perm filename COUGH[DOC,TOB] blob
sn#050510 filedate 1973-06-18 generic text, type T, neo UTF8
00100 600 YORICK 15 NOV FORMAT SUGGESTION
00200 1400 YORICK 16 NOV QUESTION BRUCE RE HUMAN PROBLEM PROTOCOLS
00300 2300 BRUCE 17 NOV ADVERTISES MINSKY AND PAPERT
00400 3300 BRUCE 18 NOV FORMAT SUGGESTION
00500 3800 YORICK 24 NOV LINGUISTIC AND WORLD MANIP DIFFERENT
00600 6900 BRUCE 26 NOV CHIDES WEYHRAUCH AND ANSWERS WILKS
00700 11200 BRUCE 27 NOV FORMAT SUGGESTION
00800 11900 YORICK 27 NOV CLARIFIES LINGUISTICS POINT TO BRUCE
00900 12900 RICHARD 24 NOV QUESTIONS ALL ON REASONING
01000 14700 BRUCE 27 NOV REL BETWEEN AI AND TRAD DISCIPLINES
01100 17300 RICHARD 29 NOV ANSWERS BRUCE'S QUESTIONS
01200 19600 BRUCE 30 NOV FORMAT SUGGESTION
00100
00200
00300
00400
00500 YORICK 15 NOV FORMAT SUGGESTION Comments take form of a line with
00600 name and date (so we can refer back to each others
00700 points),followed by a justified para or paras of statements or
00800 questions to others.Perhaps contributions should also be
00900 separated by at least two lines ,so as to distinguish them easily
01000 at sight from long single contributions .
01100
01200
01300 YORICK 16 NOV QUESTION BRUCE RE HUMAN PROBLEM PROTOCOLS Bruce:
01400 your positive statement about a nonlogical way of going about
01500 solving problems(tho one that might subsequently "turn out to
01600 have " a logic),how does such a proposal really differ from the
01700 earliest AI,which I take to be Newell ,Shaw and Simons proposals
01800 for writing down human protocols for solving problems and
01900 programming them ?[I dont imply that it would be bad if that WAS
02000 what you meant].
02100
02200
02300 BRUCE 17 NOV 1972 ADVERTISES MINKSY AND PAPERT All: Have you read
02400 MIT AI memo #252 "Research at the Laboratory in vision, language
02500 and other problems of intelligence" by Minsky and Papert. It
02600 contains lots of nice ideas, though at a
02700 not-concrete-enough-to-be-implemented level. But (as some
02800 knowledge of philosophy would/should) they do constrain the kinds
02900 of programs one should write. I have a copy, and Russ Taylor has
03000 a couple of spare copies.
03100
03200
03300 BRUCE 18 NOV 1972 FORMAT SUGGESTION All: Don't forget that if you
03400 do an ED COFFEE[1,YAW] the edited copy will be on your own disk
03500 area, not on Yoricks. Use ALIAS or COPY. I say don't forget
03600 because I did!
03700
03800 YORICK 24 NOV LINGUISTIC AND WORLD MANIPULATIONS DIFFERENT Id
03900 like to try an justify briefly what I meant when I said I thought
04000 it important to distinguish linguistic and world
04100 manipulations---and that it was gratuitous confusion to mix them
04200 up.Two cases come rapidly to mind (1) in the analysis of,say,"I
04300 dropped the bottle on the table and it broke"(Winograd) it may be
04400 plausible to say that when we ask for the antecedent of "it" what
04500 we want is the answer to the question "did the bottle break".
04600 OK,but that translation is not always so plausible:in the case of
04700 "In order to construct an object,it usually takes a series of
04800 drawings to describe it" Here,the reader sees immediately that
04900 the 2nd. "it" refers to "object" not to "series".In this case it
05000 seems to me far less plausible to transform the problem to a
05100 problem about the real world----the common sense version is that
05200 here is a LINGUISTIC problem,answering the question "what is the
05300 antecedent word for the 2nd.'it'?" (2) when translating the
05400 sentence "he killed her out of pity" into French,a speaker knows
05500 that "outof" is translated by "par" and not by one of the other 5
05600 or 6 alternative renderings, because "pity" is a feeling
05700 notion-I'm deliberately putting this in the vague terms of common
05800 sense---but that is the hard piece of info that a
05900 speaker/translator actually uses. Now,in linguistic terms it is
06000 relatively easy to represent this in terms of the constituent
06100 concepts of the concept "pity".But how does a strong-minded
06200 real-world man deal with this situation?Does he really call in
06300 some fact axiom "Pity is a felling"!?! What kind of fact about
06400 the world is that?Anyone who thinks its a fact,should ask
06500 themselves what kinds of empirical test would confirm/disconfirm
06600 it.I.E. its not a fact but a disguised statement about our use
06700 of words---so why not face up to that and represent the matter
06800 appropriately by conceptual codings?
06900
07000
07100 BRUCE 26 NOV 1972 CHIDES WEYHRAUCH AND ANSWERS WILKS BRUCE 26 NOV
07200 1972 Where are Weyhrauch's thoughts? First a reply to Yorick of
07300 16 Nov. My proposal does not differ from NSS, except that they
07400 had, and still have, the following funny notion. They think that
07500 because computers are "good" at symbol-manipulating tasks, we
07600 should study humans doing that sort of thing and try to implement
07700 that. For example see Newell and Simons new book. So they never
07800 get into the real-world stuff which I am interested in, and which
07900 I believe underlies everyting else. For example, I can
08000 understand objects only when I contain some procedures which are
08100 about lines meeting and parallelism etc, and it is the existence
08200 of these routines which makes it possible for me to do geometry.
08300 The chances of discovering these routines by trying to make a
08400 program that does geometry are much lower than the chances of
08500 discovering them by studying vision. There are no "quick hacks"
08600 for vision, but probably enough for geometry to give one the
08700 illusion of progress for some time. Though there are no "quick
08800 hacks" for anything really. Just that there are better ones for
08900 geometry. And thus more misleading. I am in fact writing
09000 protocols for a system to do vision, but of course they differ
09100 from the NSS stuff in that they are an attempt to display
09200 reasoning that is unconscious. I don't know how to program them
09300 either!
09400
09500 Now to Yorick of 24 Nov. OBVIOUSLY there are linguistic facts.
09600 The routine for "it" has to know about sentences, as well as
09700 about things more concrete. I am not convinced that the universe
09800 of facts (NOT the facts of the universe, note!) can be split into
09900 three viz. linguistic, concrete and the rest. Are you? On to
10000 par/out_of. OBVIOUSLY there have to be facts about language in
10100 any system that will use language. So the concept of pity, call
10200 it /pity/, will have to be connected to "pity", and to /feeling/.
10300 And /out_of/ will have to know about /feeling/ and "par". If
10400 everything which you cannot pick up and throw has to be handed
10500 over to the linguists because it is a 'concept` then God help us.
10600 Thank you and good night. What a mixed metaphor on l10000!
10700
10800 The name of the philosopher at the AI institute of Notre Dame is
10900 KENNETH SAYRE.
11000
11100
11200
11300
11400 BRUCE 27 NOV 1972 FORMAT SUGGESTION Another format suggestion =
11500 COMMAND. When you have done your bit, replace line 1 of the file
11600 with >>> <name> <date> <line # where your bit starts> so that one
11700 can find the latest thing and so on.
11800
11900
12000 YORICK 27 NOV CLARIFIES LINGUISTS POINT TO BRUCE Bruce of 26 Nov
12100 (part 2) misinterprets the thrust of my point to which he was
12200 replying----I wasnt trying to secure jobs for linguists God
12300 forbid,I want to see them all out of work anyway,but only to
12400 establish that there are other facts than those about the
12500 nonlinguistic world.I seem to have overkilled that point,but I
12600 did have the feeling in the Coffee House that it was being
12700 denied!
12800
12900
13000
13100 RICHARD 24 NOV 1972 QUESTIONS BRUCE ON REASONING, AND EVERYONE
13200 ELSE Bruce: I still don't feel comfortable with your answer to
13300 my original question. What kind of "reasoning" do you propose
13400 that is not related to some calculus for making deductions? The
13500 study of (or notion of) the validity of this reasoning is surely
13600 in the traditional realm of logic. Tom: I thing both Bruce and I
13700 were surprised by your opinion(?) that the solution of several
13800 individual AI type problems can add up to a general solution. I
13900 would like to explore this question as it certainly severly
14000 influences HOW we do (or should do) research. Yorick: You don't
14100 escape. I hope we can pick up on your comments about the
14200 separability of semantic analysis from other problems. All: I
14300 would be interested in knowing what results (either particular
14400 achievements or ideas) you each consider 1. Historically
14500 Important to AI and 2. Of current interest. That is sort of a
14600 survey of what has happened in thies field (AI or representation
14700 theory) that is likely to be of permanent interest.
14800
14900
15000 BRUCE 27 NOVEMBER RELATION BETWEEN AI AND TRAD DISCIPLINES All:
15100 here are some more questions, though much more vague than
15200 Richard's i.e. so vague that we can argue over the meaning of the
15300 questions, yet alone their answers. >>What is the relation
15400 between AI and the traditional studies of intelligence viz.
15500 philosophy, psychology and linguistics? In the following senses
15600 →→→→→1 Would a knowledge of X help me to make an intelligent
15700 program directly? →→→→→2 Indirectly? →→→→→3 Is it a good
15800 heuristic to ignore ∀X (gross misuse of logic there!) because it
15900 would take so long to sift through X to find something of value
16000 that in that time you could have discovered it from first
16100 principles? [ Analogy: not much point in ploughing through Roman
16200 arithmetic if you have just invented zero and the arabic
16300 notation] →→→→→4 Now the other way round. Would a working robot,
16400 that is with natural language input etc, have any effect on
16500 practitioners of X? I suspect taat philosophers would be
16600 unaffected, psychologists helped quite a bit, and linguists
16700 mostly wiped out. But →→→→→5 What is the chance that we can
16800 develop a robot that is psychologically interesting before the
16900 psychologists get into the right methodology and start telling US
17000 what to do? →→→→→6 How confident are you in "mere hacks"? For
17100 example, do you believe that there is some hack for the Mars
17200 lander that is genuinely useful to NASA but genuinely boring to
17300 AI people? Similarly for French translation, Yorick!
17400
17500
17600 RICHARD 29 NOV 1972 ANSWERS BRUCE'S QUESTIONS In answer to Bruce
17700 immediately above. →→→→→1 No!! I don't think so. There is even
17800 less evidense to believe that X has anything computationally
17900 significant to say even granting that they have clear "ideas".
18000 →→→→→2 Yes. At least in some cases they can tell you what not to
18100 do. →→→→→3 In general I think that it is important to think
18200 through unsolved problems (or those with disputed "solutions")
18300 without being greatly influenced by peoples previous attempts at
18400 the problem. Studing another persons blind ally's before you have
18500 a collection of your own is probably a waste of energy. Of course
18600 that doesn't mean you should work in ignorance of other peoples
18700 work, just that if it has not obviously suceeded then you should
18800 be skeptical of it. GOOD ideas eventually shine through. In
18900 addition with respect to X most of their work was pre any
19000 knowledge of any complex but "well-understood" and manipulatable
19100 objects like computers. Thus our whole experience is different
19200 from theirs. →→→→→4 Really??? →→→→→5 99 44/100% →→→→→6 Yes
19300 clearly there are already many industrial "robots" which are
19400 boring.
19500
19600
19700 BRUCE 30 NOVEMBER 1972 FORMAT COMMAND CHIDING YORICK Yorick and
19800 others: This file is intended to be APPEND ONLY so remarks once
19900 entered should not be removed. THIS MEANS YOU.
20000
20100
00100 100 YORICK 4 DEC REPLIES TO RRICHARD AND BRUCE
00200 10400 YORICK 5 DEC QUESTION TO TOM ABOUT SENSE OF MODEL
00300 11100 YORICK 5 DEC SUMMARY OF MACAYS (sic) ARGUMENT FREEWILL
00400 14600 YORICK 5 DEC YW'S SELF ADVERT OF ANTI-PHILOSOPHER NOTE
00500 15800 YORICK 5 DEC STATES A THEORY OF PREFERENCE SEMANT. VS TP.
00600 29300 BRUCE 7 DEC LIST OF SAYRE BOOKS
00700 31100 BRUCE 8 DEC SAYS WILKS CONFUSES FORMAL LOGIC AND REASONING
00800 32600 YORICK 8 DEC AGGRESSIVE CHALLENGE TO BRUE
00900 36800 BRUCE 8 DEC POOH-POOHS YORICK'S CHALLENGE
01000 37800 BRUCE 21 DEC QUESTIONS ON NATURAL LANGUAGE
00100 YORICK 4 DEC REPLIES TO RICHARD AND BRUCE Richard(24 Nov):Turing
00200 for starting AI(1950) with a proposal;and then Weizenbaum for
00300 showing how much easier it was to satisfy the proposal than
00400 anyone had thought.Also,and above all,Wittgenstein for making it
00500 clearer than anyone else had that our thought processes were very
00600 ill mapped by conventional logic ----I mean our problem solving
00700 thought processes, not our dreams and wishes,everyone knew about
00800 them-----this was a prerequisite for Newell,Shaw and Simon style
00900 AI----before the new orthodoxy [?] set in.
01000
01100 Richard(29 Nov)(→→→→→→3):I disagree about" most of x"-----at
01200 least in that ,in all those disciplines X, there were people
01300 making important use of notions like "clear and definite
01400 procedure" long before the first computers:the behavioral
01500 psychologists;Vienna Circle empiricist philosophers etc.
