Hacker News Reader

Comment by zaroth | Hacker News Reader

With all the horsepower directed at machine learning, why does it seem like text autocorrect hasn’t improved in the slightest?

View on HN

Autocorrect is fine for me in iOS 14. But Machine learning can also lead to regressions or edge cases. Here a fun rant about autocorrect in iOS 13:

https://www.smh.com.au/technology/iphone-autocorrect-driving...

The good:

> Machine learning has enabled the iPhone to correctly predict oddly spelt suburb names like Prahran and Woolloomooloo, to read the screen of an article and guess the surname or product I'm about to Google, and to suggest the spelling Bryan over Brian, because the former is a close friend.

The ugly:

> For at least two years now, my iPhone has insisted on autocorrecting Sydney to SYDNEY, every single time I type the word. After weeks of troubleshooting that included resetting the keyboard and uninstalling any transit apps I thought might be to blame, I realized the culprit. I have a few hundred contacts on my phone from a previous employer, all with UNSW SYDNEY as their address.

What should happen here with a few _hundred_ entries of "SYDNEY"? I guess it should be unlearned after the user entered a few times (once? two times?) by hand "Sydney" and user entered text privileged over other text corpora (even dictionaries?). But then you have other problems again like autocorrect learning potentially wrong spelling idiosyncrasies of the user.

> What should happen here with a few _hundred_ entries of "SYDNEY"?

I'd say, change the case in the database to proper "Sydney" and have whatever writes a piece of physical mail do the transformation to the correct format?

Because gboard predicts email addresses, and I have several Haskell-related mailing lists in my whitelist, gboard always tries to autocorrect "Haskell" to "haskell-mailing-list@..." when I am trying to text my pal about Haskell questions, so I get messages like (because I sometimes just swipe and send, without checking the message, like a moron):

"Hey, do you know in haskell-mailing-list@something.org if the Prelude..."

So, now I've turned off EVERY bit of AI bullshit everywhere, and I deal with mistakes myself.

I think the number one fix for autocorrect is adding an "I'm an adult, no I never mean duck" toggle

Could even call it a ducking toggle.

Its general purpose hardware, which without a good model doesn't do much.

The keyboard is "good enough" and Apple is going to capture the market or retain users with a better ecosystem of apps. So they're investing in making it easy for other developers to build things using ML on the iPhone.

It could be personal preference, but I've found that since iOS7 or so it's actually become worse. Since then, I've found autocorrect and other text input related actions like selection and moving the I-cursor around became much more frustrating to do. There have been various tweaks since then, but it's never felt as seamless as it did in earlier versions of iOS.

One really silly bug (can't call it anything else) is the insistence to capitalise Apple product names like in their marketing speak.

When I'm writing in my native language, I might type a phrase like "dit is echt the next best thing" or "ik breng je naar Brussels Airport". My iPhone will 'correct' this to NeXT and AirPort.

I keep asking myself the same questions every time i read about modern (apple) hardware specs: what are they going to do with all this ML-dedicated hardware ??

I don't see anything smart with my iphone, nothing has changed since the introduction of Siri. I still can't use it seriously for anything else other than to set the alarm clock, and not a single app i download or use seem to do anything special needing ML running on the client side.

For folks heavily using their iPhone as a camera and photo album, there are lots of ML-aided things going on. You can search through photos by text without having to have previously annotated your photos, it automatically builds albums that contain people it's decided are important to you, it automatically pulls out the "best" photos for you which really helps if you like to take tons of photos, it can automatically create entire slideshows for you. This is probably just a subset of the things to which ML is applied in one app on the device.

Fully agreed on the Siri shortcomings, though! But, I'm not sure how much of that work is done on device with the neural engine vs in the cloud. Not like that matters when the end result is so bad.

I agree. I’m amazed how well my favourite IDE - Webstorm does spell correction. If you add keyboard shortcuts it’s just delightful.

Question about iOS I have is if it’s packed with vision models, why can’t I as a developer just feed an image and get tags / text?

As a developer, you can. The APIs exist (https://developer.apple.com/documentation/vision/vnrecognize...), seem simple to use (https://www.hackingwithswift.com/example-code/vision/how-to-...) and appear to perform decently (https://heartbeat.fritz.ai/comparing-apples-and-google-s-on-...; compares with Google’s on-device OCR)

So, finally, they don’t step on the toes of third-party developers, and you still complain? ;-)

Thanks, was looking for it few weeks back and found you'd need to load up your own models.

Reply on news.ycombinator.com