On Fri, 8 Sep 2006, Randal, Phil wrote:
Score appropriately, train your Bayes well, and the false positives
should diminish.
FUZZY_OCR gives crazily high scores to certain things.
One point per matched keyword, I believe. I've seen FUZZY_OCR,
by itself, give scores as high as 24.00.
Here's the distribution from one of my log files, as a matter
of fact:
score count
----- -----
4.00: 21
5.00: 9
6.00: 6
7.00: 4
8.00: 4
9.00: 15
10.00: 7
11.00: 7
13.00: 1
14.00: 1
24.00: 1
As you can see, 24 only happened once, but 9, 10, and 11 are
very common.
So yeah, false positives should diminish some, but there is
no way a BAYES_00 is going to make up for a score of 11.
Personally, I think the scoring strategy for FUZZY_OCR needs
to be revamped...
- Logan
