GAND benchmark exposes gender bias in machine translation via contrastive attribution
New dataset reveals how MT systems default to stereotypes when gender cues are ambiguous.
Deep Dive
Machine translation systems continue to produce gender-biased translations. To understand how these systems translate gender when cues are absent, researchers present GAND, a gender-ambiguous natural data benchmarking resource of English source sentences. They translate a subset into two grammatical gender languages, add contrastive translations, and use feature attribution analysis to identify which source words drive gendered translations of an ambiguous referent.
Key Points
- GAND dataset contains over 1,000 English sentences with ambiguous gender referents, translated into two grammatical gender languages with contrastive pairs.
- Feature attribution reveals that profession-associated words (e.g., 'engineer' vs. 'nurse') drive gendered translations even when context is neutral.
- Accepted at EAMT2026, the resource is open-source and designed for reproducible bias auditing in commercial and open-weight MT models.
Why It Matters
Helps developers and enterprises detect and fix gender stereotyping in AI translation tools that serve millions daily.