Form Recognizer - Continued training

Question

I have used tools similar to form recognizer prior (Automation Anywhere's AI tool), and in them, you could review the executed extractions, correct anything it got wrong, and then use those corrections to further train the model. Over time, it got better and better (at least in theory).

With this tool, I cannot find a way to do that. so a few questions:

1. Is there a way to see what data has been processed by the tool?
2. Is there a way to correct what it extracted, and add that to the training set?

Thank you,
John

Answer 1

@John McTiernan You can lookup the processed files from labeling tool in the blob container that was configured with the tool. The files that are generated when the tool runs OCR are .ocr.json and .labels.json for labels that are created from the tool. *.fott files correspond to the project configuration that is created from the labeling tool. You can copy these files to another blob container or setup another labeling tool with a different storage connection with these files if required.

For the second part of the question, you can edit the extracted files from the backend and use them. But, in most cases the ability of the tool to label the image or file from the browser should be suffice.

Every time you make a change either from the tool or the files there will be an option to retrain and this will create a new model that can be updated in your application to extract data as required.