[Corpora-List] Off-the-shelf coreference packages

[Mark Sammons]

Hi, All.

Craig Pfeifer said:
>
>You could try the coreference module from the U Illinois Cognitive
>Computation Group:
>
>http://cogcomp.cs.illinois.edu/page/software_view/18
>
>one caveat : "To train the coreference classifier, you will need
>annotated training data such as the LDC's ACE 2004 corpus. "

Just in the interest of clarity: the downloadable jar already comes
with a trained classifier, so you should be able to use it as-is without
training it.

Regarding the types of errors it makes: it is somewhat conservative,
according to precision/recall/f1 statistics, in that it tends to make errors of 
link omission rather than of adding spurious links. However, it still adds 
*some* spurious links. 

It also anecdotally performs better on medium-long texts (i.e. more than
just one or two sentences) than on short texts, presumably because 
additional context helps.

I hope this helps.

Regards,

Mark 
Mark Sammons
Principal Research Scientist
Cognitive Computation Group
University of Illinois,  Dept. Computer Science
(217) 265-6759  mssammon at illinois.edu

_______________________________________________
UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
Corpora mailing list
Corpora at uib.no
http://mailman.uib.no/listinfo/corpora