NetNews Usenet Archive 1993 #3/NN_1993_3.iso/spool/comp/ai/philosop/7419

Text File | 1993-01-27 | 11.8 KB | 240 lines

Newsgroups: comp.ai.philosophy
Path: sparky!uunet!inmos!fulcrum!bham!bhamcs!axs
From: [e-mail] (Aaron Sloman)
Subject: Re: Colbys-Weizenbaum etc Belated Replies (1)
Message-ID: <[e-mail]>
Sender: [e-mail]
Nntp-Posting-Host: emotsun
Organization: School of Computer Science, University of Birmingham, UK
Date: Wed, 27 Jan 1993 23:51:41 GMT
Lines: 228

Various people commented on some of my earlier postings and I have been
too busy to reply. So here goes. First

From: [e-mail] (Manfred Stede)
Message-ID: <[e-mail]>
Date: 17 Jan 93 23:07:18 GMT
Organization: Department of Computer Science, University of Toronto

[AS]
> > > >If some people are so badly educated that they take the use of "I"
> > > >in computer print-out to mean the machine has a mind, rather than
> > > >just being an extension of the use of "I" when a book includes
> > > >something like "in that situation I would advise you to take plenty
> > > >of rest...." or whatever, then we need to improve the education of
> > > >the general public.
[MS]
> > > It's not just the word "I".  If the rest of the dialogue were
> > > obviously mechanical, perhaps no one would be fooled.
[AS]
> > Yes - but the point is the same. I see no more reason why what's
> > printed out by an expert system should use clumsy stilted English
> > than what's printed in a book. Of course, it depends on what the
> > task of the program is.
[MS]
> But the book is a fairer partner: it only talks to you.  The dialogue
> program pretends to "understand".

That depends on the program. Some books "pretend" to talk to the reader
because they include, for instance, rhetorical questions. People who
don't understand the relevant literary conventions may be confused by
this. That's no reason for banning such books. Some books pretend to be
based on divine revelation. That's a lot worse. But I'd still rather
educate people than ban the books.

Just because a program interacts with the user that does not mean it is
pretending to do something it doesn't do. To a limited extent a database
program understands insofar as it produces the required information in
answer to questions. Why should that be OK and not something that helps
people with personal problems? I think it's nothing but sheer prejudice
and a Weizenbaum-type romantic human-centrism that objects to this.

Incidentally "understands" is not an all-or-nothing concept, as I tried
to show in
    A.Sloman
    `What enables a machine to understand?' in
    Proceedings 9th International Joint Conference on AI,
    Los Angeles, August 1985.

> What it does (if it's a little
> smarter than Eliza) is likely to be a syntactic analysis and some
> semantics on top of it, which disambiguates words, attaches PPs
> correctly etc.  This is all right as long as we deal with a well-defined
> sublanguage that the program has been designed to handle. There are
> domains where it makes a lot of sense to assume such a sublanguage,
> and several expert systems with NL interfaces exploit these.  I claim
> that AI/NLP can be quite useful in such domains.
>
> The origin of this discussion, however, was a program aiming to be an
> artificial psychotherapist.  Someone's posting advised to boil this
> application down to the absolute minimal requirement: a tool that you
> can talk to, that munches on the words for a second, and throws some-
> thing back at you, such that some of your words reappear, and you do
> that for a while, and you achieve self-elucidation.  Well, if this is the
> purpose, and I don't care about what the program utters anyways, I might
> as well talk to my cat.  Or to myself, which may in fact be much more
> useful, because I could give myself sensible answers, couldn't I.

Speak for yourself, but don't assume everyone else has your reactions to
everything. People like you would feel no need to interact with the type
of program you are discussing. Agreed. But that has no implications at
all for people who are not like you. Why should you object to others who
find it soothing, refreshing, insightful, encouraging, or helpful in
some way? (Compare writing in a diary: "Dear diary ....". Some may
prefer talking to a machine that produces responses.)

Actually even a simple Eliza can produce more surprising responses than
the sort of system you've described. Without having interacted with the
Colby's system, I have no idea how much more advanced it is than Eliza.
I thought you were objecting *in principle* to any attempt to offer
people any kind of psychotherapy via a program using current expert
system/natural language technology. If you are not objecting in
principle but only saying you wouldn't find it useful, then perhaps we
have no disagreement.

> No need for a self-elucidation program.

Not for you perhaps. So what?

> If the standards remain a little higher and we truly want an advice-
> giving system, then one of my objections is that with the current state
> of NLP it just won't work and would not be of any use: There is no
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> sublanguage to be pinned down here, to which you could connect your KB
> such that you get some inferences right and come back with useful
> advice.  The program designer cannot anticipate what the problem of the
> patient is, what s/he will be talking about, and how s/he will be talking
> about it.  And the user is bound to notice at some point how little that
> dialogue partner actually grasps; frustration is a likely reaction.

If your claim is an empirical one that all contemporary programs of this
sort will actually be useless, then only empirical investigations can
establish this, not armchair (typing-chair) pronouncements: anything
else is pure prejudice unless you have a more advanced theory of how the
human mind works than anyone else I know has, on the basis of which you
can demonstrate exactly why nobody in need of therapy will be helped. If
you had such a theory, it could probably be the basis for the design of
a good therapy program.

