Building efficient and effective OpenQA systems for low-resource languages

Emrah Budur, Riza Özçelik, Dilara Soylu, Omar Khattab, Tunga Güngör, Christopher Potts

Abstract: Highlights•We show OpenQA is feasible in low-resource language contexts without gold labels for training.•Key enabler of OpenQA in low-resource languages: weak supervision and unstructured data.•We demonstrate that only a few hundred gold examples suffice to evaluate OpenQA.•Growing knowledge sources impact OpenQA results based on retrievers’ noise handling capability.•We release SQuAD-TR, a large scale Turkish QA dataset derived from SQuAD2.0.