Thread Subject: Neural Network to generate English letters

I'm designing a neural network to predict english letter,
based on user's typing history. Can I produce the output
coding in ASCII (is there some examples?). It seems not
easy to standardrize the input and output.

Another question is, what do you think about input, I mean
what kinds of user's performace should be considered at the
side of input? Thank you!

"Jun Li" <arrangers@googlemail.com> wrote in message <g1jro8
$m3$1@fred.mathworks.com>...
> I'm designing a neural network to predict english letter,
> based on user's typing history. Can I produce the output
> coding in ASCII (is there some examples?). It seems not
> easy to standardrize the input and output.
>
> Another question is, what do you think about input, I
mean
> what kinds of user's performace should be considered at
the
> side of input? Thank you!

At the risk of upsetting Walter again, I would have thought
that you were trying to do something almost impossible. I
would have thought that there are very few rules which
define which letter follows a previous one (i before e
except after c seems to be one of the few).

I would have thought that you would be better off working
on a word basis. From what I understand, a well educated
English speaker will only have a few thousand words in
their active vocabulary. I would guess that monitoring
someone's typing and comparing this with a machine readable
dictionary, you could generate a histogram showing the
extent of their use of the English language, from which
monitoring their typing you could make a good guess at what
they are trying to type. For example you can search through
virtually all my work and will never find that I use the
word "aardvark" (whoops the truth of that statement bites
the dust;-).

I would have thought that such an approach would have been
ideally suited for neural net techniques, and wouldn't be
surprised that something very similar is used in predictive
text systems built into a modern mobile phone and PDA.

Regards

Dave Robinson

In article <g1k2hl$r6n$1@fred.mathworks.com>,
Dave Robinson <dave.robinson@somewhere.biz> wrote:
>"Jun Li" <arrangers@googlemail.com> wrote in message <g1jro8
>$m3$1@fred.mathworks.com>...
>> I'm designing a neural network to predict english letter,
>> based on user's typing history.

>At the risk of upsetting Walter again, I would have thought
>that you were trying to do something almost impossible. I
>would have thought that there are very few rules which
>define which letter follows a previous one (i before e
>except after c seems to be one of the few).

'q' is always followed by 'u' except in imported words
of arabic origin. 'j' only ever occurs at the beginning of
a syllable: intra-syllable 'g' is used instead.
'gh' at the beginning of a syllable ('ghost',
'ghoul') is of different origin than 'gh' at or near the end
of a syllable ('tough' or 'thought') and follows a different
(much smaller) statistical distribution. 'x' always ends its
containing syllable, except for leading 'xy'. Etc..

No, there is no fundamental problem in using neural networks to
predict english letters. Shannon's fundamental work on information
theory was based upon a letter guessing-game, in which people were
shown a small number of letters in English words and asked to guess
the next letter; the average number of guesses that they needed
(over a large number of trials over different people) was only about 2.2.
This suggests that "not so horrible" prediction could be done with
as little as 3 NN layers, with additional layers adding to the accuracy.

--
  "The study of error is not only in the highest degree
  prophylatic, but it serves as a stimulating introduction to the
  study of truth." -- Walter Lipmann

Many thanks Dave and Walter. It seems my question is very
clear. (You see, It's funny a non-english man is trying to
predict english :-) ).

Some apps have been developed, such as using PPM algorithm
etc. But for me, there are two problems: result is not
accurate, and they cant correct typing errors. So if we
take one of those apps as NN's input, let nn deal with
these two problems. after epoches, NN could generate a
better solution.

Then, please look back my questions. Thanks


Ps, this are certain rules behind two consecutive letters
of universal english. if you check 'bigram'


Jun Li