Words with similar paths depend on a language model to disambiguate
Aliases: swipe disambiguation · confusable shapes · path collision
What it is
Different words can trace almost the same polyline on a full keyboard: neighbouring letters, a shared prefix, or a one-key difference in the middle. Shape matching returns a band of near-equidistant candidates; which one wins depends on a language model—prior context, frequency, words the user recently used. A glide is not “this line equals this word.” It is “this line, here, most resembles which word.”
Why it happens
Template distance (often dynamic time warping on the key sequence, or a geometric comparison of polylines) ranks words into a neighbourhood band. Geometric gaps inside the band can be smaller than ordinary hand jitter, so shape alone flips pairs such as this/that, about/above, night/right. The language model supplies a dimension the shape channel lacks: which word is more likely in this slot. After the channels are multiplied (or log-added), the ranking stabilises. Without a language model, gliding degrades to “nearest polyline wins,” and a tremor replaces the whole word. With one, the cost is that the same path yields different words when context changes; users who memorised “I slid this shape” experience the system as capricious. Disambiguation happens after lift and before the word lands, so what people see is already the chosen one.
Studying it
Build geometrically close minimal pairs, place them in different priors (some biased to word A, some to B), and compare top-1 accuracy with and without a language model. Independent variables include template distance, strength of the prior, and personal frequencies; dependent measures include top-1 hits, rate at which the target is only second, and whether users notice the swap. Testing glides on an isolated word list underestimates the model’s contribution and its risk of rewriting user expectation.
Where it stops holding
Out-of-vocabulary strings have no template; a strong language model still has nothing to vote for. A tightly constraining prior (a frozen phrase) can overpower shape and pull a clearly off path onto the phrase’s word, which looks like an unsolicited rewrite. With personalisation off, disambiguation falls back to general frequencies, and dialect and jargon lose systematically to frequent near-shapes.
Applying it
- After lift, keep the second and third near-shape words in the candidate bar so a shape-right, context-wrong case can be retapped in one hit instead of deleting and re-gliding the whole word.
- Accept top-1 on near-pairs with realistic priors; isolated-word hits do not represent in-context behaviour.
- If the same path yields two unwanted words in two contexts, check whether the language model is overpowering shape before drawing the keys larger.