00100	1. There seems to be an implicit idea that each word has a finite number
00200	of senses that could be listed and that if the correct sense of each
00300	word were listed, the sentence would be understood or at least it
00400	could then be automatically translated.  I would like to know where
00500	you stand on this point.
00600	
00700	2. You are unfair to the Air Force MT program.  Please find out
00800	(both for yourself and for me) what the current state of the
00900	Wright Field translation system is.  What it does and how much
01000	translation it does.  Is it still being improved or has it reached
01100	a static state?  Is the program portable, i.e. could we get it.
01200	
01300	3. I think you should reserve opinion on the correctness of this
01400	official epitaph.
01500	
01600	4. My own position on MT is rather different from the "AI position"
01700	that you give.  First of all, one should separate the question of
01800	learning a language as children to from the question of already
01900	knowing it.  In my opinion, much of the motivation for having
02000	machines learn a language inductively comes from the difficulty
02100	in understanding what it is that the language knower knows.  If
02200	we understood this better even for rather simple domains, it would
02300	be a matter of convenience whether we put our own knowledge of
02400	language in by hand or have the machine learn it.  I think it will
02500	be easier to put the information in than to simulate the learning
02600	environment of a human from birth through school and college including
02700	formal education and interaction with family and peers.
02800	
02900		Secondly, my position is that the knowledge we cannot yet
03000	express to the machine is, first of all, knowledge of the world
03100	rather than knowledge of language per se.  This is why I like
03200	Winograd's thesis; it imbeds its knowledge of language in a
03300	knowledge of the world sufficient to successfully perform its
03400	tasks.
03500	
03600	5. You exaggerate the success of ELIZA.  In the example you give,
03700	non of the gross inadequacy of ELIZA is exhibited.  Turing's test
03800	really should specify that the questioner who is attempting to
03900	detect the machine should not be naive, but should be generally
04000	familiar with the current state of AI if not with the idiosyncrasies
04100	of the particular program.
04200	
04300	6. You say there is no complete solution to resolving the ambiguity
04400	of "the bar is closed, because the barman is sick" and cite the
04500	example where the bar is across the barman's driveway.  This may
04600	be true if we look for the meaning of a word in a text, but I
04700	would argue that words and sentences are uttered in situations and
04800	not in contexts.  You could construct many examples where knowing
04900	the context of a sentence does not resolve ambigutity, but where
05000	knowing the physical situation does.
05100	
05200	7. The meaning of "psoriasis" cannot simply be "a chronic disease of
05300	the skin", because there are other chronic diseases of the skin.
05400	Dictionaries do not always give the complete meanings of words.
05500	To give the complete meaning of "psoriasis", it would be
05600	necessary to describe the symptoms, the microbe that causes it,
05700	and what it does to the tissues.  Even then, the meaning would be
05800	relative to a certain public state of knowledge.  Improved knowledge
05900	might lead to the disuse of the word as reflecting a confusion of
06000	several conditions.  (To assign meanings to words, one must have
06100	not only metaphysics but also some epistemology).
06200	
06300	8. Whenever someone quotes Godel's theorem for philosophical purposes,
06400	I get suspicious, and in your case, the suspicion seems justified.
06500	That truth of propositions cannot be decided in general by proof is
06600	obvious, and Godel didn't have to prove it.  What Godel prove is that
06700	even in certain cases where the truth of all propositions is 
06800	determined by the meaning of the terms, there does not exist a
06900	formal system with a recursivel enumerable set of axioms and
07000	computable rules of inference in which all true propositions are
07100	provable.  It is not clear that you make any use of this result.
07200	In fact, it seems that you don't understand the result, because
07300	you use the word "analytic" to mean "decideable" whereas Godel's
07400	theorem essentially asserts that for any formal system of arithmetic,
07500	there are analytic sentences undecideable in that formalism.
07600	
07700	9. In the sense in which the terms are used in logic, there
07800	are two versions of Godel's theorem, a semantic version and
07900	a purely syntactic version.
08000	
08100	10. Oh dear, you really have to learn more mathematical logic.
08200	Why do you say "survey", when the logicians have shown how
08300	to define the truth in certain theories, especially those
08400	expressed in first or higher order predicate calculus where
08500	there exist well known notions of interpretations of sets
08600	of predicate and function symbols in terms of which truth is
08700	defined.  The truth of formulas of arithmetic is defined with
08800	respect to a "standard" interpretation of the predicate and
08900	function symbols of arithmetic.  The provability of classes
09000	of true sentences is then a definite if not always decideable
09100	question.
09200	
09300	11. It's not a square, it's a rectangle.
09400