Keep a correspondence between original ink and recognised text so people can fall back to the handwritten original to check

What it is

If recognition results and leftover ink can be aligned—this ink is that character—people who doubt a reading can tap back to the original to check, or restore that span as handwriting without losing the page. Correspondence is an index for post-editing, not the recognizer itself, and not “the text can be edited”. Without it, checking the original is an eyeball search through a whole page of ink.

Why it happens

Alignment is usually a box per character, or a map from each glyph to a stroke-id span. When a segmentation hypothesis changes, the map must follow, or tapping this character highlights the wrong ink. After commit, a keyboard edit can break the map: newly typed characters have no ink and should be marked “no original”. Revert swaps the selected characters for a rendering of their ink and marks the character layer uncommitted or deleted. A page that only stores one ink bitmap and one plain string, with no index, cannot revert locally. Storage costs more than text, and is often worth it for legal and note checking. Correspondence also drives two-way highlight—“tap a character, see ink; tap ink, see characters”—a visual anchor for proofing.

Studying it

After recognition, have people find a deliberately wrong character, comparing two-way highlight with two unconnected layers. Then test whether local revert restores only that span.

Independent variables: character–stroke map present, whether the map updates after the user edits segmentation, how keyboard inserts are marked. Dependent measures: time to find the original of a wrong character, neighbour damage on revert, map-break rate after edits.

Where it stops holding

Offline scans that never had a path can only align to bitmap fragments, less precise than online. Reflow moves boxes; a map in screen coordinates will point wrong—use document coordinates. Encryption or export to a format that does not know the map (a raster PDF) drops revert; the export option should say so.

Applying it

• On commit, write a map from characters to ink spans; rewrite it after the user edits segmentation.
• Highlight ink from a character and characters from ink; revert applies only to the selection.
• Mark keyboard inserts as having no original, so revert cannot fake one.
• How to check: tap a wrong character, its ink should light. Revert that character; neighbours should not become ink. Type a few characters in the middle, tap an old character; highlight should still land on old ink. Warn on export to a map-less format that revert will be lost.

Related

• Same group: C5.15.1 Recognised text stays editable like typed text · C5.15.3 Batch correction after writing costs more than per-character confirm · C5.15.4 Reflow after convert can wreck spatial layout
• Adjacent: C5.14 Handwriting Recognition · C5.05 Handwriting Recognition and Correction
• Search: ink alignment · revert to handwriting · glyph mapping