[Heartlogic-dev] Re: do we need to keep 'em coming?

[William L. Jarrold]

On Mon, 24 Nov 2003, Joshua N Pritikin wrote:

> On Mon, Nov 24, 2003 at 12:56:07AM -0600, William L. Jarrold wrote:
> > On Sun, 23 Nov 2003, Joshua N Pritikin wrote:
> > > What is it about a submitted story that is rate'able or interactive?
> >
> > E.g. the following (from recent real life, names have been changed to
> > protect the innocent)....
>
> I'm still a little bit uncertain.  Let's see if I got this right -

Good.  I.e. Thanks for asking for clarification.  Under time pressure,
I departed from my usual overly careful email style wrote it fast
and skipped over many details.  But, you did a most excellent job
in decoding my "telegraphic" writing.

>
> Here is the cue:
>
> > Jean, Kathy, Rachel and Bob were hanging out.  Jean had recently had a
> > sex change.  Jean, Kathy and Bob had known each other for a while.  In
> > fact, they knew each other from before Jean's sex change.  Shortly after
> > Jean and Rachel were introduced, Kathy said something to Rachel clearly
> > implying that Jean had recently had a sex change.

Basically, yes.  In the dissertation there were "scenario cues".  A
scenario cue has two parts, an overriding goal and a situation.
(Did the Glossary make it into the draft that you had?) At some point in
the future, for some items we might want to vary this format.  E.g. one
thing we will want to do is remove the overriding goal part and see how
that effects things.  During my defense (back in March) Bruce asked a
question something like, "Might you be interested in how goals interact."
So, one experimental condition for future studies might be whether or
not subjects (abreviated as SS) see an overriding goal or not.  Another
condition might be whether or not there are two overriding goals.
Or N overriding goals.  I should see precisely what Bruce said during
my oral.

Ug, there are so many bewilderingly different directions to take this work
and each new condition adds many more SS we'll need in order to test a
given hypoth associated with that condition.  Our hypoth are best checked
with virgin subjects.  I doubt we'll have a steady stream  of virgin
subjects.

I think I am going to start keeping a to do list to track everything.  I
know you have a proposed project list thing on the website too.  Should
I merge mine with that list?  Maybe, lets see how things evolve.

Anyway, although we have scadzillions diff experimental directions to go
in, our subjects will get bored fast.  That is why I sorta, only
slightly sorta, lean in the direction that we should allow the slightly
experienced subject the privelge of contributing their own scenarios.
The reason I say slightly experienced is that we want to use their
virginness towards special high priority things and towards sensitive
things....One high priority thing is the replication of the dissertation
on the www....I envision we'll say somethign to the effect of "Hey, dude,
if you do this kind of boring survey contaiing 9 or so items, don't
worry, it will get slightly less boring soon, but you have to pay you
dues first, okay?".

We should try to sit down with folks and watch them fill out the survey
get their feelings as they fill the thing out.  There are different
folks we want to sample...E.g. computer nerds and autistic spectrum types.
For the latter I have some sources...1) temple grandin herself - we need
to approach her very carefully, we want to establish a good communication
channel 2) simon baron cohen and any autism researcher (ditto) 3) I went
to a conference for CASD (a new autism communitiy in austin) and made
several contacts.

>
> Here is the affect:
>
> > Jean was offended.
>
> Since the affect is subject to experimental ablation,
> we limit the affects to one of happy, indifferent, or sad.
> Correct?  (You seem to imply this later in your message. I'm
> just double-checking.)

I suppose I'll say, yes, correct.

Strictly speaking, here's what I'll say.  In the diss, the valence of
each appraisal was exactly one of "happy", "sad", or "indifferent."
To say "Jean was offended" we are adding a new "valence", i.e. offended.

The reason I say "valence" rather than valence is that, I *think* valence,
in the literature, is by defintion a scalar quantity (i.e. one which has
pos and neg).  If we've got all sortsa emotions stuffed into the range of
the "valence" slot (e.g. all sortsa emotions such as happy, sad,
indifferent, offended, embarrased, pitying, schadenfreude...), then the
arg2Isa (sorry, I'm forcing cyc lingo on you) of valence (in other words
the range) is not really a scalar quantity any more, but more of a vector
or, who knows, maybe even a tensor.

