Pierre Thomson <[EMAIL PROTECTED]> wrote:
> [...]
> I'm not saying Bayes isn't working most of the time, but it
> does seem possible to craft sentences that skew it strongly
> towards ham.

Oh, and I should clarify that I don't think it's as much of a problem for SA as
for "pure bayes" solutions. Since SA uses a variable number of OTHER rules for
detecting spammish characteristics, I find that the combination is quite
effective. Other rules detect the spam, those are fed into sa-learn, and
eventually bayes figures out that other traits mark those messages. At least
based on my small setup.

>  The faked sentences in the original posting
> contained plenty of hammish words, but there were a few that
> would eventually end up as spam markers with training.  For
> example, the improperly hyphenated "calms-down" and
> "stands-still", as well as "caw" and "binocycles".

Again, as long as I don't start subscribing to literary lists, I think I'm safe
with SA. :)

One of the reasons I've been going through all the gyrations with procmail,
antivirus and SA is to try running a couple head-to-head. I hope to configure
bogofilter and a couple of other bayes-only tools and run them in parallel.

- Bob

Reply via email to