Beta Handwriting Feedback

mikelove

皇帝
Staff member
We just launched a new Android build on beta channel (can sign up for betas via our Play catalog page) with a new handwriting recognition algorithm which you can enable in Settings / Input Methods / Handwriting; it's based on / should perform similarly to the algorithm from Google Translate and Gboard (and for that matter pretty much any other neutral network based Chinese handwriting recognizer, e.g. the one built into recent versions of the iOS system Chinese handwriting keyboard).

New recognizer will also appear shortly on iOS with our next update there, which we'll submit to Apple whenever they start accepting updates for iOS 15. (the reason Android gets it in a beta is because we also had to make a couple of other major changes - most importantly, migrating to Google's new in-app billing system - which we were nervous about releasing without additional testing, while on iOS the changes outside of this are pretty minor)

It's currently set up to only output one character at a time but it's theoretically able to do entire words, we just wanted to start off with the closest possible comparison with the current system.

Anyway, we're very curious whether people find it better or worse than the current system - we have no plans (nor any legal obligation or strong technical argument) to take away the current system for people who already own it, but we're debating a bunch of ways we could roll out this new system (or something like it); skip it altogether, offer it to free users with the old one as an optional paid upgrade, offer it to everyone by default and only keep supporting the old system for people who already own it, etc.
 

Shun

状元
I installed it right away on a Pixel 4. I noticed that there is a slightly longer delay, maybe one tenth of a second instead of two hundredths of a second with the old algorithm, until the first character guess appears. But this doesn't matter to me. It certainly is a great feature to have the algorithm adapt to the user over time by way of machine learning/a neural network. For that, I am not the ideal tester, since I got a hit rate quite close to 100% with the old algorithm, too.
 

mikelove

皇帝
Staff member
Sorry, it's not adaptive, the templates are fixed and don't update based on usage. It's a neutral network in the sense that it's a newer type of AI algorithm; our current recognizer is a highly refined version of the same sort of old-fashioned template-matching handwriting AI they were using in the Newton days.
 

Shun

状元
After some more testing, I can now confirm that it is more reliable. The following characters did occasionally require me to enter them two or three times with the old recognizer, but are child's play for the new one:

日、目、白、自、艺、凡
 
Last edited:

mikelove

皇帝
Staff member
Thanks! (and yes, cases like this are exactly why people are using neural networks for Chinese handwriting now - very hard to teach an old-school statistical system to recognize subtle and unimportant-in-other-circumstances differences)
 
This may not be the best place for me to post as I'm referring to the AI handwriting recognition on iOS, but...

Overall it seems very reliable, but I get good results on the old recogniser, so it's hard to know which works better overall.

The times the new system gives the wrong character do seem more repeatable though, to the point where trying several times can still get the same wrong result. For example: even in my best handwriting, 开 very frequently comes up as 井, and 夺 as 寸.
 

mikelove

皇帝
Staff member
Thanks! (Perfectly good place to post, yes - it’s in this forum because we launched it on Android in a beta release before we launched it on iOS)
 

jurgen85

榜眼
I've tried it for a week or so and it is usually on par or better (as with 日 etc.) but twice in a row it recognized 躲 as simply the 身 component, where one would expect something with a comparable amount of strokes/components like 躺 or 軀.

This seems like an edge case since I haven't had any systematic issues. Both times it happened were in flashcard tests, so maybe I was too quick on the "recognize" button and it couldn't recognize the rest of the character in time? Would that have happened with the old system (too)?

Edit: another issue is 壺 which tends to be interpreted with an extra stroke to be 壼.
 

Attachments

  • Screenshot_20211022-155226.png
    Screenshot_20211022-155226.png
    180.9 KB · Views: 115
Last edited:

mikelove

皇帝
Staff member
I've tried it for a week or so and it is usually on par or better (as with 日 etc.) but twice in a row it recognized 躲 as simply the 身 component, where one would expect something with a comparable amount of strokes/components like 躺 or 軀.

Most likely it would have actually reported 身 and 朵 in that case if we let it output more than one character; it's not exclusively tied to single characters. Our plan is to plug in a list of common side-by-side characters and automatically convert the output to a single character in those cases (but then we may also add an option to write more than one character at once).
 

jurgen85

榜眼
I tried drawing 身 and 朵 at varying distances from each other and it works just like how you say, so it seems I need to work on my handwriting. ✍️

I wonder how to handle when both options are legit, like 好 vs 女子. The engine probably knows best?

If you do add multi-character input it would make sense to disable it automatically if a flashcard only has one character. Sounds promising either way!
 

mikelove

皇帝
Staff member
I wonder how to handle when both options are legit, like 好 vs 女子. The engine probably knows best?

I think we're going to default to 1-character-at-a-time, but then also have an option to do multiple characters in which case you'll either have to trust the engine or pick your preferred option if you see both.
 
Top