|
Project Name: SCHISMA: SCHouwburg Informatie Systeem
Abbreviation: SCHISMA Start date:
January 1, 1994 End date:
February 1, 1998 Project Description:
In the SCHISMA project we investigate different possible architectures
and techniques for building natural language dialogue systems. It
serves as a motivation and testbed for various approaches and
theories. SCHISMA is both the name of the project and the name of our
prototype system: a dialogue system for theatre information and
booking. The acronym is derived from the Dutch SCHouwburg Informatie
Systeem. SCHISMA started as a joint research project of PTT Research
and the Parlevink group of the University of Twente. The Twentse
Schouwburg provides us with the necessary information to fill our
database. The SCHISMA project has lead to publications, a prototype, a
corpus and automatically generated grammars, work on spoken dialogue
and various other activities related to dialogue systems, such as a
virtual theatre and graduation work by students.
We believe that natural language dialogue systems have several
advantages over command-languages or graphical user interfaces.
Natural language is natural to the user; no need to learn an austere
command language or master the graphical interface controls. When
using speech, it is even more natural. Moreover, speech technology
opens the possibility to use the telephone for automatic information
services. The use of dialogue, i.e. an interactive exchange of
meaningful utterances, makes it possible to correct misunderstanding
and recognition errors. In dialogue, circumstances and previous
utterances play a role: the dialogue context. The context can be used,
not only for disambiguation, but also to make information access more
easy to use and more effective. Users can continue to ask queries
about the previous topic. The system can remember user characteristics
it has asked before.
The disadvantages of natural language interfaces have to be carefully
dealt with. For instance, natural language interfaces do not naturally
indicate the limits of the capacities and vocabulary of the system.
Users will often be irritated by a missing functionality or unknown
word. Such problems can however be overcome by a careful design of the
systems `prompts' and thus manipulating the users perception of the
system.
The application chosen for our prototype is a theatre information and
booking service. The reason we choose theatre information, is that the
domain is richer then the usual train- or airline information
services. We could add all kinds of general knowledge about
Shakespeare, Balanchine and so on. However, when restricted to the
theatre schedule - what performance plays at what day - it can be
constructed just as easily as the public transport information systems.
The second interesting feature of our application, is that it involves
reservations. The user may reserve tickets for a particular
performance. From the perspective of the system, a reservation
requires a unique performance, a number of tickets and information on
discount, preferred seats, as well as the name and identification
number of the user. The user however, would first expect detailed
information on the price and seats before giving his consent. We like
to look at a reservation transaction as a promise of the user to pick
up the tickets half an hour in advance and pay the price, given that
the information the system provided was correct. In future we hope to
extend the prototype into a system that is able to give advice or help
the user make up his or her mind.
Building a dialogue system requires a corpus. The corpus gives
insights into the behaviour of users and the typical expressions used
in a particular domain. With a corpus the coverage of the lexicon and
grammar can be evaluated and improved. In a Wizard-of-Oz experiment we
collected a large sample of dialogues. The corpus is in Dutch and has
been annotated and described.
Hugo ter Doest has generated a
probabilistic unification grammar from the corpus. The corpus was
tagged with syntactic categories and superficial structure.
Unification constraints and probabilities were derived. The grammar
was then tested on `seen' and `unseen' data from the same domain using
a probabilistic left-corner parser for PATRII unification grammars.
For more information see here or the list of publications (Ter Doest,
1998).
Schisma version 1.0 was build by Danny lie, Bas Leijdekkers and Eric
Barten. The parser is based on a rewrite approach. The system has a
simple agent architecture. For a more extensive description, see the
list of publications (Lie et al 1997).
In the SCHISMA project we investigate natural language dialogue
systems in general. Dialogue is a natural way of accessing
information. However, natural interaction is not necessarily
restricted to speech or keyboard input. We are interested in other
forms of communication as well. One such possibility, is the spatial
metaphor: the user can make a virtual walk through the theatre. Inside
the theatre information can be found at various locations. Information
may be presented in the form of brochures, posters or in the form of
an artificial agent Karin that is able to answer questions about
performances and make reservations.
A virtual browser based on the VRML technology was developed by
Richard Selles and Eduard Selles (Hogeschool Enschede, HIO). It
contains the floor-plan of the Muziekcentrum building and various
types of information related to theatre and music. At the moment, we
are connecting the Karin agent interface to our SCHISMA prototype.
More information on the Virtual Theatre is here.
In several student projects we acquired experience in building
telephone-based dialogue systems for transactions, especially ticket
reservation. Our experiences with the development toolkit,
Speechmania, and some of our recommendations and conclusions can be
found at Dialogue Systems for Reservation. |
The following HMI-member(s) is/are coordinator of this Project
Dirk Heylen
Here you can find the publications
|