01600
01700 AI CONSTRUCRTS DO NOT ACTUALLY SOLVE PHILOSOPHICAL PROBLEMS
01800 Bruce(27 Nov →→→→6 and Richard 29 Nov →→→→→6):Isnt there a
01900 difference here :an industrial robot working by hacks may well be
02000 boring yes;but would a thinking talking one be so too?Surely
02100 not,because boringness is a function of what they do,not of the
02200 hacks involved. May I be boring here at some length,by
02300 connecting this point with Bruce (27 Nov,→→→→→4) and agreeing
02400 that philosophers would be unmoved ,and try to show why,and why
02500 they would be right, at least against a naive AI person (no names
02600 but they ∃ in the project I promise you) who thinks that
02700 traditional philosophical problems can be SOLVED by robot
02800 building.
02900
03000 AI people often use the following argument implicitly or
03100 explicitly:the problem of mind is now solved because we can
03200 create an intelligent human-like robot,and if anyone asks about
03300 its "mind" we can simply display its blueprints and program in
03400 complete and full answer to such a question.
03500
03600 First one should notice that this argument is hypothetical:there
03700 is no "can create" really now anymore than there was when men of
03800 Descartes time postulated a wholly clockwork,though
03900 convincing,homunculus.The argument now is exactly where it was in
04000 the seventeenth century(give or take a few technological bits and
04100 pieces) and its real force is "if I brought this thing in here
04200 and it was intelligent etc.etc. how could you go on talking about
04300 ts mind after Id shown you all the blue prints/programs etc.?"
04400 (20th.C.readers read "programs" :17th.C. readers,"blueprints".)
04500
04600 If the automaton displayed is to be said to have a mind,and the
04700 sense that its going to do so well that seeing the program is
04800 really going to shut the questioner up in the way required,then
04900 the performance is going to have to be a whole lot better than
05000 the old conversation down the teletype fooling us for 10 minutes
05100 with talk about the weather,the world or what it feels about the
05200 mafia.
05300
05400 What its going to have to do,too satisfy people like me anyhow,
05500 is to discuss the relation between its mental states and its
05600 program in the way that a human can ,say,discuss his mental
05700 states with the brain surgeon who is sticking a probe into his
05800 head. This for me is the crucial argument,and Ill call it the
05900 NECESSARY AUTHORITY argument (or Wilks' test!).
06000
06100 That is to say,if its operator says to it at any point
06200 "monitoring your program I see youre now thinking about x",its
06300 got to be able to say "no Im not, Im thinking about y"---in just
06400 the way that the patient can say that to the brain surgeon,and he
06500 CANNOT BE WRONG when he says it because only he (the patient) has
06600 the necessary authority to know what he's thinking.If anyone else
06700 thinks he knows better that the subject/patient he must,logically
06800 must,be kidding himself.
06900
07000 Now,if the operator says ,"oh well,you cant expect behavior like
07100 that from it-----if I say its thinkin about X,then it is,because
07200 Im the final authority on what its thinking,not it.",Then the
07300 questioner is home and dry because he can now just refuse to
07400 accept the proffered automaton as human like----on the ground
07500 that it lacks a necessary feature for being considered "
07600 mind-like".
07700
07800 The operator can now perhaps go away and comeback,or envisage the
07900 possibility of coming back rather,with an automaton having the
08000 desired property: it can talk about its "mental states" in the
08100 sense of what it thinks its doing, and what its thinking about at
08200 any instant.Let us suppose it passes whatever test we impose and
08300 the questioner accepts it as mind-like.But now,paradoxically, the
08400 operators original hopes are dashed,because he cannot now display
08500 the blue prints and program and clear up the questioners
08600 mind-talk in the way he wanted----because of the very nature of
08700 the test.
08800
08900 All this is not,iI should point out, a mere crude variant of
09000 Turings proposal ,in that I specify a certain subject for the
09100 Turing test conversation.The "Wilks test" requires simultaneous
09200 real time access to the conversation and to the program as it
09300 runs.
09400
09500 This "Wilks test" stuff is a rhetorical joke of course,but there
09600 is Im sure an important point here for the naive modellist who
09700 thinks AI is settling philosophical problems in some
09800 straightforward way(Im not denying that it may affect them in
09900 other ways) .For in some sense ,to qualify as human,a system will
10000 ,in doing so,cease to give the simple read-off explication of
10100 intelligence that some people still seem to hope for after 300
10200 years of argumeent on this point.Its very important to note that
10300 the invention of computers hasnt shifted the substance of this
10400 argument at all ---because it is all in terms of "what would you
10500 say if I produced a thing that did so and so..."The actual
10600 production of the thing doesnt affect the argument one way or the
10700 other------all of which makes me think that someone must have
10800 produced my argument before too.
10900
11000
11100 YORICK 5 DEC QUESTION TO TOM ABOUT HIS SENSE OF MODEL Tom:can one
11200 press you a bit to say more clearly what you mean by a "model of
11300 the world"---it`s clearly fundamental to your view of AI yet also
11400 doesnt correspond particularly with any of the standard senses of
11500 "model". I think discussing this could be very important.
11600
11700
11800 YORICK 5 DEC SUMMARY OF MACAYS ARGUMENT ABOUT FREEWILL AND
11900 COMPUTERS
12000
12100 In the note above (4 DEC) I`m not at all trying to say that
12200 philosophical issues are independent of empirical research ;not
12300 at all.The clearest case is the traditional free-will problem
12400 which has been much influenced by developments like computers.I
12500 mentioned in passing the other mmeeting Mackay's linking of the
12600 two (and I think he's wrong----and am prepared to say why at the
12700 usual infinite length)) and then couldnt state his main
12800 point.It's as follows :(Freedom of Action in a mechanistic
12900 universe:CUP ,Eddington Lecture,1967)
13000
13100 "Suppose all the relevant facts on the workings of your brain
13200 could be made available,without disturbing it,to a computer
13300 system ,capable of predicting its future behaviour from these
13400 facts and the environmental forces acting on your nervous
13500 system.........
13600
13700 [then,given that belief changes cause brain changes[YW]]
13800
13900 "It follows that even if the brain were as mechanical as
14000 clockwork,no completely detailed present or future description of
14100 a man's brain can be equally accurate whether the man believes it
14200 or not.(a) It may be accurate before he believes it,and then it
14300 would automatically be rendered out of date by the brain changes
14400 produced by his believing it ;or(b) it might be possible to
14500 arrange that the brain changes produced by his believing it would
14600 bring his brain into the state it describes,in which case it must
14700 be inaccurate UNLESS he believes it,so he would not be in error
14800 to disbelieve it.
14900
15000 "In either case the brain description lackes the 'take it or
15100 leave it' character of scientific descriptions of the rest of the
15200 physical world,since its validity depends precisely on whether
15300 the subject takes it or leaves it"(pp.8 & 11-12)
15400
15500
15600 YORICK 5 DEC YW'S SELF ADVERTISEMENT OF ANTI-PHILOSOPHER NOTE.
15700
15800 By the way,apropos of philosophers being awful at AI questions
15900 and showing ,by writing,that they should busy themselves with
16000 something else,theres a terrible article in the current MIND
16100 called "How to tell your friends from your machines" that argues
16200 that people who use the phrase "artificial intelligence" as if
16300 they really believe in the existence or possibility of what it
16400 signifies ,are loony and should be locked up. Theres also a reply
16500 on MIND.NOT[1,YAW]---but I wouldnt want to drag you all though
16600 that.
16700
16800
16900 YORICK 5 DEC STATES A THEORY OF PREFERENCE SEMANTICS AGAINST TP
17000 At the risk of being drummed out of the group for obsessional
17100 behaviour,may I hark back to the "fell and broke example" (Yorick
17200 24 NOV).I wasnt trying so much to make a distinction (between
17300 linguistic and non-linguistic "facts") as to emphasise the
17400 importance of the former sort,within the unity of facts:an
17500 importance I think that theorem proving-oriented people tend to
17600 play down------though of course theorem proving techniques can be
17700 used equally well to manipulate both within a single
17800 symbolism,Winograd to wit.
17900
18000 May I then hark back to that example and try to make the point I
18100 now see I was trying to make,by contrasting two kinds of
18200 technique applied to that example.I may be making howlers about
18300 strategies at the end of what follows,but if I am one of you can
18400 correct me.
18500
18600 I want to contrast theorem proving(TP) and "Dictionary
18700 semantics"(DS) approaches to Winograd's "fell and broke " case
18800 and to argue that the two approaches are different (TP people
18900 agree with this ,since they dont like what they conceive DS to
19000 be),and that ,though one of the motivations of the example was to
19100 show the need for TP,and hence the inadequacy of DS type
19200 approaches,the reality is different,and that what I call
19300 "preference semantics" (my form of DS) may even be superior in
19400 some overall sense.
19500
19600 The example,to refresh memories ,is in outline as
19700 follows.Consider "I dropped the bottle on the table and broke
19800 it".We do not know ,goes the argument,the antecedent of "it" at
19900 this point.Moreover the following two sentences should cause use
20000 to decide the question differently :"Bring me the hammer"(i.e.
20100 its the table) ;"Bring me the broom"(i.e. its the bottle).In
20200 either case we need facts about the real world expressed in axiom
20300 type form,plus as TP strategy of some sort for manipulating them
20400 in order to settle the matter.
20500
20600 So much for the TP point of view.(I may have misrepresented
20700 Winograd's particular version,but that doesnt matter too much for
20800 the moment) I disagree in a number of ways:(i)I do not think that
20900 the antecedent of "it" is ambiguous at the end of the first
21000 sentence.I think it is clear at that point that the bottle
21100 broke,and that anyone who meant otherwise would have said "I
21200 dropped the bottle on the table and the table broke".
21300 (ii)moreover this is settled by "preference semantics" in that in
21400 the information attaching to "break" we find that it prefers
21500 subjects marked as RIGID and FRAGILE in some appropriate
21600 way.Preference here merely means that if it cant find such a
21700 subject it'll take whatever its given ,just as people do (just as
21800 humans usually do all the talking,but if someone says "the
21900 machine told my weight" we accept it ,rather than rejecting
22000 it,whereas if one says "the chair adjourned the meeting" a human
22100 subject will be preferred AND CAN BE FOUND ,given an adequate
22200 dictionary entry for "chair")
22300
22400 Now I agree we may have been told some long preliminary story
22500 about someone intending to break the fragile little dolls tables
22600 in Marys dolls house---thats another matter. But I think it is
22700 absurd and untrue to reality to say sentence #1 is ambiguous in
22800 any sense. Now if we get sentence #2 about a hammer,this perhaps
22900 should cause us to unpick the decision for "it=bottle".But that
23000 fact doesnt argue for TP ,for any method of analysis would
23100 require to know why that should be requestioned in the face of
23200 new information ----after all if THAT decision is up for question
23300 so is absolutely ANY decison in language processing (= hopeless
23400 nightmare).
23500
23600 But if the decision is up for question then again DS can cope
23700 perfectly well given that in the semantic information for
23800 "hammer" we have a preference for repairng or making ,and for
23900 acting on RIGID,NONFRAGILE THINGS or SUBSTANCES. In the case of
24000 instruments we normally include information about their preferred
24100 use (though people can of course use hammers to kill with). Given
24200 this preference,and given the cue to unpick,then the second
24300 sentence with hammer will indeed lead us back to it=table,since
24400 only it is RIGID AND NONFRAGILE.With "broom" in second sentence
24500 there is no new preference to overthrow "it = bottle".
24600
24700 Note here first that it is easy to specify preferences,though I
24800 would not know how to begin to specify the appropriate axioms and
24900 boundary conditions for a TP approach to the example.
25000
25100 Note also,and more importantly, that the information content in
25200 the two approaches is different. The rule "Prefer a
25300 nonfragile,rigid object of repair" is cleary rarely misleading.I
25400 would claim that it can never be wrong----only that an example
25500 maker or careless speaker can mislead a hearer,so as make
25600 possible a correct choice with the preference rule ,only to have
25700 to see it contradicted later.That is the case in this example I
25800 believe and it is indeed a very difficult situation to diagnose.
25900
26000 But what is the TP alternative:it seems to require such "axioms"
26100 as "fragile objects are never repaired with hammers",or in more
26200 canonical form "Only non-fragile things are repaired by hammers".
26300 But this is just not a true statement about the world----- cast
26400 iron is fragile but can be carefully beaten with a hammer:many
26500 other cases could be thought up.So then if TP uses normal
26600 categorical axioms then they will frequently be inductively
26700 false----in a way that preference semantics will not ,for it does
26800 not commit itself so far.For example if in the present case it
26900 was a cast iron bottle dropped---then preference semantics and TP
27000 would chose the table and neither would be WRONG for the choice
27100 would be arbitrary.But TP is now left with an inductively
27200 unreliable axiom in its working inventory of axioms that will
27300 mislead it when it comes to analyse "He mended the iron bottle
27400 with the hammer and then put it away in the closet". This case
27500 will give preference semantics no trouble ,but ought to ring some
27600 kind of "contradiction bell" in the TP system.
27700
27800 A few additional observations on the example:I do not think that
27900 the preference semantics,as I have described it,is in any sense
28000 probabilistic.Yet,to achieve the same force in the TP view of
28100 things probabilistic "axioms" would be necessary such as "Usually
28200 hammers repair only RIGID NON-FRAGILE THINGS".Yet,surely, no one
28300 has the least idea how to handle axioms in such a form
28400 computationally.A form like "Usually etc." has the virtue of
28500 being true in this case,but no one would have any idea how to
28600 take advantage of that truth.
