FamilySearch's AI is improving with time, but I've seen similar issues there with the Canadian records. HTR is just hard for software.

FamilySearch is always a great place to start! They've digitized a lot of Connecticut records.

... and the having to pay to present thing.... :(

Here are a few of mine from last year: connecticutroots.org/2024/11/28/s...

The challenge really lies in getting across that the model that works for - say 1800-1810 - won't work for 1890, even if the record is the same and the boilerplate is similar. The language has evolved.

The Napoleonic Code specified a certain word order for a certain time period. It's definitely possible to train for a translation that would mimic that word order. It's essentially a variation of the "model translations" genealogists used to use. T

For context - I'm a translator.

A lot of genealogy records are functionally "boilerplate" in that you'll see the same and/or similar phrasing used across examples. If you can limit the usage enough, you can train for it. But the challenge becomes limiting the usage. Genealogy often forgets that language evolves; meaning changes.

And then even if you do manage to train for translation, what are you training for? Contemporary language? The form of the language used in 1810? 1870?

Most of what's being developed for genealogy right now is focusing first on transcription, so that things can be read with OCR or HTR. The problem is that transcription and translation are functionally two different skills, and transcription is much harder to teach a machine.

There's been an effort for a long time to get actual programs on translation into general genealogy education programs. The usual reason for it not happening is "there's not much interest." Meanwhile "how to use AI for translation" is getting lectured on... It would be nice if it could change.