Decipher cannot remember the value I trained it to capture

[John_GilbertDu1]

On our case, there is an invoice number that must be captured, but no matter how many times we selected that value/region (which the position is fixed) in Data Verification stage, Decipher's Machine learning doesn't seem to take effect. the next time we import the same file, the Invoice number still can't be captured and remains blank

has anyone here encountered the same and was able to resolve it?

[ArunGandhi]

Hey can you try this instead: https://nanonets.com/blog/invoice-processing-using-blueprism-rpa/

Basically have seen issues with Decipher which is why I'm presenting an alternate solution.

[HowardBlake]

John I am seeing exactly the same behavior as you have described.  

Is there anyone who has an understanding of how to get Decipher to retain the Capture information used when selecting a field value in a document.  I am also reimporting the same document multiple times and yet Decipher fails to capture the field I keep selecting when verifying the batch in the Decipher UI.

[Ben.Lyons1]

Hi John and Howard,

It's quite tricky to answer this without seeing your exact scenario, so I will do my best.

Firstly, I would ask that if you have enabled the ML capture model in Document Type, that this be disabled during the initial training. Decipher has a native training mechanism that can provide a strong degree of accuracy and the ML model can be enable once this has been brought up to speed.

If the value is on a strongly colour background this may be causing the OCR engine some difficulty, so for example, white text on a dark background can make it more difficult to extract the text. This might be indicated where Decipher does not create a clear region around the text you're looking to read.

Does the text have a "header" near by such as "Invoice Number: 123456"? If so have you added this to the Sample Headers field in your DFD? This helps train Decipher on its location.

In addition you can add "StrictPosition=On" to the Misc Parameters column for the fixed field, this will inform Decipher that it only needs to look in the specified location. It should then read whatever it finds there.

If training has been carried out multiple times on this document and changes have been made to the DFD, it can take Decipher a little more time to retrain the DFD. Deleting the DFD will not delete the training data, as it's not linked (there are good reasons for this, but difficult to cover at a high level).

Let me know if any of this helps, or if you can provide any more context I'll see how else I can help.

Thanks

Ben

[HowardBlake]

Ben

Many thanks for your response - it is appreciated.

I am actually using the BP Training sample invoice (Busted Knuckles Motorsport) to help me develop a BP process whilst I wait for the client to send me their pdf form that will drive the ultimate process. So this helps me answer your first few questions:

Q1. Is ML capture model in Document Type enabled?

A1. Yes


Q2. Is the value on a strongly coloured background.
A2. Nope - Example invoice has black text on white background

Q3. Does the text have a "header" near by such as "Invoice Number: 123456"? If so have you added this to the Sample Headers field in your DFD? This helps train Decipher on its location.
A3. I have NOT tried this but will do so - thank you

Q4. Have I added "StrictPosition=On" to the Misc Parameters column for the fixed field, this will inform Decipher that it only needs to look in the specified location. It should then read whatever it finds there.
A4. I have NOT tried this but will do so - thank you

Q5. If training has been carried out multiple times on this document and changes have been made to the DFD, it can take Decipher a little more time to retrain the DFD. Deleting the DFD will not delete the training data, as it's not linked.
A6. I have made multiple changes to the DFD as part of my own training to learn how the different actions work within Decipher (formulae etc) so this may have an impact as you suggest.

If none of your suggestions help I think I will create a new DFD and ML and effectively start again to see if that improves the situation - although doing this feels like a cope out and not something I am that happy doing as I cannot do this when implementing in anger on a client site.

Thanks again

[Ben.Lyons1]

Hi Howard,

Thanks for the response.

I appreciate the training information could use a little more detail and I hope to be making a lot of new course materials, tutorial videos shortly.

If you're still training and you want to 'reset' the training. You can go to Admin>Training Data and select Delete Training Data. This will then take up to an hour to clear the cached training data and will be like new.

I don't recommend this where you have multiple people in the same environment and are using Decipher with production processes.

Thanks

Ben

[HowardBlake]

Thanks Ben

I did as you suggested - deleted all existing training data and then initially disabled the ML Selection in my Document Type.  The positive is that Decipher is now always returning a value for all fields.  

My annoyance now is that Decipher continues to always return a value of:
1. 'BustedKnuckles Motorsport' - please note the lack of space between 'Busted & Knuckles' despite me always correcting this, by editing the DFD template or by resizing the field in the invoice. 
2. A hash for Invoice No field despite me having an 'Exclude Values' rule for this field stating '#'. 

I have run the same process through Decipher about 25 times and each result is the same for 'BustedKnuckles Motorsport' & '#1234' for Invoice No.

I have run Decipher initially without ML enabled and then run it another 20 time with it enabled and always the same result.

Am I missing something fundamental here or doing something stupid?

Thanks Howard

[Ben.Lyons1]

Hi Howard,

I think you may have found a couple of bugs there. But I've put together a DFD that works for me on that document, give it a try and see if it works for you.

I'm please to say that they have been resolved in v1.2, which is due out shortly.

Regards

Ben

[HowardBlake]

Thanks again Ben - 1st class support!!

I will try your DFD and also look forward to getting my hands on 1.2.  At least it is not me going mad or doing anything stupid