An underrepresented language has comparatively little presence in digital data, computational resources, or artificial intelligence training material. That imbalance can leave speakers with less reliable access to language technologies. section 3.3
Large language models often perform worse when a language uses unfamiliar scripts, grammatical structures, or cultural conventions. Limited data can produce outputs that are grammatical but culturally inappropriate or confusing. section 3.3
The gap appears in several forms:
- Different linguistic features create distinct difficulties: Turkish and Finnish combine morphemes into complex words, while Mandarin uses pitch to distinguish meaning. section 3.3
- Cultural meaning can be missed through honorifics in Japanese, gendered nouns in Romance languages, and context-dependent pronouns in many Asian languages. section 3.3
A measurable performance gap
GPT-4 scored above 85% on English MMLU tasks but only 62% for Telugu, illustrating how benchmark performance can vary sharply across languages. section 3.3
