search

aws-samples / amazon-textract-textractor

Analyze documents with Amazon Textract and generate output in multiple formats.
Apache License 2.0
404 stars 145 forks source link

Detected in EXPENSE_ROW but not as ITEM #385

Open arsher-b opened 3 months ago

arsher-b commented 3 months ago

The expected item did not parse as an item, but it exists in the expense row. We expected “SURF CB 6+1” to be parsed as an item so that it could be recognized as the item name.

image

Receipt Used: photo_2024-08-06_21-27-04

athewsey commented 3 months ago

Unless I'm mistaken, this is an issue with the detection within Amazon Textract itself, rather than the processing in the Textractor library right?

It looks like the preceding record should've captured the whole "Zonrox\nLemon 1000ml", but instead the newline caused Textract to treat "Lemon 1000ml" as the item for the following record. I tentatively think fixing this behaviour might be outside the scope of the Textractor library to fix, because it'd probably require fine-tuning a receipt-parsing ML model.

Today, Amazon Textract's expense model doesn't support fine-tuning - so I think it would be a question of whether these kinds of errors are common enough to be worth you A) documenting examples & raising it as a support case with the service team, and/or B) implementing a post-processing model using open-source like HuggingFace or similar, to try and edit the results?