naive bayes algorithm in ifile?

another idea i'm toying with is making a (portable) standard-prolog
implementation of naive bayes for (email/usenet) text classification.  the
free prologs have improved much over the years, and i want to know if a prolog
implementation is fast enough.

given n categories, t[i]; i {1..n} tokens per category, m[i]; i {1..n}
messages per category and for every token a record (age, c:i); i {1..n}, could
somebody please give a simple, english description of the algorithm needed to
classify a message?  i need to understand how token ageing can be used to keep
the database small, containing only the tokens that contribute the most to
classification and dropping the rest.

do i really need floating point operations or can i get away with integer
arithmetic?  could rational numbers be a better solution?

  clemens