AI Builder vs Text Recognition

[Anonymous]

Hello,

I am using the AI Builder to attempt to extract data from our forms, which have extensive tables and are mostly filled out by hand. During the table tagging process, there are many cells where the hand written text is not recognized at all. There are no bounding boxes and I cannot draw one.

However, using if I use text-recognition in a flow, the hand written text is correctly identified.

There are other instances where the AI builder sees text, but stops the bounding box early. The text-recognition flow catches it all and delivers the appropriate results.

I was under the impression that both of these services used the same tools to identify and OCR text. Why can one "see" the text while the other can't? Is there a way for me to force the model to see the text? 

Thank you.

[plarrue]

Hi @Anonymous!

Thanks for reaching out

In the choose information to extract, did you combine Field and Table? or only Table?

Thanks!

[Anonymous]

Thank you for replying. 

I have tried both. Initially I was using the table feature. However, once that didn't work, I tried it again with fields. It didn't resolve the issue.

Thank you

[Anonymous]

In an attempt to understand this more, I've focused on the sections that cause the most issues, which are simple numbers.

I've selected "Show detected words". You can see that one a single "1" is identified. If any "1" is found in this format from my other samples, they are always seen as a "-". Never an "i", "L", or "1". However, sending this through the text recognition flow, I get get the following results: 1 2 2 - 1 - . Not perfect, but closer.

While I can potentially understand the text not being recognized, what truly confuses me that that the two applications perform the analysis differently.

[antoinec · Accepted solution]

Hi Coweringomega, we're in the process of updating models we use in AI builder to match the latest and greatest from Form Recognizer. Deployment will kick-off next Monday with all regions updated by end of September / early October.

In the future we'd like to get to a state where users could opt-in to using newer preview versions of models as they get released by the Form Recognizer team to provide both Generally Available models by default for users looking for stability and Preview models for users looking to leverage some of the upcoming capabilities.

I hope this helps a bit 🙂