ERIC WILLIAMS

Chu-Carroll and Carberry

As I read the list of giants upon whose shoulders the authors stood, I was 
mightily impressed.  Then I noticed that though Clark was mentioned, his 
buddy Horvitz was not.  I was further dismayed by this upon reading about 
the evidential support system they designed.  It's reasoning under 
uncertainty - Horvitz's specialty!  Worse yet, it's not nearly as nice.  I 
really dislike the discrete certainty measures (strong, weak, etc).

RW: Footnote 11 may be relevant.

I was also unsure how the logical propositions were constructed from dialog 
(ie On-Sabatical(x)).

RW: I'm not clear either, but discussion on p. 388 may be helpful.

CORE seemed awfully complicated to me.  I don't know if I was just too 
tired to absorb its thoroughness, or if they're really trying to swat a fly 
with a Buick.

RW: It seemed quite complicated to me as well, leading me to try to reduce the
cognitive load by synthesizing the works at a high level, suppressing the (voluminous)
detail.

Zukerman and Litman

Interesting historical summary.  If I wasn't so tired, I might be able to 
form questions about it.  Oh well, c'est la vie.

RW: I used it to select the Kashihara et. al and Chu-Carroll & Carberry articles. 
I don't think the Zukerman & Litman piece is 'merely' historical: it is both analytical and
prospective.

Walker, et al

Mmmmm...gigantic paper number 3...eyes crossing, brain turning to mush....

Interesting ideas.  At the risk of beating a dead horse, I'd like to point 
out that this, too, would mesh well with Horvitz's work.   I was also 
reminded of Diane's reinforcement learning research.

RW: Please elaborate: are you thinking of a BN approach applied to SPG or SPR?

It's a good thing I lost the other paper cuz I'd be too stupid from 
exhaustion to read it. ;)

RW: Although I read Kashihara numerous times, and that helped, I can't say I reached clarity about
it in detail.

================================

ANTONIO ROQUE

"Natural Language Processing and User Modeling" p. 151 discusses the need
to study which additional advances in machine learning and reasoning under
uncertainty are needed to support new capabilities in natural language
systems.  Given the warning on p. 149 about intuition not being enough to
motivate work, how should we go about deciding which advances to pursue?

RW: My copy isn't numbered: what section/subsection are you referring to?

================================

AMY SOLLER

What hard-core evidence do we have that tailoring dialog to the user via a 
user model in the specific ways that have been done in the described 
systems (which seem to very quite widely)  is useful? What are the 
potential consequences if the predictions about the user's knowledge or 
inclinations are incorrect?

RW: I'm not in a position to provide an authoritative response to the first question, and much
depends on what one is willing to count as "hard-card evidence." It is certainly possible that the
payoff from user-modelling could be quite small in domains where the variability of user
characteristics is low and/or the the effect of such variability on some perfromance metric is
negligible. In general, however, there is (hard-core?) evidence that adequately individualized
responses lead to greater system use.

================================

ALAN BERFIELD


Zukerman & Litman

Has the community embraced the issues in this paper? Has any work been
started yet that attempts to address them?

RW: This question is for Diane, I think.


Kashihara et al.

They don't mention much about the ordering and interaction between the
three task stages. For example, if the follow-up stage (stage 2) reveals a
misunderstanding, would a system go back to stage 1 again or would the
further explanation be a new cycle after stage 3 completes? Also, are they
forcing all 3 to always happen?

RW: I agree that more discussion about ordering/interaction would have been helpful. It appears to
me that the system does not attempt to recognize misunderstandings as such, but rather assesses the
match between the evolving Understanding Structure of the student and it's own co-evolving
Explanation Structure. My guess is that an impasse is recognized by a lack of change in US. I'm also
guessing that an impasse may lead to a follow-up question from student/system concerning a
component/complement or a restriction negotiation of the ES by the system if the system inquires
about a different component/complement. If by stage 3 you mean restriction negotion, I think it is
necessitated only by certain follow-up clarifications by the student.

===============================

THERESA WILSON

Zukerman & Litman - "Natural Language Processing and User Modeling"

This was a good review article that proposed a number of questions
of its own and challenges to future research.  One challenge that
the authors raised was evaulating the contribution of user models
and natural language systems in general.  As the authors point out,
"generally accepted guidelines for the evaluation of user models
and natural language systems are still forthcoming."

I'm interested in what the class thinks some of these guidelines
for evaluation should be.


Chu-Carroll & Carberry - "Response Generation in Planning Dialogues"

In this paper, the authors give a detailed look at the types of
responses typically used in a collaborative negotiation, and they
develop a planning system for working with these responses.
Has a similar sort of planning been applied to argumentation
and non-cooperative dialogues?

RW: Stefanie is the expert on CBR.

=======================================

STEFANIE BRUNINGHAUS

Walker et al.: 

What is the tradeoff in developing a trainable system? It seems that
getting labeled training data is the bottleneck here, but where is the
breakeven point in terms of effort for building a rule-based system
compared to generating enough training data for a trainable system? 

RW: I believe that, in this instance (see p. 12), the labeling is done by the SPG, so that labeling
itself is not the bottleneck (though created the SPG may be).

Chu-Caroll & Carberry: 

Would case-based reasoning be applicable here? The authors already
discuss that unsatisfactory plans are not completely thrown out, but
rather repaired (or, one could say, adapted), suggests that CBR could be
a good idea. 

RW: Being completely without knowledge about CBR, I wonder what the cases would be, how they would
be acquired by the system, and how it would aid with selecting a justification chain.

--