Include test documents in extraction accuracy calculations

The purpose of extraction accuracy is to ensure that your extraction results are as confident and correct as possible when your solution is in production. This is where extraction accuracy calculations can help by comparing the extraction results on a test document with an expected value provided by you. If formatting and validation is configured for a field, the extracted value is formatted and validated before it is compared to the expected value.

When adding test documents for extraction accuracy, include test documents that have both confident and unconfident field data. This enables you to identify documents where extraction needs improvement. Fields with low extraction confidences indicate that more extraction training may be needed.

If you include test documents with perfect extraction results only, you can generate a perfect extraction accuracy report. However, this does not accurately reflect the results you see during production, nor does it help you improve your extraction results. That is why it is ideal to include documents with green, orange, and grey extraction results.

Use the procedure in this topic to mark a document for use in the extraction accuracy calculations.

You can calculate extraction accuracy for both configurable and predefined document types.

Before you begin

Before you can include a test document for extraction accuracy calculations, ensure that you have performed the following steps.

Enable extraction accuracy.
For each document type, all training documents must return confident extraction results for all fields.
For each document type, have several test documents to test your extraction and extraction accuracy results.

Procedure

Select the  Fields menu.
Select a document type and collapse all other document types to help focus on a single document type at a time.
Scroll down to the test documents and select  Test extraction for one of the documents.

Extraction is performed for all fields on that document under the Fields tab, including the confidence for each field.

Not all test documents are suitable for use in extraction accuracy calculations. Do not include documents that are stretched or skewed, or have image artifacts that interfere with extraction results.
For the extracted test document, perform the following steps.
1. Select the Use for extraction accuracy checkbox.
  
  Additional settings are listed for each field.
2. For each field in the document, in the Extracted field, select  Copy to Expected.
  
  Skipping the copy step returns results that highlight issues that may not exist. If you do not copy the Extracted value to the Expected field, the extraction accuracy calculation flags this difference as an extraction error, even if the extraction results are very confident. It is very important to perform this copy step, so the extraction accuracy calculation has the necessary information it needs to identify genuine extraction issues in your solution.
  
  The extracted value is copied to the Expected field.
  
  The extracted value is also copied to the clipboard.
3. If there are any discrepancies in a copied value, select the Expected field and edit its value as needed.
Repeat steps 2 to 4 for additional test documents.
Optional. Repeat these steps for other document types.

Next steps