Open mcdurdin opened 4 years ago
Potential resolutions to #5025 lend credence to this position. Things get a bit funny for dictionary-based word-breaking at times when the extra-short words aren't included:
Neither I
nor am
is currently in the MTNT model, so merging consecutive unmatched chars ends up treating "iam" as a word.
I'm not completely clear about what, specifically, this issue is requesting.
What is clear:
Parts I'm less clear on:
but the wordlist author should be able to configure the model to not present them as options for completion
minCompletionLength
) that would cause shorter words to be "skipped over" when making predictions?(although presentation for correction is still helpful)
Suppose that we only allowed words of length four or greater as predictions ("for completion") for the following cases.
Example case 1: Suppose the text tue
has been typed. Would the
be considered a valid predictive-text suggestion here? u
neighbors h
on the QWERTY layout, so it would be a reasonable correction.
Example case 2: ti
has been typed, which naturally leads to time
and similar suggestions. i
is correctable to o
, though - do we offer to
as a correction if it's sufficiently likely?
I wonder if the primary desire is to avoid listing the really short words on empty context? I am not sure. @dhigby your thoughts?
Is your feature request related to a problem? Please describe. One, two and three letter words (perhaps other lengths?) should not be excluded from a wordlist, but the wordlist author should be able to configure the model to not present them as options for completion (although presentation for correction is still helpful).
Suggested by Doug Higby.