A I Indexing
I wonder how long it will take the A I to learn how to do this well. From what I'm seeing doing name review, it's doing a very poor job.
Answers
-
I think letting AI do indexing is a bad idea. I had a few today that repeated the name; like James James, then the surname when I could clearly see that it was just James. Human eyes are better.
4 -
Reading handwriting is clearly something AI is getting very good at, and that's very helpful in a FS context.
And very well structured, accurate and predictable content (especially a clearly printed form that applies to all images and that results in one record per image) is very susceptible to automated indexing because the structure can be overlaid accurately on the content.
But there's a Neil Armstrong sized leap between any of that and indexing a City Directory, for example, because of the amount of human brain power needed to impose the structure (i.e. the Project Instructions) on the content before even starting indexing.
It is not clear to me how far FS has got with this element, but a stage where AI is an aid only, and can learn from the human, feels to me to be called for, at least at the beginning of each project.
In the meantime, presumably normal PIs continue to exist for AI indexing, because human reviewers need them. I guess the next stage would be for those PIs to be based on the rules AI has come up with for getting from content to metadata. The Indexing community's review of such PIs would be an essential check on the process imo.
@SerraNola would be v interested in your opinion on all this should you ever have a spare minute!
0 -
Currently there are uncountable indexing and transcription errors on FS. Places with the same or similar names conflated as the same place even when they are 1000 miles apart.
I imagine because the original humans didn't actually read what was written on the record. or the name or the date of the record collection.
Handwriting is always going to be an issue, but the printed name on the record set is accurate.
And we also endure places transcribed as personal names.
Example a local directory that lists the names of the people in a town or village.
The entry says Smith, William and then place name and address, the transcription has the place name inserted between Smith and William.
Such as Smith, Whoville, William.
If AI can fix this mess even just with printed not handwritten records I say go for it.
0 -
Monica Ross_1, the name review I've been doing is all in cursive and AI isn't doing well in capturing the whole name, is grabbing words that aren't names, and spells wrong (although humans do that, too).
0 -
I agree with @barbaragailsmith1, AI isn't doing well in capturing names and doesn't read cursive writing well at all. Additionally, we need to teach humans the basic of reading and writing better in order to accurately index such things as censuses and obituaries whether the originals are typewritten or in cursive. A knowledge of basic geography is also in order — the continents haven't moved more than a few inches in thousands of years.
1 -
The handwriting problems you are all encountering make me realise that my comment 'reading handwriting is clearly something AI is getting very good at' was unhelpful. AI is getting good at producing transcriptions from handwriting - there, the context is available to the user and therefore misinterpretations of names don't necessarily prevent the results from being useful. A misinterpreted name in AI-generated metadata has far more impact.
0 -
I agree that it seems AI is definitely struggling with many aspects of the names. It commonly gets the given and surnames reversed. I have even noted many errors in transcription of not only cursive names, but legible typed names. I find that hard to believe that it cannot distinguish typed letters.
1 -
AI cannot distinguish typed letters that are not part of a person's name because it doesn't read the column headings and neither do the human reviewers who check them the first time they come thru name review. Sorry, but we "ancients of age" see these same names come up over and over again
1 -
Sinceramente, a mi me gusta más el método de antes, pero encuentro que mi relación con la IA, en lo que me tocado revisar, ha cometido bastantes errores, los mismos que han mencionado en los otros comentarios, en mi caso para nacimientos, en los que se mencionan padres y abuelos, los intercambia pone un abuelo como padre y viceversa, no reconoce otros nombres, pone fechas y paises que no están en la imagen, y en lo que me tocó me colocan una sola página y el registro se completa con una página anterior o posterior que no podemos visualizar, por lo tanto el dato del declarante o padre del niño no se puede ver, aparece la firma al final pero no siempre es legible o firma otra persona, y no se si a ustedes les ha pasado pero entre editar, resaltar el nombre y tratar de completar varios errores, cuando le doy guardar me sale un cartelito que dice que el tiempo ya pasó, que hay que actualizar y volver a hacer todo de nuevo, así no se puede hasta límite de tiempo tenemos ahora, no cuento con la velocidad de la IA obviamente, pero entorpece bastante, frustra y así me parece que no muchos van a querer colaborar, a mi me decepciona, extraño el familysearch de antes, pero igual voy a seguir colaborando en lo que pueda, aunque el rendimiento no va a ser el mismo, yo creo que tienen que analizar y reflexionar más si lo que realmente se quiere para family es "CANTIDAD" o "CALIDAD", sí, optaron por la cantidad, pero con muchos errores, yo creo que muchos más de los que como humanos podemos cometer, y espero no haber ofendido a nadie, pero es lo que siento más un poquito de decepción….
1 -
Google Translate (not its greatest effort I suspect) of Miriam's comment:
Honestly, I like the old method better, but I find that my relationship with the AI, in what I have had to review, has made quite a few mistakes, the same ones that have been mentioned in the other comments, in my case for births, in which parents and grandparents are mentioned, it exchanges them, puts a grandfather as the father and vice versa, it does not recognize other names, it puts dates and countries that are not in the image, and in what I had to do, they put a single page and the record is completed with a previous or later page that we cannot visualize, therefore the data of the declarant or father of the child cannot be seen, the signature appears at the end but it is not always legible or another person signs, and I don't know if it has happened to you but between editing, highlighting the name and trying to complete several errors, when I click save I get a little sign that says that the time has passed, that it is necessary to update and do everything again, so it cannot be done until the time limit we have now, I do not have the speed of the AI obviously, but it hinders a lot, frustrates and it seems to me that not many They are going to want to collaborate, it disappoints me, I miss the familysearch of before, but I will continue to collaborate in what I can, although the performance will not be the same, I think they have to analyze and reflect more if what is really wanted for family is "QUANTITY" or "QUALITY", yes, they opted for quantity, but with many errors, I think many more than we as humans can make, and I hope I have not offended anyone, but that is what I feel most a little disappointed...
