Yes it's oddly not great in places. It seems to suffer from "YouTube syndrome" where it gets hooked onto something you searched for/saved once. Just because I once entered "Aldi" manually, now that always seems to have an extremely heavy positive weighting, even though I've only actually wanted it a very small number of times and "also" is far more likely as it is used at a 20:1 ratio, and I'm endlessly deleting "Aldi". Trivial Bayesian type things aside, even basic linguistic analysis, let alone an AI superpower, could estimate when "Aldi" and "also" are each appropriate in context from earlier words.
Another example: what was more likely after "20:1": "ratio" or "radio"? Get it tighter^Wtogether!
> Just because I once entered "Aldi" manually, now that always seems to have an extremely heavy positive weighting, even though I've only actually wanted it a very small number of times and "also" is far more likely as it is used at a 20:1 ratio, and I'm endlessly deleting "Aldi".
I'm using Google Pinyin Input, which is discontinued. (I use it anyway because the replacement input methods are worse. Don't ask me why they stripped useful functionality in order to replace it with nothing.)
It has swipe input, but as far as I can tell it isn't possible to add words to the swipe input dictionary. However, if that were possible, your problem wouldn't exist - the way the system works is that you swipe something, the system's first guess appears in the text input, and a bar of suggestions appears over the top of the keyboard. If the first guess was wrong, you can select from the suggestion bar, and the word in the text input will be changed to whatever you selected.
So I'm kind of bemused at the idea that you need to be deleting wrong guesses.
Ok, changing it then, bit^Wnot felting^Wdeleting (though since much of the time the wing^Wwrong guess isn't even there, but rather it's plurals and verb forms and si^Wso on, I do often need to refocus away from the import^Winput fitting^Wfield, pause, scan, conclude the right word isn't there, then delete anyway, so my muscle memory is to go for the delete key). It's still a suboptimal experience because the keys for the next weird^Wword like punctuation are at the bottom and the wing^Wwrong guesses mean you have to scan the list at the top.
These ^W marks are all organic mistakes that it really shouldn't be making so many of considering the word prevent^W placement[1] and statistics and thousand upon thousands of units^Wunless^Wit's^Winputs.
Maybe it's the never^Wnerve damage in my diggers^Wfingers making me sort^Wsuper clumsy, or being a lefty makes it wise^Wworse sommelier^Wsomehow, but I do remember it being better, far better, in the past. And it never, ever, send^Wseems to learn (I don't have Google services so maybe it needs to phone home to learn?)
It seems possible to add words: you type then^Wthem n^Win and choose then^Wthem from the bar m^W.
But thank you for so kindly telling me I'll^WI'm once-unimaginable^Wimagining[2] it.
[1] take this one for example: "considering the word prevent" isn't grammatical, but "considering the word prevent^Wprevent^Wplacement (argh again! Twice! And prevent^Wplacement want^Wwasn't in the list, it was prevention, preventing and prevents) is.
[2] "once-unimaginable" is one I put in by typing it once as it has little concept of compound Shane^Wadjectives. It seems to completely hallucinate the "once-un" to prioritise a previous manual unit^Winput. But I can't actually swipe to get it again, no matter how carefully I go. I get "once-imagined".