Proper names and rare words fail at a much higher rate
Aliases: swipe OOV · out-of-vocabulary glide · name swipe
What it is
Glide recognition depends on word templates on the key plane. Proper names, abbreviations, neologisms, and deliberate misspellings sit outside the lexicon and have no template. The path is pulled onto the nearest in-vocabulary near-shape. The failure is not a slightly crooked glide. The system returns a completely different frequent word, and the user has to delete and tap. Failure rates on this class sit well above those on ordinary in-lexicon words.
Why it happens
The template library is a closed set. An out-of-vocabulary string still lies geometrically near some in-lexicon template—person names near common verbs or nouns, product names near close frequent spellings. The language model’s prior on OOV items is near zero, so even a small shape distance loses to a frequent neighbour. Glides on OOV items therefore show a systematic “rewrite into a dictionary word,” not “unrecognised, so keep the raw path.” Users who do not know the string was OOV experience a failed gesture; those who do abandon gliding in advance. Both reactions send this class back to tapping, and the throughput advantage of gliding does not exist for them. Adding the item to a personal lexicon later helps only the next identical string; the first attempt still fails.
Studying it
Measure top-1 glide hits and tap-fallback rate separately on frequent in-lexicon words, rare in-lexicon words, and an explicit OOV set (names, brands, internal codes). Independent variables include whether dynamic add-to-lexicon is allowed and whether a fallback candidate “interpret keys along the path as letters” is shown; dependent measures include the rate at which OOV items are rewritten into in-lexicon words, and which channel first succeeds. Testing glides only on news text hides OOV failure.
Where it stops holding
Names already in a personal lexicon and glided many times approach in-lexicon behaviour. Passwords and one-time tokens should not be glided at all. Morphologically rich languages may treat a “rare” surface form as an in-lexicon stem plus an inflection, with failure rates between in-lexicon and OOV. Handwriting or speech offer other channels for OOV items and sit outside the glide-template mechanism.
Applying it
- Beside the glide result, offer a candidate that interprets keys along the path as letters, so an OOV item need not become a wrong dictionary word and then be deleted.
- On name, address, and brand fields, weaken gliding by default or prefer tapping, and enable gliding for a word only after the user explicitly adds it.
- The acceptance set must include proper-name samples from target users; hit rates on news words do not represent glide quality in the product.