Custom extraction model can’t improve the analysis result of the table

Question

Custom extraction model can’t improve the analysis result of the table

Yin Li 0

I am using Neural Model to train Custom extraction model to read pdf.
However, there are some problems in reading dynamic table. after analysis one of the pdf, the result show one field has been separated to next row. This case appeared in only one column so the next row only contain the value of the column. Other Columns display null in the same row.
I have tried to add the document to label data and redraw the label region to train the AI.
Even I put more copies to label data to label that the field is in same row. It still detect the field into two rows.
How can I do to improve the model?

Aryan Parashar 3,380 Reputation points Microsoft External Staff Moderator

2025-12-04T10:40:15.41+00:00

Hi Yin Li,

I understand that working with highly dynamic tables can be challenging, and I appreciate the effort you have already invested in exploring solutions. Even with variable structures and changing headers, it is still feasible to train a custom model to extract the necessary information.

For your reference, the following video provides best practices and guidance for effective custom labeling:

https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-labels?view=doc-intel-4.0.0#video-custom-label-tips-and-pointers

When dealing with document variations, it is important to evaluate which approach best suits your scenario. You may either:

Train a single custom model that accounts for all variations, or

Train multiple custom models tailored to different document formats and then create a composed model. A composed model allows you to combine up to 100 trained models under a single model ID, enabling automatic selection of the most appropriate model during analysis.

Here is supported documentation:

https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/how-to-guides/compose-custom-models?view=doc-intel-4.0.0&tabs=studio#create-a-composed-model

Let me know if you have further queries.

Your answer

Aryan Parashar 3,380 Reputation points Microsoft External Staff Moderator

2025-12-04T10:40:15.41+00:00

Hi Yin Li,

I understand that working with highly dynamic tables can be challenging, and I appreciate the effort you have already invested in exploring solutions. Even with variable structures and changing headers, it is still feasible to train a custom model to extract the necessary information.

For your reference, the following video provides best practices and guidance for effective custom labeling:

https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/train/custom-labels?view=doc-intel-4.0.0#video-custom-label-tips-and-pointers

When dealing with document variations, it is important to evaluate which approach best suits your scenario. You may either:

Train a single custom model that accounts for all variations, or

Train multiple custom models tailored to different document formats and then create a composed model. A composed model allows you to combine up to 100 trained models under a single model ID, enabling automatic selection of the most appropriate model during analysis.

Here is supported documentation:

https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/how-to-guides/compose-custom-models?view=doc-intel-4.0.0&tabs=studio#create-a-composed-model

Let me know if you have further queries.

Share via

Custom extraction model can’t improve the analysis result of the table

Your answer