Add a mixed_trg option for corpus pairs #766
Description
Some projects that don't make a practice of creating back translations. When there's a need to review their vernacular text with an outsider (e.g., a translation consultant), the effort to create a back translation is challenging. And, since the team doesn't have an existing base of back translation text, these projects are not good candidates for current drafting techniques.
As a method for bootstrapping a new back translation of a book X, existing published translations in the back translation language can be used. The model would be trained by pairing books different from the vernacular project with the corresponding books from one of the published translations. Then, the trained model would be used to translate book X into the back translation language.
Ideally, the books of the vernacular translation would be evenly spread across multiple published translations, so that the back translation does not follow a specific published translation too closely. The current book-by-book method of distributing the books among the published translations is cumbersome and likely less optimal. A random distribution, similar to the mixed_src option for randomly selecting from multiple sources, would be preferred.