Cory, (01)
Before getting into the details of your note, I'd
like to emphasize the point that the the idea of
concentrating on vocabularies for UF and minimizing
the detailed axioms is only part of the story. I
definitely do *not* advocate throwing away axioms.
But I believe that putting them into a single monolithic
lump is too inflexible. Even Cyc has subdivided the
axioms in multiple microtheories, but there's more
structure needed for interactions among modules. (02)
> Similar to your suggestion of modularity, perhaps
> the common framework is itself modular in such a way
> that conflict is explicitly allowed for between its
> modules. There is a tendency to think of these systems
> monolithic statements of Truth, and while some have had
> that intent, it may not be necessary for our purposes. (03)
Yes, I very strongly agree. (04)
> If we instead think of each of these ontological modules
> as "speech act"(s), an assertion by an individual, group
> or authority at a particular time. We can have a framework
> for dealing with a system of these modules that may or may
> not be in conflict. (05)
Yes. Inferencing and information flow are related, but
they have different requirements. In communication, the
most prominent features are the speech acts and vocabulary.
I published an article that emphasized speech acts and
their relationship to modularity: (06)
http://www.jfsowa.com/pubs/arch.htm
Architectures for Intelligent Systems (07)
This article was strongly influenced by John McCarthy's
Elephant 2000 language, which is based on speech acts.
The article introduces the Flexible Modular Framework (FMF),
in which interacting modules pass messages among themselves,
each one with a tag that indicates the speech act. For
McCarthy's original paper on Elephant, see (08)
http://www-formal.stanford.edu/jmc/elephant/elephant.html
Elephant 2000: A Programming Language Based on Speech Acts (09)
A quotation from that article: "Human speech acts involve
intelligence. Elephant 2000 is on the borderline of AI, but the
article emphasizes Elephant usages that do not require AI." (010)
My colleague Arun Majumdar implemented a version of the FMF
while he was living in Palo Alto and had a number of discussions
with McCarthy about Elephant, speech acts, and related issues
of logic and reasoning. (011)
> The speech act asserting a set of statements forms one dimension
> of context for those statements. A set of statements is valid
> within a context. There are, of course, other forms of context
> such as authorities/political, physical or situational. All can
> govern how and when statements within that context are to be treated. (012)
The word "context" has many senses. The speech act indicates the
purpose of the communication: tell, ask, command, promise, commit,
authorize, deny, etc. The basic three -- tell, ask, command --
have been implemented in computer languages for years, but the
others are important for on-line business transactions. (013)
Another sense of context involves the subject matter, which brings
in specialized axioms or programs that deal with the information
content of the message. The kind of axioms required depend on
both the subject matter and the speech act. One of the basic
principles of the FMF is that a module should be able to determine
quickly whether it is capable of handling a message, and if not,
where to send it. (014)
> Using something like wordnet makes a lot of sense for informal
> (but still useful) connections, but I don't see how it works
> for formal ones. (015)
You can think of the modules as having different levels of expertise.
Some modules may be highly specialized, and others could behave
like a receptionist, who isn't an expert in anything, but who
knows enough about the subject to route a message to another module
that can handle it. The FMF also supports various blackboards, on
which a module can post messages that are retrieved associatively
by other modules that are looking for things to do. (016)
Inside a module, there might be a highly specialized and optimized
program, or there could be an inference engine that does reasoning,
which could transform a message and generate new messages for other
modules. So even for the same subject domain, different modules
could have different axioms (or compiled programs) that do different
things. (017)
This approach does not rule out monolithic systems, if they are
needed -- you could take all of Cyc, put a wrapper around it,
and it would look like a module. But you could also split the
microtheories or other axioms into different modules that would
deal with specialized subjects in specialized ways. (018)
> Given such contextual modules you can determine what modules are
> in conflict with others, and perhaps provide logic to reduce or
> eliminate the root causes. (For example, the assumption that time
> is the same for all participants - which works just fine in the
> context of earth systems but not in the context of space flight). (019)
That's an example where different modules could use the same
vocabulary about space and time, but different sets of axioms
for reasoning about them. Even for travel, you need different
axioms and ways of thinking about driving a bus, guiding a
Martian rover, sailing a ship, running a railroad, or planning
a space mission. One set of axioms definitely does not fit
all applications, even when there is a lot of common vocabulary. (020)
> Almost all human abstractions seem to be highly contextual, yet
> most logics don't have the mechanisms to deal with it due to
> the monotonic restrictions. (021)
I certainly agree. OWL, for example, is an extremely limited logic,
which doesn't even support full FOL. It might be adequate to handle
the limited kinds of axioms needed for WordNet, but certainly not
for nonmonotonic, modal, higher-order logics that interface with
programs that do specialized computation. (022)
> One of the goals for this is to have a wider net for capturing
> knowledge, much of which is expressed in ways that are imprecise,
> lacking in detail and contradictory. (023)
Knowledge acquisition, design, problem specification, and many
related issues involve very different issues that are orthogonal
to the communication/inferencing dichotomy. That is one more
reason for breaking up the system into multiple interacting modules,
each with different kinds and levels of expertise. (024)
> So, in summary, can we find a way to "admit all" instead of
> least-common denominator by applying context to statements as
> speech acts from various communities or authorities? (025)
I believe that it's possible, and I also believe that a monolith
is not sufficiently flexible to handle the job. Modules, along
the lines we are discussing are a prerequisite, but there's still
a lot of R & D to do in order to determine the best way to divide
up various tasks among interacting modules. (026)
John (027)
_________________________________________________________________
Message Archives: http://colab.cim3.net/forum/ontac-forum/
To Post: mailto:ontac-forum@xxxxxxxxxxxxxx
Subscribe/Unsubscribe/Config:
http://colab.cim3.net/mailman/listinfo/ontac-forum/
Shared Files: http://colab.cim3.net/file/work/SICoP/ontac/
Community Wiki:
http://colab.cim3.net/cgi-bin/wiki.pl?SICoP/OntologyTaxonomyCoordinatingWG (028)
|