When Semantic Overlap Is Not Enough: Cross-Lingual Euphemism Transfer Between Turkish and English
New study shows semantic overlap isn't enough for AI to detect culturally nuanced euphemisms across languages.
Researchers Hasan Can Biyik, Libby Barak, Jing Peng, and Anna Feldman published a paper analyzing cross-lingual euphemism detection between Turkish and English. They categorized Potentially Euphemistic Terms (PETs) and found a transfer asymmetry: semantic overlap alone doesn't guarantee positive AI model transfer, especially from low-resource Turkish to English. Performance can degrade for overlapping terms, challenging assumptions in multilingual NLP. This work was presented at the EACL 2026 workshop on NLP for Turkic Languages.
Why It Matters
Highlights a critical blind spot for global AI systems that must navigate sensitive, culturally-specific language.