Azure Form Recognizer - Random data extraction

Question

Hi!

Initial Situation:

We are using Azure AI Custom extraction model for data extraction from different forms.

A trained model with 42 data files as the data basis has a fix position for the data type: pickup_date.

The pickup_date is configured as a sub type: date dmy. This value is always positioned in the middle of the (trained) documents (see reference 1). We achieve an overall accuracy of 95.2% for this value (see reference 2).

Issue:

With several documents identically to the trained documents, Azure is recognizing random data for the pickup_date value (see reference 3); recognizing data in the upper part of the document which was never trained. Also including regular string data into an date value dmy.

Question:

1. Why is Azure recognizing this random data despite perfect accuracy in trained model as well as perfect documents provided - identically to trained data set?
2. Why is Azure keeping a confidence of 50+% for a pickup_date value in the format dmy with random string data?

Since this has quite significant impact on our operations, I appreciate any suggestions for solutions.

Reference 1:

Reference 2:

Reference 3:

Thanks in advance and best regards

Oskar