topic RE: How to improve accuracy in case of dynamic rows in Product Forum

How to improve accuracy in case of dynamic rows

kiranb25 — Mon, 03 Jul 2023 15:09:00 GMT

Hi,

I have a PO document(pdf) where the address data is dynamic in rows, Address may be in 3, 4 or 5 lines. In DFD we are assigning with multiple lines for the address field and trained with all format files in sufficient number (~100). But every time a new batch is loaded, decipher able to read only a few documents(40%) with correct address. for others either it is extracting less rows of data or more rows.

Eg: Decipher is reading only 3 lines in 5 lines of address data in a single block, or 4-5 lines in 3 line address.

refer below screenshot:

Any suggestions to improve the accuracy?

Thanks ,

Kiran.

------------------------------
kiran b
------------------------------

RE: How to improve accuracy in case of dynamic rows

Ben.Lyons1 — Wed, 05 Jul 2023 08:49:00 GMT

Hi Kiran,

Decipher should be able to handle that type of dynamic field. What version are you using? And what are the DFD settings for the field? Are you using a Capture ML model?

Thanks

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

RE: How to improve accuracy in case of dynamic rows

kiranb25 — Wed, 05 Jul 2023 10:28:00 GMT

Hi Ben,

Appreciate your response, I am using Decipher 2.2ver. DFD settings for this field is Assignable, Multiline and Miscellaneous parameters are set to multiple. I am not capturing any ML model for this Document type.

As of now, decipher able to identify very few documents with correct address.

I will try capturing ML model for this document type.

Any other suggestions.?

Thanks,

Kiran

------------------------------
kiran B
------------------------------

RE: How to improve accuracy in case of dynamic rows

Ben.Lyons1 — Thu, 06 Jul 2023 07:00:00 GMT

Hi Kiran,

I wouldn't create a capture model just yet, as that's more for optimisation than initial training.

Why are you using a misc parameter? This doesn't look like something that's required for reading this field.

Have you followed the document training best practices from our online help pages?

Thanks

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

RE: How to improve accuracy in case of dynamic rows

kiranb25 — Thu, 06 Jul 2023 08:59:00 GMT

Hi Ben,

Thanks for the suggestion,

Coming to the first question, anything that improves the accuracy to extract the dynamic row data apart from training more files with best practices.?

Thanks,

Kiran

------------------------------
kiran B
------------------------------

RE: How to improve accuracy in case of dynamic rows

Ben.Lyons1 — Wed, 26 Jul 2023 06:58:00 GMT

Hi Kiran,

If for each layout the address appears in the same place, you could set it as a static field using the misc parameter "StrictPosition=On". You'll need to restart your training and create a region large enough to capture the longest address.

Essentially this parameter will treat the field as if it never moves and will read anything that's in the box, though this will be specific to each layout i.e. vendor A's address doesn't have to be in the same place as vendor B's.

Thanks

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

RE: How to improve accuracy in case of dynamic rows

kiranb25 — Wed, 09 Aug 2023 11:43:00 GMT

Hi Ben,

The layout is changing as per address (no of rows), So after using "StrictPosition=On" decipher is reading the extra lines in case of 3 rows of address. We have tried to use other options like "Semantic=On" assuming will consider outline and stop reading the unwanted data. Still not able to achieve.

Appreciate any other suggestions.

FYI the pdf loos like below:

4 rows;

3rows:

5 rows:

After training the above formats , if we have batch with all the 3-address format of pdfs. Our target is to read only the address (Ship to)correctly.

Scenario 1: For some(3 lines of address file) decipher reading the values of ship to po as well considering it as 4, 5 lines.

Scenario 2: For 4 lines of address file, it is reading either 3 lines or ship to po as well considering it as 5.

Scenario 3: For 5 lines of address file, it is reading either 3 or 4 lines.

I hope you got the issue,

Note: We have connected with decipher support team, but the issue still persists.

Thanks,

Kiran.

------------------------------
kiran B
------------------------------

RE: How to improve accuracy in case of dynamic rows

Ben.Lyons1 — Wed, 09 Aug 2023 11:50:00 GMT

Hi Kiran,

Are all these addresses known? Could you create a list of them?

