> In these sensitive parts of the coder, you
> obviously want to use as much context as
> possible, but if you use too much your
> statistics become too sparse and you start
> making big coding mistakes.

In other words, the context quantization function
is a lossy compression function, but unlike
lossy media coders, we have a well defined
quality metric here (entropy).

> For example something that hasn't been explored
> very much in general in text compression is
> severely assymetric coders.

I guess you just missed it completely.
In text compression, there're preprocessors, like
And in special cases, like Hutter Prize, people
put a lot of work into selection and arrangement
of words in the dictionary.

Also there're quite a few projects with
parameter optimization pass.
For example, see
There's a "o" processing mode which builds a
model profile for given data samples (which
includes context masks and counter update
constants and the like).

There're also many other projects with a similar
approach (eg. epmopt and beeopt), and most of my
coders are made like that, just that it makes
more sense to use in the development stage than
in public utilities.