COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	⊂ommon[f83,jmc]		Common sense reasoning is not natural language reasoning
C00014 ENDMK
C⊗;
⊂ommon[f83,jmc]		Common sense reasoning is not natural language reasoning

Points made at a discussion with Patrick Suppes and Joseph Almog

1. There is no such thing as natural language reasoning.  When one
person accepts another person's natural language argument (except
in a narrow purely logical or mathematical domain), the reasoning
was not based on "natural language rules of inference".  Almost
always there were unmentioned premisses expressing facts about
the common sense world.  These facts are not represented internally
in natural language, and so the reasoning process did not simply
apply rules to natural language premisses to get natural language
conclusions.

2. In order to build robots or other intelligent programs, we need
to determine what these facts are and how we can conveniently
represent them.  This is a different problem than that of how humans
represent them internally, although the two problems interact.

3. Languages of mathematical logic are still good candidates for
representing facts about the common sense world provided we are
willing to use ontologically rich languages.  For example, we may
want to use propositions and individual concepts as objects, and
we may want to use all the set theory of ZF when this is convenient.

4. Formalizing common sense physics and psychology is essentially
different from the formalizations of physics that have taken place,
because the non-monotonicity is unavoidable.  In the axiomatizations
of physics, the non-monotonicity is all pushed out into the English
surrounding the logical and mathematical formulas.  For example, the
fact that  s = 0.5 g t↑2 only applies if air resistance
and other forces present are neglibible and that we only take into account those
forces we know about is pushed out into the English.  This trick
won't work for common sense reasoning, because neat complete domains
cover too little of common sense reasoning.

	I am optimistic that my circumscription formalism will handle
a large part of the non-monotonicity required for
axiomatizing common sense, but there may be other problems I don't
know about besides non-monotonicity.

5. The emphasis on natural language is keeping philosophers and
cognologists (informaticians) (AI people) away from formalizing
common sense facts.  Thus Perry and Barwise worry about what
verbs permit formulations like "I saw him go" but ignore the
problem of expressing the facts about seeing such as those
about occlusion, transparency and lighting requirements.  Barwise
told me that everything he had said in his 1979 seminar about
vision applied equally well to hearing.  This was disappointing,
because a common sense database would require distinct facts
about the different methods of observation.

Relevant to Joseph Almog's 1983 October 3 lecture:

6. Almog discusses Fregeans and neo-Fregeans.  I find myself to be
a micro-Fregean in the following sense.  A simple-minded theory of
sense and denotation has been shown by various philosophers to be
inadequate to represent all the phenomena of ordinary language.
They then argue about whether Frege's ideas can be patched up and
if so how to do it.  From the AI point of view the original ideas
are attractive in that they point to a formalism that may be useful
for doing more than present AI systems do with logic.  For example,
they may permit a reasonable formalization of an interesting set
of reasoning about knowledge, belief, intention, promising, wanting,
etc.  An appropriate AI goal is to build as much of a usable formalism
as possible, accepting the fact that it may be incomplete and rejoicing
in the fact that we can now do more than previously.  For example, we
might proceed as follows:

Start with a first order theory  T0  constituting presuppositions of
the group by whom these meanings are to be jointly understood.  Include
also some of the metatheory including at least  denotes('John,John).
Much more will be wanted.

The beliefs of the members of the community are represented by extensions
T1,  T2,  etc. of  T0.  It may even be helpful to consider "members
of the community" within a single person in order to represent points
of view he is entertaining.

	We then say that two terms  t1  and  t2  have the same sene if
T0 infers t1 = t2.

	Actually  T0  is a circumscriptive theory, and this explains why
many apparent identifications of sense vanish when one entertains
new possibilities.  This remark is very important!!  It can explain
many puzzles about meaning and why when philosophers consider additional
possibilities they are forced to split hairs far beyond what people ordinarily
think of and beyond previous philosophical theories.

7. While Frege's idea that anaphoric sentences are replacable by
sentences with definite descriptions may be false in general, there
certainly is a large and useful part of natural language in which
this substitution may be made.  Presumably this corresponds to
situations in which there is no problem about definite descriptions
made by one person being unknown to another.

1983 December 22

	Here are some examples of apparent natural language arguments.
They include premisses, reasoning steps and and conclusions that one
person presents in natural language, spoken or written, and to which we
presume the other person assents.  Nevertheless, they aren't arguments,
because they depend on suppressed premisses that are not expressed in
natural language by either party to the dialog.

1. "If you don't make those forms stronger, they'll burst when you
pour concrete into them."

2. "If you don't invite Mary, she'll be angry unless you explain
to her that the dinner is a working one."

	If the suppressed premisses were universally accepted by English
speakers, we could consider them as part of the rules of inference of the
language.  However, in the above cases, some of the suppressed premisses
are facts about the particular situation, shared by the participants to
the dialog but not put in words.  In the second example, we have the
participants shared opinion of how Mary reacts to hearing about dinners to
which she is not invited and to explanations thereof.  I suppose that
sharper examples could be found, but these ought to make the point.

	From the AI point of view, it is important to understand why
these facts are not expressed in natural language and to figure out
how to express them in the memory of a computer and to use them in
connection with facts that are so expressed.  From a biological point
of view, we must remember that language evolved quite late, and that
animals could remember and use many kinds of facts about the world
long before our ancestors evolved language.  Therefore, we must
expect that the "lower layers" of intellect still exist and play
the same role they played pre-linguistically.  Moreover, our linguistic
access to them is limited and varied.  We can expect them to be
just as varied in content and accessibility as the mechanisms for
pain discussed in Daniel Dennett's "Can a computer feel pain?".