Challenge index · Evidence reviewed 6 September 2026
Evidence overview
64 languages
African languages evaluated in AfroBench, published in 2025.
Ojo et al. · AfroBench · Findings of ACL, 2025 ↗Data scope: A research benchmark covering 15 tasks and 22 datasets; it does not represent every African language or every real-world service.
What the evidence tells us
Ojo and colleagues find substantial performance gaps between African languages and high-resource languages across most evaluated tasks. Results also vary with available language data. Their study makes evaluation in the intended language a necessary starting point for discovery.
Read the underlying source ↗Axum’s research agenda
Axum seeks interfaces that understand the language people actually use. Research should establish where code-switching or local terminology causes service failures, then test whether a new technical method improves completion of useful tasks. Commercial development depends on reliable performance and appropriate rights to contributed language data.
These are proposed research directions. Where research leads to an invention, patentability and commercial viability remain to be established.
Research directions
Code-switched agricultural services →
Yoruba–English and selected South African language pairs
How well can a service understand requests that switch languages during a conversation?
AfroCS-xs · ACL 2025 ↗Local-language information retrieval →
A selected language and public-service workflow
Can people find accurate information using the terms they use locally?
AfroBench · ACL 2025 ↗Selected literature
A starting library of published studies and institutional evidence. Source types and dates are identified below; inclusion does not imply endorsement of an Axum product.