Re: SUO: Finding an upper ontology
Pierre,
I find most of your suggestions compatible with what I am
proposing. Therefore, I am baffled that you can't see that
our goals are quite compatible. Your view of ontology as
a scientific enterprise is fine with me, and I thought that
you would be pleased with the direction I proposed.
In any case, I'll answer your questions:
PG> I'm puzzled by the claim that the issue is not only a question of
language.
> I don't see how language is assumed to be relevant in the first place.
I don't know what you are disagreeing with. What is it that language
is supposed to be relevant or irrelevant to?
JFS>I like Downes's phrase "a regime of ontological authoritarianism",
PG> I think you once said that some people on this list were comparable to
> religious fanatics. I'm glad that you would seem to be ready to attribute
> rationality to those people now.
I consider both religious fanatics and authoritarian dictators to
be extremely rational -- too rational in fact. It's their intuition
that is misguided. Everybody's starting axioms are based on a "gut
feel" or an emotional attachment of some sort. What is wrong with
fanatics and dictators is that they never readjust their axioms
to accommodate reality.
> It seems to me that you have already determined whether a single SUO
> can emerge. The real problem is the way the output of this group
> will be _built_.
I don't think that we disagree. Please reread point #2, which
most definitely is not a joke:
JFS>> 2. For starters, the only universal point we can all agree to
>> is a one-category ontology, which contains nothing but the
>> empty node at the top. That is a start, but not a very
>> useful one.
As I said, there will always be a single upper level, but it
might consist of nothing but the empty node at the top. It is
an empirical issue to determine whether a common hierarchy of
more than one node is possible. And even if we discover a larger
commonality, of perhaps a few dozen nodes, that still might be
too small to be useful. I make no assumptions, a priori,
about how large that commonality might be.
JFS> 3. The next step is to determine a set of distinctions that
>> can be used to classify entities. That is a time-honored
>> principle since Socrates and Plato. I describe that
>> process in Ch 2 of my KR book and summarize it on my
>> web site: http://www.jfsowa.com/ontology/
>
> Just a small clarification question: would you really be happy with an ontology
> which resembles your lattice of categories of beverages or is it an ad hoc
> dummy? (BTW the link for your tutorial on lattices is broken)
That is an example (which I extracted from a paper by one of the
FCA proponents). I corrected the URLs, and you can go to the
math & logic paper to see how the method works. It always generates
a lattice from any collection of distinctions (which need not be
binary). If the shape of the lattice is lopsided, that is often
a clue that the starting set of distinctions is the result of
an incomplete or inadequate analysis of the original data.
JFS>> 4. The combination of all distinctions that are truly universal
>> -- i.e. those that apply to any entities whatever -- form
>> a lattice, as I summarize in the opening page of my web site
>> and as the Formal Concept Analysis people have supported
>> with a great deal of theory, practice, and software.
>
PG> I don't see how distinctions can form anything but trees, honest.
That depends on whether the distinctions are ordered or unordered.
If the distinctions are completely independent (i.e., they can be
determined in any order), all possible combinations are permitted,
and they generate a perfectly symmetic complete lattice. Leibniz was
the first to propose that method for his Universal Characteristic.
The result of his method was a complete lattice, but Leibniz
wasn't the one to invent that term.
If there is a strict linear ordering where each choice depends on
the answer to the previous choice, then you get a tree (which could
be converted into a rather uninteresting lattice by tying all the
leaves of the tree into a single bottom node). In any case, the
FCA method (and the software that supports it) will always generate
a lattice from any collection of distinctions.
JFS>> 7. One example of an incompatible distinction is the 3D vs. 4D
>> representation of space-time, which everybody has been citing.
>> The question of how many such incompatible distinctions there
>> are is an empirical issue that can only be determined by
>> analyzing the data.
>
PG> Which data? Do you mean that you will get the distinction for a SUO form
> OpenCyc, SUMO and the like?
By data, I mean any source of inormation about any subject matter
that is being represented in the ontology. It could be a raw,
unanalyzed description, or it could be the result of somebody
else's analysis -- such as SUMO, Cyc, etc., or the sources from
which those developers took their axioms.
> The distinction you are speaking of, you can't expect to find them in a sample
> of ontologies, as broad may be while remaining manageable, especially when the
> elements in the sample bear so much similarities. You acknowledge yourself
> (don't you) that particular theories will follow a path in the tree (of
> distinctions) or maybe semi-lattice of theories or whatever. So you hope to
> find in OpenCyc and SUMO a number of choices which indicate a number of
> alternative and build your lattice of theories based on these alternatives?
> Again, it seems to me a loosy way to go. Distinctions should come a priori,
> especially if your reference is Plato and not Lenat.
The distinctions can come from anywhere. Many of them have been well
documented in the literature of philosophy, linguistics, psychology,
etc. Others are implicit in the documentation of whatever subject
matter is being represented, and somebody who is trained in conceptual
analysis must dig them out.
If the SUMO and Cyc developers have done a good job, their documentation
should indicate what distinctions they were trying to represent. I am
not assuming that they have always done a good job, but I am also not
assuming that their work is useless. I am willing to take anybody's
previous analysis as a potentially useful starting point. But I am
always trying to determine whether a better analysis can be done.
JFS>>Summary: I believe it is possible to *discover*, not *legislate*,
>
PG> Lesgislation = philosophical analysis?
> Discovery = software engineering exegesis?
Legislation is most definitely not analysis. Legislation is an
attempt to force one's opinion on everybody else -- independently
of whether the analysis is complete or inadequate.
Discovery is the result of the best possible scientific methodology.
The people who do the discovery should be well informed about what
has previously been done (i.e., they look at everything, including
previous attemtps, such as SUMO and Cyc), but they should also have
a thorough training in the best methods of analysis and knowledge
representation. And I regard the work that has been done on IFF
an example of good work that the analysts should know. And FCA,
which I illustrate in the math & logic paper, is one of the methods
that is included in IFF.
What other methods of discovery do you recommend? In recommending
IFF, I certainly am not excluding other methods. And I believe
that the IFF developers would also be happy to learn of any other
methods that you or anyone else might suggest.
John Sowa