28700
28800 Again,even in conventiional form,is there much idea how to handle
28900 such bodies of axioms ,outside the
29000 complete,sledgehammer-to-crack-a-nut ,strategies like
29100 resolution?Is there not something a trifle odd in the present
29200 situation where TP men like to dismiss DS men as
29300 "heuristic"(opprobiously that is) when any translation of a
29400 "hammer axiom" into PC would be heuristic(in the sense of
29500 requiring an arbitrary choice between the many ways of renedering
29600 it formally); and then any subsequent use of the formal version
29700 of the hammer axiom would also be heuristic in that there are no
29800 obvious strategies for using it together with the existing body
29900 of axioms.Moreover,it will also be heuristic,as I have argued,in
30000 the sense of being quite likely inductively false. In the case
30100 of the preference semantics ,however,there is no strategic
30200 trouble in this sort of example,in that there is no call at any
30300 point to search any body of semi-axioms.The information required
30400 is available at the point of its application .
30500
30600 Im not trying to say with all this that there arent examples
30700 where general inference rules will be needed over and above what
30800 can be attached to the dictionary entries of words.Only that DS
30900 goes a lot further than some people think.
31000
31100
31200 BRUCE 7 DEC 1972 LIST OF SAYRE BOOKS Wish you guys would keep
31300 your stuff justified. The books by SAYRE are the following
31400 Sayre, Kenneth M. 1928- Recognition: a study in the philosophy
31500 of AI (1965)
31600 Q335 S38 Consciousness: a philosophic study of minds and
31700 machines (1969)
31800 Q335 S37 (in CS too) Plato's analytic method (1969)
31900 B398A6 S28 ~~~with Crosson F. J. (eds) The modelling of
32000 mind: computers and intelligence (1963)
32100 Q335.5 S2 (in phil,cs and eng too) Philosophy and
32200 cybernetics (1969)
32300 Q315 P5 Where are Binford's thoughts? (This worked last
32400 time!)
32500
32600
32700 BRUCE 8 DEC 1972 SAYS YORICK CONFUSES FORMAL LOGIC AND REASONING
32800 It is quite clear from the above pieces by Yorick that he does
32900 not understand the "world model" approach which Tom and I believe
33000 in, nor how sophisticated the (for want of a better phrase)
33100 Winograd type stuff can become. It is a BIG mistake to confuse
33200 this withe theorem-proving which is very inflexible and indeed
33300 prone to the errors he mentions. I will expand on this later -
33400 though most of it came out at the meeting on 6 Dec. In
33500 particular I think that Charniak's thesis would be of use in
33600 showing Yorick that his ideas on what can be done with "usually
33700 hammers..." are NAIVE, and that indeed it is his dictionary
33800 semantics which are inflexible, as he really admits when ......
33900 and so on.
34000
34100
34200 YORICK 8 DEC AGRESSIVE CHALLENGE TO BRUCE Bruce:look,its just no
34300 good calling someone naive,when theyve produced a detailed
34400 argument which you havent attempted to answer in detail. Dont you
34500 see that it's the detail of the argument that counts-----if what
34600 I say in the long hammer argument is wrong or naive at certain
34700 points---where are they?Why is it confusinhg to assimilate the
34800 Winograd and TP approaches for the purpposes of the points I want
34900 to make?. The onus is on you---you dont seem to realise
35000 that.Shrugging your shoulders:acting as if its not worth the
35100 bother to point out opponents errors:acting as if the received
35200 wisom of AI shows I must be wrong-----none of these will do.And
35300 pretending they will is one reason why the intellectual standard
35400 within AI is so abysmally low.
35500
35600
35700 Its true I dont understand your world model approach---but thats
35800 hardly surprising since neither you ,nor anyone else to my
35900 knowledge ,has made any attempt to state it in a succinct way.
36000 Its clear what it ISNT:it isnt formally hard like the set theory
36100 approach;nor does it take the problem of understanding natural
36200 language seriously ,as I would claim I do,as well as some others
36300 (not Winograd though). By that I mean that your approach centres
36400 on the formalism-programming language (PLANNER or PC),and doesnt
36500 really ever consider what it would be like to go from language as
36600 she is writ to that. But what IS it?It clearly centres on facts
36700 and their interrelation ,but what are these relations to
36800 be---apart from some blanket reply like " anything we find we
36900 need and can get into PLANNER?"If you accept that reply ,can YOU
37000 still call ME naive?
37100
37200 Nor do you have any monopoly on world model--for every approach
37300 has something that could be called that --but what exactly
37400 distinguishes yours and gets over certain well known
37500 difficulties?To start with what about the example I give-----how
37600 does your system in that case pullin the relevant "axioms" from
37700 its presumably enormous body of them?Look up "hammer" and bring
37800 in everything with that word in ?(youd be joking) How do you
37900 avoid retaining a false premise for future use IF you have solved
38000 the case?What strategy do you have to handle those ahmmer axioms
38100 to compete with my simple working one?In what cases would mine go
38200 wrong ,where yours wouldnt?Why dont you answer line by boring old
38300 line?
38400
38500
38600 BRUCE 8 DEC POOH-POOHS YORICK'S CHALLENGE My short thing above
38700 was intended as a statement of my view, not as an explanation of
38800 them. That I said I would do below, and I will. But I do think
38900 we are getting to the heart of our mutual misunderstandings and
39000 will soon be able to separate differences in terminology from
39100 those in belief. [All beliefs are stated in God's Great Assembly
39200 Language, but we speak in the Cobols and Fortrans of .....]
39300
39400
39500 BRUCE 21 Dec QUESTIONS ON NATURAL LANGUAGE Well this file is
39600 getting unmanageably long and we should give some thought to
39700 somehow making it a file of summaries and pointers to some
39800 place(s) where more details are stored. All but Tom and Arthur
39900 attended at least some part of a dreadful conference on
40000 set-theoretic semantics for natural language organized by John
40100 Wayne, sorry I mean Patrick Suppes last week There seemed to be
40200 no real communication going on between the different groups
40300 there, and the following thoughts were generated during the most
40400 boring parts. →→Yorick. I would be interested to see how your
40500 system splits up into knowledge of French, English and (!)World.
40600 For example, how easy would it be to turn it into a
40700 French-to-English translator? Into a question-answering system?
40800 →→The World. What is the problem of understanding natural
40900 language, or of finding a semantics for it. What would a
41000 scientific theory of language look like? Is there some
41100 associated concrete goal? Could a program that "spoke" a
41200 (natural) language be this goal. Could such a program fail to
41300 be/embody a theory of language? Are there linguistic problems
41400 that it would not address? To what extent would linguists be
41500 wiped out by such a program - would they all turn into
41600 philosophers?
41700
41800
00100 100 YORICK 8 JAN ANSWERS BRUCE 21 DEC ON NAT LANG
00200 MORE SPECIFIC ANSWERS TO BRUCE QUESTIONS
00300 SCIENTIFIC LINGUISTICS
00400 13900 BRUCE 9 JAN HAS WILKS SEEN THE LIGHT?
00500 14900 YORICK 11 JAN SAW LIGHT YEARS AGO
00600 17000 ARTHUR 24 JAN FIRST THOUGHTS
00700 22300 YORICK 25 JAN ARTHUR AND LUDWIG
00800 26100 ARTHUR 31 JAN FORMAT SUGGESTION
00100 YORICK 8 JAN ANSWER BRUCE 21 DEC ON NATURAL LANGUAGE Your
00200 question helped me see how much the apparent dispute between us
00300 is misunderstanding.Specific answer to the question to me:one can
00400 point to aspects of my system ,but there is and can be no hard
00500 language -world distinction ,since of course facts about language
00600 are in the end facts about the the real world (ex:"Spanish
00700 startts questions with an inverted ?")
00800
00900 Aspects are best pointed up by looking at files.A dictionary
01000 entry for a sense of "drink" ,for example,ougth to be unpackable
01100 so as to express the facts that drinking is an act of putting
01200 fluids into the mouth by the owner of the mouth,or some such.That
01300 surely is not in dispute. So,what is in dispute?So far,only the
01400 format of the information.Clearly, Sandewall or someone like
01500 him,would understake to express what I just said in his PC
01600 format.Thats fine,no problem ,I admire his faith that doing so
01700 will enable him to plug it into some prooof system to make use of
01800 it an justify all that work, but no dispute in principle,about
01900 the expression of the meanings of the sense of words that is
02000 (except for well known difficulties about mass and time and so
02100 on)
02200
02300 Again wheni disputed with Tom about the meaning a complex concept
02400 like "the United Nations"---the gist of his answer was that it
02500 was essentially a corpus of FACTS ABOUT the UN.Here I have more
02600 of a grumble,and it connects with disputes weve have at the
02700 coffee shop about the expression of science within the expression
02800 of meaning.There is it seems to me a distinction that common
02900 sense makes between explaining the meaning of a word and giving
03000 facts about a thing.It is importantt and any intelligent system
03100 will have to have access to the diistinction.The fact that its
03200 not easy to express in the "meaning is facts" view of life is an
03300 arguemnt against the adequacy of that view.Take the case of
03400 "water":I suspect Tom thinks that part of he meaning of that word
03500 is the fact of physics that it freezes to ice at 32%F.But that is
03600 not so ----demonstartion:there are thousands of speakers of say
03700 Swahili in the Congo unaware of the existence of ice.Can we say
03800 without absurdity that THEY DO NOT KNOW THE MEANING OF THE WORD
03900 "Maji".Clearly not,they are simply ignorant of some fact of
04000 physics about water.(end of demo)
04100
04200
04300 So,to return belatedly to Bruces question,of coourse the formula
04400 for "drink" can be interpreted to yield the fact that fluids are
04500 what is drunk ,or that drinking concerns a body part.And if and
04600 when my system is plugged to a question answerer as I intend, and
04700 any one is silly enough to ask suchh questions then those are the
04800 answers he should get from the formulas. (Formulas interpreted as
04900 FACTS ,that is)Answer to "Where is my book" should clearly come
05000 from somewhere else,and why not?The basis common sense
05100 distinction I am appealing to(explanations of meaning vs.facts
05200 about things) cannot be simplly laughed off.
05300
05400 To cycle back a little:the inadeqaucy of the work on natural
05500 language by Ai people lies else where(and independently of the
05600 fact that Sandewalls PC could express much of the meaning of word
05700 senses that is----though I doubt if Winograds could)
05800
05900 The inadequacies are here:PC coding renders the meanings of
06000 messages in an unnatural way. "John struck Mary" contains
06100 amessage ,and it is badly represented by any predicate reducible
06200 to "struck" along with a couple of arguments---because the
06300 message isnt one about an act of striking(nor is it about John
06400 ,and so is not we;ll represented by the predicate "John" and a
06500 couple of argeuments either). I try to do this with template
06600 message forms matched on as semantic patterns.
06700
06800 Other inadequacies:all such work,Sandewall,Winograd et al,has no
06900 theory of languuage in two other senses.(1)It has no theory of
07000 content or information---in thee sense of preferring
07100 representations connected in certain ways to others(as do my
07200 representations with the aid of their principle of
07300 choice)---anything that fits the parsing rules is ok for them,and
07400 thats crazy.They could never in principle hope to cope with
07500 evereyday realities like ambiguity and metaphor.
07600
07700 (2)It is hard to see how,even in principle,such logic oriented
07800 systems could cope with real sentences like any one on this
07900 page.There is no possibility of it---- no glimmer of light how to
08000 parse its complex information into the primitive structures they
08100 use. Basically,and in spite of Winograds use of a Halliday
08200 linguistic grammar,they both require that the input comein
08300 looking like the format its going into. And,as Wittgennstein
08400 said of a simple referemntial language he' made up,this IS a
08500 language ,but one quite different from ours.We should at least
08600 acknowledge the existence of ours in AI I think.
08700
08800 MORE SPECIFIC REPLY TO BRUCE QUESTIONS I think that AI ought to
08900 drive linguists out of work:when a proper Ai paradigm for
09000 language has been established they ought to see this.Why AI is
09100 ideally fitted for this is that, because it can deal with systems
09200 that can be wrong and can reexamine their inferences ,it can deal
09300 with the inferences required to understand natural language.
09400 Linguists are still locked inside the metamathematical paradigm
09500 they inherited via Chomsky:that of a simple derivational proof
09600 system that simply generates strings from bodies of rules:no
09700 possibility of being wrong,and absurd becaiuse the set of wff
09800 sentences in a real language is not well defined as the
09900 logical-linguistic paradigm requires that it should be.
10000
10100 Doing this wont destroy ALL linguits of course,only Chomsky and
10200 after.Linguistics has always had perfectly proper and traditional
10300 comparative and classificatory concerns.
10400
10500 But to do this,and create a real theory of language,rather than
10600 just a coding scheme,AI must give up a couple of bad
10700 habits:(1)the "Queen of the Sciences" thing---assuming that to
10800 deal with language you must express all the information
10900 expressible in it -----Tom believes this I think (random
11000 thought:how would you keep out the BAD theories like Phrenology
11100 and Astrology?How about Religion?)I think Ai must recognize its
11200 task as expressing 20th C.common-sense knowledge.This was one
11300 trap that ,to its credit,linguistics never fell into---it was
11400 well aware of the difference between a dictionary and an
11500 encyclopedia.
