Re: [NTLK] Apple Seeks Handwriting Recognition Expert - We Can Dream, Can't We?

From: John Charlton (johncharlton_at_mac.com)
Date: Fri Aug 26 2005 - 17:23:23 PDT


On Aug 25, 2005, at 4:23 AM, Larry Yaeger wrote:

> Extending Inkwell
> to Kanji or the like would be terrific, but very difficult, due to
> the dramatically different form of the text, though I've long thought
> that substituting radicals for characters and whole
> pictographs/ideographs for words might work fairly well.

I've been meaning to ask you about that...

Actually, didn't know who to ask, let alone that you would be on the
list.

I'd imagine (and I know just enough to be dangerous about Chinese and
neural nets) that it wouldn't be much more difficult than roman,
simpler in some ways.

Since each character is a neat block and there's typically less
overlap than you might see in sloppy handwritten english it would be
easy (and when I say easy, that would be for someone like yourself
who has done it) to pick out each character. Also you don't need to
worry about word spacing, as there is none.

On top of stroke count which should be pretty consistent (one of the
principal ways characters are classified, handwritten characters take
some short cuts, but again those are common and consistent) you have
stroke order, which again is taught and follows rules. And since a
character is usually a mix of sub-characters if you can pick out any
of those you limit your possibilities again. (i.e. finding the "tree"
radical on the top with a good degree of confidence narrows the
search down considerably)

There are subtle differences with some characters, just like English
writing, and you'd learn to cope. For example I have trouble between
u&n, and j&y, and try to exaggerate to improve my score. No
complaints there. Again Chinese could have an advantage here because
the relative frequencies of characters are more exaggerated than the
26 characters in English, adding weight to a common character over an
obscure one.

So you're left with a lot of information to make a decision on for
any particular character. Which is what you'd expect given that there
are many characters to choose from and a person has to recognize it
as well - you need the information there in a fairly unambiguous way.

How much time and how many people does it take to train a neural net?
Is it even possible to train it to your own writing from scratch and
have good recognition? I imagine that to release a product that works
from the get-go you need a large population's information. Recruit
the Chinese Newton Users Group? After reading the recognition topics
in the Programmer's Guide I'm inclined to give it a go - in a small
way, perhaps with a flash card game where you're scored on how well
you draw the character. When I need a break from my real job. (Which
is, like, two years ago.)

I would have thought that some clever Japanese or Chinese would have
tackled this already. Seems like a natural. There are many people who
won't take to writing on a roman keyboard, (there are many different
input methods, none intuitive) how else will they use this
technology? Anyone have any experience with the Japanese input
methods for the Newton? (Anyone want a Chinese keyboard driver while
I'm working on the Stowaway? Please say no...)

You're comments about writing on a graphics tablet are dead on - my
experience exactly with the Wacom, why not use the keyboard? Only if
it's a word or two and you're already comfortable with it. But for a
proper tablet computer (like the Newton) that you carry around
there's no place for a keyboard and the ease and accuracy of
handwriting input is paramount. For the right market it's a godsend,
pity that wasn't recognized (no pun intended) with the Newton, such a
head start on the competition. (Like seven years and counting.) If
Apple are about to introduce a tablet, a Kanji/Chinese handwriting
recognizer that worked as well as Rosetta would be the killer app I'd
say. Think they could sell a couple in Japan or China?

And finally, Egg Freckles. There's a form of humour in China that
westerners know nothing of that plays off the many homonyms and the
fact that you can read a note right to left or left to right and get
different meanings. Perhaps errors of recognition might lead to a
whole new genre?

Just thinking,

jc

-- 
This is the NewtonTalk list - http://www.newtontalk.net/ for all inquiries
Official Newton FAQ: http://www.chuma.org/newton/faq/
WikiWikiNewt for all kinds of articles: http://tools.unna.org/wikiwikinewt/


This archive was generated by hypermail 2.1.5 : Sat Aug 27 2005 - 01:00:02 PDT