Re: >H Eliza

[goetz@zoesis.com (Phil Goetz)]

Transhuman Mailing List

> Transhuman Mailing List
> 
> Phil Goetz wrote:
> 
> > I argued, partly but not complete tongue-in-cheek, in a post called
> > "Believable stupidity", that people are more like Eliza than they are like
> > logical information processors.
> 
> Phil, would you mind re-submitting that post or forwarding it to me?  If your
> right, perhaps we can expect believable agents sooner that we might otherwise?
> 
> Paul

I wouldn't go that far -- really, please; there are already plenty of people
saying "We don't need to make agents smart, we just need to make them
/believable/".  I don't agree with that.  But here's the post.
It was posted to rec.arts.int-fiction



 			BELIEVABLE STUPIDITY

I. We Are Eliza

Every year the Loebner competition pits different natural-language
understanding strategies against each other, as teams of researchers
compete to see whose computers can hold conversations the most like humans.
Every year the sophisticated programs with complex knowledge representations
and reasoning and planning strategies lose out to some variant of Eliza,
a program Joseph Weizenbaum created in 1966.

Eliza works basically like this:  The computer has a list of templates,
which are generalized patterns that sentences can match.  Each template
has a transformation which specifies a way to rearrange and add things to
or subtract things from any sentence that matches that template.
Every time you give it a sentence, it finds the first match in the
list, does the transformation, and prints the result.  No memory,
no understanding, nothing but stimulus-response.  For example,
you might have these templates and transformations:

	Template				Transform

  I think $1.				Why do you think $1?
  I $1 don't know $2.			You sound uncertain.

People have puzzled for years over why it is that Eliza still simulates
human conversation better than sophisticated knowledge representation
and logical reasoning.  Three decades of research in computational
linguistics has not given us an answer.  A few hours of conversation
at your local bar, however, suggests an answer:  People are more
like Eliza than they are like sophisticated knowledge representation
and reasoning systems.

A lot of the people I've met can be modelled pretty well with Eliza.
Consider the following script:

	Template			Transform

  You could $1.			Don't tell me what to do!

  Yes, but $1.			Why are you always criticizing everything I say?

  $1 is so cool.		Like, you are just sad.  $1 is lame.

  $1 lower taxes $2 |
  $1 our children $2 |		Hey!  I'll vote for him!
  $1 this great nation $2

  $1 Alaska $2.			Did I ever tell you 'bout the time I was
			up in Alaska hunting moose with Clem?  Think it was
			four, five years ago, that year we had the frost
			that killed all Molly's orange trees she'd worked so
			hard on.  I told her the damned things wouldn't grow
			this far north.  So anyway, there we were in
			Alaska...

  $1 Clinton $2.		Bill Clinton is the biggest goddamn
			hypocrite in the U. S. of A.  Supposed to be out
			leading the country and here he's having floozies
			go down in him in the oval office.  God damn.
			It's no surprise them damn Japanese are whupping
			our asses.  Hey, Mel, pour me another.


II.  Stupidity Makes the World Go Round

I was thinking about how to write Eliza scripts or something else for
an interactive drama system that could generate responses that would
convince the user that the system understood.  Then it occurred to me:
Most in the drama in my life doesn't come from the intelligence and wisdom
of the people around me, but from their stupidity.  Person A says one
thing.  Person B filters it through their prejudices and preconceptions
about person A, adds in a few assumptions to make it all gel better,
then gives person C a summary.  Eventually word gets back to person A,
who concludes person B is maliciously making up lies about A to further
certain evil intentions that person A has always suspected person B of
harboring.  Instant drama.

But -- as we know from interactions with unsuccessful computer programs --
just any old stupidity won't work.  Simple ignorance or failure to
understand results in programs that say "Huh?" and "I'm not in the mood
to talk right now" a lot.  To create drama, we need to push the stupidity
envelope, to create agents that are aggressively stupid.  Whenever
they are confronted with a user action which makes no sense, they
should make assumptions until it makes sense.

The trick is to produce /believable stupidity/, by making our agents be
stupid in the same ways that people are stupid.  To do this, think about
the misunderstandings you've had with other people, and what caused them.
From my experience, I suggest the following rules for

			How to Be Stupid

1.  Have preconceived notions about every other character:  a label,
a category, an expected way they will act, a model of their goals and
motivations.  If character A does something so out of line with character
B's categorization of A, B must have a set of other categories and
expectations and quickly squeeze A into one of them.
Specifically, for each character, have ready to hand a list of suspicions
about that character which can be used to explain any action.

2.  Assume that everything the other character does is directly related
to you.

3.  Never attribute to forgetfulness, incompetence, coincidence, or a lack
of information what can be explained by hatred, dishonesty, or conspiracy.

4.  When confronted with conflicting information, never perform inference
to find a belief whose alteration could resolve the conflict.  Only
question one of your own beliefs if it is directly contradicted by
physical evidence.  To do otherwise would not be believable.


Here is an imaginary interaction with an intelligent agent:

  Player:  Moe, what is the weather like?

  Moe:  I don't know what the weather is like.

  Player:  Moe, where is Curly?

  Moe:  Curly is in the foyer.


Here is the same scenario played out with an aggressively stupid agent:

  Player:  Moe, what is the weather like?

  Moe:  Why do you want to know?

  Player:  I want to take a walk.

  Moe [matching on "I $1 walk $2."]:  Hah!  You want to go see Larry and
tell him what you learned from listening behind my door.  But I'm on to you!

  Player:  Moe, where is Curly?

  Moe [matching on "Curly" in the context that the "tell Larry" suspicion
is active]:  So Curly's in on it too, is he?


The second interaction is clearly more dramatic.

The nice thing is that these kinds of responses can probably be
implemented partly with Eliza-like scripts, which just match a few words
in a sentence and assume the rest does what it is supposed to.
There will be some knowledge representation requirements, but it shouldn't
amount to more than remembering which labels an agent has put on each
character, providing enough reasoning capabilities (or perhaps simply a
list of cues) to confirm each suspicion, and a list of which suspicions
are currently active and which have been confirmed.  The ability to
retract confirmed suspicions should not be necessary.


Thoughts?  Template matches?

Phil

"Give a man a fire and he's warm for a day --
set fire to him, and he's warm for the rest of his life."
***************************************************************************
* If you need to unsubscribe, send email saying "unsubscribe transhuman"  *
* to majordomo@logrus.org, without the quotes, from the SAME email address*
* as the one you subscribed under! Yes, it is case-sensitive. Don't blame *
* the admin for your spelling errors.                                     *
*                  Please email all technical problems to                 *
*               owner-transhuman@logrus.org, NOT to the list.             *
***************************************************************************