January 2024 (9 months)
SURE-GB: Identifying Gender Bias in Machine Translation
SURE-GB is an innovative project focused on identifying and addressing occupation-related gender biases in machine translation (MT), particularly in translations between English, French, and Greek. Machine translation systems, widely used in our daily interactions, can inadvertently perpetuate societal biases. SURE-GB aims to tackle this issue by developing an automated service that detects three types of gender bias in MT:
- Under-representational bias: When gender assignment in translations reflects the low representation of a specific gender in linguistic corpora, even if it's accurate to reality.
- Stereotypical bias: When gender assignment reinforces harmful stereotypes, for example, translating "professor" as masculine despite the presence of many female professors.
- Algorithmic bias: When bias arises from technical shortcomings in MT systems, unrelated to actual gender representation in data.
Key Features:
- Curated Knowledge Graph: We are building a knowledge graph that encodes standardized data for occupations based on sources like the European Labour Force Survey (EU-LFS), the International Classification of Occupations (ISCO), and sociolinguistic statistics.
- Machine Learning Toolkit: Our toolkit uses the knowledge graph to automatically detect and categorize gender biases in translations, providing actionable recommendations to improve accuracy and fairness in machine translation systems.
Why It Matters:
Gender bias in machine translation doesn’t just reflect societal prejudices—it can reinforce them, contributing to deeper inequalities. By identifying and categorizing these biases, we aim to raise awareness and establish guidelines for more equitable translation practices.
Video Presentation: https://www.youtube.com/watch?v=YAkWg0ciiw8&ab_channel=UTTER-UnifiedTranscriptionandTranslation