Azure AI Document Intelligence
An Azure service that turns documents into usable data. Previously known as Azure Form Recognizer.
1,508 questions
Azure Cognitive Services Form Recognizer Prebuilt Receipt Model
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(9.6, 92%, 10%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3ECS%3C/text%3E%3C/svg%3E)
Chris Sampson
1
Reputation point
This is probably a question more focused at the development team and developers using the solution to backend their systems.
I have been playing around with the Form Recognizer and building a test front end that uses the extracted data for later processing. On completion of the AI Process, the idea is that a user will compare and correct the receipt data recognised against the image of the receipt. (Mismatches/Incorrectly recognised data seems to be higher based upon quality of the initial image scan.)
The process of using the AI to do the hard work and following with a user check should produced quite accurate results and might well be akin to a double entry approach. Any corrections that are made by the user are stored in the database as well as the initial scanned data. I am sure that this approach will ultimately be deployed by many organisations in the future given we are still stuck on receipts for now.
One of the benefits here might be that there is an opportunity to further train the pre-built models using this data.
Is there a feedback loop available or planned where this corrected data can be pushed back for further AI model training to built greater accuracy for all users?
{count} votes
-
romungi-MSFT 43,656 Reputation points Microsoft Employee2021-03-04T06:32:45.523+00:00 @Chris Sampson The current feedback loop mechanism for model training or retraining that is just introduced is for bulk training scenarios using data factory and datalake. This might not be the exact solution for the scenario described above but we will certainly pass this feedback to our team and check if there is a plan to provide a more simpler way to achieve this.
-
Chris Sampson 1 Reputation point
2021-03-04T06:56:19.423+00:00 From my implementation perspective, a very simple approach might be to add a route on the WebAPI that would allow posting back the updated data (as JSON) is the same model format it was originally provided. (Note I am using the WebAPI for my purposes).
Hopefully my feedback helps.
Sign in to comment