question

KuntalVarun-1860 avatar image
0 Votes"
KuntalVarun-1860 asked GiftA-MSFT commented

Getting low accuracy on two fields after labelling using the tool, Form Recognizer, Custom Label API, please help.

I need help with recognition of two particular fields- "credit date" and "credit type". Getting low accuracy (training ~30%) after labelling and even lower on the test set (~10%). The receipts along with the two mentioned fields look like this (highlighted):

30712-1.jpg


30540-2.jpg

I am using Custom Label API after labelling, tagging and training.
I think as these two fields appear at different places relative to other fields due to different number of entries in in different receipts.
Is there anything that I can do to improve these fields accuracy.


azure-cognitive-servicesazure-form-recognizerazure-computer-vision
1.jpg (142.3 KiB)
2.jpg (79.0 KiB)
· 2
5 |1600 characters needed characters left characters exceeded

Up to 10 attachments (including images) can be used with a maximum of 3.0 MiB each and 30.0 MiB total.

Thanks for reaching out. Have you tried using the sample labeling tool? Here are some tips to help improve accuracy. You can also use the sample labeling tool to generate label files.


0 Votes 0 ·

Actually tested using the sample labeling tool and applied those tips.

0 Votes 0 ·

1 Answer

GiftA-MSFT avatar image
0 Votes"
GiftA-MSFT answered

Thanks for following up. We recommend that you label each value on the form, for example date_p1, date_p2, date_p3 as the label names for each value. If your forms have tables with varying numbers of rows, label at least one form with the largest possible table. Furthermore, it seems like you're trying to label a single model for different form types. As every hotel invoice has a different structure, we recommend you create a model per hotel vendor with at least 5 sample invoices from that hotel, and then compose those into a single composed model. In some scenarios, table extraction doesn't happen automatically so manual tagging is necessary, however, we plan to support more complex scenarios as the service evolves. Hope this helps!


5 |1600 characters needed characters left characters exceeded

Up to 10 attachments (including images) can be used with a maximum of 3.0 MiB each and 30.0 MiB total.