Migration Humanitarian Health Collective

Translation systems for low-resource languages are usually trained once, deployed flat, and never adapt. That is exactly where the quality breaks down. Different communities use the same nominal language differently, with their own register, terminology, and edge cases. A single static model serves all of them equally badly.

For a humanitarian-health context, equally badly is not an acceptable floor.

We built a translation system paired with an online tuning loop. Community members can mark up edits and corrections on the translation surface itself, in place. Those corrections are not discarded after the session. They are routed back into community-specific tuning, so the model that serves a given community converges toward that community’s actual usage.

The tooling around the loop is explicit about what gets sent where, with the goal that a community using the system can reasonably understand how their feedback shapes their model.

The result is translation quality that compounds with use, on a per-community basis. The longer a community uses the system and feeds back corrections, the better it performs for them, instead of regressing to a generic average. That is the property the project was set up to prove out, and it is the one that matters in the field.