[Aleader-dev] Re: 2 papers of interest

[William L. Jarrold]

On Mon, 21 Jul 2003, Joshua N Pritikin wrote:

> On Sat, Jul 19, 2003 at 11:18:02PM -0500, William L. Jarrold wrote:
> > Two papers.  One is about testing and has some simple emotion inference
> > paths in CycL.  Another is about emotion KR methodology and probably
> > has some emotion inference paths.  Would love to hear any comments.
> > Might help with your understanding of how to teach Cyc about emotion.
>
> I mostly understand how to teach a KR system about emotion.
> What really helped was looking at the code you wrote for KM.
> Seeing code that works is a hundred times better than some
> design proposal & hand-waving.  I still haven't finished
> reading the code or the KM manuals but I will eventually.
>
> I guess my reaction to your papers is that I don't consider
> the KR part of the the project to be the most interesting part.
> I know I need to do KR, but what is interesting to me is the
> performance of the emotion predictor.  I believe I am really
> breaking new ground.  I feel like I am getting better
> repeatability with this model than with other models I've seen.
> Even recognizing the importance of repeatability seems to be
> innovative, relative to the literature I have read.

Reliability is jargon in psychometrics that refers to pretty much
the same thing to which you are referring.

It should not be hard to studies that have looked at the reliability of
judging emotion in pictures of people's facial expression.

It seems to me that one problem with the scale of your problem (looking
at raw slices of film rather than just e.g. faces) is that any one
character looks like they are experiencing several emotions at once.

>
> I admit that this is all subjective.  As the author, I am
> probably the person least trusted to make these kind of
> subjective judgments.  But there it is.
>
> Given that you've authored a bunch of papers on the KR of
> emotions, it seems to me that you are more interested in the KR
> issues than in a particular cognitive model of emotion.  Would
> it be correct to say that you don't care about the cognitive
> model, per se, but rather how to do a good KR?

Yah, that might be true.  What would really rock my world is an AI that
could infer the emotions felt by characters in stories as well as a 4
year old.  Right, I don't care so much how it does it, what theory is
inside.  I just want the damn thing to work.

> Even if this is
> your perspective and interest, I think you will find that the
> Aleader model is already well suited to KR.  To contrast, you
> had to modify and narrow Ortony to make it tractible for KR.

More precisely, to make it tractable for a small short term KR project.
I believe that KR can encompass nearly all of OCC, but it will take
zillions of person years of effort to get there.

> Furthermore, I speculate that you looked at a lot of even less
> suitable cognitive models before selecting Ortony.

Actually I just saw an article in AI Magazine that reviewed OCC.
That was maybe around 1993.  I focused on OCC because it seemed good
enough, the AI community was behind it.  So many people have told me
(including myself) that I need to focus that I put on blinders and
marched forward.

>
> So whether you are interested in the KR aspect or a model
> that feels good, I think the Aleader model is a good choice.  ;-)

Good.

>
> > Thanks for your emails today.  More later (hopefully Sunday, my time).
>
> --
> Victory to the Divine Mother!!         after all,
>   http://sahajayoga.org                  http://why-compete.org
>