Abstract
Most existing tools for detecting gender bias in Artificial Intelligence are designed exclusively for the English language, creating a significant bottleneck in AI ethics for global applications. This paper introduces a novel multilingual dataset and a sensitive, information-theoretic metric, $S_{JSD}$, to uncover gender stereotypes in Masked Language Models (MLMs) across diverse linguistic structures. By evaluating models like mBERT and xlmR across five languages, the study demonstrates that gender stereotypes are a universal phenomenon in MLMs and that traditional binary scoring methods often fail to capture the nuance of bias in non-English settings.
Introduction
Research into NLP bias has historically focused on English, often utilizing simple templates such as "[Person] is a [Occupation]." However, these templates are difficult to translate accurately due to linguistic variations. In German, for instance, the word for "doctor" and its associated article change based on gender (der Arzt vs. die Ärztin). In gender-neutral or "pro-drop" languages like Finnish and Thai, the "he/she" distinction central to English datasets may disappear entirely.
Furthermore, standard metrics like those used in the CrowS-Pairs (CPS) dataset rely on a binary "winner-takes-all" approach, checking only which sentence in a pair is more likely. This binarization discards significant information from the model's probability distribution, making it particularly difficult to detect bias in smaller, non-English datasets.
Methods
To address these challenges, the researchers proposed two primary innovations: a name-based gender-swapping strategy and the $S_{JSD}$ metric.
Gender Swapping via Names
To ensure gender signals survive translation in pro-drop or gender-neutral languages, the authors replaced pronouns with "monosex" names—names that are culturally and unambiguously associated with a specific gender (e.g., "Robert" vs. "Olivia").
The $S_{JSD}$ Metric
Instead of a binary choice, the authors utilized Jensen-Shannon Divergence (JSD) to measure the "distance" between a model's predicted probability distribution and the ground-truth token. The $S_{JSD}$ metric represents the difference between this distance for a stereotypical sentence versus a non-stereotypical one. This results in a continuous score between -1 and 1, preserving far more information than binary metrics.
Datasets and Models
The study utilized a refined version of the CrowS-Pairs dataset, consisting of 212 high-quality, human-translated sentence pairs for English (En), Finnish (Fi), German (De), Indonesian (Id), and Thai (Th). The researchers evaluated several Masked Language Models, including multilingual models (mBERT, xlmR) and English-specific models (BERT, RoBERTa, ALBERT).
Results and Key Insights
The $S_{JSD}$ metric yielded negative scores across all tested languages and models, indicating a consistent preference for stereotypical associations, such as linking "doctor" with male names.
Key findings include:
- Universal Bias: Multilingual models exhibit systematic gender-biased behavior in non-English languages similar to documented patterns in English.
- Metric Sensitivity: $S_{JSD}$ proved more sensitive than traditional metrics. For example, on the German dataset, the standard CPS metric failed to show bias (scoring near 50), while $S_{JSD}$ clearly identified it.
- Model Comparison: The xlmR model, trained on larger web-based datasets, generally showed higher levels of gender bias than mBERT, which was trained on Wikipedia. This suggests that larger, less curated training sets may amplify stereotypes.
- Subword Complexity: In many languages, models predict subwords (like the gendered suffix "##in" in German) rather than whole words, which can skew bias measurements if not properly accounted for.
Discussion and Limitations
While the $S_{JSD}$ metric provides a more robust tool for bias detection, the study is currently limited to a male/female binary. Representing non-binary identities in languages with strict grammatical gender remains a significant challenge. Additionally, while the current dataset scale is sufficient for identifying general bias, larger datasets would be required to compare specific cultural subtypes of bias. The authors emphasize that these tools should be applied with an understanding of local cultural nuances, as stereotypes vary significantly across the globe.
Summary updated