cancel
Showing results for 
Search instead for 
Did you mean: 

Machine Learning didn't enhance accuracy?

MohamedSaheb
Level 3

Hi all,

We have set up ML for a Document Type that recently went to Production with a success rate of about 55% (55% of documents had all fields captured correctly by Decipher).

Our target threshold is 75%

We pushed 1000+ documents through manual verification with ML enabled. However, our first unattended run post ML is at a 52% success rate (lower than before ML was implemented)

Has anyone else experienced this, and how would you go about enhancing the accuracy?

Thanks!



------------------------------
Mohamed Saheb
------------------------------
7 REPLIES 7

BenLyons
Staff
Staff

Hi Mohamed,

What was your accuracy prior to adding the ML capture model?

Thanks



------------------------------
Ben Lyons
Senior Product Specialist - Decipher
Blue Prism
UK based
------------------------------
Ben Lyons Senior Product Specialist - Decipher SS&C Blue Prism UK based

Hi Ben,

It was roughly 55% before enabling ML...

Thanks!



------------------------------
Mohamed Saheb
------------------------------

Hi Mohamed,

That indicates that the DFD configuration probably needs some work, the capture ML model won't be able to improve that. Ultimately it's working from the same set up as the default rules-based model.

I would suggested reading through our best practice guidance and seeing if that helps improve things.

Thanks



------------------------------
Ben Lyons
Senior Product Specialist - Decipher
Blue Prism
UK based
------------------------------
Ben Lyons Senior Product Specialist - Decipher SS&C Blue Prism UK based

Thanks for the response, Ben.

We've spent quite a bit of time configuring the DFD. I've attached a snap of the target document type - we are reading some data fields within the red highlight ribbon.

One document will return 100% correct data. Another document might return blank fields throughout. And others will return incorrect fields. These documents are all bulk scanned (as a batch) into a folder which is then fed to Decipher so there is not much variance there.

Any further thoughts?



------------------------------
Mohamed Saheb
------------------------------

Hi Mohamed,

It would be difficult to comment without seeing the DFD. If you'd like some direct support for your use case, you may be entitled to some Expert Connect time. This would give you some time with our professional services team to work on your use case, this can be accessed via the Customer Support menu on the portal.

Check with your account manager what your entitlement is for Expert Connect.

Thanks



------------------------------
Ben Lyons
Senior Product Specialist - Decipher
Blue Prism
UK based
------------------------------
Ben Lyons Senior Product Specialist - Decipher SS&C Blue Prism UK based

Hi Mohamed, 

Do you have any improvement on this? Could you let us know what you did to improve?



------------------------------
Pranav
------------------------------
Pranav

Hi Pranav,

We had a consultation with a Blue Prism expert regarding this but unfortunately he was unable to improve things either. I assume the document we're using is not suitable for the current version of Decipher.

AWS Textract handles this document like a breeze, but it's a premium product so the costs are premium too. If your document volumes are low then it would be worth checking out.

Good luck!



------------------------------
Mohamed Saheb
------------------------------