From co-creation to technical bias detection methods: an interdisciplinary showcase from the BIAS project

Area of Science:

• Computational Linguistics
• Sociolinguistics
• Artificial Intelligence Ethics

Background:

• Machine learning models and word embeddings can reflect and amplify societal stereotypes.
• Existing bias detection methods primarily focus on English, failing to capture language- and culture-specific nuances when applied to other languages.
• Direct translations of English bias benchmarks overlook unique linguistic and cultural contexts.

Purpose of the Study:

• To introduce novel metrics, BIAS-WEAT and BIAS-SEAT, for detecting biases in word embeddings and language models.
• To address the limitations of English-centric bias detection by developing methods for Dutch, German, Icelandic, Italian, Norwegian, and Turkish.
• To investigate how language models embed and reproduce context-specific biases.

Main Methods:

• Developed BIAS-WEAT and BIAS-SEAT metrics tailored for bias detection in non-English language resources.
• Utilized insights from co-creation workshops with native speakers to identify real-world biases in a hiring context.
• Translated these identified biases into technical evaluation metrics applicable to general-purpose language resources.
• Applied metrics to evaluate biases in language models across six different languages.

Main Results:

• Demonstrated that language models embed and reproduce biases specific to their linguistic and geographic contexts.
• Showcased the inadequacy of direct translation approaches for capturing cross-lingual and cross-cultural biases.
• Highlighted significant differences in bias manifestation across the studied languages.

Conclusions:

• Culturally grounded approaches are essential for effective bias detection in language models.
• The developed metrics provide a framework for assessing and mitigating language-specific biases.
• Future research should focus on creating diverse, culturally relevant datasets and evaluation tools for AI fairness.