AFAIK, sphinx log-base is used to enhance the dynamic range of likelihood
scores. Does scaling scores enhance accuracy in any way? One positive effect I
can see from the scaling is all math can be transformed into the 'int' domain.
But do the numbers need to scaled by sphinx-base values set closer to the
default of 1.0003? Does anyone have experience on this, or could share their
insights? What is the suggested range over which acceptable results can be
obtained, and what are the trade-offs involved?
cheers,
- H
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
AFAIK, sphinx log-base is used to enhance the dynamic range of likelihood
scores. Does scaling scores enhance accuracy in any way? One positive effect I
can see from the scaling is all math can be transformed into the 'int' domain.
But do the numbers need to scaled by sphinx-base values set closer to the
default of 1.0003? Does anyone have experience on this, or could share their
insights? What is the suggested range over which acceptable results can be
obtained, and what are the trade-offs involved?
cheers,
- H
Well, if you really want to collect those breadcrumbs, you probably will be
interested in checking
https://sourceforge.net/tracker/?func=detail&aid=1205946&group_id=1904&atid=1
01904
In general, I don't think there is something to gain there. Anyway, it's just
a computation.