TL;DR
This paper introduces WEFE, a novel framework designed to standardize and compare different fairness metrics for word embeddings. By abstracting away metric-specific inputs and outputs, WEFE enables a unified evaluation of how well various metrics capture biases related to gender, ethnicity, and religion, revealing that gender bias is more consistently measured across metrics than other forms of bias.
What Problem Does This Paper Address?
Word embeddings, the numerical representations of words used in many Natural Language Processing (NLP) tasks, are known to inherit and amplify societal biases present in the text data they are trained on. This means embeddings can inadvertently associate certain attributes (like professions or traits) with specific demographic groups (e.g., gender, race, religion). While numerous "fairness metrics" have been proposed to quantify these biases, a significant challenge arises from their diverse input requirements and incompatible output formats. This makes it difficult to directly compare the effectiveness of different metrics or to understand how they relate to each other. Furthermore, it's unclear if models that perform well on fairness metrics also perform well on downstream NLP tasks. This paper aims to bridge this gap by providing a unified framework to evaluate and compare these fairness metrics.
Methods
The core contribution of this paper is the WEFE (Word Embeddings Fairness Evaluation) framework. WEFE tackles the input/output heterogeneity of fairness metrics by introducing an abstract view.
- Queries and Templates: WEFE defines a "query" as a pair consisting of a "target set" (words representing social groups, e.g., "man," "woman") and an "attribute set" (words representing traits or occupations, e.g., "programmer," "homemaker"). A "template" specifies the size of these sets (e.g., (2, 2) for two target sets and two attribute sets). Fairness metrics are then associated with specific templates they can handle.
- Unified Input: WEFE allows for the creation of "subqueries" that satisfy a given metric's template, ensuring that all metrics receive inputs in a format they understand, regardless of their original design.
- Ranking Aggregation: For each fairness metric, WEFE calculates a score for each embedding model across all relevant queries. These scores are then aggregated (e.g., by averaging) to produce a single fairness score per model per metric. This aggregation function must be consistent with the metric's inherent ordering of "less biased."
- Correlation Analysis: WEFE generates a matrix of rankings, where each column represents a fairness metric and each row represents an embedding model. This matrix allows for the calculation of correlation coefficients (e.g., Spearman) between the rankings produced by different metrics and across different bias dimensions.
This abstraction allows WEFE to encapsulate, evaluate, and compare diverse fairness metrics in a standardized manner.
Datasets
The paper does not introduce new training datasets for word embeddings. Instead, it focuses on evaluating existing, publicly available pre-trained word embedding models. The models used in the case study include:
- conceptnet-numberbatch
- fasttext-wiki-news
- glove-twitter
- glove-wiki-gigaword
- lexvec-commoncrawl
- word2vec-gender-hard-debiased
- word2vec-google-news
The paper also utilizes pre-defined query sets for evaluating bias, drawing from existing work on gender, ethnicity, and religion. These query sets are constructed using attribute words related to pleasantness, unpleasantness, intelligence, appearance, sensitive topics, occupations, and sentiment.
Key Insights and Ideas
- Standardization is Crucial: The core idea is that by abstracting away metric-specific inputs and outputs, WEFE provides a much-needed standardized way to compare fairness metrics. This allows researchers to move beyond anecdotal evidence and conduct systematic comparisons.
- Gender Bias is "Easier" to Measure: The study found that fairness rankings for gender bias tend to correlate well across different metrics. This suggests that current metrics are relatively consistent in identifying gender-related biases.
- Other Biases are More Elusive: In contrast, correlations between metrics for ethnicity and religion bias were significantly weaker. This indicates that measuring and comparing biases beyond gender is a more complex challenge, requiring further research into metric design.
- Fairness vs. Performance: The paper investigates the relationship between fairness rankings and performance on downstream tasks (using the Word Embeddings Benchmark - WEB). The findings suggest that there is no clear correlation between an embedding's fairness score and its performance on tasks like word similarity and analogy. This implies that debiasing might not necessarily improve general utility and vice-versa.
- Debiasing Effectiveness: The study observed that the debiased version of word2vec (word2vec-gender-hard-debiased) consistently outperformed its non-debiased counterpart across all fairness metrics for gender bias, reinforcing the impact of targeted debiasing methods.
Key Results
The paper presents a case study where seven pre-trained word embedding models are ranked using four fairness metrics (WEAT, WEAT-ES, RND, RNSB) across three bias dimensions (gender, ethnicity, religion).
- Gender Bias Correlation: Spearman correlation matrices showed high positive correlations between fairness metrics when measuring gender bias.
- Ethnicity/Religion Bias Correlation: Correlations for ethnicity and religion bias were substantially weaker, indicating less agreement among metrics for these dimensions.
- Model Performance: For gender bias,
word2vec-gender-hard-debiasedandconceptnetconsistently ranked high. The debiasedword2vecmodel generally outperformed its non-debiased version. - Fairness vs. WEB Performance: Rankings from WEFE (fairness) and the Word Embeddings Benchmark (WEB, performance) showed no clear correlation. For instance,
lexvecwas poorly ranked by WEFE but in the middle for WEB, andword2vecand its debiased version swapped positions between WEFE and WEB rankings. - Specific Model Behavior: Models like
glove-twitter,glove-wikipedia,lexvec, andword2vec-google-newswere rarely found in top ranking positions for fairness across various metrics.
Limitations and Future Work
- Subjectivity of Queries: The authors acknowledge that the design of queries (target and attribute sets) is a subjective aspect of the framework and can impact the obtained rankings. They suggest that defining a standard set of queries for the research community would be a valuable future direction.
- Metric Aggregation: Choosing an appropriate aggregation function for fairness scores can be non-trivial, especially for more complex metrics.
- Future Research: The paper calls for more research into developing fairness metrics that can consistently rank embeddings for dimensions beyond gender, particularly for ethnicity and religion. Further investigation into the trade-offs between fairness and performance on downstream tasks is also encouraged.
The authors have released WEFE as an open-source toolkit to facilitate further research in this area.
Summary updated