On 3/3/11 3:05 AM, James Kosin wrote:
Jorn,
I performed the tests with the CONLL 03 data. 1.5.0 didn't have the
converters; so, I used the converted data from the 1.5.1 releases.
That is fine.
I trained both from the same set of training and tested on both sets of
CONLL 03 data for the English data.
Results below.
In my opinion we should train with only 100 iterations,
and test also for the individual name types.
That way we would do the test 5 times.
The model training output was identical; however, the models themselves
had minor differences.
The models contain a version, that is different, and if I remember correctly
also a time stamp.
Jörn
