[Date Prev][Date Next][Thread Prev][Thread Next][Chronological Index][Thread Index]

cross-entropy reduction is not enough

Hi all:

>From Martin Rajman to Lin Chase:

Computing cross entropy is interesting, but is there any evidence showing
that it correlates with the accuracy of the systems?

My response:

For evaluation confidence annotations (at any level), simply measuring
reduction in cross-entropy is absolutely not enough.  In my Granada
papers on this topic I give detailed examples indicating that application-
specific measures are very important.