Character-level recognition differs from word-level recognition
Aliases: per-character recognition · whole-word recognition · handwriting unit
What it is
Once handwriting must become searchable, re-editable characters, the recognizer has to choose how much ink is one decision: a character (or letter), or a word/phrase. Per-character nails correction to one glyph; whole-word uses a lexicon and language model to decode a span of ink, and a mistake is often a whole-word substitute. That is unit choice inside a coupled recognise-and-correct loop, not the online/offline algorithm itself, and not editing ordinary text after the result is committed.
Why it happens
A per-character path segments then classifies; each slice yields candidates and a confidence, so a correction UI can list them beside that ink. A whole-word path hands a run of ink to a decoder; lexical constraints resolve single-glyph ambiguity (Latin c/e, similar Chinese forms). When it is right, interruption drops; when it is wrong, the blame boundary grows: the user sees a wrong word and cannot tell which stroke was mis-cut. The unit also sets waiting: per-character can emit at end-of-stroke or end-of-character; whole-word waits for a word boundary (space, timeout, punctuation), so feedback is later. Languages differ in natural unit size. Chinese is visually character-sized, so per-character matches writing rhythm; English cursive joins letters, and hard letter cuts split n into r+i, so whole-word is often stabler. The loop’s cost is “notice the error + fix it”: per-character notices early and fixes small; whole-word notices late and fixes in one go. Without a correction entry near the error, both units collapse into after-the-fact proofreading.
Studying it
Same writers, same copy, compare per-character decoding with word/sentence decoding. Log time to first visible result, where the first error appears, and operations to fix one error.
Independent variables: decode unit (character / word / sentence), lexicon on/off, how word boundaries are cut. Dependent measures: character or word error rate, delay from lift to glyph, correction steps, correct characters damaged by a whole-word substitute.
Split the scenario: names and codes are hurt by a lexicon; ordinary sentences often need fewer fixes at word level.
Where it stops holding
Languages without spaces cannot cut words on blanks; whole-word then needs a language model or an explicit break. Rare characters, proper names, and formulae break the lexicon; a fallback to per-character must exist. Committing a whole-word result with no in-word edit locks the unit choice. Live “emit as you write” with whole-word makes the last letters jump at word end, which fights people who proof the latest character.
Applying it
- Pick a default unit from content: continuous prose can be whole-word; fields, names, and codes stay per-character, with a switch to the other grain.
- Keep per-character candidates even under word decoding so a fix can replace one glyph rather than topple the word.
- Show the current unit’s result beside the ink, not a distant string after the whole passage.
- How to check: short sentences with proper names and ordinary short sentences; count fixes and collateral damage. Measure time from finishing a character to seeing that character (or its word) and confirm the unit matches the feedback rhythm.