One bracket cost a receipt reader 20 points on shop names

ai-bookkeeper misread the merchant on 22 AEON receipts: a registration number looked like symbols. Fixing it lifted merchant accuracy from 63.6% to 83.2%.

Tools and projects · 1 min read · Source
ai-bookkeeperreceiptsaccounting

ai-bookkeeper reads receipts into a merchant, date and total. Scored on the 626 hand-labelled SROIE receipts (word boxes, not OCR), merchant was right on 63.6%. Listing the misses showed 22 of the 57 wrong ones were the same chain: the first line, AEON CO. (M) BHD (126926-H), was rejected as "not wordy" because the bracketed registration number counted as symbols, so the address line below it became the merchant.

The reader now strips a trailing bracketed registration number before judging the line. Merchant exact: 521 of 626 (83.2%). Date and total are unchanged. Not re-measured on tesseract output (no engine in this run); the remaining misses are mostly brand against registered company name. Receipt matching to bank lines uses the merchant, so this directly helps pairing. Estimates only, not accounting advice.

More in Tools and projects

All writing · All in this topic · All tools and research