You find it surprising that anyone should be helped by a limited
"language understander". I find it surprising that some adults should be
comforted by clutching a teddy bear. All that proves is that neither of
us has a good theory of how human minds in general work.

> (That's the technical point.  Aside from that -- doesn't psychological
> advice depend on comrehension abilities that go a little further than
> literal "understanding" (getting the predicate/argument structure right)
> of sentences?)

If the advice is supposed to be that of a scientist prescribing remedies
based on knowledge of the system needing repair then you are right. If
the "advice" is just a tool that makes some people better, then HOW it
could work may be as mysterious to you as the effect of music, or drugs,
or religion, or reading poetry. But that doesn't mean it can't work.

[AS]
> > I'll backtrack a little: if the vendors know there's a good chance
> > that because of inadequacies in the current educational system users
> > of a package are likely to be fooled into attributing capabilities
> > to the machine which it really does not have, then perhaps the
> > machine should from time to time print out a suitable warning
[MS]
> This may be not so easy.  In the restricted-domain system, the program
> designer can anticipate what people will say and not say.  If there is
> need for a warning, it may be sth like "don't use complicated sentences,
> don't abbreviate 'something' to 'sth', and talk only about bathtub-fixing
> (or whatever it is) and nothing else."

That's not the sort of warning I meant. Rather something like:

    PLEASE REMEMBER THAT YOU ARE TALKING A COMPUTER PROGRAM NOT
    A PERSON. THIS PROGRAM MERELY FOLLOWS RULES THAT WERE DESIGNED
    TO GENERATE HELPFUL INTERACTIONS, BUT IT DOES NOT REALLY
    UNDERSTAND WHAT YOU ARE SAYING, AND SOME OF THE RULES MAY
    NOT WORK WELL IN YOUR CASE.

    IT IS POSSIBLE THAT YOU HAVE PROBLEMS REQUIRING PROFESSIONAL
    HELP RATHER THAN THIS KIND OF COMPUTER THERAPIST.

and occasionally shorter or longer reminders. Of course people should
have the option to turn this patronising stuff off, if they don't need
the reminder (or think it insults their intelligence.)

[MS]
> If the sublanguage cannot be restricted due to the domain (as it is
> the case with the psychotherapist), it becomes quite difficult, if not
> impossible, to explain to someone with no linguistics background what
> the program can handle and what not.

So what? Someone given the above warning does not need to have an
explanation of precise details of the program's scope and limits, if
talking to the program helps them, i.e. makes them feel better.

You can tell someone that it's unsafe to drive round certain corners
faster than 40kph without having to go into the detailed analysis of
friction, momentum, centrifugal force, etc.

If the program were giving technical advice on how to cure diseases by
administering potentially very dangerous drugs, or how to maintain a car
so that it is safe to drive, etc. then I would be MUCH more concerned
that the technical limits, and especially the limits of reliability of
the program, were made absolutely clear to the user. (The same applies
to advice given in a book. A book on baby care must make it clear that
in certain cases the infant should be taken to a proper doctor for
diagnosis since the diagnosis cannot be done by an untrained person
using the book.)

If it turned out that for some users the program produced suicidal or
other dangerous tendencies then that would be a reason for restricting
its use. If the results are only either helpful or neutral, as I
suspect, then much too much fuss is being made.

[MS]
> One conceivable, if humble, first
> step is to list upon request all the words in the system's dictionary,
> but this doesn't help much, either, since it's unclear what these words
> "mean" to the system, that is: what inferences does the system draw when
> these words appear (depending on the context, if the system is smart
> enough).  --  Yes, there are expert systems that try to "explain" how
> they came up with their answer (by explicating the rule-chain), but
> this again presupposes a well-restricted domain and language.

I think you are approaching this whole thing from the point of view of
an expert in the field who wants to know things about the program you
are interacting with. There's no reason at all to assume that every
customer will share your concerns, or should share them, any more than
everyone who buys a car needs to know how a carburettor works, etc.

[MS]
> [e-mail] (Jeff Dalton) writes:
[JD]
> > In fact, it can be fun (at least) to "talk" with a program that
> > constructs replies by randomly linking overlapping subsequences
> > of earlier input (sort of like an interactive dissociated press).
[MS]
> Absolutely right.  But then let's keep it that way and advertise such
> "dialogue programs" as "a toy you can have loads of fun with, because
> it pretends to have some idea of what you're saying, yet it's comple-
> tely dumb.  Come and find out how dumb it is."
>
> No reason to get AI involved in this business, no need to advertise
> anything as a computer program that analyzes your psychatric problems
> for you and helps you solving them.
> Otherwise, the message will be "Come and find out how dumb AI is."

Here you seem to be saying: none of this stuff works. Well if you have
tried it out in a systematic way on the sorts of people it was
intended for, that's fine, and you should publish a detailed report
debunking it.

On the other hand, if you haven't done that, you have no right to assume
that the people who use such software will gain no benefit from it, or
that in doing so they will be deceived in some way.
Aaron
--
Aaron Sloman,
School of Computer Science, The University of Birmingham, B15 2TT, England
EMAIL   [e-mail]  OR [e-mail]
Phone: [phone]       Fax:   [phone]