11600
11700 Secondly,it must rethink its bias infavor of superficial problem
11800 solving. Consider Winograds linguistic examples about finding the
11900 referent of "it" ----say the well known "women of Berkeley"
12000 example,right at the beginning of his thesis .This is an example
12100 where HUMAN READERS FIND IT DIFFICULT TO SEE THE RIGHT
12200 ANSWER----i.e. it is aproblem FOR THEM,unless they happen to live
12300 in Berkeley and there are many millions living outside that city
12400 who can hardly be said not to understand the English Language
12500 because of that. (Cf.The COngolese and ice)
12600
12700 What this artificiality of exampples,oh so carefully cooked
12800 up,and the silly prejudice in favor of gamesy problems ,comes to
12900 is that it doesnt help at all towards the discovery of a theory
13000 to idientify referents in real caeses of "it" :ones where a
13100 normal speaker of the language SEES NO PROBLEM AT ALL. There is a
13200 huge non sequitur here that solving the problem cases will
13300 thereby solve the real caes.Not so.
13400
13500
13600 SCIENTIFIC LINGUISTICS. Cannot exist ,and the pretensions of
13700 linguists to such should be understood by now as as resting bad
13800 arguments (see my book,or AIM157,which also,incidentally,
13900 contains nearly all the points I was trying to make about there
14000 beeing serious prexisting senses of "model" and "theory" that
14100 cannot just bee gaily ignored by AI people).
14200
14300
14400 BRUCE 9 JAN 1973 HAS WILKS SEEN THE LIGHT? At last!! When I read
14500 lines 68600-68800 of Yorick's last piece I did a wild yell
14600 "YEEEEOUAAAAIIIIIOOOOO". Here is one of the key points I have
14700 been trying to make for the last ALEPH-0 time-instances and which
14800 I wrote up with Pat Hayes in our memo BEING MADE BY YORICK.
14900 Sandewall cannot do this sort of thing, nor can any
15000 straightforward extension of Winograd's stuff do it, though it
15100 has I think been done by Charniak and is certainly a very
15200 important part of the PLANNER subculture.
15300
15400 YORICK 11 JAN SAW LIGHT YEARS AGO Glad to have given Bruce
15500 pleasure,tho I have the illusion I saw the point years ago and
15600 have been plugging it both abstractly(MIND,1971) and with a
15700 system-----in particular the earlier weaker version of my system
15800 at SDC that could make guesses about the possible sense of a word
15900 when it encountered a word sense that wouldnt fit what it
16000 expected.I put that in "linguistic mode" there,tho I suppose it
16100 can be translated,inadequately Id guess,into fact mode for those
16200 who will only accept their info that way.
16300
16400 On second thoughts maybe it cant be translated-----some word
16500 sense shifts can be so abstract as to be meaningless in the naive
16600 language of facts.For exammple,a communist begins to use
16700 "democracy" to mean not the rule of the majority by vote but rule
16800 by the decisions of the party's central committee. This is a
16900 clear shift of word sense in the common sense use of that
17000 phrase.But how can we call this the "addition of a fact to the
17100 system" in any useful sense of that phrase What sort of "fact" is
17200 "democracy is rule by the decisions of etc." Of course not ,its a
17300 definition of a word sense-----a dictionary entry if you like---
17400 masquerading as a fact.
17500
17600 ARTHUR 24 JANUARY FIRST THOUGHTS It seems to me that no_one
17700 on my side of things( I include people like Bruce, albeit kicking
17800 and screaming) is trying to deny that any significant language
17900 understanding system, be it natural or artificial, can get along
18000 without a dictionary in some sense (possibly even a
18100 stright-forward one!) The question is rather how we are to define
18200 the "primitives" of this dictionary. Words like `democracy', as
18300 Yorick rightly points out, are not themselves facts, but on the
18400 other hand, the dictionary entry for such an abstract word must
18500 surely, at some remove,refer back to `real' facts.
18600 I gather people were rather disappointed by the Suppes
18700 Conference on Set-theoretic semantics. Although I wasn't there, I
18800 can guess that the reason for this disillusionment was the same
18900 as the feeling that Yorick seems to have about 'gamesy' problems.
19000 I find myself feeling rather ambivalent here. On the one hand,
19100 the Wittgensteinian referential semanticist (pace Yorick!) in me
19200 feels that there is some value to be gained from consideration of
19300 toy problems, insofar as they force into theopen our prejudices
19400 and hidden intuitions. On the other hand, having over the last
19500 few days watched a reductio ad absurdum of the `axiomatic'
19600 position in John McCarthy's attempts to formalise the Blind Robot
19700 problem, I've come to the conclusion that we really can't carry
19800 the gamesy approach too far. As YAW would say: we really can't
19900 go looking for problems where no problems exist. Look what
20000 happened to philosophy when it fell into that trap - it's still
20100 reeling from the results! But (and it's a big but), `category
20200 mistakes' and such like aside, we in Ai are obviously not going
20300 to get anywhere if we bite off more than we can chew.
20400 I feel this last sentiment also applies to linguists'
20500 attempts to construct semantics for natural languages. The work
20600 of people like Montague, Davidson and Lewis (incidentally, I
20700 really think people should look at David Lewis' excellent article
20800 on 'General Semantics' in the Davidson/Harmon volume of
20900 "Synthese") was aimed at developing a way to deal with the very
21000 simplest kinds of meaning constructs - those that make direct
21100 reference to real-world(physical,geometric) concepts. The work
21200 of the developmental linguists (Bierwisch,the Clarks, et al.)
21300 shows conclusively that perceptual entities are the earliest
21400 linguistic primitives that a child acquires. Nobody can deny that
21500 as he matures the child uses this primitive referential semantics
21600 to construct a a more connotative system. Isn't it a bit like the
21700 way old Ludwig Wittgenstein saw things : the primitive
21800 structures, the "pictures of facts" SHOW their meaning directly,
21900 while the complex sentences constructed from them only SAY their
22000 meaning indirectly.
22100 I should like to see peoples opinions o the value of
22200 thinking about toy problems and their associated philosophy. Is
22300 that whole approach worthwhile? Should we, for the sake of making
22400 at least some progress, restrict ourselves to problems that are
22500 fairly easily expressible in PC or whatever( assuming that there
22600 are ANY!), or should we cast ourselves adrift in a sea of real
22700 world troubles, held up only by the rotten hulk of the "Planner
22800 sub-culture"? (sorry about the mixed metaphors, folks).
22900
23000 YORICK 25 JAN ARTHUR AND LUDWIG I have this irresistible urge to
23100 enter into a form of dispute unknown in AI(and possibly rightly
23200 !):what Bruce would probably call the Mr.Footnote program.In this
23300 case the "was Ludwig Wittgenstein a referential semanticist?"
23400 question [all those whove said "who cares" already,type E now].
23500
23600 I dont see how Arthur can say yes to that because he probably
23700 knows the references better than I do ,let alone the cliche
23800 version of LW doctrine (the meaning is the use) ,and they dont
23900 say referential semantics to me,at least not in any ordinary
24000 sense of that phrase. The key paragraph of Witters I cant number
24100 is (a propos of Augustines ref sem) "this describes a language
24200 system but one quite different from ours" .
24300
24400 Also theres Philosophical Investigations Part
24500 I,$$13,26,27,30,31,32,38 ,43 to give a few early ones,from which
24600 I type just a couple of key sentences to fill up my day and a bit
24700 more of this file:
24800
24900 Augustine describes the learning of human language as if the
25000 child came into a strange country and did not understand the
25100 language of the country;that is,as if the child already had a
25200 language only not this one.($$32.I.e. referential explanation is
25300 only ok if you know the meaning of the word ALREADY)
25400
25500 For a large class of cases---though not for all---in which we
25600 employ the word "meaning" it can be defined thus:the meaning of a
25700 word is its use in the language. And the meaning of a name is
25800 sometimes explained by pointing to its bearer.($$43 The last
25900 sentence isnt even a ref.sem. doctrine for the smaller class of
26000 cases, because he says the meaning is EXPLAINED by pointing to
26100 etc.He never says that IS the meaning.As I understand ref sem it
26200 says (1) the bearer is the meaning in general,and (2) hence
26300 Wittgenstein's smaller class(the one in which the bearer is at
26400 least RELEVANT to questions about meaning) is the larger class
26500 ,and is to be differently explained as well.)
26600
26700
26800 ARTHUR JANUARY 31 FORMAT SUGGESTION
26900 May I suggest that in future we paginate this file, so
27000 that we dont have to spool the whole thing for the benefit of a
27100 few pages at the end? BUT DO NOT RENUMBER THE LINES....
27200
27300
00100 100 BRUCE 1 FEB NEW FORMAT - IMPORTANT
00200 700 ARTHUR 1 FEB PLEASE-NO WITTGENSTEINIAN EXEGESIS!
00300 7000 TOM 3 FEB REPLIES TO YORICK
00400 26300 YORICK 5 FEB REPLIES TO TOM(AND SIDESWIPES ARTHUR)
00500 37300 ARTHUR 9 FEB ONCE MORE UNTO THE BREACH, CRYING "GOD!!!!"
00600 42500 YORICK 9 FEB A LAST LOOK ROUND THE ESTATE
00700 50800 BRUCE 15 FEB LET US FIND OURSELVES
00800 62200 YORICK 16 FEB THAT MAN WANT TALK TALK OR MORE PUSH PUSH
01000 63600 BRUCE 19 FEB AI JOURNAL ISSUE ON WORLD KNOWLEDGE
01100 TOM Feb 20 volleys in flight at the cease-fire of the pop-philosophy war
01200 69600 YORICK 24 FEB AS I WAS SAYING WHEN I WAS SO.......
01300 →→→→→→→→ end of month - go to page 9 for March ←←←←←←←←
00100 BRUCE 1 FEB 1972 NEW FORMAT Page 1 of the file is the contents
00200 list for November 1972, and page 2 is the content(!) for that
00300 month. And so on. Read up the SOS commands for dealing with
00400 multi-page files. When you add something, update the contents
00500 list for the month, starting a new one if you are the first of
00600 the month.
00700
00800
00900
01000 ARTHUR FEBRUARY 2 PLEASE - NO WITTGENSTEINIAN
01100 EXEGESIS(THERE ARE
01200 CHILDREN ABOUT!)
01300 I don't really think that a dispute with Yorick over what
01400 Wittgenstein really did or didn't say is what is important(even
01500 if it were it would probably bore the pants off everyone else,
01600 and possibly off Yorick too.As they say, there is none more
01700 zealous than a converted heretic!)
01800 What surely is important are the kinds of arguments that
01900 my ersatz-Ludwig would use in this situation.
02000 I don't think that I disagree with the spirit of what
02100 Yorick says about the pointed-to-object being the explanation of
02200 the meaning, as opposed to being the meaning per se. But my
02300 agreement is predicated upon my assumption that there really is
02400 NO difference between explanation and meaning - consider the case
02500 of an electron. Actually when we refer to "an electron" we are
02600 referring to its place holder in our atomic theory, rather than
02700 to any concrete entity . Indeed there seems to be considerable
02800 doubt as to whether we can ever "know" an electron directly. My
02900 claim is that the same is true for all kinds of other individual
03000 terms occurring in the language: after all, we all know (and
03100 some of us even believe) that the objects of our direct
03200 acquaintance are always in a sense only ghosts of what we take to
03300 be the "real", "concrete" objects. I would want to say that the
03400 objects that we know directly are MODELS of their real-world
03500 counterparts.
03600 This makes an interesting connection with the general
03700 issue viz. what kinds of things do we understand models to be;
03800 can we agree on some standard interpretation of this term; of
03900 what heuristic value can they be in AI. I (and YAW) would be very
04000 interested to get TOM's views on this - when he talks aout
04100 models, what precisely does HE mean. Information please!
04200 A postscript on "meaning/use". My own view, for what its
04300 worth (cries of "nothing!" from the gallery),is that "use " here
04400 means something rather closer to UTILITY rather than UTILISATION.
04500 That is to say, the role of language(obviously) is to convey
04600 information about the state of the world to the hearer. The
04700 information conveyed to the receiver serves to restrict the
04800 alternative states of affairs which could exist at that time --
04900 it can be interpreted by the receiver only with reference to HIS
05000 own model of the world around him.
05100 Compare this view with traditional "speech act " theory a
05200 la John Austin:
05300 "we isolated...the performative utterance as not, or not
05400 merely, SAYING something , but DOING something"
05500
05600 It is fairly easy to show that when we are dealing with
05700 quantified sentences, we cannot in a straightforward way compare
05800 them to reality in the way that we can in the case of atomic
05900 sentences. Rather we must attempt to construct a model in which
06000 the sentences can be imbedded, and compare these models to
06100 reality. This leads directly to the conclusion that:
06200 "what is important .... is not the way that words "mean"
06300 or directly refer to their objects of reference, but rather that
06400 it is in the use of words IN GIVEN SITUATIONS that they have
06500 their meanings".
06600 My suspicion is that, if you ignore the stupid work of
06700 people like Pat Suppes, and look at the work on indexical
06800 expressions from Saunders Peirce through Dana Scott to Richard
06900 Montague, you will see at least an idea of how to go about
07000 considering real language sentences.
07100
07200 TOM COMMENTS ON YORICK JAN 8
07300
07400 Surely, one can't be serious in thinking that there is a meaning
07500 for water for us who know ice, without the knowledge that water
07600 transforms into ice at low enough temperature. And even
07700 without thermometers, we have a sense of what low enough is.
07800 Surely, the meaning of water must also change depending on
07900 geographical accidents, historical accidents, and the state of
08000 one's own ignorance. The fact that ice is not in the Swahili
08100 experience would suggest to me that their meaning for "Maji" does
08200 not include ice. It would not suggest that the MEANING OF WATER
08300 is something which is common to all human experience, a least
08400 common denominator.
