AI Back Translation generator - protecting references and USFM tags

Scripture Forge
1

Our team has begun using the AI tool to draft our back translation. One thing we’ve noticed is that the algorithm spits out odd things in cross references (\x … \xo … \xta … \xt …\x*) and parallel passages lines (\r …). In cross references, it also changes the USFM tags (consistently, but not to anything it’s ever been fed from the training set of texts).

It seems the tool could be improved by cordoning off material in those tags and letting them be handled by a less opaque (i.e., non-ML) system, since properly done cross references and parallel passage lines are completely non-creative.

I’d be happy to correspond off-site with a developer to give detailed feedback and some ideas for what I have in mind. Thanks.

2

This just came to our attention a few days ago. We are working on it and will be preserving anything in an \x or \r (or \fig) instead of trying to “translate” it.

3

Hey Josh - thanks so much for your careful explanation. I’d be quite interested in connecting with you to get the detailed feedback you mentioned, which would be very helpful as we think about ways to improve the drafting in this area.

Could we follow up through email (michael_martin @sil.org) to set up a conversation? Thanks!