Azure Document analysis custom extraction model failing to recognize '.'(Period).

Question

Azure Document analysis custom extraction model failing to recognize '.'(Period).

BIGWORKS Support 0

When training the custom extraction model within the Document Intelligence Studio, the default Optical Character Recognition (OCR) engine appears to exhibit a limitation in accurately identifying the period character when it precedes numeric digits. To put it simply, I am not able to select the preceding period while labelling the documents.

The image below illustrates the mentioned issue.
User's image

Conversely, the OCR engine accurately recognizes and processes the period when it functions as a decimal separator following a numeric value.

User's image

This limitation is hindering the accurate labeling of fields containing such instances (e.g., ".789").

Any insights or potential solutions would be greatly appreciated.

Saideep Anchuri 7,040 Reputation points Microsoft External Staff Moderator

2025-05-07T01:39:12.5166667+00:00

Hi BIGWORKS Support

The issue with Document Intelligence Studio’s OCR engine failing to recognize periods preceding numeric values during document labeling stems from how OCR systems interpret context. Typically, a period is identified as a decimal separator following numbers (e.g., 123.45), but when placed before numeric values (e.g., .789), the system often misinterprets or overlooks it. This challenge arises due to predefined processing rules within OCR models. To address this, several immediate workarounds can be applied, such as manually adjusting labels by completing the initial labeling without the leading period and adding it afterward while verifying numerical consistency. Alternatively, leading decimals can be temporarily represented as 0.xxx during labeling and converted back to .xxx format in post-processing to maintain data integrity.

For more effective technical solutions, a custom training approach can enhance model performance. Developing a dedicated dataset that includes varied cases of leading decimals with different font styles, sizes, and high-quality images improves recognition accuracy. Additionally, pre- and post-processing techniques, such as automated validation scripts and regex-based corrections, can help identify and resolve missing periods while maintaining audit logs for tracking changes. Standardizing document templates with consistent spacing, ensuring uniform font styles, and optimizing image quality before processing can further improve OCR reliability. Quality assurance measures, including double validation of critical fields, checklist-based decimal point verification, and systematic accuracy reviews, provide structured error management.

Long-term recommendations involve enhancing the OCR model with advanced engines better equipped to handle leading decimals and exploring specialized solutions tailored for financial document processing. Process optimization through standardized labeling guidelines, predefined templates for decimal formats, and clear escalation procedures ensures a smoother workflow while reducing errors. Regular team training on handling edge cases reinforces accuracy and adaptability in OCR implementations.

Kindly refer below link: : Azure Document Intelligence Documentation,

overview-ocr

quick-reference

Thank You.
Ravada Shivaprasad 130 Reputation points Microsoft External Staff

2025-05-08T18:09:30.0166667+00:00

Hi BIGWORKS Support

Did you get any chance to check the response.

Thank you!
BIGWORKS Support 0 Reputation points

2025-05-09T07:37:17.19+00:00

Hi Saideep Anchuri

Thank you so much for the reply.

As you mentioned about training a custom model, the issue I am facing is actually while training a custom extraction model. Before labelling the documents, I am supposed to run the analysis button, after which I can select the data for labelling. Image below is the after-analysis of the document.

If you observe the image below, the highlighted text is all we can select for labelling. The period character is not highlighted, and I am not able to select it for labelling.

There is no issue if there is a number preceding the period character. Please let me know If I am missing something here.

Thank you.
Ravada Shivaprasad 130 Reputation points Microsoft External Staff

2025-05-12T18:01:32.55+00:00

Hi BIGWORKS Support

The issue with Document Intelligence Studio's OCR engine failing to recognize periods preceding numeric values during document labeling arises due to context-based processing limitations within OCR models. Specifically, during the analysis phase, the system does not highlight period characters when they precede numbers (such as in .14 Miles E), despite correctly recognizing periods that serve as decimal separators following numeric values. This challenge is a result of predefined recognition patterns within the Document Intelligence Read OCR model, which operates at a higher resolution than standard OCR systems but still adheres to context-sensitive processing rules. Several practical workarounds can be applied to mitigate this limitation.
One approach involves manual label adjustments, where the initial labeling is completed without the leading period, and it is added afterward while verifying numerical consistency. Another method is temporarily representing leading decimals as 0.xxx during labeling and converting them back to .xxx format in post-processing to preserve data integrity. Additionally, optimizing datasets by incorporating cases of leading decimals with various font styles, sizes, and high-quality images can improve recognition accuracy.
Pre- and post-processing techniques, such as automated validation scripts and regex-based corrections, also help identify and resolve missing periods while maintaining audit logs for tracking modifications. To enhance OCR reliability further, standardizing document templates with consistent spacing, uniform font styles, and optimized image quality before processing can reduce recognition errors. Quality assurance measures, including double validation of critical fields, checklist-based decimal point verification, and systematic accuracy reviews, provide structured error management. In the long term, improving the OCR model with advanced engines designed to handle leading decimals and exploring specialized solutions for financial document processing could result in better recognition performance.

Reference : Azure AI Document Intelligence.

Thanks
SeethaRam266 0 Reputation points

2025-05-13T10:13:30.7066667+00:00

Hi Ravada Shivaprasad

I cannot follow the manual label adjustments since no evidence confirms the consistent presence of decimal values from the document providers.

For example, there might be values like '14 Miles SE' where the distance is 14 but not 0.14 Miles.

Regarding the other solutions provided, I am facing the issue while training a custom model for extraction. And I don't have control over the OCR engine to follow/modify pre- and post-processing steps, as it is performed automatically.

Please provide any workarounds with steps to proceed further.

Thanks.

Your answer

Ravada Shivaprasad 130 Reputation points Microsoft External Staff

2025-05-08T18:09:30.0166667+00:00

Hi BIGWORKS Support

Did you get any chance to check the response.

Thank you!
BIGWORKS Support 0 Reputation points

2025-05-09T07:37:17.19+00:00

Hi Saideep Anchuri

Thank you so much for the reply.

As you mentioned about training a custom model, the issue I am facing is actually while training a custom extraction model. Before labelling the documents, I am supposed to run the analysis button, after which I can select the data for labelling. Image below is the after-analysis of the document.

If you observe the image below, the highlighted text is all we can select for labelling. The period character is not highlighted, and I am not able to select it for labelling.

There is no issue if there is a number preceding the period character. Please let me know If I am missing something here.

Thank you.
SeethaRam266 0 Reputation points

2025-05-13T10:13:30.7066667+00:00

Hi Ravada Shivaprasad

I cannot follow the manual label adjustments since no evidence confirms the consistent presence of decimal values from the document providers.

For example, there might be values like '14 Miles SE' where the distance is 14 but not 0.14 Miles.

Regarding the other solutions provided, I am facing the issue while training a custom model for extraction. And I don't have control over the OCR engine to follow/modify pre- and post-processing steps, as it is performed automatically.

Please provide any workarounds with steps to proceed further.

Thanks.

Share via

Azure Document analysis custom extraction model failing to recognize '.'(Period).

Your answer