08500
08600 I agree that predicate calculus is an inadequate language for the
08700 expression of linguistic knowledge. I am not convinced that
08800 yours is adequate either. Winograd's as programmed is also
08900 inadequate, and as a principle, it seems none. The grammar
09000 part seems to be traditional; I found it too heavily dependent on
09100 grammar.
09200
09300 I want to know what you mean by MEANING. My own sense of meaning
09400 is that an orange is an orange. The meaning of a particular
09500 orange is the orange itself. Now we can't possibly keep an
09600 orange in our head, so that we have a structure of descriptors
09700 (always with the possibility of referring back to the original or
09800 a specimen of the class (go buy an orange) to enrich the
09900 description) but the important part of meaning is the reference
10000 to examples. Also, it must encompass the possible experience
10100 with that object or class. In some cases, that may be a
10200 considerable body of knowledge, and that is the meaning of the
10300 word, concept, or whatever. What can we possibly have but facts
10400 (in the broad sense of relations and references among concepts)?
10500 Thus a meaning is a model, which we can change by reference to
10600 the real world (experiment); of course that reference depends on
10700 models.
10800
10900 TOM 4 FEB ANSWERS ARTHUR 2 FEB
11000
11100 I would entirely agree with your statement that the objects that
11200 we know directly are MODELS of their real-world counterparts.
11300 That is their purpose. Language is nothing but a low quality
11400 link from one's models to another's. We have certain
11500 descriptive elements and certain modelling elements (assume,
11600 suppose, and all the imperatives used in that sense). People's
11700 models are not at all identical, but there is something in
11800 common. Most of my discussion is centered, however, on what
11900 structure models should have. Apologies to Yorick on the use of
12000 the word model. I am equally appalled by his devaluation of the
12100 word theory to a little word which is applicable to the products
12200 of sociologists and ambitious engineers. read on. But what I
12300 mean by model is a structure, since knowledge is tightly
12400 interwoven. Add to that, computational structure, since what I
12500 want to do is compute. Amend that to read, computational
12600 structure which mimics the world, since what I want to compute
12700 is: given an identification between some models and
12800 observables of the real world, can I explain the changes of state
12900 of selected observables by their connections with my models.
13000
13100
13200 TOM 7 jan become 2 feb models,learning
13300
13400 I have been called to task to say what I mean by models in the AI
13500 sense, that it is substantially different from the several
13600 meanings of model used by logicians and by social scientists. In
13700 logic, a theory is the set of axioms (eg field axioms) while a
13800 model is an object which satisfies the theory (eg a particular
13900 field). It is an interesting question as to whether a theory
14000 has a unique model. A way of thinking is that formal theory
14100 hopes to have thinking as a model; or physics theory hopes to
14200 have the world as a model (rather that the theory has a model
14300 which closely approximates the world). All this from Weyrauch. In
14400 social sciences, the word theory is horribly misused. Theory is
14500 used for any trivial explanation: one fact, one "theory". I
14600 would instead call this paraphrasing facts. Theory as it should
14700 be used (say I) is a compact analysis of fundamentals which
14800 predicts a coherent and extensive set of measurable phenomena
14900 (emphasize fundamentals,predict, extensive, measurable). Thus,
15000 a "theory" for each phenomenon is not theory at all. There are,
15100 in this sense, very few fields with any theories at all. Probably
15200 AI has none, although a few extremists would say that predicate
15300 calculus with resolution is a theory of reasoning.
15400
15500 The word model is taken to mean a structure of relations between
15600 concepts. A model can be thought of as the substructure to
15700 a theory; that is, a theory is an analysis of some model. Such a
15800 model can be called analyzable. A model which is adequate and
15900 analyzed is what I mean by theory. The meaning of adequate is
16000 that it predicts a coherent and extensive body of measurable
16100 phenomena, and that it be fundamental; ie that its externals be
16200 measurable, while its internals be directly connectable to the
16300 externals. (One school of physics maintains that a theory
16400 should only involve measurables, not any internal, unobservables.
16500 That is almost always not the case.) As an example, consider
16600 psychology of personality. There are observables: for example,
16700 the score on some arbitrary exam. There is no predictable
16800 connection between the measurables and internals such as
16900 aggressivity, etc. Often models are taken because they are
17000 analyzable (linear economic models) and not because they are
17100 adequate. This is frequent in engineering and social science.
17200 There are models which are unanalyzable, because the relations
17300 are too complex for mathematical analysis, or because the
17400 relations are not well-defined, or because the model is
17500 incomplete. A phenomenological model may describe the results of
17600 a coherent set of measurements, without any sense of describing
17700 unrelated phenomena at all. This is a type of curve fitting, and
17800 hasn't any relation to any fundamental structure. It may have
17900 utility for experimentation, engineering, or functioning as a
18000 biological creature. But phenomenological models are just a form
18100 of paraphrasing facts.
18200
18300 These meanings of model are about what I mean by model in the AI
18400 sense. It is a computational structure which has several
18500 purposes and features. It has a symbolic description of
18600 internal structure, ie intent, and symbolic and procedure
18700 description of external behavior, ie input/output behavior. An
18800 example which I discussed is our model of gravity. Yorick might
18900 code (STUFF FALL DOWN) as form in which gravity might appear;
19000 this is a particular form of axiomatization of gravity. What
19100 must any model be for gravity. If I tilt my head, then I see that
19200 things fall in the direction of gravity as measured by gravity
19300 sensors in my ears. Thus the model involves sensing and relating
19400 phenomena to this sensed reference system. If I drop a stone
19500 down a well, then I can say that the well is deep. Thus I have
19600 the ability to make a crude calculation of the time required for
19700 falling from a given height. I also know that it does not make
19800 any difference from which horizontal position I drop the stone,
19900 that is, the time of fall depends only on the vertical fall. I
20000 know also that a feather takes much longer to fall than a brick.
20100
20200 This is intended to go much further; that knowledge contains many
20300 such models and they reflect our informal knowledge of physics,
20400 chemistry, psychology, geometry, etc. There are many of these
20500 microtheories. They are not the physics, chemistry, etc that
20600 we learn in school. They are instead, the knowledge of such
20700 phenomena which we have without schooling; perhaps formal
20800 knowledge alters these microtheories, but they are not limited to
20900 physicists and mathematicians or people with formal schooling.
21000 Such microtheories are the meaning to semantics, and
21100 understanding takes these for granted.
21200
21300 I would like to say something about learning and how it is
21400 possible. There are usually discussed two forms of learning which
21500 might take place. They are really all the same in concept, but
21600 they differ in methods and proponents enough to be
21700 distinguishable. The first might be called deduction: we know
21800 some things to be true and we use rules of reasoning to arrive at
21900 everything which is true in consequence. The second is induction:
22000 the observer watches the world about him and keeps track of
22100 things and draws conclusions about this world. There may be a
22200 teacher who guides his conclusions, and the pupil may build up
22300 interesting models of world situations, as in Winston's valuable
22400 work. Or the models can be trivial and the emphasis on keeping
22500 track as in statistical learning.
22600
22700 I emphasize yet a third form of learning, which is the only one
22800 capable of really learning enough to handle world experience. In
22900 the above types of learning, the observer was passive and things
23000 happened to him. My contention is that experimental learning is
23100 necessary. By experimental learning, I do not mean statistical
23200 learning, but I mean a system which combines building of models
23300 with some descriptive mechanisms which affect the type of models,
23400 combined with some simple experimental paradigms, and inference
23500 rules. (A triple: descriptive mechanisms, experimental paradigms,
23600 inference rules). Let me explain all these things. Building of
23700 symbolic models is obvious, but empty without specification.
23800 But the descriptive mechanisms determine what sort of models are
23900 built up. This amounts to classifying the type of problem and
24000 what the model depends on (what variables, what is the context).
24100 The experimental paradigms are a set of experiments for obtaining
24200 a set of relations for the model structure in terms of other
24300 models: how to measure a quantity. But most of all, the
24400 experimental paradigms contain the independence hypothesis:
24500 assume that variables are independent and experiment by varying
24600 each separately; patch the procedure by adding some interaction
24700 when necessary. It is my contention that the only way to deal
24800 with the complexity of phenomena in the world is the experimental
24900 paradigm. That is, to separate variables under the control of
25000 the observer. It is his purpose to manipulate the world anyway.
25100 This is the learning paradigm as I see it. mModels are of
25200 crucial importance in this endeavor.
25300
25400 Yorick has said that AI should avoid the pitfall of the "Queen of
25500 Sciences". I contend that it should avoid the pitfall of
25600 thinking that there is a shortcut to intelligence without being
25700 "Queen of Sciences". This does not mean that we must extend
25800 the frontiers of chemistry, astronomy, physics, neurophyisiology,
25900 etc. We must, however, incorporate the common sense knowledge
26000 (a vast structure) in all of us, and the many models included
26100 there. Again, I would ask, what does understand mean? Or
26200 translate. To translate, I contend, requires understanding. And
26300 that is precisely because there is in language an enormous
26400 reference into that structure of knowledge. Some things, like
26500 poetry, are more difficult to translate, because of the intimacy
26600 of the contact with social and personal context (which is the
26700 least shared of our models), as well as the great gaps left
26800 unexplained. Of course, people can understand and translate
26900 those things for which they have adequate model systems, and they
27000 don't understand other things. They may be able to translate as
27100 word strings, other material. If so, the results may be
27200 sometimes adequate and sometimes laughable. It does seem that
27300 there is reading and reading, and that the Swahili with the
27400 limitations that you have burdened them with, can read some
27500 things and not others. So what?
27600
27700 Yorick, what do you mean by meaning?
27800
27900
28000 YORICK 5 FEB REPLY TO TOM So what indeed!WE all agree these days
28100 that translation requires understanding; the only dispute is
28200 about what the word "understanding " means there.All our
28300 positions in this dispute are,I suspect,circular.Tom says that
28400 people can translate those things for which they have adequate
28500 model systems,in his sense of those words,and cannot those for
28600 which they dont etc.I cannot see that that is any more than a
28700 partial definition of what Tom means by the phrase
28800 "adequate....etc",the whole thing is circular because if a system
28900 set up by Badman translated without having some set of facts that
29000 Tom formerly considered essential for A[dequate]M[odel]
29100 S[ystems],he would then say ,oh well ,so in this case only a part
29200 of the REAL AMS was required, but watch out next time Badman!
29300
29400 I do indeed mean what Tom feared about the meaning of
29500 "water"---if we found a substance like water in all respects
29600 except that it didnt freeze at all,we would still call it water
29700 wouldnt we[doubt that!?] -----and probably add "tricky water" or
29800 something.It would still be water wouldnt it----thats my point
29900 exactly.
30000
30100 I am prepared to go a long way with Tom and admit,as Fodor and
30200 Katz claimed to have shewn, that for any fact F ,a sentence can
30300 be produced that cannot be understood without knowing F .OK,but
30400 Im not sure that this tells Tom's point of view as strongly as he
30500 may think-----all it proves is that any system can be fooled. If
30600 Tom sets up an AMS tomorrow,with all the facts he think
30700 important, someone will come along with that extra fact F and
30800 there'll be a sentence S. it doesnt understand----then itll be
30900 just like he says of my system, some things it understands and
31000 some it doesnt.So what!
31100
31200 There's no dispute here about the need for knowledge to
31300 understand: only a question of how much and how to organize
31400 it,and how to extend it where necessary in the fact of awkward
31500 F's.
31600
31700 Its clear that I think that a lot less knowledge will get you
31800 along with translation understtanding than TOm does: moreover I
31900 think it should be largely conceptual understanding not
32000 superficial facts.MOreover,as Ive said before,I think it should
32100 be organized nondeductively ad have precisse suggestions for how
32200 to do that.When Tom talks of "models" as structures of
32300 facts/explanations---I do not know what structures he has in
32400 mind.I really dont.I know to some extend wat systems Tom
32500 considers inadequte,but not what would do better for him.
32600
32700 Tom may well come back on my quoting of his "so what",and say
32800 well,mine will clearly do more,even if no system is or can be
32900 provably correct in all cases. I then reply,not clearly at
33000 all:Charniaks seems to me a system closer to Toms desires
33100 (structures of facts ,deductively organized) and I doubt its
33200 ability to do better even if programmed.I just do not think
33300 deductively organized chains of facts are going to solve
33400 understanding problems;except for those examples set up too prove
33500 the applicability of such methods.i.e. I feel(Ive probably said
33600 this before on this file) that if an example REALLY REQUIRES such
33700 deductive chains (and they are uniquely imposbale as Charniaks
33800 arent) then probably an average person cant understand the
33900 example either,because people are notoriously bad at deductive
34000 problems,but understand more or less alright.
34100
34200 If Tom emans by mdel a structure of reasoning quite oother than
34300 this,then lets see it.THEN Ill know what he means by model.
34400
34500 A few last philosophical desiderata for this.What do I mean by
34600 meaning,Tom asks. When someone asks me the meaning of X I give
34700 explanations till he's happy or shuts up.I rarely if ever point
34800 to anything;I often refer to dictionaries because what they
34900 contain (not pictures in British dictionaries) maps more or less
35000 onto what I mean by explanations. From some of the things he
35100 wrote I think Tom accepts this explanation view of meaning. But,I
35200 would argue strongly,the view is prima facie different from 2
35300 other views of meaning he also seems to accept.