Thanks

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

RE: How to improve accuracy in case of dynamic rows

kiranb25 — Wed, 09 Aug 2023 12:09:00 GMT

Hi Ben,

Most of the address are in US region. We cannot create a list since they are customers, can be new address.

Can we use any regex to identify as address as a whole...?

Note: The provided addresses are manipulated in the above screenshot.

Thanks,

Kiran

------------------------------
kiran B
------------------------------

RE: How to improve accuracy in case of dynamic rows

Ben.Lyons1 — Thu, 10 Aug 2023 09:20:00 GMT

Hi Kiran,

I've got some Regex that validates the full address, but unfortunately I've not been able to get the region training to consistently identify the full address.

But you can try using "([\s\S]*[\r?\n ]*){3,5}[0-9]{5}$" .

I've raised this query with the development team and hope they will come back with some further advice shortly.

Thanks

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

RE: How to improve accuracy in case of dynamic rows

kiranb25 — Thu, 17 Aug 2023 09:07:00 GMT

Hi Ben,

I tried using the above-mentioned regex still not able to achieve. We are trying different possibilities as well.

Any update from dev team regarding the issue? Please let me know if any.

Thanks,

------------------------------
kiran B
------------------------------

RE: How to improve accuracy in case of dynamic rows

Ben.Lyons1 — Tue, 22 Aug 2023 09:12:00 GMT

Hi Kiran,

Apologies for the delay. The development team have indicated this scenario is not well catered for in Decipher 2.2, though it may be worth trying to set a higher threshold for the ML match. If you set the misc parameter "TemplateMinMatchPercent" to a higher value (default = 60), this may help separate more similar scenarios.

Aside from that, there are new features being added to Decipher 2.3 which better cater for this scenario.

Thanks

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

RE: How to improve accuracy in case of dynamic rows

kiranb25 — Thu, 24 Aug 2023 09:00:00 GMT

Hi Ben,

Thanks for your response,

We are trying to train the data as suggested along with few other misc parameters as below:

BorderTable=On
BottomStop="SHIP TO PO"
TemplateMinMatchPercent=80
StrictMode=1

We are still in training phase, But we have observed that decipher is not able to identify the table details in the same pdf document which we used to get the details before.

QQ: Does the above misc parameters makes to ignore the other table details in the same document? Any suggestions?

refer the below screenshot for pdf templet.

Thanks,

------------------------------
kiran B
------------------------------

RE: How to improve accuracy in case of dynamic rows

Ben.Lyons1 — Fri, 25 Aug 2023 07:33:00 GMT

Hi Kiran,

BorderTable might not work correctly for this example as there aren't borders between the rows.

BottomStop only works with single words, it can't be used with phrases and is for use in tables only. "Ship to" appears above the table in this example, worth being mindful of the impact this might have.

TemplateMinMatch will require some trial and error, because we don't record the % match for a respective page. So train 1 document, then upload 1 you want it to treat as a different layout. If it recognises it and captures more information that the DFD would automatically, increase the match % and restart the batch at the capture stage.

StrictMode is a different option for table recognition, it's worth trying if you've otherwise been unsuccessful. It takes at least 3 documents for it to be trained.

Thanks

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

RE: How to improve accuracy in case of dynamic rows

kiranb25 — Mon, 28 Aug 2023 09:22:00 GMT

Hi Ben,

Thanks for the response,

Have removed the border table, bottom stop parameters and trained the data again. Still, we see the issue with capturing the ship to address with different lines (read with 60%), but able to read the table data like before after removing the misc parameters.

Thanks,

------------------------------
kiran B
------------------------------

RE: How to improve accuracy in case of dynamic rows

Tejaskumar_Darji — Thu, 31 Aug 2023 09:33:00 GMT

In my experience also I've seen this issue where multiple lines in ship to and bill to address do not work consistently. I was using a multiline flag but still, it was not extracting the full address in all layouts. Also when the bill to and ship to is empty it starts picking a value from some other field that is adjacent to it. Can not keep adding exclude headers and exclude values for all kinds of layouts.

------------------------------
If I was of assistance, please vote for it to be the "Best Answer".

Thanks & Regards,
Tejaskumar Darji - https://www.linkedin.com/in/tejaskumardarji/
Technical Lead
------------------------------