|
ESSLLI 2001 Course
The information state approach to dialogue management: theory and
implementation
Lecturers: Robin Cooper and David Traum
(with assistance from other TRINDI consortium members: Johan Bos, Staffan Larsson, David Milward)
Type: advanced course
Section: Language and Computation
Description: The EU project TRINDI (Task Oriented Instructional
Dialogue) has developed an approach to formalising
computational theories of dialogue processing using
information state update. The approach leads to straightforward
implementation of dialogue management models based on
these theories making it possible to experiment with
different kinds of information state and update rules and
enabling the rapid prototyping of systems. To this end
the project has developed a toolkit called TRINDIKIT and
a number of implementations to illustrate both the kit
and theories of dialogue processing. The project has
resulted in a book manuscript which will be used as the
textbook in the course. We will give an overview of the
general approach as well as presentations of the
information state theories we have investigated and
systems we have implemented. If feasible, we would like
to give some hands-on experience in using the TRINDIKIT.
The following lectures will be given (references are to the
TRINDI book draft):
- 20th Aug, 11-12.30
The information state approach to dialogue
management (David Traum)
Reading: Chaps. 1, 2.1-2.2
Introduction to the TRINDIKIT (Staffan Larsson)
Reading: Chap. 2.3-2.5, App. A; slides
- 21st Aug, 11-12.30
GoDiS: a dialogue system based on questions under
discussion. (Staffan
Larsson)
Reading: Chap. 3; slides
- 22nd Aug, 11-12.30
The Poesio-Traum theory of dialogue and the EDIS
implementation. (David Traum)
Reading: Chap. 4
- 22nd Aug, 11-12.30
MIDAS: a dialogue system based on DRT which
exploits first-order theorem proving (Johan Bos)
Reading: Chap. 5
Robust semantic processing in dialogue systems (David Milward)
Reading: Chap. 8
- 24th Aug, 11-12.30
Towards flexible and adaptable dialogue systems: work on the
SIRIDUS and D'HOMME projects (Staffan Larsson, David
Milward)
Discussion (slides)
Literature: The TRINDI book
The toolkit and some of the systems we are discussing are downloadable
from TrindiKit
web page where you will also find the manual for the kit.
Prerequisites: Basic knowledge of computational linguistics,
computation, and computational semantics (including AVMs,
logical inference, DRT and pragmatics). Some previous
knowledge of dialogue management would be helpful.
Course requirements: If you wish to obtain credit for the
course we suggest that you carry out a small project involving one to
two weeks of work during the three months following the course. Here
are some suggestions for the kind of project that could be done. If
you wish to do this please mail both Robin Cooper (cooper@ling.gu.se)
and David Traum
(traum@ict.usc.edu) and discuss your plans with us. On successful
completion of the project we will be able to provide you with a
document saying that you have taken the course and describe the nature
of your project. This can possibly be used to help you obtain credit
for the course at your home university but it is up to you to arrange
this.
Proposed TRINDI course projects:
- Write a short paper (5-10 pages) discussing how part of your
favourite theory of dialogue might (or might not) be formulated
using TrindiKit formalisms.
- Take one of the existing systems provided in the TrindiKit
download and add some new dialogue behaviour to it.
- Take one of the existing systems provided in the TrindiKit
download and (begin to) port it to a new domain or language.
(The result should be a working system which we can evaluate,
even if it covers very little.)
- Add an existing module or resource that you have worked with
before but which is not part of the TrindiKit. Examples are
speech recognition and
synthesis modules, NL interpretation and generation modules,
reasoning module, planning module, parser, database resource
interfaces (e.g. SQL, XML), device interfaces etc. It may be
something you have written yourself or something that is
available off the shelf.
Implementation projects that can potentially be included in future
releases of the TrindiKit are particularly welcome.
|
 |