From: Eric Armstrong <eric.armstrong@eng.sun.com>
<snippage/>
> There is an interesting possibility that there is. Looking over the
> Traction offering and comparing that with IBIS concepts led to the
> minor epiphony that the simple act of categorizing information nodes
> according to some (agreed upon) schema is in essence a knowledge-
> abstraction process. I'm convinced that Traction is absolutely on the
> right track with respect to categorization -- IBIS is an
> easily-definable subset of their system. Where they fall down is with
> respect to document hierarchy, but they've made a contribution (to my
> thinking, at least)
> with respect to categories.
>
Kindly take the time to elucidate that which convinces you.
> What is interesting, here, is the concept that the whole "knowledge
> management" domain exists in the realm of the categories, where
> "documents" are found among the information nodes. If it makes sense
> to think of knowledge management in those terms, then we can conceivably
> apply some interesting abstract manipulations to "knowledge", where
> knowledge means a common (or possibly standard) set of categories, and
> where the underlying information is unique to each domain.
>
Standard categories: Ha!
Lakoff wrote a book called "Women, Fire, and Dangerous things," which, as I
recall, were the primary categories of some aboriginals. The Maoris include
reproduction in Earth Science. They do this because they see the
cycle of life and always bury a placenta next to a tree. So much for
logical categories.
> For example, the category "argument for" can be applied to information
> nodes in a biological sciences domain, or to one in an art analysis
> domain. The category is a form of meta-data that is independent of the
> information content.
>
> Now, given a standard set of categories, it might be possible to begin
> describing category-relationships. That would produce the property of
> abstract reasoning, that was independent of the problem domain.
>
> I keep thinking in terms of "implies". If there is some way to add the
> meta-data "implies" in the category space, then automated reasoning
> becomes possible.
>
> Example: at the initial writing, node A is written, as well as node B,
> with the "implies" attribute linking the two. Later, someone adds C as
> an implication of B. The system can now deduce that A implies C --
> regardless of the information content contained in the nodes.
>
> Perhaps the "category" for such a system is "implication". Categorizing
> B as an implication then requires pointing to A, to identify the node
> from which B was derived. The symettric relationship can then also be
> added -- call it "motivator", or some such. (If there is a logic term
> for it, I've forgotten it.)
>
> There might also be categories for preconditions, requirements, and
> what have you, all of which would allow for fairly sophisticated
> reasoning engines to be built on top of the fundamental structures.
> [There are also evaluations -- of node content as well as the logic
> employed...]
>
> The interesting point to all this is that the "DKR" becomes a layer
> of abstraction built on the OHS, where the categorization-capability
> is already built into the OHS.
>
Actually, as I see it, the DKR is an API that the OHS supports. Thusly,
most of the use cases should discuss what one wants to do at the OHS, and
DKR use cases will define that API.
In all of the above, reference is made to categories, relations,
implication, and so forth. It seems to me that mankind has been discussing
this since Aristotle, maybe before. It also seems to me that nobody has
achieved "the solution tres grande." It further seems to me that nobody ever
will. Therefore, we must take pains to define just what the DKR is intended
to represent and manipulate, then give it our best shot. Lenat has made an
enormous effort along these lines with CYC. There is even a public domain
version of CYC evolving. This all stems from the fact that Eurisko, powerful
as it was, never went very far simply because it lacked common sense. Guha,
McCarthy, and lots of others are making careers trying to figure out how to
represent common sense. Guha is not quite the champion of RDF, McCarthy
remains the champion of a variety of logical formalisms.
Automated reasoning is well documented. It does work for limited domains. In
fact, just about anything will work when operating in a sufficiently
constrained domain. DKR, however, wants to take on the universe, and
everything. Not particularly constrained, IMHO.
Humans talk in qualitative terms for normal conversation. We will use
qualitative descriptions of probabilistic issues (often, seldom, etc), we
use qualitative descriptions of fuzzy issues (tall, fat, ...), and we use
crisp terms to describe other things (absolutely, never, ...). IMHO, the
game is to invent a KR scheme that lets us partially automate handling of
all kinds of representations. Zadeh has recently (at KR 2000) proposed
something he calls "precisified natural language." (PNL) This appears to be
a highly constrained natural language. Telling jokes with PNL would not be
easy, but describing the evolution of an hiv infection would.
Daphne Koller at Stanford has developed what appears to be a seamless
integration of bayesian and description logics. One of her students, now at
Harvard, is adding linguistics to that mix. Maybe, just maybe, they are on
to something we need to understand better.
I think that I am saying that the DKR warrants a deeper look at KR than is
suggested by appeals to categories, relations, and logic. We are trying to
represent things which are complex. Newtonian mechanics and reductionist
thinking will not get us there. Indeed, there may not exist an atomic
structure capable of supporting our dream.
------------------------------------------------------------------------
Now the best and coolest websites come right to you based on your
unique interests. eTour.com is surfing without searching.
And, it's FREE!
http://click.egroups.com/1/3013/3/_/444287/_/956851629/
------------------------------------------------------------------------
Community email addresses:
Post message: unrev-II@onelist.com
Subscribe: unrev-II-subscribe@onelist.com
Unsubscribe: unrev-II-unsubscribe@onelist.com
List owner: unrev-II-owner@onelist.com
Shortcut URL to this page:
http://www.onelist.com/community/unrev-II
This archive was generated by hypermail 2b29 : Thu Apr 27 2000 - 09:14:52 PDT