COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	lenat@mcc.com,guha@mcc.com
C00014 ENDMK
C⊗;
lenat@mcc.com,guha@mcc.com
comments on Mid-Term draft
1990 Mar 26
I have notes on the paper to the end, but I have an appointment
now, so I'll type in the rest later.  These only go to p. 27,
and I may have some overall comments.  I think it's a good paper.

%cyc[w90,jmc]		Comments on Cyc: A Mid-Term Report by Guha and Lenat

Notes for comments:

p1
contrast context and microtheory

What are the disadvantages of identifying an object and the event
of its existence?
	a. predicates that apply to one and not the other.

I took actor to mean performer.  Didn't learn otherwise till p. 27.
On p. 28 we have agent.  What is its relation to performer.

p3
I suggest a one sentence explanation of the phrase ``ontological
category''.  Many in AI are vague about what ontology means.

There should be a few more examples of pre-scientific knowledge here.

re: machine learning.  From what experience would the machine learn?

p4
	omit qualifier `more healthy'.  It looks like unsupported
bragging.

I think the formalisms used by philosophers are still too weak for
common sense.

p5
	A useful domain to test Cyc's ability to understand are the
short biographies in the Micropedia.

p6
How about some positive remarks about frames?  What is it that
frames do well?  My present opinion is that much information can
be handled with frames.  They just need to be used in a framework
of more powerful expressions.

p7
I'm pleased by the abandonment of numerical certainty factors and
the reasoning given for it.

The phrase "inferential capabilities" is somewhat vague.  It should
be explicit that it isn't mere verifying conjectures but includes
finding x and y such that P(x,y).

p8
It might be worthwhile to use ZF and not merely first order logic.
In particular the ability to form the set of entities having a 
property is useful.

I don't know that clean and simple semantics and computational
efficiency are fundamentally at odds.

p9
Whether there is a penalty for interacting at EL depends on what the
user says.

p10
Say problem solver and not just theorem prover.

It seems to me that you aren't using "valid arguments" in accordance
with the usage of mathematical logic.

Assuming all abs false might produce many contradictions.

p11
A little detail about how arguments are formed and identified
would be nice.  Actually you might get by with a promise here referring to
section 4.8.1.

You might mention Tarski in connection with leaving out
True('p) iff p.

Is the quantifier inside or outside the scope of  ist  in

ist((not supported (x) implies falls(x)), NTP)?

What about balloons?

p12
In
sigma x KB → KB
it would be more standard mathematically to use a symbol for the set
of sentences, e.g. capital sigma.

Does the KB remember enough so that an immediate unassert would
restore the previous KB?

p13
Can sigma contain free variables?
If so we get an analogy to Prolog that should be discussed.

→ previous five FI function, not six

I didn't understand n/40 vs. 2n at all.

p15
Here we have second order sentences.

p17
4.3.3 It looks like the control info is associated with rules not mere
assertions.

Does the KB or some definite part of it consist of ground literals?

p18
The nature of  deny  might be mentioned earlier.

p19
	Some of the occurrences of "ontology" really refer to Cyc's EL
language.

Section 5.1 should be rewritten to describe first the kinds of things
there are before getting into detailed discussions and motivations.

p20
	[11] is a whole book, so the reference should be to a page.

"folds into itself" isn't clear.

The last two paragraphs are rather futuristic.

p21
	replace "english" by "linguistics and philosophy".

should be.
"So, rather surprisingly, the two properties are extensionally equivalent"

From the middle of the page it seems argumentative rather than expository.

p22
I'm somewhat skeptical about the dividing up of events, but I have
no specific conundrums yet.
 
p23
It seems to me that Roger at the restaurant should be regarded as a
rich entity so that we don't want to mention all its components.

Here's a new idea.  Regard it not only as a rich entity but one
with indefinitely expandable boundaries.  The idea is that the name
of a concept like ``Roger at the restaurant'' is a pointer into
a memory structure.  Suppose we regard this structure as not having
explicit atoms or else regard the structure as continuing into the
property lists of the atoms and from the list structures on the
property lists into the whole of memory.  It seems to me that human
memory is like that.  Moreover, the world is like that.  How far
into the structure relevance extends is conditional.  We can make
up a scenario in which the price of tea in China is part of the
story of Roger at the restaurant.  I want to think further about
how to regard such indefinite entities.

Anyway it seems to me that the example given of the slots of Roger
at the restaurant is rather arbitrary.  Some of the slots given
might be omitted and others included depending on what actually
occurred at the restaurant.

p24
``concave'' has no obvous meaning here.

p25 
I don't think associating nominal intervals with events will
help much with the frame problem.  If something happens they end
early, and that has to be provided for.  The usual methods of
handling the frame problem correspond to taking the nominal
interval as infinite.

p26
LogCause seems to be a form of "strict implication" as proposed
by Lewis.  Mostly (p strictimp q) can be replaced by N(p → q)
but not always.  A remark about the relation of  LogCause
to modal logic would be appropriate here.

p27
I agree with the idea of not having primitive actions.

It isn't clear that the solution offered to concurrent actions
is epistemologically adequate, i.e. that it permits using the
information that will actually be available.  I don't have
time to elaborate this point now, but I'll return to it.

p28
Is "agent" the same as "performer"?

p29
should be
One of our recent reports [5] ...

I think it's a mistake to define all attitudes in terms of three
primitive ones.  Cyc shouldn't have primitive attitudes any more
than it has primitive actions (p.27).

p30 The reasoning mentioned at the top of the page doesn't normally
go through reasoning about a desire to sleep at home.  It short-circuits
that, because it's common sense knowledge that people normally
sleep at home, e.g while the desire to do so may be a cause, we
normally don't need to refer to the desire.

Maybe "The formalisms presented" should be weakened to "The formalisms
discussed".

These considerations have LED us (typo that a spelling checker won't catch)

p31
I think we can do better with space than just a collection of
non-interacting specialized microtheories.

I don't think contexts and microtheories are the same.

p32
I'm not convinced that the distinction between Substance and
IndividualObject is unnecessary.

p35
I still dream that dream, i.e. think a substantial part of both
technical and common sense knowledge is use-neutral.

That's all I have noted in my copy of the article.  I may have
some general comments later.