I am currently working with Document Intelligence (formerly known as Form Recognizer) on a custom model and am encountering a problem when processing multi-page documents. Specifically, the issue involves tables that span multiple pages. Problem: When a table entry starts on one page and continues on the next, the continuation part on the second page is not captured. This results in information loss, whereas complete table entries that fit within a single page are correctly recognized. Visual Example: Page1 Header 1 Header 2 Row 1 Row 1 - Text Row 2 Row 2 - Beginning of longer Text Page2 Header 1 Header 2 Row 2 - End of Text from Row 2 Row 3 Row 3 Using Document Intelligence the part of Row2 - End of Text from Row 2 would be lost. Question: Is there a specific configuration or approach I could use to ensure that table entries continuing across pages are fully captured? Has anyone else experienced this or can recommend best practices how to cope with it? PS: I have found this Thread before, but it wasn't of much help unfortunately.

Issues with Handling Multi-page Tables in Document Intelligence (Custom Model)

Kohl, Konstantin 0

I am currently working with Document Intelligence (formerly known as Form Recognizer) on a custom model and am encountering a problem when processing multi-page documents. Specifically, the issue involves tables that span multiple pages.

Problem: When a table entry starts on one page and continues on the next, the continuation part on the second page is not captured. This results in information loss, whereas complete table entries that fit within a single page are correctly recognized.

Visual Example:

Page1

Header 1	Header 2
Row 1	Row 1 - Text
Row 2	Row 2 - Beginning of longer Text

Page2

Header 1	Header 2
	Row 2 - End of Text from Row 2
Row 3	Row 3

Using Document Intelligence the part of Row2 - End of Text from Row 2 would be lost.

Question: Is there a specific configuration or approach I could use to ensure that table entries continuing across pages are fully captured? Has anyone else experienced this or can recommend best practices how to cope with it?

PS: I have found this Thread before, but it wasn't of much help unfortunately.

dupammi 6,560 Reputation points Microsoft Vendor

2024-04-22T10:19:22.5733333+00:00

Hi @Kohl, Konstantin

Thank you for using the Microsoft Q&A forum.

Tabular fields in Document intelligence support cross page tables by default. To label a table that spans multiple pages, you need to label each row of the table across the different pages in a single table. As a best practice, it's recommended to ensure that your dataset contains a few samples of the expected variations. Please follow this documentation: https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept-custom-label?view=doc-intel-4.0.0#tabular-fields

Multi page tables: When tables span multiple pages, label a single table. Add documents to the training dataset with the expected variations represented—documents with the table on a single page only and documents with the table spanning two or more pages with all the rows labeled. https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept-custom-label?view=doc-intel-4.0.0#create-a-balanced-dataset

you can create table tags by selecting Add a new table tag. Specify whether the table has a fixed number of rows or variable number of rows depending on the document and define the schema.

In this kind of cross-page table scenarios try with Custom Neural model for best results.

I hope this helps. Thank you.
Kohl, Konstantin 0 Reputation points

2024-04-22T14:13:26.6+00:00

Thank you @dupammi .

Unfortunately this does not anwer my question. What you descriped works well if the data of one row does not span over 2 pages. In this case it works well just like you descriped. Maybe it was not clear so i created an example PDF where not only the table but also the data of one entry spans over 2 pages. Sample Document with Table.pdf

The description of Emil in the example table would be missing the second part from the second page (red circled part in screenshort) while all other entries would be sucessfully extracted.

Does this make my problem more clear?
dupammi 6,560 Reputation points Microsoft Vendor

2024-04-23T08:48:33.4033333+00:00

Hi @Kohl, Konstantin

I see, thank you for the additional information and sorry for the trouble you are facing. In this case, I would recommend you raise a support case through Azure portal.

I hope you understand. Thank you!