35400
35500 (1)the reference vieew(comes in crude and sophisticated[1,AJT])
35600 versions.I cant understand Toms sentence "The meaning of a
35700 particular orange is the orange itself".I dont think oranges have
35800 meaning----except possibly and derivatively as symbols in
35900 Prokofieff operas. Woord strings have meanings,and I believe they
36000 are ALWAYS other words ----cf.Quine and Wittgenstein passim on
36100 "the inscrutability of reference".It is a profound and enduring
36200 myth that we mean by pointing ---we can never do that in fact,at
36300 least not unambiguously,and without the whole weight of the
36400 meaning being carried by the language and assumptions we
36500 share.See Quine ad nauseam on trying to know WHAT a savage is
36600 pointing to as he says "Gavagai". Tom sometimes seems to admit
36700 this when ,for example,he says that "we know only models
36800 directly".
36900
37000 (2)The meaning is facts view.Depends how widely you take
37100 "facts".Much meaning is explained by sentences like "Meanings are
37200 what words refer to"(false in this case),and "Facism is the last
37300 stage of monopoly capitalism"(false again). Theses are not facts
37400 ,in the ordinary sense of that word.Anyone who thinks they are
37500 should then ask himself how he would check up on their truth or
37600 falsehood.Most sentneces on this file are of this sort.AJT thinks
37700 that all such senteces are ultimately REDUCIBLE TO elementary
37800 facts.An interesting thesis,but it is a philosophical
37900 position,and not self evidently true.In the common sense sense of
38000 "fact" those sentnces are not facts.If by "fact" Tom simply means
38100 "any assertoric sentence",then ok,but,as he wouldsay,so what?!
38200
38300 last thoght after rereading Tom: T says meaning=model:Y says
38400 meaning=explanations.May be no real difference here except
38500 "model" suggests the structure of the explanations is KNOWN.I do
38600 not think it is known,by Tom or anyone else: so I see
38700 "model=explanations + ASPIRATIONS ".The aspirational mode may
38800 raise the morale of the troops,but I dont see it does anything
38900 over and above that.
39000
39100 ARTHUR FEBRUARY 9 ONCE MORE UNTO THE BREACH, CRYING "GOD!!!!" 1.
39200 Now I, and I suspect Tom, am completely confused by what Yorick
39300 means(if you'll forgive the nasty word) by "explanation". I get
39400 the impression, possibly (i hope) incorrectly, that YAW
39500 "explanations" are rather like semantic markers a la Fodor &
39600 Katz. That is to say, a never ending regression of (possibly
39700 recursive) pseudo-explanations. You are quite right in saying,
39800 Yorick, that my thesis about ultimate reducibality to elementary
39900 referring locutions is a philosophical position - what I fail to
40000 understand is why you find this derogatory, and further why you
40100 think its truth SHOULD be self-evident. Surely all of us in this
40200 discussion are putting forward philosophical positions. We may
40300 care to support these with empirical evidence, but that is in a
40400 sense peripheral. Speaking of such nasty empiricism, Richard and
40500 others may be interested in the large number of studies on child
40600 language corpori which are available, and in studies relating
40700 development of language with development of perception. I can
40800 point you at some of this stuff if you want. 2. Surely objects DO
40900 have meaning in and of themselves in a very direct and crucial
41000 way. If I say to you "you are sitting on a bomb which is about to
41100 explode", the WORD "bomb" in this sentence is in itself
41200 insignificant - it is the ACTUAL BOMB which is about to blow you
41300 to smithereens. As the old saying goes "sticks and stones may
41400 break my bones, but NAMES can never hurt me". Unless of course
41500 the NAME is carved in a block of granite which I drop on your
41600 head. Words only act as POINTERS to the objects which themselves
41700 have significance in my life. Tom, and I, would say that they act
41800 as models for the real thing(without saying that that is all a
41900 model is). 3. I find it amusing that Yorick thinks of
42000 referential semantics as a "profound and enduring myth". After
42100 all, is not the case that myths generally endure because they
42200 contain some element of truth? 4. I think it is fairly
42300 clear how to decide upon the truth and falsity of the "Meanings
42400 are what words refer to" and "Fascism..monopoly capitalism "
42500 sentences . One would stand around and watch for the effects of
42600 their utterance upon observers. For instance, one might say that
42700 the response of hysetrical laughter to the former sentence which
42800 would emanate from YAW would be an indication of its undoubted
42900 truth! But seriously though, folks. I am , rather against my
43000 will, being forced to act as advocatus diaboli for the speech act
43100 theorists at this point. Consider the revealing title of John
43200 Austin's book "How to DO things with words"; the sentence I
43300 quoted above in my note of February 1 continues:
43400 "....as NOT a TRUE OR FALSE report of something."
43500
43600 Incidentally, for a very interesting discussion of the
43700 differences and similarities between the "formal logical
43800 semanticists" and the "theorists of communication intention", see
43900 Davidson's "Truth and Meaning", and Strawson's "Meaning and
44000 Truth". I think that Peter Strawson rather overdoes the
44100 distinction between the two groups, but it is food for thought,
44200 and I would like to get other peoples opinions. Consider also
44300 John Searle's "Speech Acts".
44400
44500 YORICK 11 FEB A LAST LOOK AROUND THE ESTATE.
44600
44700 notes on arthur's numbers immediately above: not up your usual,
44800 arthur! 1)No,"explanations" nothing to do with dear old departed
44900 Kodor and Fatz ------who are quite quite inadequate (we all
45000 agree,except Winograd,perhaps!?) -----if I had been talking at
45100 the formal-expression-of explanation -level ,I would have
45200 proferred my own,which I consider superior by far.But I wasnt,and
45300 by explanations I meant the ordinary language sense of
45400 word----i.e.more or less what is found in dictionaries.
45500
45600 No,of course,I didnt mean "is a philosophical position" as a
45700 derogatory dismissal,merely as a marker or set of parentheses
45800 meaning "this is an arguable assumption which we mya acccept or
45900 reject as we choose,and justify doing either,of course,at
46000 whatever length we find necessary(or is accepted by SOS!!"
46100
46200 As to the thesis in question,I ve never believed it since John
46300 Wisdom convinced me that the sentence "England declared war on
46400 Germany" is not reducible without loss of meaning to ANY
46500 conjunction of sentences about individual Germans and Englishmen.
46600
46700 2)Cant work this section out.I still think that to say an object
46800 has meaning is to joke or pun,and I still find the last senence
46900 about models inscrutable,and as I ve said before prima facie
47000 inconsistent with the "words refer to objects(physical)" view
47100 ---I may be just dim about this point,but I still think a lot of
47200 quite low level clearing up has to go into the Tom-Arthur view(if
47300 they ARE the same)before its comprehensible to the unwilling to
47400 understand,like me.
47500
47600 Afterthought:to be precise ,if words refer to physical objects
47700 they do not refer to models,whatever "models" may be.If words are
47800 models (last sentence I think),then words do not refer to
47900 them(selves?),moreover I dont think you can really have meant
48000 that, because I already know you think models are set theoretic
48100 constructs.I do not think MY words (though I cant speak for
48200 others) can refer to THOSE things,becasue I cannot understand
48300 them,and yet seem to get on with words more or less ok(hyacinth
48400 laughter from the side) .I know the old argument about:your
48500 introspections are no guide to how your brain really works,but I
48600 dont think applies here.The present situation is more like
48700 ,physics can tell me what wavelength of light is associated with
48800 my use of "red",but it never tell me(never,never) that my use is
48900 not really associated with redness,because the presented
49000 sensation here is the primary datum.Similarly with my use of
49100 words in the circumstances in question:the speakers authority is
49200 final on what his refer to,if he is able to refer at all.(tho it
49300 is not,as every Wittgenstinian knows,final as to what he means in
49400 any OTHER sense of "mean".)
49500
49600 In an earlier contribution you actually said,I think,we only have
49700 direct access to our models of the world.That just CANNOT BE TRUE
49800 in any sense at all of the words in that sentence.(just think
49900 about it again,slowly.)
50000
50100
50200
50300 3)not so,astrology,alchemy,phrenology,haruspication,tea
50400 leaves.....
50500
50600 4)I cant go along with this "behavioral analysis and test of
50700 philosophical views".I dont think its even a starter so dont
50800 really want to go on with it,let alone burst out into
50900 "hysetrical(sic)" laughter (would if I knew how!!!).
51000
51100 Afterthought:I am not saying there are not proper arguments about
51200 the truth of such statements;they go on all the time,here and in
51300 better known publications.What is clear is that such
51400 justification procedures are not behavioral and not proof or set
51500 theoretic in ANY SENSE at all There is no adequate analysis of
51600 the structure of real arguments,outside classical Rhetoric and
51700 some reaches of modern psychiatry.I did PhD thesis on argument in
51800 metaphysics with a working imitation (model,anyone!!?] of bits of
51900 some arguments ----that imitation could be expressed set
52000 theoretically of course,but that would prove nothing any more
52100 than doing the same for geometry shows anything about the nature
52200 of real geometric proofs.
52300
52400
52500 5)good refs(though Ive always thought of the Strawson as
52600 supporting my position in this quibble.Would also suggest
52700 ----supportuing yours I think----Quines early and good stuff on
52800 "Truth by Convention",in a memorial volume to Whitehead.
52900
53000
53100 BRUCE 15 FEB LET US FIND OURSELVES
53200 It seems to me that the discussion is spiralling inwards
53300 and will shortly disappear up its own termination. What do I
53400 mean? Well, the stuff in the file has two components. The first
53500 is technical, and is concerned with making robots. It is about
53600 questions such as "how much does a robot have to know about the
53700 world vs. about language", and I think that in fact we are all in
53800 agreement about this. There are some misunderstandings, but
53900 these are all relatively unimportant, and it is truly amazing how
54000 much steam has been generated by mere hot air.( Ouch!). The
54100 second component is about things like "what is meaning?". But if
54200 we are agreed on the first component, then the second is really
54300 rather pointless - I really don't care which part of the robot's
54400 hardware/software is the embodiment of the meaning of the word
54500 (for example) "like". If the robot can talk about liking, and
54600 reason about it, and seems to "understand" the word in the sense
54700 that someone else does, then I am happy. There would no doubt be
54800 700 Ph.D. theses on whether the thing was dictionary-oriented or
54900 not, and whether it was Popperian or Carnapian and so on, but
55000 these disputes would be restricted to crabby old philosophers and
55100 their pupils.
55200
55300 So why do we talk philosophy and psychology at all? The
55400 answer is that although we are agreed about the types of
55500 structures we need for our robot, they are nowhere near
55600 specific/well-defined enough to be implemented. Part of the
55700 steam in the Yorick vs (Bruce ∪ Tom) disputes is generated by
55800 Yorick's thinking that the models we are proposing are
55900 well-defined but that he does not understand them. Or that he
56000 does - for example in his seeing Planner and predicate calculus
56100 as two faces of the same (evil "deductive") coin. But really
56200 the requirements we have put forward are incredibly vague. We
56300 look to philosophy to find MORE DETAILED ideas about the kinds of
56400 structures that could be in our robot. And of course we are
56500 disappointed when we find lots of philosophers are even vaguer
56600 than we are!
56700
56800 So what do I think the group should be doing? There
56900 should be two aspects. One is the technical side mentioned
57000 above. Can we sharpen our ideas of design criteria for
57100 intelligent programs? How sensible are various approaches that
57200 are being/have been taken? What are the (generally unstated)
57300 theories of AI that people would have to have in order to
57400 rationalize the work they are doing? I think we are all pretty
57500 much agreed on the facts in this area, but discussion might
57600 stimulate some interesting ways of looking at/classifying things.
57700 In particular I think we might discuss the world knowledge vs.
57800 language knowledge issue, perhaps thinking about writing programs
57900 which can SPEAK a language. OK if you have lots of semantics you
58000 can make your machine understand "It with woman push push man
58100 want" but I hope that at least British robots won't talk like
58200 that.
58300
58400 This "technical" side might be called the philosophy OF
58500 ai, whereas the second aspect is philosophy AND ai. Here we
58600 should talk about how psychology and philosophy interact with AI.
58700 How can we USE philosophical results? And conversely, how will
58800 the making of a robot affect philosophy? Would such an event have
58900 any bearing on the mind-body problem? Does philosopy tell us that
59000 certain approaches to AI will fail? Or that the methods may be
59100 classified in some way? I don't know if there will be any
59200 differences among us here - it is an area that we have not
59300 touched on enough, probably because it is a very difficult one.
59400 Can we firm up philosophical (or even psychological) theories
59500 enough to make concrete statements about robots (or statements
59600 about concrete robots) with them. For example, what would the
59700 Austin-Strawson-Grice approach to a natural-language
59800 understanding program be? Note that (like everyone else
59900 interested in philosophy) Yorick and Arthur are pretty
60000 schizophrenic about philosophy in that their philosophy does not
60100 enter into their programs, and their writings are clearly split
60200 between the practical(cough)and the philosophical. Pat Hayes is
60300 the worst at this - despite lots of talk about modal logic and
60400 "Davidson on mass terms" he has no more idea about how to program
60500 intelligence than I have. This split is a bad sign - is there
60600 any mutual feedback between phil-wilks and robo-wilks? Getting
60700 into philosophy in classical philosophy's terms is probably a
60800 waste of time (for AI people at least) because the aims and ways
60900 of thinking are so different from ours, and digging in in our own
61000 terms is so difficult that it has often been said (e.g. above)
61100 that it is quicker to rediscover any of anything you need rather
61200 than dig through that subject trying to decide what is useful.
61300 This is also said of psychology and linguistics, but not of
61400 mathematics. We might discuss that.
