Murray Altheim wrote:
Dan Corwin wrote:
I feel [1] is a decent spec on most of the role types [CG's] require,
and is certainly ample for an informal debate on this thread about
what future changes or extensions (if any) we'd like to make to it.
Are you talking about specifically Figure 3, "Thematic roles as subtypes
the four types of participants", that kind of thing?
Yep. Under ontology A, we can model the same case role types,
changing only the ways constraints get expressed for each of
the (19) thematic (case) role types of CGs, defined at [1] and [2]
This role ontology forms the bulk of "Conceptual Graphs". WORDS just\
wraps its own PSIs around these public definitions, so as to "snapshot"
today's published version of those concepts. Then it adds ..
the (27) constraint types and values at [4], standing in for TMCL
the (50) part-of-speech types at [4] for case-role constraints to cite
The rest of ontology A covers TM constraints in general under the
WORDS paradigm. This is bigger in scope than CGs, but WORDS can
readily absorb the CG constraints - mostly just a domain - with no
change in meaning.
And [4] lets *your* ontology types be cited as constraints, whereas
under [1] they can must be John Sowa's. This seems a needed upgrade.
[3] http://www.icsi.berkeley.edu/~framenet/ (browse its examples)
Patterns under ontology B effectively define verb types, which help
a WORDS script expand like a templates into an instance of CLAUSE.
But nobody can yet make *any* such verb lexicon a standard, nor
can CGs - as that is just too big a job. In practice, many people
will do in many small steps of only a few verbs at a time.
I must admit (somewhat sheepishly) that in looking down the list of
people on the FrameNet site, I'm hardly familiar with any of them.
My guess is they are mostly local grad students, looking to pick
up some extra cash and/or glory as computational lexicographers.
BTW, another NLP group doing similar work is PropBank. I suspect
it is staffed the same way, but this link may give details:
http://www.cis.upenn.edu/~ace/ace_feb02_files/v3_document.htm
Teams like Framenet and PropBank work slowly. Lexicography is
painstaking, so just be patient. They do build quality models.
Any pointers on what are the best papers, best summaries, etc.? If
we're going to work on this together I'd like to get up to speed on
some of the same things as you've looked into.
Google on James Allen, who literally wrote the (text)book on NLP.
Or just find it in your local library. Focus on semantic processing
areas (versus syntax), and be aware that Allen is well known for his
definitive ontologies of tense and speaker-relative time.
[4] http://www.lexikos.com/psi/words/
This looks suspiciously similar to the ontology I've been building
in Ceryle... some of the Things certainly are identity matches (at
least at first glance).
Great minds do think alike :-) But here, please glance again.
Technically, I'm modeling "part-of-speech" (POS) for word-meaning
topics viewed as data structures, NOT the type of their subject.
Philosophically, this is very slippery turf, but I do it so that
the WORDS type hierarchy can stay almost 100% independent of its
client's, and well hidden within obscure "POS" occurrences.
I'm going to have a wiki up in a week or two (I hope), and perhaps
we can devote some space there to discussing this, such that the
results of our discussions gradually build up some documentation.
Timing wise, that is fine. Meanwhile, I'm traveling so much of late
I barely have time to read email, much less answer it.
So please forgive response delays; I have another trip tomorrow.
Does that sound like a reasonable plan of attack?
See my concurrent reply to Jack on this.
Cheers,
Dan Corwin
| 
