How to get better results extracting from tables in Custom Extraction.

Fred Vasquez Pinheiro Gomes 0

Hi everyone,

I'm working on a custom extraction model using Azure Document Intelligence and encountering some challenges with table parsing. Some of my documents contain tables where the columns are very close together, and in some cases, there is no clear separation between columns. This often leads to merged cells or incorrect data extraction from the tables.

I've tried adjusting the model's labeling and training, but still struggling. Does anyone have advice on how to optimize models for better accuracy when extracting tables? Are there specific techniques or configurations that could improve the model's ability to distinguish between closely spaced columns?

Any guidance or recommendations would be greatly appreciated.

Thanks in advance!

Fred

santoshkc 9,400 Reputation points Microsoft Vendor

2024-09-18T12:54:29.89+00:00

Hi @Fred Vasquez Pinheiro Gomes,

Thank you for reaching out to Microsoft Q&A forum!

Improving table extraction accuracy, especially with closely spaced columns, requires a strategic approach. One key step is to add more varied examples of complex table structures during training, focusing on problematic layouts where columns are too close together. Preprocessing the document before analysis can also be beneficial- enhancing the contrast or sharpening the image can make the table boundaries more distinct, which helps the model detect column separation more accurately. Additionally, you could manually adjust the document by separating columns where needed, ensuring the input is better structured for extraction.

Post-processing the model's output is another useful technique. You can handle cases where cells have been merged incorrectly, automatically splitting them based on the expected structure.

Moreover, leveraging the Document Layout API can give you more granular control, as it provides detailed visual analysis of tables based on font size, alignment, and other layout cues. Finally, checking for skewness in scanned or photographed documents and correcting the orientation before processing can ensure the tables are properly aligned, further improving extraction accuracy.

For more info: Build and train a custom extraction model.

I hope this helps. Thank you.
santoshkc 9,400 Reputation points Microsoft Vendor

2024-09-19T17:24:22.4166667+00:00

Hi @Fred Vasquez Pinheiro Gomes,

Following up to see if the given response was helpful. Thank you.
santoshkc 9,400 Reputation points Microsoft Vendor

2024-09-20T14:39:32.7566667+00:00

Hi @Fred Vasquez Pinheiro Gomes,

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution, please do share that same with the community as it can be helpful to others. Thank you.
Fred Vasquez Pinheiro Gomes 0 Reputation points

2024-09-23T15:02:37.86+00:00

Thank you for your response and suggestions! In some of the templates our product will be processing, clear column separation isn't always present and manually adjusting those documents before analysis isn’t a feasible option for us right now, as they're defined by external vendors.

Regarding your point on leveraging the Document Layout API for more granular control, could you clarify what you meant? Using Template build mode instead of Neural? and how does it provide detailed visual analysis of tables based on font size, alignment, and other layout cues? Additionally, could you explain how detecting skewness and correcting the document's orientation before processing could help with table extraction accuracy in cases where the document lacks clear column boundaries?

Any further insights on these topics would be greatly appreciated!

Thank you for quick response.
santoshkc 9,400 Reputation points Microsoft Vendor

2024-09-24T17:32:59.2+00:00

Hi @Fred Vasquez Pinheiro Gomes,

Thank you for your follow up query.

Using the Document Layout API provides more detailed insights into a document's structure, such as font size, alignment, and spacing cues, which can help distinguish closely spaced columns. It doesn't require fully switching to Template mode, but Template mode can offer more control in handling repetitive table structures. Correcting skewness in scanned documents helps the model by ensuring straight rows and columns, reducing errors in table extraction and preventing merged cells.

I hope this helps. Thank you.
Fred Vasquez Pinheiro Gomes 0 Reputation points

2024-09-24T18:17:26.6566667+00:00

Thank you!
MicahlS Sa 0 Reputation points

2024-09-25T18:29:14.57+00:00

I’ve dealt with similar issues when working with closely spaced columns in tables using Azure Document Intelligence. Here are a few suggestions that might improve your model's accuracy:

Preprocessing: If possible, try pre-processing your documents to enhance table structure clarity. This can include increasing the spacing between columns or adding lines that make column separation more evident. If you’re able to adjust the source documents before extraction, this can help the model distinguish columns more effectively.

Region Labeling: When training your model, ensure that you’re using precise region labeling for each column. If the columns are too close, the model may interpret them as a single region. You could try breaking down the table into smaller labeled sections and see if that improves the accuracy.

Table Structure Configuration: Look into configuring the table structure recognition settings. Azure Document Intelligence allows some customization to better identify table boundaries and column breaks. You may need to adjust settings related to cell boundaries, merging logic, or how the model handles grid recognition for closely spaced tables.

Multiple Passes for Extraction: Consider running multiple passes with different extraction models or combining table extraction with form extraction. Sometimes breaking down a table into parts or using a hybrid approach for different sections can yield better results.

Post-Processing: You might also want to implement post-processing logic after extraction to clean up or restructure the data. This could include splitting merged cells or adjusting column positions in your data based on certain rules.

Document Layout Changes: In cases where document layout flexibility exists, slightly adjusting the layout by increasing column spacing or using borders could make a significant difference in extraction accuracy.

It might take a combination of these methods to achieve the best results, but refining your model’s labeling and exploring these options should help improve its performance.

Share via

How to get better results extracting from tables in Custom Extraction.

Your answer