Recognition and correction

How errors enter a corpus

Optical character recognition (OCR) proposes text from print or images; handwriting text recognition (HTR) does the analogous work for handwriting. Both can confuse similar shapes, skip faint characters, join lines, split words or misread page furniture as text. Human transcription has its own errors. A later correction may change the searchable text without changing the underlying witness.

The Chinese Text Project distinguishes some OCR-derived transcriptions in its formatting conventions and links many electronic passages to scanned source pages. A digital character seen in a hit list is therefore a prompt to check the image when an argument depends on its exact identity.

An audit sample

Sample hits from the beginning, middle and end of the result list, and include unusual spellings rather than only neat examples. Compare each to the scan or facsimile where available. Classify an apparent mismatch as recognition error, transcription choice, legitimate variant, editorial correction or unresolved. Keep the platform's reading and your proposed correction in separate fields.

False negatives are harder to see: search for a visually similar alternative or inspect a small set of images in the relevant series. If no images are available, state that the reading was checked only against a digital transcription.

Do not erase the trail

If a tool displays a corrected text, cite the correction and the source reading separately. TEI's distinctions among unclear, gap and supplied text are useful models for keeping uncertainty explicit, though a particular project may encode it differently. For a damaged papyrus, an unresolved stroke is not an OCR mistake merely because it defeats an exact search.

Create a correction log with document identifier, line, image link, original digital string, proposed reading, reason and confidence. Do not incorporate a correction into a count until it is reviewable.

Practice

Audit ten hits for a short string in an image-linked collection. Give the number whose reading you could verify, the number with a visible discrepancy, and the number without a usable image. State how the verified subset changes your claim.