|
TRINDI (Task Oriented Instructional Dialogue) is concerned with the
use of natural language in human-machine interaction that enables the
human to make choices in the performance of a certain task (task oriented instructional dialogues). As one
example of this we consider Route Planning as a basic scenario. The
application of the research to other scenarios (such as Machine Repair
and Software Help) is also being considered. The emphasis of the
project is on finding an appropriate way to represent the information
which is exchanged in a dialogue. In doing this the project builds on
results from a previous LRE project FraCaS (A Framework for
Computational Semantics). This is a prerequisite for the
development of technologies for machines with more understanding than
those presently available in dialogue technology.
The project began work 1st February, 1998. One of the first tasks was
to characterize and evaluate the three scenarios and define their role
in the project. This gave us a focus to do a brief survey of
currently available technology, looking at a number of dialogue
management systems and dialogue design toolkits. We needed to develop
a metric for evaluating such systems in terms of their ability to
handle information exchange, that is, to quantify the degree of
understanding present in a system or allowed for by the tools in a
design kit.
We needed to find a computationally viable way of representing the
process of information exchange in the kind of dialogues we are
concerned with and to test this by annotating sample dialogues and
instructional texts.
Final versions of deliverables on these topics are scheduled for
February, 1999 after the completion of the first year of the project.
This work will position us for the building of a computational model
and implementation.
We are mainly concentrating on the Autoroute Scenario (of the three we
are concerned with). It is a route planning task for which we have a
working (keyboard input) dialogue system at SRI Cambridge.
The application that actually plans the route (`Autoroute') is a
commercially available PC software package distributed free by many
suppliers bundled as part of Windows. The
Autoroute scenario is typical of those for which
implementations of spoken dialogue systems are commercially
attractive: the task involves acquiring certain pieces of information
- too many to be displayed as menu items - and then optionally some
further details which can then be used to interrogate an application
information system.
Several projects within Europe have been looking at on-board
navigation aids as an attractive and useful feature for the car of the
future, and some simple commercial voice-driven navigation aids are
already available. Voice operation has obvious advantages in the
driver's situation.
In surveying current dialogue systems and toolkits we focussed on how
and to what extent they are sensitive to the current dialogue context,
and how and to what extent they can use dialogue in furthering the
user's objectives.
The commercially most advanced of the systems we looked at, Nuance's Dialog Builder application, is a means
for declaring who can speak on the basis of generating or recognizing
predefined utterances. Making a proactive contribution requires one
to have aims and objectives relating to the conversation based on
representations of information already exchanged. The Trains system is the most complex of those we
looked at in this respect and the system continually attempts to infer
the user's overall plan.
The choice currently is between robust systems with little or no
attempt at understanding and knowledge intensive systems which are
limited by their brittleness. There is a need for robust technology
enhanced by representations of information that will aid systems to
make proactive contributions in limited task oriented dialogues.
Why do people engaged in a dialogue say what they say when they say it
and what effect do their contributions have on the information
available to the participants in the dialogue? These are central
questions for the automation of dialogue.
Machines need to have simple strategies that simulate aspects of human
dialogue behaviour. The strategies need to be simple enough to be
computationally robust and yet complex enough to make the dialogues
coherent and not too irritating for the user. The simplest strategies
involve just producing particular utterances in sequence or directly
computing a response on the basis of the preceding utterance from the
user. A common more sophisticated approach is to treat dialogues as
games and utterances as instantiations of
a limited set of moves (such as question,
answer, clarification, request, confirmation, etc.). Allowable moves
can be specified in terms of transition networks.
Our approach is to enhance this conception by associating with
dialogue moves representations of the
information that the dialogue participants have. The main
effect of an utterance is to change this information in some way, and
the information is used by the participants to decide what to do next.
Natural language is multiply ambiguous, and disambiguation is one of
the hardest and most expensive tasks in language
understanding. However, disambiguation is not always required. Often
the information relevant to the current state of the dialogue is
common to all the readings of a dialogue contribution and it is not
necessary for the dialogue participant to resolve the ambiguity. Tools
for the representation of underspecified
information about meaning and techniques for disambiguation on
demand were explored and developed in the FraCaS
project and we are investigating how these tools can be used to
enhance the computational model of information state revision.
Even with underspecified interpretation, classical interpretation
techniques lead to brittle systems which are unusable on free natural
language input. We are investigating how robust
techniques like finite state technology can be applied to or
combined with semantic processing. For example, it may be possible to
characterize transitions over information states with finite state
machines.
The project has created an International
Consultation and User Group (ICUG). Currently the group has 25
members (11 of whom are from industry). We plan to increase the
membership during the second year of the project. We welcome
unsolicited expressions of interest in joining this group.
All members of the ICUG will be invited to the two public workshops
which the project will organize at which project work and plans for
the future will be presented. These workshops will include invited
speakers external to the project and members of the TRINDI ICUG will
be invited to comment on the progress and future plans of the project.
The first of these workshops has been scheduled for 6th May, 1999 and will be held in conjunction
with Amstelogue'99
(Amsterdam Workshop on the Semantics and Pragmatics of Dialogue),
Amsterdam, 7th-9th May. We have chosen to do this in order to maximize
the number of people who will be able to attend our workshop.
Selected members of the ICUG will be invited to consortium meetings
for which their expertise is of particular relevance and some members
of the ICUG will be invited to work with the project on particular
tasks.
In addition project participants have made regular contributions to
international specialist conferences during the year.
Deliverables associated with the work described in this report will be
completed in February, 1999 and will be made publically available on
the web and in hard copy.
Work scheduled for the second year of the project (beginning February,
1999) includes the building of a dialogue move engine to illustrate
how representations of information states can be associated with
dialogue moves and assist in the determination of next moves. We will
also present the theory that underlies this and give an overview of
the practical possibilities for ways of representing information
states and the kind of information that can be included.
Progress will also be underway towards results that are to be
presented in the final six months of the project in February and May
2000 including prosodic clues to information structure and the
representation of focus in information states.
Further information about the project can be obtained on the TRINDI web
page or from the project coordinator:
- Coordinator
- Robin Cooper
- Email
- trindi@ling.gu.se
- Tel.
- +46-31-773-2536
- Fax
- +46-31-773-4853
- Mail
- TRINDI Administrator,
Department of Linguistics,
Göteborg University,
Box 200,
S-405 30 Göteborg,
Sweden
|
