Low-resource language
A language with little digital text available for training, so models handle it less well.
Many languages with millions of speakers, such as several in Africa, South and Southeast Asia, have relatively little text online. Models trained on web data see little of them and perform worse. Researchers address this by collecting and digitising text, recording speech, building translation corpora, training dedicated tokenizers and creating benchmarks. India's AI4Bharat and the government's Bhashini mission have gathered datasets for the 22 languages of the Constitution's Eighth Schedule, and African groups such as Masakhane do similar work.