Researchers have introduced AmchiBias, a novel benchmark designed to measure stereotypical bias within Goan identity groups, utilizing a dataset of minimal pairs in both English and Konkani. The benchmark evaluates five multilingual encoder models, revealing that while models show high bias for pan-Indian groups when queried in English, their performance in Konkani is near chance, indicating a lack of linguistic and cultural competence for hyperlocal identities. This work highlights a significant gap in low-resource multilingual NLP evaluation for specific community identities. AI
IMPACT Highlights a critical gap in low-resource multilingual NLP evaluation for hyperlocal community identities.
RANK_REASON Academic paper published on arXiv detailing a new benchmark for NLP bias. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →