When we ask if a word embedding model is "biased," we are usually asking how closely it associates certain social groups (like gender or race) with specific concepts (like career or science). To answer this, researchers use metrics like the Word Embedding Association Test (WEAT). However, a growing body of evidence suggests that these "rulers" are highly unstable. The perceived fairness of a model often depends less on the model itself and more on the specific mathematical choices made during the measurement process.
The Sensitivity of the Mathematical Lens
Measuring bias requires two main components: a way to measure the distance between word vectors (a similarity metric) and a way to summarize those distances across a group of words (a descriptive statistic).
Azarpanah et al. demonstrated in Measuring Biases of Word Embeddings: What Similarity Measures and Descriptive Statistics to Use? that these choices are not neutral. By testing combinations of similarity measures (like Cosine, Euclidean, or Manhattan distance) and aggregation statistics (like mean or median), the authors found that the detected severity of bias can fluctuate wildly. A model that appears biased under one combination might appear perfectly fair under another. This sensitivity means that current techniques may not provide a definitive assessment of a model's fairness, as the "signal" of bias is often drowned out by the "noise" of the measurement methodology.
The Problem of Metric Inconsistency
If our metrics were reliable, different tests should generally agree on which models are the most and least biased. Unfortunately, this is rarely the case.
In Interpreting Pretrained Contextualized Representations via Reductions to Static Embeddings, Bommasani et al. identified a "dramatic inconsistency" between different methods used to estimate social biases. This lack of consensus makes it nearly impossible to trust the validity of any single estimator. Elsafoury et al. further explored this in A Comparative Study on Word Embeddings and Social NLP Tasks, finding that different metrics yield such varied results that researchers cannot even establish a clear hierarchy or ranking of embeddings based on their inherent biases. When metrics disagree, it becomes difficult to determine if a new training algorithm actually improves fairness or if it simply optimizes for one specific, arbitrary metric.
Evaluating the Evaluators
Because we cannot take bias scores at face value, recent work has shifted toward "meta-evaluation"—evaluating the quality of the metrics themselves.
Spliethöver et al. introduced a method called Bias Silhouette Analysis (BSA) in Bias Silhouette Analysis: Towards Assessing the Quality of Bias Metrics for Word Embedding Models. Their analysis revealed a troubling reality: many existing metrics and word lists struggle to reliably distinguish between models that are intentionally biased and those that are not. The effectiveness of these metrics varies significantly across different types of bias and different embedding architectures.
This complexity is echoed by Sesari et al. in An Empirical Study on the Fairness of Pre-trained Word Embeddings. By comparing 15 different pre-trained models, the authors found that the choice of metric must be carefully matched to the specific type of bias under investigation. There is no "one size fits all" measurement; instead, the community must move toward a more nuanced, context-dependent approach to quantifying fairness in language models.
Key Foundational Work
Measuring Biases of Word Embeddings: What Similarity Measures and Descriptive Statistics to Use? (2021) — contributed 5 concept elements to this direction.
Interpreting Pretrained Contextualized Representations via Reductions to Static Embeddings (2020) — contributed 3 concept elements to this direction.
Bias Silhouette Analysis: Towards Assessing the Quality of Bias Metrics for Word Embedding Models (2021) — contributed 3 concept elements to this direction.