61500
61600 So I've said what we SEEM to be doing, and what we OUGHT
61700 to be doing, though like the robot's insides the latter is not
61800 specified enough to let us just go and do it - to decide topics
61900 worth discussing is as hard as discussing them! But please let us
62000 have less of the "meaning of meaning" discussion sort of thing
62100 ("The word model is taken to mean a structure of relations
62200 between concepts" indeed. That is worse than most of the stuff
62300 you read in 1950's cybernetics books. ). I also think that
62400 there should be less of the "do read Strawson on Individuals it
62500 will blow your mind and alter your programming style
62600 immeasurably" stuff too - the latter really means "I think
62700 Strawson on Individuals has something to say to AI people, but
62800 I'm damned if I know what it is - maybe you will find out for
62900 me". As usual with DBA it is all anticlimax, so I will close
63000 with a few puny ideas for discussion areas. **1. To what
63100 extent will robots might be/have to be like people. Does your
63200 body shape effect how you think. Should robots have human
63300 emotions? What would be the use of that? Should they have
63400 "judgement"? **2. Yorick puts Planner and predicate calculus
63500 in the same bag. So does Richard, I think. (Unlikely
63600 bedfellows, you say?). Both think that these approaches to
63700 intelligence involve making "proofs" -Richard thinks this a good
63800 thing, Yorick thinks this makes them useless. I'd like to get
63900 straight why I don't like proofs, but DO like Planner, and why
64000 the "conceptual understanding" Yorick wants is just what his
64100 so-called "deductive chains" in e.g. Charniak's work, are about.
64200 There is some X of which Charniak's stuff is an example - this is
64300 the "Planner subculture" thing and I think Yorick has got the
64400 wrong X. Anyway, I'd like to know Richard's views on (a) formal
64500 systems in AI, and (b) JMC's views.
64600
64700 [The above work is entirely fictional, and any similarity between
64800 characters therein and any person living or dead is entirely
64900 unintentional]
65000
65100
65200 YORICK 16 FEB THAT MAN WANT TALK TALK OR MORE PUSH PUSH? Applause
65300 for Bruce`s last comments and suggested topics. Can I go on a bit
65400 about his points concerning AI and second hand reach-me-down
65500 philosophy that fills up too much of our files.I dont think this
65600 is any sort of accident,or mere perversity on the part of some of
65700 this group.I think that most of the real issues in AI just are
65800 philosophical issues whether or not AI people realise that
65900 -----all the discussions since 1600 of robots and their possible
66000 abilities and necessary limitations (all those worth reading that
66100 is) have gone on in philosophy publications---and that is still
66200 true.Id agree with Bruce,if he said it,that most of modern
66300 philosophy is hair splitting rubbish,but the robot area is one of
66400 the exceptions.
66500
66600
66700 BRUCE 19 FEB AI JOURNAL ON "WORLD KNOWLEDGE"
66800 The AI Journal is having an issue on representation of
66900 world knowledge, for which papers have to be submitted by August
67000 this year (approx.). Richard is going to try and get his ideas
67100 in shape for this, and so am I. Perhaps everyone else is too -
67200 after all this discussion we MUST be highly qualified - so
67300 tomorrow let's discuss the possibility of doing something
67400 together. My idea is that in July everyone who wants to shows
67500 what he has done with a view to putting it all together. Maybe
67600 this is too socialist. As for content, I'd like to concentrate on
67700 thinking about problems of AI, rather than philosophical problems
67800 in AI (which, as Yorick (push push) says above, our technical
67900 knowledge of computers etc does not really help us to solve,
68000 Scriven and others notwithstanding).
68100
68200 TOM Feb 20 volleys in flight at the cease-fire of the
68300 pop-philosophy war It seems that much discussion turns on points
68400 taken for granted at first exposure long ago: impossibility of
68500 direct reference. That is such a trivial thing it is useful for
68600 brilliant argumentation, but is best forgotten lest it impede
68700 conversation. Only philosophy has the time to make profundities
68800 of such trivia. Of course, we cannot point to a thing. We can
68900 suggest a set of experiments (look, touch, listen) which have
69000 reproducible results and we can store those results in a coherent
69100 way. To a certain extent we can communicate the results of
69200 experiments to other people, but not very accurately. At best we
69300 say "Look, now, there. Forever after, I mean something like that
69400 when I say orange." There are enormous difficulties in saying
69500 "similar", but we share the same meaning of similar, so saying
69600 "forever after" works. Let me repeat "a model is a computational
69700 structure which mimics the world", ie a model is intended to
69800 allow thought experiments. Thought experiments are safer,
69900 quicker, and more economical than blundering along. I do not
70000 understand what Yorick means when he says explanation. Does it
70100 mean substituting word strings for word strings? If so, how does
70200 the poor soul who receives this treatment know when to shut up,
70300 ie be satisfied with the explanation (if he perceives that is
70400 happening, he should shut up immediately, of course, and seek
70500 better company). And do you deny entirely reference in language?
70600 Do I or do I not have a structure "that orange" which refers to
70700 the particular orange which I have brought for lunch? There are
70800 two argumentative copouts to avoid:
70900 not all things are referential;
71000 reference cannot be infallible.
71100
71200 Clearly, many words or phrases do not refer to objects; they
71300 refer to models, which refer to other models. Some models have
71400 reference to objects. A robot must have a belief in objects in
71500 the world (in the above sense). This might seem just a useful
71600 self-delusion. It seems more fundamental in that it corresponds
71700 to a discrimination about validity of types of knowledge;
71800 touch/pain and manipulation seem primal-they do not seem
71900 equivalent to vision, hearing, etc. Solipsism does not seem very
72000 important; for a solipsist to function in a world like mine, he
72100 must use descriptions rather like mine, with all that entails. So
72200 reference is not infallible, but it is consistent and predictable
72300 (I actually can find that orange).
72400
72500
72600 YORICK 24 FEB AS I WAS SAYING WHEN I WAS SO..... AA LITTLE MORE
72700 ON PHILOSOPHY AND AI I simply disagree with Bruce above about the
72800 relations between "tough, good little " ai ,and "vague sloppy old
72900 "philosophy.Bruce says that maybe philoosphy could be suggestive
73000 if only it could be "firmed up".This is all topsy-turvy:its
73100 philoosphy that is precise and ai that needs firming up,as he
73200 puts it. Bruce`s criteria of firmness are all wrong:writing
73300 programs is indeed firm,but)can be firm and totally
73400 pointless,just like copying out the p phone directory by hand or
73500 something equally definitive .Philosophical analysis can become
73600 trivial,Id be the first to agree,but never quite as trivial as
73700 mindless programming.
73800
73900 I dont think there is a split between X the programmer and X the
74000 philospher,or rather there need not be.Programs and systems
74100 generally need to be justified iin abstract and general
74200 terms---so what,no schizophrenia there!!
74300
74400 Philosophy has been far more precise and to the point over the
74500 centuries about meaning,minds,robots,understanding etc.etc than
74600 20 years of ai people ----of course,how could it be
74700 otherwise.Thats what worries me about Bruces phil&ai vs.phil of
74800 ai distinction.I dont think ther is any phil of ai ----discussion
74900 in ai,if and when it happens,is almost always 2nd.hand
75000 philosophy,badly done,just as almost all theology always has
75100 been.There are no intellectaul questions to which ai has
75200 contributed a thought(ID LOVE TO BE REFUTED HERE).The ad nauseam
75300 assumption of many ai discussions `suppose a robot walked in here
75400 an d behaved jsut like wilks etc.,has been practically exhausted
75500 fotrr centuries.Descartes and Leibniz made the same assumption in
75600 their discussions----the nature of THE ASSUMPTION has not been
75700 altered in the least by the invention of the computer.
75800
75900 Also I do think the models [i give in on this word) that would be
76000 constructed by a humean or by a reatuionalist would be a proiori
76100 different:just as ,given his presuppositions about possinble
76200 worlds and modal logic,AJTs models shoud be different from thoose
76300 of people withh different metaphysico-logical beliefs.i.e.
76400 beliefs at this level do make a difference ,pace Bruce.
76500
76600 Again,as I tired to argue the other day,not very successfully,I
76700 thinkk philosophy can play an excellent negative role in saving
76800 us from conceptual errors aabout notions like understanding----
76900 still think some of the group think that there is an essence of
77000 understanding----a feeling of rightness that cant be
77100 wrong-----and that this is a mistake.Only people who have this
77200 belief can say ----ah well yse,Xs system performs with language
77300 ok,but it doesnt really understand it you know,its just a big
77400 kludge.
77500
77600 QUERIES 1)someone please tell me why the theorem proving language
77700 PLANNER ,as used by Winograd and co,is not a theorem proving
77800 language ---i.e.why it is significant ly different from using
77900 thee PC in a heuristic way,without trying to plug in a resolution
78000 prover for instance. And ,why as Bruce says ,is doing this like
78100 the preference methods I advocate where no truths are proved at
78200 all,by any method.
78300
78400 2)Would AJT please give us some more on set theraoretic models:in
78500 particular justifying what I think he believes,which is that
78600 pre-"semantics" logic wasnt really doing anything much ----i.e.
78700 all the stuff on deduction from 500 BC to Godel was a mistake in
78800 some sense. And if it wasnt,then why does he say deduction needs
78900 a semantics at all?
79000
79100 (Im sure these are foolishly expprssed points,but those who know
79200 better can probably translate them into real questions before
79300 answering them.)
79400
79500 REPLY TO TOM
79600
79700 Tom you say that clearly many words and phrases rfere not to
79800 objects but to models----sorry,but that isnt at all clear to the
79900 majority of the human race,including most of its best informed
80000 members.Its sounds just like non-common-sense jargon.Why do
80100 things have to REFER at all to be
80200 meaningful----why,why,why???They are all right as they are,you
80300 see.
80400
80500 As to explanations,yes I really do mean,explanations.Moreover,you
80600 rightly say,on that view,how could the one explained to ever have
80700 a definte shut-up-point----and youre dead right,he doesnt,and all
80800 that there is such a point(a definite,logical,satisfaction point
80900 ,as it were) is utterly wrong.Explanations can ALWAYS be pushed
81000 on further -----look at this file,for Gods sake----there is never
81100 a determinate stop point ---just as a painter always COULD put
81200 another stroke on a painting,but at some point he nmerely
81300 stops,so with explanations.There is always the possibility of an
81400 infinite regress,which does not stop with any first terms or
81500 principles,because dictionaries are ultimately circular.
81600
81700
81800 →→→→→→→→→→→→→→→→→→eom...eom...eom←←←←←←←←←←←←←←←←←←←←←←←←
00100 100 BRUCE 2 MARCH FEARS WILKS
00300 500 JMC 5 MARCH ON BRUCE 27 NOVEMBER EFFECT OF ROBOT APPEARANCE
00400 2200 JMC 5 MARCH ON YORICK 4 DECEMBER WEIZENBAUM DIDN'T SATISFY TURING
00500 4000 JMC 5 MARCH ON YORICK 5 DECEMBER EFFECT OF BEHAVIORAL PREDICTION
00600 4900 JMC 5 MAR ON YORICK 5 DEC WHAT BROKE NOT MAINLY LINGUISTIC PROBLEM
00700 6700 JMC 5 MAR ON MEANINGS IN YORICK 8 JAN
00800 8500 JMC 5 MAR ON BRUCE 15 FEB INWARD SPIRAL
00900 9600 JMC 5 MAR ON YORICK 24 FEB IMPRECISION OF PHILOSOPHY
01000 ARTHUR 7 MARCH I SPY A CATEGORY MISTAKE (I THINK)
01100 11300 JMC 20 MARCH ON ARTHUR'S 7 MARCH SPYING A CATEGORY MISTAKE
00100 BRUCE 2 MARCH FEARS WILKS
00200 OK so just becos yorck is big an tuogh he think he are
00300 the majoraty of the humn rais. OK I fere him but one da he will
00400 FAL.
00500
00600 JMC 5 MARCH ON BRUCE 27 NOVEMBER EFFECT OF ROBOT APPEARANCE
00700 Bruce asks, "Would a working robot, that is with natural
00800 language input etc. have any effect on the practitioners of X?
00900 A robot that came from Mars without blueprints or program or
01000 permission to be taken apart would merely be a seven day wonder
01100 except perhaps for a few philosophers who would be surprised. If
01200 the blueprints and program were available, but incomprehensible
01300 because of size, the situation wouldn't be much different. If
01400 the robot had "evolved" from a simple system in a simulated
01500 environment and the simple system and the environment were
01600 understandable, this would be more informative even if the final
01700 form of the program were not understood. If the robot had been
01800 designed by someone who understood what world-view to give it and
01900 how to program the relation among language, observation, goal
02000 structure and heuristic procedures, then all the fields would be
02100 decisively affected by the ideas involved.
02200
02300
02400 JMC 5 MARCH ON YORICK 4 DECEMBER WEIZENBAUM DIDN'T SATISFY TURING
02500 Weizenbaum didn't satisfy Turing's criterion, because I
02600 can tell anyone how to tell the behavior of current programs from
02700 human behavior. The fact that psychiatrists and random members
02800 of the ACM were fooled by PARRY shows that people are more
02900 careless than might have been imagined and also that they don't
03000 realize how incredibly bad current programs are. Example: If the
03100 questionee says the Mafia is after him and when asked how he
03200 knows replies that he noticed a strange car parked by his house
03300 that is a human; none of the present programs are capable of that
03400 much reasoning. If you think the programmer might specifically
03500 have anticipated being asked what the evidence is, ask a few
03600 questions like when, how often what color, did you get the
03700 license number, and the program's incomprehension of these
03800 matters will be clear. Bruce: there is no need to ask it about
03900 its mental states; its incomprehension of elementary causality
04000 and purpose will be enough; of course, it won't be able to do the
04100 mental state questions either.
