Issue with Custom Neural Model in Document Intelligence for Table Extraction

Question

Issue with Custom Neural Model in Document Intelligence for Table Extraction

Alex Dominici 0

I am having issues with a custom neural model for table extraction in Document Intelligence. In particular, during labelling the model sees two different words from two cells as one. So, I am not able to label correctely the dataset, which is causing inaccuracies in the data extraction process.
Is there any way to label the two words separately?

VasaviLankipalle-MSFT 18,676 Reputation points Moderator

2024-02-07T22:11:26.98+00:00

Hello @Alex Dominici , Thanks for using Microsoft Q&A Platform.

Is it possible to share the API version you are working on and how are you labeling the table data?

Have you tried selecting the auto label button? With this you can then edit the values as needed: https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept-custom-label-tips?view=doc-intel-4.0.0#auto-label-tables
Alex Dominici 0 Reputation points

2024-02-08T08:17:10.1966667+00:00

Hi @VasaviLankipalle-MSFT ,
I am using Azure.AI.FormRecognizer 2023-07-31 3.1 General Availability ( for Custom Neural Model).

The problem is upstream, in the 'label data' section when I load the image and execute 'run layout' the model sometimes interprets two handwritings of two different cells as being the same. (perhaps because they are too close together, but still in two different cells). To answer your question, I created a dynamic table where I associate each cell field with what is being extracted. The problem is that I cannot associate each cell's contents as two of them are taken as one. (and yes, I've already seen auto label button, but the result is the same because the problem occurs first.)

1 answer

Your answer

VasaviLankipalle-MSFT 18,676 Reputation points Moderator

2024-02-07T22:11:26.98+00:00

Hello @Alex Dominici , Thanks for using Microsoft Q&A Platform.

Is it possible to share the API version you are working on and how are you labeling the table data?

Have you tried selecting the auto label button? With this you can then edit the values as needed: https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept-custom-label-tips?view=doc-intel-4.0.0#auto-label-tables
Alex Dominici 0 Reputation points

2024-02-08T08:17:10.1966667+00:00

Hi @VasaviLankipalle-MSFT ,
I am using Azure.AI.FormRecognizer 2023-07-31 3.1 General Availability ( for Custom Neural Model).

The problem is upstream, in the 'label data' section when I load the image and execute 'run layout' the model sometimes interprets two handwritings of two different cells as being the same. (perhaps because they are too close together, but still in two different cells). To answer your question, I created a dynamic table where I associate each cell field with what is being extracted. The problem is that I cannot associate each cell's contents as two of them are taken as one. (and yes, I've already seen auto label button, but the result is the same because the problem occurs first.)

Answer 1

VasaviLankipalle-MSFT 18,676 Moderator

Hello @Alex Dominici , I can understand your issue. As a temporary workaround, try with the latest 2023-10-31-preview to see if that helps.

If you still face issues, please raise a support ticket under the Azure Portal.

Will definitely share this feedback with the product team.

I hope this helps.

Regards,
Vasavi

Share via

Issue with Custom Neural Model in Document Intelligence for Table Extraction

1 answer

Your answer