Delivering machine learning advisory in low-resource languages.

The last mile of an ML system is linguistic

Most of the machine learning pipeline for crop disease advisory is language-neutral: an image arrives, a classifier scores it, a treatment recommendation is looked up. The failure point is the final step, where a probability distribution has to become a sentence a farmer will read, trust and act on. If that sentence is in the wrong language, or in the right language but the wrong register, everything upstream was wasted compute.

Hausa is spoken by tens of millions of people and is still poorly served by language technology. Terminology is the first obstacle: plant pathology has a settled English vocabulary and an unsettled Hausa one, where the local name for a disease varies by region and sometimes covers several conditions that the model distinguishes. Building the advisory means building the glossary, with agronomists and extension officers, before any text generation is worth discussing.

Templates before generation

It is tempting to hand the output problem to a large language model and ask for fluent Hausa. We deliberately do not. Advisory text is a safety surface: it names chemicals, doses and timings, and an error is not a typo, it is a ruined harvest or a health hazard. Our advisories are assembled from reviewed templates, written and approved in Hausa by people qualified to approve them, with the model selecting and filling rather than composing. The result is less fluent than generated prose and far more controllable, and every sentence that can reach a farmer has been read in advance by a human who understood its consequences.

Confidence has to survive translation

A classifier's confidence score means nothing to most readers in any language. The honest move is to translate confidence into the advisory's structure, not its adjectives. High-confidence identifications carry a direct recommendation. Mid-confidence results are phrased as a shortlist with distinguishing signs to check, in effect recruiting the farmer into the diagnosis. Low-confidence results say plainly that the image was not enough and route toward an extension officer. The system never dresses a weak classification in strong language, because the reader has no independent way to detect the mismatch, and the first bad recommendation ends trust in all the rest.

The interface is part of the language

Delivery conditions shape the writing as much as vocabulary does. Advisories are read on small screens, often shared devices, sometimes read aloud to others. Sentences are short. One recommendation per message. Numbers are stated in units in local use, not the units the agronomy literature prefers. And the advisory travels with its date and its subject image, so that when it is forwarded, as useful advice always is, it does not detach from the context that made it correct.

None of this is glamorous machine learning. All of it decides whether the model's accuracy ever becomes a field outcome. Language delivery is not a localisation step bolted on after the science. In advisory systems, it is the science.