[topicmapmail] Conceptual Graphs are Step 6

[Dan Corwin]

Jack Park wrote:

 > ps: where are the references you cite below?

I just posted this on the inferencing thread.  It leads to all of
the definitions and links I have been using, and also cites this
post and TSC as illustrations of type D extensions:

   http://www.infoloom.com/pipermail/topicmapmail/2004q2/006090.html

I'm a frames guy from way back, and I think naturally in their terms.
Discussions of logic, I confess, often seem pretty alien turf, but you
appear quite comfortable below with the blend.  So positive feedback
from you on ontology C, especially near [8], would be very reassuring.

To All -

I am comfortable believing that ontology C does "work" at some level,
because all it does is model CGs - their graph topology, not what it
represents.  Notice that CGs themselves do nearly the same thing in
ontology A, when they effectively model the parse tree of a clause.

In theory, logic cares about truth, not meaning, so all this should
work fine for inference engines, *provided* that the rules they use
are written around the same conventions as CGs: A-B-C types only,
plus other characteristics of the roleplayers for roles A.

But please complain here if you disagree, or see anything else that
jumps out as wrong about my posted May 25 summary.

Thanks,
Dan Corwin

Jack Park wrote:

> I think that there is a legitimate marriage between topic maps and more 
> than one class of inferential structure. I personally don't think that 
> raw XTM is particularly suited to inferencing, though Eric Freese has 
> been showing some pretty slick stuff with XTM. I would like to think 
> that there is some underlying representation scheme which, depending on 
> the type of fetch or access, yields an XTM structure, a CG structure, 
> plain old frames (POF), or whatever. Using PSIs in that structure 
> satisfies the need for identity, no matter what the structure might look 
> like.
> 
> I cannot escape the intuition that the TMRM, as I am watching it 
> metamorphose from the gem of an idea hatched by Steve and his friends, 
> to a structure suggestive of the, shall I say, canonical representation 
> scheme about which I just spoke.
> 
> Recall that Doug Lenat was pretty much "frames all the way", until 
> others got involved in Cyc, and then you got assertions, logical 
> statements, and a dozen or more different kinds of inference engines,. 
> If I had all the time in the world, and didn't have bills to pay, I'd go 
> to my grave chasing this wild-assed belief that, with assertions, we can 
> pretty much say whatever we want, and we can pretty much perform the 
> inferences we want with what we just said, and we can have topic maps at 
> the same time from the same effort.
> 
> TSC was originally just a rational reconstruction of Eurisko. Nothing 
> more than that, except that, whilst it was a POF system, it also brought 
> to bear the formalism of QP theory, which posits, in my own translation, 
> actors, relations, and states, all tied together with process rules 
> which roam about in episodes (akin to time slices of the history of some 
> process), looking to fire and have side effects, making new episodes.
> 
> That's not unlike what you could do with conceptual graphs, and even 
> topic maps, if you like. A process rule is like a template with unbound 
> variables. You could have association templates in XTM; why not have XTM 
> templates with unbound variables, looking to mate? It just seems that 
> there may be a bazillion ways to crack this puppy.
> 
> Ultimately, I'd like to do all that with a constrained natural language. 
> I built a template matcher for TSC and it got used by some kids to do 
> science fair projects, was used in a high school biology class to let 
> students query a QP knowledge base about immune responses to bacterial 
> antigen, and my daughter, at age 7, used it to invent her own version of 
> a taxonomy of living things. That turned out to be somewhat prophetic; 
> she and her brother coauthored a chapter in _XML Topic Maps_ on the 
> Linnean taxonomy as a topic map.
> 
> TSC was originally written in Forth on a Macintosh. System 6. Remember 
> that?
> TSC has recently been transliterated into Java and is slated to be open 
> sourced when I finish the triple-store backside and the tuplespace 
> server which it needs as an agenda manager in a distributed, 
> seti-at-home-like environment. It needs to be distributed because I 
> anticipate blooming enormous models of organisms when it is running.
> 
> 
> That's what I think I know at this time in history. Your mileage may vary.
> Cheers
> Jack

