Answered

AI Builder misses text after a colon in document processing

(1) Share

Report

Posted on by bleupen

Hi!

We're using a custom AI model to process PDF documents that our clients sent to a mailbox. Each PDF document has 7 pages with a lot of fields that need to be filled in. One of the fields is a filenumber with the following format XX:000000

So this can be for example TD:123456. Some clients fill in the form a bit different, so sometimes it's being filled in as TD: 123456 or TD 123456.

When we train our model with this values and test the model after successful training, it seems impossible for the model to identify the full value when this is being filled in as TD: 123456. The only part that's being extracted is TD:

The other possible values are being detected correctly. It looks like it's struggling with the colon (:).

The accuracy score for the field is 99% and in the quick test it returns a confidence score of 99% as well, despite it's not detected correctly.

What can we do with the AI model to get this value correct?

Thanks in advance.

Categories:

AI Builder

I have the same question (0)

All responses (3)

Answers (1)

Verified answer

bleupen 22 on at

Like (1)

Report
Copy link

Link copied!

The issue is resolved, there was to less difference in the documents trained in the different collections. Therefore the wrong collection was used and our trained documents in the collection weren't even used. The collection being used can be found in the Power Automate Flow action. Thanks for your help!

Was this reply helpful? Yes No
bleupen 22 on at

Like (0)

Report
Copy link

Link copied!

Hi plarrue,

Thanks for your repsonse.

We've added more documents with this filenumber, but getting the same results. There are 8 documents in this collection now. There are some more fields where a colon is used, and the AI consistently stops extracting/recognizing all text after this character.

Thanks again!

Was this reply helpful? Yes No
plarrue Moderator on at

Like (0)

Report
Copy link

Link copied!

Hi @bleupen

Thank you for sharing your use case. One way this can improve is by adding more documents for training to your model.

Could you please try by retraining your model with more documents where the field is TD: 123456 and see if you have a better result?

Thanks!

Was this reply helpful? Yes No