Matthew Haines (haines@ISI.edu) wrote:
> Even when I sat down with the students and asked them to reconcile their
> differences, they eventually just threw up their hands and declared that
> either (1) the word was ambiguous, or (2) neither sense carried the exact
> meaning of the word. The second response I attribute to attempting to
> shoe-horn Japanese into English sense divisions. But the first case often
> was simply a problem of demanding too much specificity out of a word.
One of the possible advantages of WordNet in this context is that higher-level
categories can be chosen as sense/semantic labels if the lower-level
categories are too specific. To take Ted Dunning's case of "stock" as an
example, WordNet 1.4 distinguishes sense 5,
stock, caudex => stalk, stem => plant organ => plant part => part => natural
object => object, inanimate object, physical object => entity
from sense 6
stock => plant part => part => natural object => object, inanimate object,
physical object => entity.
Distinguishing these might well be a case of demanding too much specificity,
but there is an alternative: the taxonomic structure of WordNet affords the
possibility of tagging an instance of "stock" with the synset {plant_part}.
More generally, it seems to me that sticking rigidly to the traditional notion
of "sense discrimination" --- assigning labels like BANK1 and BANK2, or just
immediate synsets in the WordNet context --- might be leading us down the
garden path. What seems far more relevant is an expression of semantic
categorization for a word (of which sense disambiguation is a proper subset),
at the appropriate level of abstraction for the task one is interested in.
Philip
----------------------------------------------------------------------------
Philip Resnik E-mail: philip.resnik@east.sun.com
Sun Microsystems Laboratories Work: (508) 442-0841
Two Elizabeth Drive Fax: (508) 250-5067
Chelmsford, MA 01824-4195 USA
----------------------------------------------------------------------------