> Here is the justification:
>
> > Why?  It would be similar to a case like if Jean had recently had breast
> > augmentation surgury.  She would not want Kathy to blurt out a private
> > sort potentially embarrassing fact....Kathy should have had the social
> > common sense, the ability to mind read that Jean would be offended
> > by Kathy's revealing Jean's sex change.
>
> OK.
>
> And we want to allow people to attach alternate appraisals
> to the same cue, correct?

Sure.  We want both. That is we want alternate appraisals and we
want believabilty judgements of those alternate apprasials.  And
we want to subject each alternate appraisal to valence manipulation.
(And someday we'll even want to manipulate the explanation too).

If my thesis about generativity is correct, then there are many many
alternate appraisals.  Sure, we can let our subjects stay happy by
generating many of them.  But the best thing is if our AI models
are able to keep up.  So, the best situation would be to encourage
"developmentally appropriate" alternative appraisals.  That is,
appraisals that require us to make simple modifications to the current
state of the AI model(s).

Mostly this is wishful thinking.  Perhaps it is worthwhile to explore
it a little just to be prepared.

>
> For example, "happy" because "Jean is a sex-change activist.
> She likes to make sure everyone knows about her sex-change
> when they first meet her.  She is a leading political
> advocate of sex-change legislature."
>
> Of course I am mixing up the cue & appraisal, but that
> is a confound we are likely to see anyway.

Right.  Sometimes people add stuff to the reason (recall that
reason is the justification for the valence.  I.e. reason is a
part of the appraisal) and it is a fair addition.  Other times,
people put an absurb amount of supposition into the appraisal.
Where the boundary lies is fuzzy....I suppose the boundary
shifts as the model gets more sophisticated.  What we like are
useful embellishments to the the justification.  Right now our
model barely knows about an ontology of fruits and vegetables.
Thus, adding an embellishment like "What Tracy really wanted
was fruit." poses a developmentally appropriate challenge
to us modelers.

>
> OK -- I am ready to code this.  Just confirm that my
> understanding (above) corresponds to yours.

Yup, I think it does.

>
> One more question - User submitted cue/appraisals will be
> segregated from KR derived cue/appraisals, correct?
> Or do you want to mix everything together somehow?

I'm not sure what you mean by segregated so I'll answer
for both possible meanings...Well, certainly we and our computers should
know what is human versus computer generated.  That was
the first meaning of segregated.

Now for the second meaning.  As for our users, if I
was forced to make an either or choice, I'd say lump them
together.  Why?  Because we don't want our users to be
biased....E.g. we don't want our users thinking thoughts
like this, "Oh this one is human generated, so sure, I
can see it as being believable.", or, "Oh, this one is
computer generated, no wonder I went, 'huh???' whe I read it.
It is certainly unbelievable."...Now, maybe after a user
"earned his/her keep", e.g. by responding to a good number
of survey questions, then they'd be given "the keys to the
kingdom", i.e. the privilege of being able to see which
items that they just viewed were human vs computer generated.

>
> > Now, what is interactive?  Well, I mean interactive only in a very broad
> > sense.  Interactive in the sense that the survey takers will have a chance
> > to chime in about the type of survey items that they think others should
> > be answering.  If surver takers are able to have more of a voice, they
> > will feel more involved, more passionate about it and will answer
> > more survey queries.  That is what I mean by interactive.
> >
> > Does that make sense?
>
> Yah, I think so.
>
> And eventually we can do a KR model to achieve simulational
> adequacy, with this material as a regression test or something.
>

And who knows, maybe the slashdot flavor of users will be able
to contribute useful rules, useful ai models.

> And there might be clinical applications, as you mentioned -- cool.
>
> --
> A new cognitive theory of emotion, http://savannah.nongnu.org/projects/aleader
>