Definition of "recognition rate"?
John Manuel
few things. For instance, just what do people mean when they talk about
the Newton's "recognition rate"? Are they talking about the percentage
of words or the percentage of letters that are recognized correctly? Or
does the recognition rate just turn out to be the same in both cases?
I'm trying to get an impression of just how tiresome spelling
correction on the Newton will be once it and I have plateaued in our
efforts at training one another.
----
John R. Manuel office: (603) 646-2723 fax: (603) 646-3856
Thayer School of Engineering, Dartmouth College, Hanover, NH, 03755 USA
Patricia Crane Wells
John,
I have done some very comprehensive studies of this issue for my masters
thesis and thought I'd pass along some of the numbers for you.
These experiments were done with the Names application, so bear in mind
how troublesome proper names are. The average recognition rate - here
meaning the word was recognized completely correctly, case and everything
- was 82% for the initial attempt. Another 8% of the words were in the
alternative word list. An additional 3% of the words were correctly
recognized using the letter-by-letter option, and another 1% were in the
second alternative word list. Overall, 94% of the words were recognized
without having to type them in. This is assuming that all names are
added to the dictionary.
The average time to correctly enter a name into the Names app by
handwriting was 2:35 (min:sec). I also timed how long it takes just
straight typing with the keyboard, 2:25, and with Graffiti, 2:46.
My thesis focuses on improving these recognition percentages and entry
times. I added proto-pickers (filled with the 4 most recent values) and
auto-filling for the remaining address lines to the built-in Names
application. The application which does this is called cardFile++ and is
available at newton.uiowa.edu in the submissions directory. When using
cardFile++ the recognition error rate falls from the 18% in the Names
application to 3%, and the entry time drops from 2:35 to 0:58.
If you have any further questions or comments, please email me at
pwe...@eecs.wsu.edu or post them here.
Patti
----------------------------
Patricia Crane Wells
School of EECS
Washington State University
Pullman, WA 99164-2752
(509) 335-2215
pwe...@eecs.wsu.edu
Tom Harrington
<pwe...@eecs.wsu.edu> wrote:
[...]
: The average time to correctly enter a name into the Names app by
: handwriting was 2:35 (min:sec). I also timed how long it takes just
: straight typing with the keyboard, 2:25, and with Graffiti, 2:46.
This long, just for a name? What kind of names did you use?
Even if you can only hit one character/second on the keyboard,
2:25 translates to a 145 character name. Even if I type my full
middle name I've only got 25 characters in my name, including
spaces. Does "enter a name" include more than just a name, or
is there some other explanation?
--
Tom Harington ------------ t...@rmii.com ----------- mech...@aol.com
-IF YOU ACT LIKE A DUMBSHIT THEY'LL TREAT YOU AS AN EQUAL. For more-
---info, email me or send $1 to The SubGenius Foundation, P.O. Box---
----------------------140306, Dallas, TX 75214-----------------------
"And, in the end, the slack you take, is equal to the slack you make"
David W. Griffin
> My thesis focuses on improving these recognition percentages and entry
> times. I added proto-pickers (filled with the 4 most recent values) and
> auto-filling for the remaining address lines to the built-in Names
> application. The application which does this is called cardFile++ and is
> available at newton.uiowa.edu in the submissions directory. When using
> cardFile++ the recognition error rate falls from the 18% in the Names
> application to 3%, and the entry time drops from 2:35 to 0:58.
>
What's a proto-picker? Perhaps I'm dense, but I don't understand how
you improved recognition percentages.
-----------------------------------------------------------
David W. Griffin | Martin Marietta Corp
grif...@escmail.orl.mmc.com | Orlando, Florida
(407)826-3697 | Posts are my opinions only
John Manuel
Patricia Crane Wells <pwe...@eecs.wsu.edu> writes:
> These experiments were done with the Names application, so bear in mind
> how troublesome proper names are. The average recognition rate - here
> meaning the word was recognized completely correctly, case and everything
> - was 82% for the initial attempt.
I understand that the recognition rate depends somewhat on the Newton
being used and how well it and the user have been trained. What version
of the OS were you using and how thoroughly trained was the Newton (and
the user) that you were testing?
----
John R. Manuel office: (603) 646-2723 fax: (603) 646-3856
Patricia Crane Wells
grif...@sunny.dab.ge.com writes:
>
OK. Lots to answer.
When I am talking about entering a name, I mean an entire 'Name' entry:
First, Last, Title, Address, City, State, Zip Code, email, and Phone
Numbers.
Yes the Newton gets better with use. You train the Newton, the Newton
trains you. I have been using a Newton for about a year. However for
the experiments, before each test case I would do a hard reset which (I
believe) wipes out all of the learned handwriting recognition. I used
an MP100 running 1.3 OS for the experiments. The 82% figure I quoted,
does not include letter-by-letter recognition attempts.
And finally, proto-pickers are the little diamonds that when you tap on
them, a list of items pops up for you to choose from. For example, the
country field in the 'Names' application. I have more like DECREASED the
MIS-recognition rate, than increased the recognition rate, I guess. It
is still extreamly helpful though. Notice how much simpler it is to
select USA from the picker than to try to write USA (especially if it is
not in the dictionary). I just added this capability to many of the
fields, along with other enhancements.
Any more questions or comments, feel free.
Patti