What the parent comment is talking about is something more extreme and I've noticed it too. It sometimes changes prior words that are valid after you have moved on to the next word. It's not correcting the word you just typed, it's correcting a previous word without any sort of feedback like you get for normal corrections- it's going backwards and changing earlier words, and then you try and fix it but the exact same correction applies automatically again over and over.

Unfortunately I can't remember any examples at the moment, it's just something that happens to me every so often. They're really irritating though, because they aren't well expressed in the current autocorrect UI (which works on the current word) and it doesn't seem to get the hint when you go back and correct it, so it keeps applying it over and over.

No, that's what I was talking about too—it needs the context of the (part of speech of) the next word to figure out in the cases I mentioned whether I actually intended the (correctly-spelled) word "wed" or the (correctly-spelled) word "we'd". It doesn't change it until after you hit space to commit the word that comes after the "wed" input.

Ah, ok. I'm pretty sure I've had it happen with more than just contractions though, like with common phrases. The real problem is that it's really hard to undo the correction. I need to start keeping better track of it so I can file a bug report... it makes it really difficult to type certain combinations of words.

If I type "its" and "it's" correctly, they are regularly changed to the incorrect spelling. If I don't bother, they are not corrected. The only way to get correct "its" and "it's" is to go back and fix them after auto-correct has screwed them up.

It's probably because the spelling is corrected based on a machine-learned model, whose corpus is likely to contain many instances where "its" and "it's" were swapped.

Which means the corpus is broken. But regular people rarely care about correct spelling, in my experience, and so I doubt corpus maintainers will care either...

You're imagining that people who make NLP corpora actually vet the text going into them? I dream of a world where people can be convinced to care that much. I'm not even talking about the scenario you suggest of filtering for proper word usage, I'm talking about filtering at all.

The corpora used for popular word embeddings are full of weird nonsense text (in the case of word2vec) or autogenerated awfulness like spam and the text of porn sites (in the case of fastText) or both (GloVe). And most people who implement ML don't care how their data is collected.

I mean, nobody expects the engineers to manually read through everything, but if the quality of the input text is significant for the quality of the autocorrect (or whatever other application you're using machine learning with), you kind of have to make sure the input is pretty good... You could for example choose datasets which is expected to contain mostly correct grammar and spelling (such as Wikipedia, books, etc.) rather than datasets which is expected to contain mostly incorrect grammar and spelling.

Or don't use a machine learning model. I honestly don't care, just don't automatically turn a correct "its" into an incorrect "it's".

Wouldn't late edition books with only corrected text be better, proofread, edited, proofread, edited, ... Google have millions of them they've assumed copyright of. Surely there's enough text there. Do they really just use random website text?? Nearly every news story I read has errors and they have style guides, trained writers, editors, etc..

Do publishers sell their published text as a mass for use in AI/ML? Like 1000 books, no images or frontispiece, etc., possibly jumbled by sentence/para/page.

That is one of the most draining things I've heard recently. Sigh. Are these open projects where someone could in principle improve them?

The best open corpus project I know of is OPUS: http://opus.nlpl.eu/

They say they welcome contributions; I don't know if they just mean new sources of text, or if this includes code for filtering or fixing their existing ones.