> 
> Dan Corwin wrote:
> 
>> Jack Park wrote:
>>
>>> Brief comment on framenet. I just did a quick scan of the site; it 
>>> looked terribly familiar. I think it's the work of Charles Fillmore, 
>>> who's early papers on case frames were referenced by Marvin Minsky in 
>>> his original paper about Frame representation systems. I'd have to 
>>> suspect that Fillmore's work predates most everything we presently 
>>> know about AI representation schemes.
>>
>>
>>
>> Yep, he's the same guy.  Influential and well respected.
>>
>>> My partner in the program The Scholar's Companion, Dan Wood, took 
>>> Fillmore's class and even implemented a case frame parser for 
>>> constrained natural languages, early in 1989 when we were working on 
>>> an SBIR project. There may be a lot to be learned from looking at the 
>>> framenet papers.
>>
>>
>>
>> One key thing I learned from it: Fillmore no longer focuses on
>> any fixed set of case frame role types - such as in [1].  So
>> the roles - in [3] - remains pretty ad hoc for ovcer all the frames.
>>
>> I sense he is not just being lazy or sloppy, but feels the role types
>> of [3] are actually sub-types of base types such as [1] (or his version
>> of it), which with modest effort could each be added by overriding (in 
>> the
>> O-O sense) some primative's rules or data for constraints, etc.  WORDS 
>> hopes
>> to support that, so roles in [3] can be derived from those in [1].
>>
>>> There may also be tons of merit in just writing de novo software that 
>>> behaves like CGs. That's what I did with qualitative process theory, 
>>> and now I'm pretty sure that QP theory can be made to behave very 
>>> much like CGs as well.
>>
>>
>>
>> I know nothing more of QP and TSC than what I was recently
>> able to find from Google and your website, but I'm interested.
>> Could you elaborate, and are they active as tools and/or R&D?
>>
>> Independently, please comment on my related reply to Murray
>> below, which addresses a "logic" ontology for CGs.  Could it
>> be used to feed TSC or QP with clause-like "facts" to evaluate?
>>
>> Regards,
>> Dan Corwin
>>
>>
>> * Murray Altheim wrote:
>>
>>>> Does [a wiki] sound like a reasonable plan of attack? The mailing list
>>>> works okay, but ideas tend to get lost into the archives to easily.
>>>
>>>
>>
>> Wiki space would indeed be nice, but ontologies A and B seem to me
>> well documented, if not always in PSIs yet, at least with subject
>> indicators like [1-3].
>>
>> The real need I see for more doc and dialog is on ontology C,
>> and related topics like Jack started above:
>>
>> To wit: how can CGs - *or any other CLAUSE-like assertions* -
>> best be packaged in XTM for processing by a remote logic engine?
>>
>> Being naive about such things, I'd guess from [2] that one really
>> only needs to define these 9 types:
>>
>>   a "C-GRAPH" type, with ontology B's CLAUSE as one sub-type
>>
>>   the six other C-GRAPH sub-types defined in [2], including
>>
>>      AND, NOT -------------------- the *only* primatives
>>      OR, IF-THEN ----------------- derived from AND & NOT
>>      PROPOSITION, SITUATION ------ sets of nested C-GRAPHS
>>
>>   a type of Part-Whole association for nesting C-GRAPH sub-types
>>
>>   a C-GRAPH attribute for TRUTH - something TMs normally ignore
>>
>> This ontology is tiny - way smaller than [1].  For each C-GRAPH
>> sub-type, it needs only a PSI, plus constraints on what other
>> types might become its role players.
>>
>> TRUTH may be harder to define in XTM, but note: we don't have
>> to *compute* TRUTH values for a typical C-GRAPH instance - some
>> FOL software could do that.  Ontology C only must define a C-GRAPH
>> network it evaluates, and set enough TRUTH values to get it going.
>>
>> Is this model on the right track for feeding inference engines?
> 
> 
> 
> 
> _______________________________________________
> topicmapmail mailing list
> topicmapmail@infoloom.com
> http://www.infoloom.com/mailman/listinfo/topicmapmail
>