I've got 500 images of paper receipts scanned and OCR as one dataset. I also have a dataset of transactions from my credit card statement including amount and date.
What model is best to match the receipts to the transactions?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Let the credit card statement be the ground truth, and the receipts be the noisy inputs. For a given line item, find the receipt with the smallest distance. If the distance is small enough, declare a match. This is a threshold you will have to experimentally determine. You can let the distance be the sum of the distances for the amount and date. A heuristic for these individual distances is simply the edit distance. A more sophisticated approach would be to model the OCR error using labeled data to determine the most likely input; print something with known text in a similar typeface to the receipts so you can learn which characters are commonly confused with one another, and thereby estimate the most likely input sequence. Going another step further, you could jointly model the density between price, store, and item, so you can recognize for example that a vegetable bought at your local grocery store probably does not cost \$100 but \$1.00
If your OCR software gives you confidence estimates over the characters and possible guesses, you can use that too.