Closed MakaraSok closed 2 years ago
The behavior is rather unpredictable. I experienced that the error would not happen right of the gate. The suggestions might keep showing after typing ស្រ . The LM would crash or stop giving the suggestions if the same sequence is typed over and again after choosing from the suggestions.
i.e. keep typing ស្រ and choose ស្រី (the first option) until the banner turns blank. See below. It turns blank after three rounds for me.
It's worth noting that errors will generally be a lot more likely for models targeting languages without proper word-breaking implementations. Our support for them is simply lacking until we're able to implement #5025.
If this can be shown to also affect models for languages with space-based wordbreaking, that would give this issue higher priority.
I am wondering if this is a performance issue -- that the prediction engine is giving up because the lookup is slow and we are near a timeout threshold and sometimes crossing it?
@mcdurdin You mentioned at one point that there were some other behaviors not currently documented here that might be related? Could you please add those details here? I'll need those in order to better pursue this issue.
There is nothing specific. I have observed, which may or may not be releated:
Hope these pointers help a bit. That last one in particular may be relevant?
- Backspace quickly until "Hello w" is left -- the banner will often (but not always) offer "the", "to", "a" instead of words starting with "w". Note that backspacing slowly will usually offer more appropriate suggestions.
Is that with a multi-tap backspace, rather than a held one?
Also, can you replicate this in iOS? Or on the "Predictive Text: robust testing" Web test page? I'm not seeing this effect on either platform in my personal tests.
deleting text (backspacing) often leaves the banner empty.
I don't see this either, and that's when backspacing as rapidly as possible, even with long words.
The closest thing I've seen was when keystroke errors were occurring behind the scene during some unrelated development/testing recently. When that happens, if memory serves, the Android keyboard tends to reset itself - so I doubt that's at fault here. Or has this changed signfiicantly - and if so, in what ways?
Is that with a multi-tap backspace, rather than a held one?
Multi-tap
Also, can you replicate this in iOS? Or on the "Predictive Text: robust testing" Web test page?
I haven't had the capacity to try and replicate on other platforms yet, sorry.
While we're not sure exactly how just yet, it seems that this may have interesting interactions with Android-specific behaviors.
@MakaraSok would you mind re-testing this one please?
Tested with Keyman for Android 16.0.84-alpha on SG S9+ running Android 10. The issue cannot be reproduced. It seems to have gone away by now.
Good stuff, closing as resolved :grin:
Describe the bug
With a fresh installed of Keyman for Android, after installing Khmer Angkor keyboard and NC Khmer Wordlist LM, start typing in the text area ស្រ to see no suggestion on the banner.
Expected behavior
The LM should give the expected/relevant suggestions on the banner.
Screenshots
Keyman for Android:
Keyboard