At 01:42 PM 6/25/2004, Pierre Thomson wrote:
Here's a debug of the Bayes portion of an SA run on an email I recently received. Individual token scores range from 0.01 to 0.91, but the final Bayes score is 0.0002!
How does the final score come out so low? I thought it was computed from the arithmetic mean of all the token scores. Or is there something else going on? (This is SA 2.63 with mostly auto-learned Bayes DB.)
No, by default it uses the chi-squared combination of the token scores, not the arithmetic mean.
You can force it to not use chi-squared combining, but even then it doesn't use arithmetic mean, it uses Naive bayesian combining.
