over time.
- Stereotype Analysis: Researchers attributed these problems partly to biases in datasets and word embeddings, but the cited studies did not provide constructive solutions.
3 Methodology
GN-GloVe extends GloVe by partitioning each word vector into gendered and neutralized components, reserving protected-attribute information in selected dimensions. Its objective preserves word proximity while restricting gender information to the gendered component, using jointly learned word categories rather than a separate classifier.
- Model setup: GN-GloVe uses GloVe co-occurrence learning while treating gender as the protected attribute.The model constructs a word-to-word co-occurrence matrix and uses center and context embeddings.
- Model setup: Each word vector is divided into neutralized w(a) and gendered w(g) components, with k dimensions reserved for gender information.The neutralized component has dimension d−k, while the gendered component has dimension k; this paper sets k = 1.
- Gender representation: The method categorizes vocabulary words as male-definition, female-definition, or gender-neutral and reserves protected-attribute information in w(g).Gender-neutral words are jointly identified during embedding learning, and their neutralized component is intended to be independent of gender influence.
- Objective: The minimization objective combines the GloVe word-proximity term with two terms that restrict gender information to w(g).The two gender-related terms use male- and female-definition seed words and two distance metrics.
- Objective: The gender-related objectives either minimize distances between the two groups or push gendered components toward specified extremes.For the second objective, β1 and β2 are set to 1 and −1, respectively.
- Optimization: The gender direction is estimated by averaging differences between predefined female–male word pairs and held fixed during each epoch's stochastic-gradient updates.Holding vg fixed reduces training computational complexity; vg is re-estimated at the beginning of each epoch.
4 Experiments
Experiments show that GN-GloVe separates gender information into designated dimensions, reduces gender stereotypes in neutral words, and preserves embedding functionality across intrinsic and downstream evaluations.
- Separate protected attribute: GN-GloVe preserves definitional and stereotypical gender associations in designated dimensions while separating them from other latent aspects.Gender-definition words separate in w(g), while associations for words such as “housekeeper” and “doctor” remain represented.
- Separate protected attribute: GN-GloVe reduces gender projections for gender-neutral profession words, whose GloVe vectors exhibit strong stereotypes.For words such as “doctor” and “nurse,” GN-GloVe projections onto the gender direction are closer to zero, substantially diminishing gender information in w(a).
- Stereotype analysis: The average gender projection is 0.080 for GloVe, 0.019 for Hard-GloVe, and 0.052 for GN-GloVe, corresponding to a 35% reduction versus GloVe.A larger projection indicates greater bias; Hard-GloVe retains less gender information, while later results compare gender-definition discrimination.
- Gender relational analogy: GN-GloVe identifies gender-definition pairs with 97.7% accuracy and outperforms GloVe and Hard-GloVe on the generalization subset.The subset contains 40 instances associated with two word pairs not used as seed words during training.
- Word similarity and analogy: GN-GloVe achieves higher accuracy on word-similarity tasks while its analogy score slightly decreases, indicating preserved word proximity.The evaluations use standard similarity and analogy benchmark datasets, with Table 2 reporting Spearman correlation and accuracy respectively.
- Coreference resolution: On coreference resolution, GN-GloVe matches GloVe and Hard-GloVe on OntoNotes while distinctly reducing WinoBias gender bias, with GN-GloVe(w(a)) reducing it further.WinoBias bias is measured by the absolute F1 difference between pro-stereotype and anti-stereotype subsets.
5 Conclusion and Discussion
The paper concludes that its gender-neutral embedding training algorithm is general, with applicability beyond the binary-gender setting left for future work.
- The proposed algorithm trains gender-neutral word embeddings and is presented as a general method.
- The method can be applied in any language when a list of gender-definitional seed words is available.
- Future work includes modeling properties such as sentiment and extending analysis beyond binary gender.