Enhancing Bias Assessment for Complex Term Groups in Language Embedding Models: Quantitative Comparison of Methods
Bibliographic record
Abstract
Background: Artificial intelligence (AI) is rapidly being adopted to build products and aid in the decision-making process across industries. However, AI systems have been shown to exhibit and even amplify biases, causing a growing concern among people worldwide. Thus, investigating methods of measuring and mitigating bias within these AI-powered tools is necessary. Objective: In natural language processing applications, the word embedding association test (WEAT) is a popular method of measuring bias in input embeddings, a common area of measure bias in AI. However, certain limitations of the WEAT have been identified (ie, their nonrobust measure of bias and their reliance on predefined and limited groups of words or sentences), which may lead to inadequate measurements and evaluations of bias. Thus, this study takes a new approach at modifying this popular measure of bias, with a focus on making it more robust and applicable in other domains. Methods: In this study, we introduce the SD-WEAT, which is a modified version of the WEAT that uses the SD of multiple permutations of the WEATs to calculate bias in input embeddings. With the SD-WEAT, we evaluated the biases and stability of several language embedding models, including Global Vectors for Word Representation (GloVe), Word2Vec, and bidirectional encoder representations from transformers (BERT). Results: This method produces results comparable to those of the WEAT, with strong correlations between the methods' bias scores or effect sizes (r=0.786) and P values (r=0.776), while addressing some of its largest limitations. More specifically, the SD-WEAT is more accessible, as it removes the need to predefine attribute groups, and because the SD-WEAT measures bias over multiple runs rather than one, it reduces the impact of outliers and sample size. Furthermore, the SD-WEAT was found to be more consistent and reliable than its predecessor. Conclusions: Thus, the SD-WEAT shows promise for robustly measuring bias in the input embeddings fed to AI language models.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.087 | 0.248 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.004 | 0.002 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.003 | 0.005 |
| Open science | 0.001 | 0.004 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".