Hello Tamy and list members,
In the various reference lists, it is likely that some of the
taggers were generated by training. For example, Brill's tagger
train on Wall-Street Journal. If your material is from a different
domain and you are going to evaluate the results, performance could
be less than the reported figures. Therefore, you might like to
try several taggers. Also refer further to the discussion on POS
tagger evaluation on this mailing list, Early 2000 as far as I remember.
In general, I think it would be in favour of the statistical NLP
community if we could gather somewhere evaluations of statistical
tools on domains and genres different from those they were trained on.
Such information can be a basis for further research on 'flexibility'
of trained systems, let alone serve as a guide for people looking
for a system for a particular need.
Yuval Krymolowski
Bar-Ilan University
Israel
This archive was generated by hypermail 2b29 : Mon Jan 22 2001 - 17:35:40 MET