3 -
@MandyShaw1 Thank you for translating @Miriam Beatriz Perez says everything: many of us agree
0 -
This new AI experiment (so far) is a FAILURE. It is yielding nothing but wasted time, frustration, and discoragement. Indexing has been a Godsend for so many people who need the opportunity to serve and are physically unable to do so in other ways. The new name and place checking activities are not emotionally rewarding and give no real sense of accomplishment. FamilySearch - You are spending a dollar just to save a dime! Is this really worth it? I have yet to hear one single positive remark or good experience reported from anyone here. (but I would love to hear someting positive - so have at it)!
1 -
In doing the Quick Name indexing A1 version I ran into several occasions when a surname was left off after correcting the initials of the given names and can't figure out how to add in the name that is left off. Also, colors are hard to see, can they be changed to something else?
0 -
@StrongViki In Quick Name Review, when only the initials have been highlighted, add the surname to the initials in the editing box.
0 -
I'm not really liking the AI at all. I like the old way better, I can get through a batch must faster doing it the old way. I can read what kind of document I am indexing, with the AI I really don't have any idea what it is. Please bring the old way back or give us a choice
1 -
Quick Review activities are not indexing; rather they are the process of correcting AI to improve its function. As for the "old way" it's already been replaced so your best bet is to work with Quick Review. I like Full Name Review and Verify Places best — or you can take a look at the following links
https://community.familysearch.org/en/discussion/117900/how-do-i-participate-in-get-involved#latest
0 -
Maybe if enough of us complain, they'll come up with a new version of the old indexing. They did it with desktop indexing. I seriously doubt it will happen, but there's always that small glimmer of hope.
My favorite verifying places. I can knock those out pretty fast.1 -
I may be the odd one. I in fact LIKE the new process. In other AI projects I've been involved with (work-related) the start can be a little rocky due to system quirks & users adjusting to a new process but, this is temporary. Users improve, the AI engine learns and engineers adapt feedback. Then comes the moment where everything falls into sync and the rhythm that follows lifts and expands progress in great strides.
Very honestly as a reviewer I've seen many human-indexed records that were no better than the AI, (and when I was a beginner I probably botched a few too). In the roadmap of IT development, we're in the learning phase and the AI engine is always in the learning phase. Human scribes produce a variety of handwriting and format of organization as different and unique as the individual creating the record. It's taken me many decades to learn to interpret different types of cursive. What the AI engine has learned about it in a very short time is simply stunning.
I like the role of coaching AI, and think it suits us. Our inputs and adjustments to each document we help index trains the AI engine to perform better. And we gain experience for our efforts.
Hurrah for Israel!
1 -
@MichaelDurwardNash I appreciate your comments. I second your opinion that AI is indeed a good thing. Right now we are in the learning phase. Just like a toddler learning to walk, there will be many ups and downs and stumbling along the way and a bit longer until the kid learns to walk and eventually run freely. I have been involved in FS indexing from the beginning and I know well the lengthy indexing and review process of the past. AI is only taking baby steps right now, but it needs trainers to show it the ropes. Our contributions of time doing AI reviews is NOT a waste of time. We are making AI smarter and faster, taller and stronger.
That being said, the Family Review tools do need more TLC. And that is something only humans can do. (See my other posts.) Right now, Family Review is only available for Spanish and Portuguese records, so many of you may not have used the review tools yet. Hopefully many of the glitches will be worked out by the time Family Review is available in other languages. I am working on an online presentation that provides workarounds for the current version. Please DM me if you want to stay on the loop.
1 -
I must say that I was completely overwhelmed by the accuracy of the AI tool in transcribing documents written in cursive, which I recently encountered when using Full-Text Search. When a Community member posted the idea (a year or two ago) of using this procedure in relation to cursive documents, I must admit I thought this rather far-fetched and something that would only happen many more years in the future.
In contrast, I have seen some appalling indexing / transcribing of records that has been carried out manually by volunteers. I know this might sound cruel, but when I have had the option to say "thank you" to the volunteer (via a link against the source), I have felt more inclined to want to offer constructive criticism of their work - given that only a person (like me) with many years experience of making searches could ever have found such a badly indexed record!
However, for the most part I can only offer great praise for the work of volunteers, without whose efforts I would never have traced a number of my family branches so far back.
I have never felt I was capable of indexing myself: yes, I have easily found the records of the individuals I have been specifically looking for when scrolling through page after page of "unindexed" documents, but haven't been able to decipher many of the other names within those documents, so would have made a thoroughly bad job if I had been involved in an indexing project of that material!
My final point concerning the "switch over" to using AI relates to the suddenness of it all. It could easily produce the impression of rather a lack of empathy towards dedicated indexers who have been abruptly deprived of their main pastime.
I work practically every day of the year within Family Tree - adding profiles and correcting relationships - and know what it would mean to me if I was suddenly deprived of that activity. Therefore, I can really empathise with all those "full-time" indexers who are suddenly having their work taken away from them. Seriously, it must be like an employee being suddenly informed they are receiving compulsory redundancy. I really hope they can find a meaningful alternative, albeit that might mean working on projects away from FamilySearch.
0
