Extracting "other" information from Azure Form Recognizer's Prebuilt Invoice Recognize output

Help Document Understanding
1

Hi all,
I did a comparison between Amazon Textract and Azure Form Recognizer. Azure Form Recognizer gives key value pairs as output and it contains these fields. This covers quite a bunch of fields. However Textract only covers very few fields (see here)

Now even though, Textract doesn’t identify many fields in a standard manner but it still recognizes key value pairs from the document well. For instance, if a document has a field P.O. # 1234567, Form Recognizer misses it but Textract extracts it and labels it as “other” fields.

I want to stick with Form Recognizer because the overall accuracy is better and it recognizes more fields which are missed by textract if they are not explicitly mentioned.
But the way Form Recognizer gives the output, I can’t find a way to extract those “other” fields other than having to parse the JSON and creating some rules. I was wondering that maybe I am missing something here and maybe Azure does provide some way to access these values directly similar to Textract.

Please comment if I need to add any more info here. TIA.

2

what I understand is you need to extract some custom fields from a form which are not getting autodetect using Recognizer?

Textract gets the unknown fields’ value but that needs to be validated and assign to a key. Best way when you have a n0n-standard for is to design a custom layout.

Form designer can recognize all the fields you need, just that if that field is not a standard one it won’t auto-recognize it, you can customize it and it will work just fine. I suggest you to have check the below video, that shows the extraction of customize fields that aren’t auto-recognized.

Extracting Data From Documents and Forms with OCR and Form Recognizer - YouTube

3

yeah I understand that we can train custom models but I have too many formats to be able to create a custom model for each. Prebuilt is working just fine.
What I am looking for is to something similar to the behavior of Textract here where it also recognizes fields it does not understand

4

Pre-built templates have the set of attributes to extract. That is static unless you customize so I believe azure form recognizer will extract the predefined entities only.

Textract has a different set of entity for pre-built models.

Both have their advantages. You can customize a bit , the one you think is giving better accuracy.

5

Could you provide some sample documents? Have you tried using Intelligent Form Extractor? Is that not ok for your use case? Are the forms not standard templates?