FairQE Framework Targets Gender Bias in Machine Translation Quality Estimation
A new multi-agent framework named FairQE has been developed by researchers to address gender bias in Quality Estimation (QE) for machine translation. While QE systems evaluate translation quality without needing reference translations, research indicates a consistent bias towards masculine forms in gender-neutral situations, often giving higher scores to translations that misalign with gender, even when it is clear. FairQE identifies gender indicators, creates gender-flipped versions, and integrates traditional QE scores with bias-reducing reasoning from large language models through a dynamic aggregation method. This adaptable approach maintains the advantages of existing QE models while correcting gender biases. The framework's efficacy was validated across various gender bias evaluation scenarios, as detailed in arXiv preprint 2604.21420.
Key facts
- FairQE is a multi-agent framework for mitigating gender bias in Quality Estimation (QE).
- Existing QE models exhibit systematic gender bias, favoring masculine realizations in gender-ambiguous contexts.
- QE models may assign higher scores to gender-misaligned translations even when gender is explicitly specified.
- FairQE detects gender cues and generates gender-flipped translation variants.
- It combines conventional QE scores with LLM-based bias-mitigating reasoning.
- The framework uses a dynamic bias-aware aggregation mechanism.
- FairQE operates in a plug-and-play manner, preserving existing QE model strengths.
- The research is published on arXiv with identifier 2604.21420.
Entities
Institutions
- arXiv