[Corpora-List] WordNet vs Ontology

[John F Sowa]

Mike,

I'm not arguing for prescriptive linguistics.  On the contrary,
note my strong endorsement of the point by Sue A. & Adam K.:
"I don't believe in word senses."

JFS
>> That is an important point:  Using the word 'ontology' as
>> a catchall term for any kind of list makes it meaningless.

MM
> it's amazing to me to see a bunch of *corpus* linguists arguing
> about what a word *should* mean...  But the discussion here has
> been bordering on prescriptive linguistics, if you ask me.

Thank you for raising that point, because it clarifies the issues.
In short, WordNet is *descriptive*, and the ontologies used in
AI and knowledge representation are most often *prescriptive*.
Unless you add an emulsifier, they mix like oil and vinegar.

Note the often quoted definition by Tom Gruber:  "An ontology
is a formalization of a conceptualization."  I don't like that
definition, and I don't recommend it -- because it is strongly
biased toward prescriptive ontologies.

For certain applications, such as software design, a prescriptive
approach is obligatory:  the chief engineer must choose the design
principles and enforce them consistently.

For empirical analysis, however, the ontology must be derived or
discovered from the documents.  The primary reason for the failure
of so many projects in formal linguistics is that they start with
an a priori ontology of the subject matter.  Then they take pride
in following Frege's principle of using the syntax to guide the way
their a priori ontology is combined to generate an interpretation.
That method is guaranteed to fail.

I keep telling them to read Wittgenstein or Adam K's article.
But they never get the point:  To understand a document in
any natural language, you must *discover* its underlying
ontology.  Nobody (human or computer) can understand a document
if they start by trying to impose an a priori ontology on it.

John

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