List of names with surname at top
I wonder if AI will ever be able to do this type of record properly… where the surname is at the top but the given name is at the bottom.
Answers
-
I doubt it as well as many other things about AI. I'm working on a similar page with Surname at the top and Given Name with ditto marks for all the 11 children —-this silly AI makes me do the same names on the same over and over and over again even though each child is listed only once. Sad because this is a full page of genealogy with parents, children and grandparents with dates all in German and found under names in the United States and shows this extended family who immigrated like mine in the late 1800's-early 1900's.
0 -
In my personal view, the term "AI" is being completely misused. AI doesn't do anything - it needs to be trained to do things and the crucial aspect is how that training is specified and designed. I was impressed with the speed of indexing of the last US census. No, it wasn't perfect, but it was incredibly better at reading handwriting than I'd ever expected a few years ago.
What's relevant to the OP's original comment is that, so far as I know, the AI indexing did not decide that each census sheet included a guy called "Federal Census". Why not? Presumably because the training specified where to look.
The sort of document that has the surname at the top needs to be identified and built into the training program. Essentially, it's only one step more complicated than having ditto marks in a census household.
The AI indexing will never cope with those documents if it's not been trained on their recognition and indexing. So it's a human based issue and not an AI issue. And let's face it, there are plenty of cases where human indexers have failed totally to recognise the document. A list of Churchings in one parish register (essentially a simple list of mothers' names and the dates that they were blessed) were treated as burials presumably because the humans doing the indexing realised that they weren't baptisms or marriages, so decided that they had to be burials because there are only 3 sorts of entry in parish registers. No, actually. This was, I emphasise, a human issue of a lack of training and knowledge. AI will simply hit the same problems as humans, and both can be fixed with more training in specific areas.
2 -
The problem I see is that we inexperienced AI people are reviewing and thus training the AI (from everything I've read here). So when inexperienced or uninformed indexers are doing this training, AI isn't going to learn very well.
2 -
AI can transcribe pretty well for full text applications where the user can see the context. That is not the same thing at all as AI indexing where it's all about generating an accurate structured metadata set, i.e. turning content into metadata.
To me the first stage is developing the AI equivalent of the Project Instructions (so, how the collection's metadata structure maps to the documents being indexed); training the AI to /use/ these instructions accurately, with, no doubt, a few iterations of the PIs, being stage 2 (while whether the AI can actually read the printed and handwritten information it encounters is a separate challenge).
What I have seen in the screenshots reviewers have posted looks like a mixture of all this, which does not feel like the right approach to me.
Stage 1 in particular needs really experienced indexers/reviewers with an in-depth understanding of the document set (including the ways in which multiple records may map to a single document) plus a full picture of how the record is broken down into personas and relationships. They also need to be 100% clear on the collection's persona and relationship metadata structure (something that is not easy to acquire while FS continue not to document collections' metadata structures for us - see also the myriad problems with 'on the fly' index editing).
(See this comment for a previous attempt to articulate some of this: https://community.familysearch.org/en/discussion/175624/a-i-indexing#Comment_592079.)
2
