Answered

AI Builder - Correct Preview Errors?

(0) Share

Report

Posted on by Salle

Hi, I have just tried to create my first application using AI Builder.
Setup:

1) AI builder trained and published - Reads scanned PDFs and recognizes text/values
2) Using Power Automate to update Excel on Sharepoint.
The Confidence is 99%. Great! Or...?

The above works as intended however I have noticed that the applications sends out wrong values when reading the scanned table (sometimes blurry image). Example of error data output; blank, skip decimal maker "0 003", add Euro sign instead of a number etc.

I was thinking that additional training of the model could help with more examples. But I noticed that the problem already during preview of the Table during the setup of the model.

Is there a way to change and correct individual values during set up of the consolidated document (5) to make the AI model more accurate?

Thanks in advance!

Categories:

AI Builder

All responses (3)

Answers (1)

Sort by

Antrod Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!

Hi @Salle ,

Indeed, adding more documents to the model training set won't help when field are not recognized properly (can happen when documents have low quality for example).
There's no way to improve this OCR detection but of course internal AI teams and industry overall keeps improving the way characters are recognized.

If you have a way to improve automatically quality of the document, it could help (sometimes just rising the contrast solves the issue).

Because AI can't be 100% accurate (yet), we recommend including human review in your automated process when you detect results could be wrong. This could be done, for example, by storing suspicious results in Dataverse and having someone review and edit them regularly either directly in Dataverse or through a Power App.
Salle 4 on at

Like
a
(0)

Report
Copy link

Link copied!

Thanks for your reply @Antrod.

So there is no way to teach the AI model that it should always read the value as a decimal number since the table has consistent decimals values. Or manually correct a faulty table preview value in the training documents setup so the model is not inaccurately trained.
Verified answer

Antrod Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!

Indeed, AI Builder is based on Azure OCR engine which has the mission to recognize the text.
AI Builder then allows to build a model that recognizes how data is structured within a document.
If the text is not well recognized to begin with, custom models created in AI Builder can't improve it.

Getting OCR work better is a permanent work that is based on training models on massive set of pre-selected data and I'm confident it can be even more powerful in the future. But we could still have some texts that would be hard to detect because they may not be obvious to humans without context neither, especially when documents are of lower quality.