Tatoeba can be a useful resource to provide community-verified translations to be used as a translation memory and as a way to correct issues from machine translation models (T351748).

Once a library/service is available to check Tatoeba translations (T351872), this task proposes the integration in MinT. MinT will use this service as a translation memory. When a translation is requested to MinT, it will be checked against the translation memory to return a matching translation if available, and responding the machine learning response for translations not availabel in the translation memory.

Additional considerations:

• Although Tatoeba is proposed as the initial source for this translation memory, it should be possible to additionally integrate other existing translation memory systems. That is, MinT may be checking Tatoeba, and the translation memory from Translate before considering the translations from the machine learning models.
• As part of the ticket it will be decided whether the request to the machine-learning models will be done in parallel (quick response time at the risk of making a useless request) or only after checking the translation memory (increasing delay).