How to train custom model for different document layouts with the same set of labels?

Question

I'm trying to understand what is the best way to train a custom model for invoices in languages not supported by the prebuilt invoice model, french as an example.

As normal we will have many different invoice layouts from different vendors, but in all of them, we will extract the same set of labels (invoice number, amount, date, vendor name, etc).

Should I create a model per vendor and compose it? If I do so, do I need to train it for all vendors, or will it work for invoices that were not trained, but use the same verbiage as trained invoices?

user_11477081 · Answer

If you are trying to get just a few fields like the invoice number, amount, date etc. You can try the Invoice pre-built and see if it extracts the data you need. It is not yet trained on French or other languages invoices and should have lower quality but might work. If you are training custom models you will need to train a model per provider and then compose all the single models into a model composed. I would recommend to take your top providers and create models for them.

user_15752834 · Answer

I got an answer from Microsoft on MS QA site, see below:
"For invoices (I believe he meant English invoices) you should use the pre-built Invoice model, no training required - https://docs.microsoft.com/en-us/azure/cognitive-services/form-recognizer/concept-invoices.
If you need to train a model and not use the pre-built than yes a model per vendor\provider and compose them. Start with the top providers so that you get more coverage."

Find more information on the MS QA Question.

How to train custom model for different document layouts with the same set of labels?

2 Answers