Using the pre-build Layout model of Azure Form Recognizer v3.0, I see a decrease in accuracy on a lab result PDF for the newer preview models, can I share this to be added to the training set?

Question

Doing a poc to validate how accurate the pre-build generic Layout models are on multipage pdf's with multiple tables in different dimensions (PDF's are lab results of chemical substances). Based on the detected text and tables post-processing is done to extract all the required data.

While using the 2022-06-30 preview model, some table data elements were not part of the detected tables:

Also other parts on this page, clearly visually boxed as table, are not recognized as a table.

Comparing this with the 2021-09-30 preview model, this is a step back, as this 'older' model sees the whole page as a single table. The blank lines between the tables are causing the confusion of the latest model I assume. However, it seems an intermediate step between being able to detect all tables separately which is a clear step in the right direction.

Would like to share this full pdf with the product team to be added to the training / testset to help to increase the accuracy of the new pre-build models. How can I do this?

Answer 1

@Devliegher, Frank The best way to submit feedback to the product team is to use the smiley icon on the top right corner of the form recognizer studio. This redirects to a page to provide feedback and documents/screenshots of scenarios that you feel the service could do better. This feedback is triaged by our product team to improve the experience and you can choose to be contacted by Microsoft for any feedback you have provided through this form. I hope this helps!!

If an answer is helpful, please click on or upvote which might help other community members reading this thread.