Re: SUO: Re: "Monolithic" Ontology
Leo, Pat, et al.,
I also sympathize with Pat's request, and I agree with Leo's response.
But I believe that it is possible to develop a compromise that satisfies
both. That is in fact, what I was proposing in my paper, "Signs,
Processes, and Language Games":
http://www.jfsowa.com/pubs/signproc.htm
The fundamental problem for NL processing is the extreme flexibility
of natural languages to adapt to any and all ways of thinking about
the world (i.e., any or all possible ontologies). Some logicians,
such as Frege, Russell, and Carnap thought that was a weakness
that had to be overcome. Frege, for example, set out "to break the
domination of the word over the human spirit by laying bare the
misconceptions that through the use of language often almost
unavoidably arise concerning the relations between concepts."
While he was still under the influence of Russell, Wittgenstein
adopted that point of view for his first book. But he later realized
that the enormous flexibility of language was not a weakness, but a
strength. I wholeheartedly agree with the later W., and the point of
my paper was to develop a framework in which we could support an open
ended number of distinct concept types (i.e., precisely formalized
predicates), which could still be mapped to the more limited number
of words available in any natural language.
Leo Obrst wrote (to Pat Cassidy):
LO> I fully sympathize with your request and understanding. However,
> what you are requesting is no less than a 'terminological
> concatenation of concepts' index, which incidentally has been a
> scourge for KR and ontologies (and symbolic processing) in the past
> and no doubt will be in the future. Not that they are not needed.
> As concepts get refined and reified, they will indeed have unwieldy
> names (perhaps, to a lesser degree if you go the pure axiomatic
> route). But the limit is of course "names" which recapitulate all
> the conceptual relations, connectives, etc. So you end up defining
> a logic of names in addition to your logic of concepts. And they may
> be nearly the same.
> There will be such entries as:
> Time
> Real-lineTime
> Integer-lineTime
> ComplexNumber-lineTime
> Real-lineTimeWithPointsAndIntervals
> .....
I agree with Leo that the naming scheme for the predicates in any
precisely formalized ontology will have to use rather long names.
An alternative, which I believe is preferable, is to use short names
that are prefixed with the name of the theory in which they are used.
For example, there might be a predicate (or concept type) "time",
which is used in multiple theories with only a four-letter name.
But when it is necessary to distinguish the exact predicate or type,
the name of the theory could be prefixed to it:
realLine.time
integerLine.time
realLineWithPointsAndIntervals.time
Within a theory, the short name could be used, but the fully
concatenated name would have to be used in global references.
In fact, this approach provides a way of bridging the NL lexicon
with the formal theories: any NL word such as "time" would be
considered the short reference that is usable only within a
particular theory. The exact "word sense" would require the short
word to be prefixed with the identifier of the theory that is
assumed (implicitly or explicitly).
LO> Of course, it may be better to have an index or catalog of
> CONCEPTUAL PHRASES (CP), but then you have to define a language and
> a lexicon for CP. I understand the need for such a beast: you/we will
> need indices in the eventual "lattice-of-theories" or "library of
> theories" to find stuff. Compare the alternatives. Decide which meets
> your need.
In the last part (Section 7) of my signproc.htm paper, I proposed
that we need such a mechanism if we are ever going to map natural
languages into any formal notation for logic. The vocabulary of any
particular NL together with the "canonical graphs" (my term for a
generalization of the "case frames" that many comp. linguists use)
determine the mappings from the NL phrases into the formalism.
LO> Again, an agreement on "primitives" would be useful, but terms are
> much looser than concepts (whatever your theories of each of those
> are).
I don't believe that an agreement on "primitives" would be any easier
to achieve than an officially standardized ontology, and I am on record
as saying that the latter is impossible.
LO> If you are also or instead asking whether it is possible to agree
> on a natural language definition of some set of primitives, that may
> be possible. But then that is the case with most non-symbolic
> standards, no? I don't want to preclude that, I just think that that
> kind of activity is a different mode for SUO.
Lexicographers have been doing that for centuries. The results are
called "dictionaries". They're useful, but not perfect, and they
never will be. I agree that they are outside the scope of the SUO,
but I also believe that a complete system of ontology must show how
to accommodate NL and support NLP.
LO> But I'd like to hear of others' responses. Perhaps I am too
> pessimistic?
No. I agree with your points, but then people have also called me
pessimistic. I agree with Alan Perlis's statement: "A year spent
in artificial intelligence is enough to make one believe in God."
But I still remain optimistic (according to my definition) because
I believe that there is a way of dealing with the mess (which I
outlined in my signproc.htm paper). But I'm not holding my breath
waiting for somebody to finish the job soon.
Patrick Cassidy wrote:
PC> John and Leo --
> You have expressed reservations about a "monolithic"
> standard ontology. I would like some clarification:
> I can easily imagine agreement on a monolithic
> **terminological** ontology, which contains every concept
> that anyone would want to work with, defined in
> consistent terms. Where two groups would like different
> definitions of some commonly-used concept, e.g. time,
> it would be necessary to use different terms to label
> the different concepts, e.g. Real-lineTime,
> ClosedIntervalTime, whatever. Different fundamental
> primitives may be needed, but the attempt to reduce
> the most basic primitives to a minimum should be
> valuable, not the least as a learning tool for those
> who come after us and have to try to understand and use
> the ontology.
This is the point that Leo and I were addressing. As I have said,
I believe that some mechanism (such as the lattice of all possible
theories) is necessary to represent everything that can be said
in any NL. The concatenated names, as I suggested above, could be
formed by prefixing the name of the theory to the short name derived
from the words of some NL.
PC> Where people are most likely to disagree is on which
> of these terms and their relations conform most closely to
> the real-world as they see it, i.e. which of the terms
> in the terminological ontology subsume certain specific
> instances of real-world objects, processes, and events.
> Where such disagreements imply logical inconsistency,
> ***then*** there would need to be a provision for
> microtheories, or a lattice of theories, or, as a
> desperate last resort, mappings between theories.
I would turn things around. Instead of considering mappings between
theories to be "a desperate last resort", I consider that approach
to be the salvation we have been seeking.
PC> Is there a reason why we can't at least try for a
> monolithic terminological ontology? I think having one
> would be very valuable to allow us to get agreement on
> the detailed definitions (complete or partial) of
> specific concepts before we engage in battle over which
> of these best describes the real world.
We already have many monolithic terminologies. They are the NL
dictionaries. The primary task of NLP is to find systematic ways
of mapping those monolithic dictionaries to the open-ended number
of possible theories.
PC> This is a very different task from just building
> a "catalog" of ontologies whose similarities and
> differences are not carefully specified.
I would never recommend such a catalog. Instead, I have recommended
a systematic lattice of theories in which the differences are precisely
specified. See Section 7 of my signproc.htm paper.
John Sowa