04200
04300 JMC 5 MARCH ON YORICK 5 DECEMBER EFFECT OF BEHAVIORAL PREDICTION
04400 The "paradox" of a prediction of your behavior affecting
04500 it is just as powerful for a much simpler system. Suppose we
04600 have a little box with two switches and two lights. One switch
04700 has the label "the red light will go on" and is connected to the
04800 green light, etc. I claim that this system is just as
04900 paradoxical as a human with regard to the effect of telling it a
05000 prediction of its behavior.
05100
05200
05300 JMC 5 MAR ON YORICK 5 DEC WHAT BROKE NOT MAINLY LINGUISTIC
05400 PROBLEM
05500 The problem of "I dropoed the bottle on the table and
05600 broke it" is only peripherally a linguistics problem. The
05700 linguistic part of the problem is solved once we have decided to
05800 settle the referent of it on the basis of whether it is more
05900 likely that the bottle be broken rather than the table. The
06000 question of under what circumstances it may be presumed that the
06100 bottle is more likely to be the object meant is a problem in
06200 knowledge of the world. This is most clear if the sentence was
06300 actually uttered and we are in a real context. If the sentence
06400 is being considered in isolation which is perfectly legitimate,
06500 more linguistic considerations are relevant. I suppose
06600 considerations of "would the speaker have said this if under
06700 such-and-such circumstances" are relevant even in the real world
06800 case.
06900
07000 In my opinion, TP needs some new gimmicks in order to cope with
07100 this problem. I have some ideas, but they are not definite yet.
07200
07300 JMC 5 MAR ON MEANINGS IN YORICK 8 JAN
07400 The concept of meaning in mathematical logic starting
07500 with Frege should be used whenever possible, because it is the
07600 easiest to handle, but unfortunately it is not always possible as
07700 the philosophers have taken pleasure in pointing out. Under
07800 this concept, the meaning of "the person I am answering" is one
07900 Yorick Wilks. Here we rely on a convenient association of his
08000 physical body, his ideas, and his appearance, and a coincidence
08100 of the ideas of these matters shared by all the discussants.
08200 When these associations and coincidences are not so clear, there
08300 is difficulty. Perhaps some concepts can only be defined in
08400 "logical context". By this I mean rules that help you
08500 determine whether a phrase refers to Yorick and what consequences
08600 a sentence containing "Yorick" will have. Since Yorick is
08700 unproblematical, the first meaning is simpler and contains the
08800 second. A phrase like "identity crisis" has at best the second
08900 kind of meaning.
09000
09100
09200 JMC 5 MAR ON BRUCE 15 FEB INWARD SPIRAL
09300 Indeed the discussion was spiralling inwards. The
09400 solution to the difficulty is to try to extend the range of one's
09500 formalism a bit at a time in the direction of expressing the
09600 information involved in the solving of more complex problems.
09700 This means using toy problems in spite of the danger that the
09800 sequence of toy problems may not terminate in the ability to
09900 solve real world problems. A criticism specific toy problems
10000 is good, but should be distinguished from a criticism of the use
10100 of toy problems in general.
10200
10300
10400 JMC 5 MAR ON YORICK 24 FEB IMPRECISION OF PHILOSOPHY
10500 I agree with Bruce that philosophy has been imprecise,
10600 but I am not sure I can make the point precisely. Consider the
10700 fact that if I know someone's telephone I can call him. In all
10800 the philosophical discussion of knowledge, there is no sense
10900 given of knowledge that will permit making that simple statement.
11000 There is an approach to it in Hayes' and my paper
11100 (unsatisfactory, I now think), but it really is a first step to
11200 understanding knowledge to demand a formalism that can express
11300 such facts and which can be used to derive their consequences
11400 such as that looking in the telephone book will help.
11500
11600
11700 ARTHUR 7 MARCH I SPY A CATEGORY MISTAKE(I THINK)
11800 I'm not sure, but I think that in his contribution of 5
11900 March, John has fallen into the old philosophical bear-trap, to
12000 whit, that of "intensional objects". He seems to think that
12100 <Idea-of-Foos-telephone number> is somehow an individual, whereas
12200 it seems to me to be merely a FACT or piece of information. What
12300 is the problem about simply having the (TH)assertion in one's
12400 world model? John seems to be allying himself with philosophers
12500 like Willard Quine, who think that "to be is to be the value of a
12600 bound variable" without realising that we need to be very careful
12700 about the way that we qunatify into various contexts
12800 (referentially opaque) We need to have constraints on the way
12900 that we identify objects of knowledge,belief and perception.
13000 There is a definite distinction (as even Russell realised
13100 )between knowledge by acquaintance and knowledge by description.
13200 As long as we have this distinction, we can get around the
13300 traditional problems like the failure of Leibniz Law and
13400 Existential generalisation. I certainly would like to see some
13500 thoughts on how to specify these quantificational constraints in
13600 a first order language. I think it can be done, but only by
13700 having fairly well deffined notion of a context, ie. of a
13800 particular world model . In this way the tradtional problem of
13900 "the possible bald man in the door" and so on are easily cleaned
14000 up in this way.
14100 Am I misinterpreting JMC / does JMC think that my claims
14200 about a trivial solution won't do?
14300
14400
14500
14600 JMC 20 MARCH ON ARTHUR'S 7 MARCH SPYING A CATEGORY MISTAKE
14700
14800 I am not sure that Arthur's criticism of 7 March applies
14900 to anything I said on 5 March. It seems to be more applicable to
15000 things in "Some philosophical problems ..." or some things I have
15100 said in class. However, I shall defend my usage not being quite
15200 sure if that is what is being attacked.
15300
15400 Consider expressions like
15500 "value(Mike,"telephone-number(Bill)",S0)". This expression is
15600 intended to be used by the program to denote what Mike thinks in
15700 situation S0 that Bill's telephone number is. The program will
15800 use such expressions in sentences like the following:
15900
16000 1. value(Mike,"telepone-number(Bill)",result(...,S0)) =
16100 "321-2175"
16200
16300 2. value(Mike,"telephone-number(Bill)",S1) ≠
16400 telephone-number(Bill)
16500
16600 3. (∀ p)(∀ q)(∀ s)(... ⊃ value(p,subst(name(q),"x",
16700 "telephone-number(x)"),result(p,look-up-in-phone-book(name(q),s)))
16800 = telephone-number(q)
16900
17000 With the aid of such sentences, a suitable reasoning
17100 program may be able to decide how to get people to do things or
17200 to reason about whether other people can get people to do things
17300 by calling them.
17400
17500 Would writing such a program commit us to a point of view
17600 about intensional objects? I think not unless we claimed that
17700 the method was useful in all cases and also settled the matter of
17800 "the possible bald man in the door". Maybe we will get to the
17900 possible bald man in the door, but my present ideas are limited
18000 to cases in which an axiom like
18100
18200 (∀ p)(∀ q)(∀ s)(∀ y)(value(p,subst(name(q),
18300 "x","telephone-number(x)"),s) = y
18400 ⊃ y = "undefined" ∨ [y is a string of 7 digits
18500 with a dash after the third possibly preceded by an area code,
18600 etc.])
18700
18800 does not produce a contradiction.
18900
19000 I am not sure what you mean by having the (TH)assertion
19100 in one's world model means to you. If you mean that the program
19200 has it in its own world model, that is very simple. There is no
19300 need even for "value". "telephone(Mike) = "321-2175"" will do
19400 fine. For expressing general facts about when a course of
19500 action will result in someone not knowing someone else's
19600 telephone number, I don't see that the PLANNER formalism suggests
19700 a particular solution.
19800
19900 To summarize: the value(p,e,s) formalism is proposed as
20000 an approach to solving a limited class of problems. Whether it
20100 will solve them remains to be seen. It would be nice to have a
20200 definition that would give us the precise way to use such phrases
20300 as "the possible bald man in the doorway" but that doesn't seem
20400 to be in the cards.
00100 Well this seems to be the end of the famous coffee file. The
00200 next page contains comments made in January on the file as it
00300 then was. Comments by a friend of Bruce's.
00400
00500 Coffee is dead - Long live coffee!
00100 COMMENTS ON COFFEE (12 Jan. 1973)
00200 by Larry Briskman, lecturer in philosophy at Edinburgh
00300
00400
00500 Yorick 27 Nov. - Of course there are other facts than those
00600 about the nonlinguistic worls; if in some OL we talk about this
00700 nonlinguistic world then clearly there are thngs to be said aboud
00800 the st. of the OL (e.g., truth). Interestingly, there are lots
00900 of things to be said about our knowledge (linguistically
01000 expressed) of the nonlinguistic world and especially about how it
01100 can grow and be improved.
01200
01300
01400 Bruce 27 Nov. - qu. 5 - clearly yes. We already, without the
01500 psychologists, know much about human capabilities. The
01600 interesting thing is that Psychologists have been so averse to
01700 actually trying to explain how we manage to do trivial tasks
01800 (perception, motor-skills, etc.) and have instead concentrated on
01900 getting ever more esoteric data. The interesting thing about
02000 trying to explain (programme?) trivial tasks are the "unintended
02100 conseq." of the programme - we may wind up being able to explain
02200 much more than we set out to explain and hence produce
02300 independent testable pred. for the Psychologist.
02400
02500
02600 Yorick 4 Dec. - a lovely argument, but I'm not so sure that the
02700 'privileged access' criterion will work. I agree that the brain
02800 surgeon must take seriously the patient's report as a possible
02900 alternative conjecture to what is going on in the patient's
03000 brain; one which might displace the surgeon's own conjecture.
03100 But I don't see why the surgeon must take the patient's word as
03200 Gospel, as Necessary Authority. But I doubt if Wilks' argument
03300 actually crucially rests upon this "Privileged Access" assumption
03400 in that simply admitting the possibility that the computer's
03500 conjecture about its present state can displace the programmer's
03600 conjecture (based upon the programme) is enough to keep the
03700 programmer from simply "displaying" the computer's `Mind'.
03800
03900
04000 Yorick 5 Dec. - Theory of Preference Semantics (PS) vs. TP. I
04100 think Yorick has some good points to make in his duscussion of
04200 the advantages of PS to TP. Firstly, I see nothing wrong with
04300 supplying the computer with semantic preferences (i.e. meaning
04400 conjectures) which are changes only if refuted by additional
04500 information. The TP approach seems to want too much - do we
04600 really wait for sufficient information to deduce the referent of
04700 an "it" before we make any guesses about its likely referent?
04800 Clearly in language understanding we are constantly guessing
04900 (viz. our uncanny ability to finish other people's sentences for
05000 them; to understand slightly ill-formed sentences; etc); we are
05100 constantly actively looking for meaning rather than waiting for
05200 meaning to be impressed upon us. The TP approach assumes a
05300 certain indecision on our part in the face of incomplete
05400 information: but this is simply not the case. More importantly,
05500 though, is the question of whether the PS and TP are really so
05600 different -- in that the same amount of information will (or
05700 should?) be stored only in different ways. In TP our information
05800 becomes additional axioms in a PC system; in PS (semantic
05900 look-up?) new dictionary entries. Also although these new axioms
06000 will often be false (a la Wilks' example) so too will our new
06100 "preference rules" often lead us into trouble on new cases - for
06200 example, spider's webs are often in need of repair yet don't
06300 quite qualify as non-fragile, rigid objects. So PS will have
06400 application problems for its preference rules where TP will have
06500 contradiction problems from its axioms. I don't see any
06600 startling difference here. Finally, to what extent is a PS
06700 system simply building knowledge of the world into the dictionary
06800 entries of words; while TP treats this knowledge as additional
06900 axioms to be used in deductions? Seemingly the information must
07000 be somewhere; so is this simply a disagreement over how best to
07100 store it and use it?
07200
07300
07400 Yorick 8 Jan. - I think Wilks is right to insist on the
07500 importance of the distinction between `explanations of meaning'
07600 and `facts about things'. Blurring this distinction is at the
07700 heart of the view (of Kuhn, Feyerabend, etc.) that Newton and
07800 Einstein (say) mean different things by `mass' since for Newton
07900 mass is invariant with velocity (conserves); while for Einstein
08000 it varies with velocity (though producing or consuming energy as
08100 a result → mass-energy conservation). These are not two
08200 competing definitions of `mass'; they are two competing empirical
08300 assertions about mass (i.e. about the referent rather than about
08400 `mass' [viz. the concept]). To blur the distinction is to turn
08500 all `factual' assertions into implicit definitions and hence into
08600 tautologies. Now what worried me about PS was that it would lead
08700 precisely to this; can Wilks show why it wouldn't?
08800
08900 More specific reply to Bruce Questions - Wilks' critique of
09000 Chomsky is very beautiful I think, but remember that Chomsky is
09100 proposing a theory of language not as she is spoke nor even as
09200 she is used, but rather as an abstract quasi-mathematical
09300 structure; as a kind of ideal. As for the point about AI
09400 thinking that a real theory of language must be capable of
09500 expressing all information expressible in language, I see nothing
09600 wrong with this. After all, would a language- understanding
09700 programme which could only express "C20th common-sense knowledge"
09800 constitute a genuine user of the language, especially if one
09900 couldn't teach it new knowledge? As for the non-sequitur that
10000 solving gamesy puzzle-cases will solve `real' cases, I agree with
10100 Wilks that it's better to try to explain (i.e. programme) the
10200 unproblematic cases first. But surely the point of the
10300 puzzle-cases is to try to get a clue (a short-cut clue) as to how
10400 best to deal with the `real' cases; and in any case we will have
10500 to deal with problem cases at some time if we want any
10600 comprehensive theory. So what needs to be argued is that, from a
10700 methodological or heuristic point of view, it is more fruitful to
10800 forget all this gamesy stuff for the time being. This argument I
10900 don't see as having been made yet.