Closed hobodrifterdavid closed 1 year ago
For NLLB specifically this is not possible since it was not trained with multiple sources.
Other models may implement multi source translation in different ways. Some encode the task in the data, while others use a different architecture.
So let me close this issue which is too general. Feel free to open another one if you work with a specific multi-source model that is not compatible with CTranslate2.
It would be neat to be able to pass 5 strings (for example, existing UI string translations, or multiple subtitle translations) to use as a translation source for NLLB, and make a best translation to the target language. The additional source languages can remove ambiguity.
I can only speculate how this would be implemented, but it would be quite unique and handy. Feel free to delete. :)