Three iOS Keyboard Changes Apple Should Make

In my personal experience, it’s become both worse at fixing legitimate typos and more aggressive about taking actual words and turning them into nonsense.

So, my understanding is that they rewrote/enhanced the engine around iOS 10(?) to use machine learning.

At first, that seemed to produce some weird bugs, particularly with capitalization. But! These days, I find that while it still sometimes gets its “correction” spectacularly wrong, most of the time, it’s way better than before those changes. In particular, it seems to have gotten smarter at grasping context beyond a single word; this resolves lots of issues, such as it having gotten smarter about whether it’s or its is appropriate.

Anyway, it’s interesting that experiences are so different here. It’s been mostly good to me!

Apple already offers a way to search for emoji on the Mac, it’s more than a little puzzling that it wouldn’t make a similar feature available on iOS.

Yeah, this is unfortunate. The emoji suggestions are pretty nice, but no substitute. ctrl-cmd-space on Mac is awesome. (Incidentally, win-. on more recent Windows is also a great start, but nowhere near as powerful.)

I’m not sure I would call it “puzzling”, though. The keyboard is already a fairly cramped and busy UI as it is. The trackpad mode, while awesome and powerful, has only made it more complex to understand (I love the feature, but I bet there are users now who wonder what is happening and how to exit that mode). Figuring out how and where to insert a search mode into that isn’t trivial, and will probably sacrifice screen estate and simplicity even further.

Machine learning is an imitator. If your vocabulary matches that of the population Apple uses to train the language model underpinning predictive text, then it will get things right.

However if you deviate from that population, eg because you like talking about esoteric cooking terms, or have a very local colloquial vocabulary, then – because Apple has increased the degree to which it prefers predictive language model suggestions over simple spellchecking – then you’ll find your input gets mangled.

What’s frustrating is that Apple has probably l’y seen it’s accuracy (customer-sat) improve as it considers English speakers to be a single homogeneous crois, rather than discretising them into several heterogeneous and equally important subgroups.

It’s a bit like the way you can get 90% accuracy on the Titanic survival prediction task by just predicting that everyone will die, since that’s the majority.

I find when discussing certain topics that autocorrect replaces correctly spelled words with nonsense, which is enormously frustrating.

Yeah autocorrect is one of the major pains of modern life. I find it amazing to thing that there’s any form of Machine learning at play because the sort of nonsense that gets suggested at times beggars belief. Maybe it’s a sign that machines aren’t good at prose or more likely the fact that machine learning has to be well implemented by humans before proving to be of any use.

The number of times I’ve seen a totally predictable word following on from another being changed to something you’d never expect to follow on from the former word just because of one erroneous letter typed. I’ve actually turned off autocorrect a few times but I usually turn it back on for the obvious corrections it does pick up but the value/annoyance balance is way closer than it should be.

The one that still bugs me is that once a word is added to your auto-correct dictionary there is no way to remove it. I've got my daughter’s name with “friend” appended to it stuck in the auto-correct dictionary with no way to remove it, and every time I type her name I get this weird double word thing offered. Ridiculous. Apparently you can reset the entire auto-correct dictionary but that is the only option.

I have noticed the "machine learning" that people mentioned above. However, unlike Bryan, I've noticed that it learns my patterns, such as when I type the military rank of somebody I send lots of emails to, it autocompletes that rank now and then suggests their last name just after. So, unlike Google's machine learning which tracks a world full of data and suggests the most common corrections, it seems that Apple's method is learning my specific habits and helping me directly.

At least the English language (and a few select others) has QuickType Predictive Input. For Swedish we are still waiting for it, it's been almost five years since it was introduced with iOS 8.