ERCIM News No.26 - July 1996 - GMD
Knowledge-based Production of Synthetic Multimodal Documents
by Melina Alexa, John Bateman, Renate Henschel, and Elke Teich
Current attempts to solve the problem of information overload
that increasingly faces users of information systems fail to provide an
adequate foundation for information presentation, appropriate to users'
information needs. Extending the multi-media capabilities of information
systems as well as the breadth of information theoretically available leaves
this fundamental issue unaddressed. The use of linguistically motivated
principles for the automatic production and presentation of infor-mation
is now enabling the construction of a new generation of generic techniques
for information access.
Complex and plentiful information requires powerful techniques for its presentation.
Some of the most powerful, flexible and user-friendly techniques known are
those of natural language. This includes not only the side of language that
all can see and hear ­p; text and spoken ­p; but also the underlying
coherent structures that motivate the hanging together of bodies of texts
(including hypertexts) and of interactions.
The multilingual language generation department KOMET of the 'Publishing
Division' of the GMD-Institute for Integrated Publication and Information
Systems is exploring the use of language technology for the automatic generation
of synthetic documents as an interface metaphor for interaction with information
systems. Based on solid experiences in the context of inter-departmental
collaborations within the 'Publishing Division' of the GMD-Institute concerning
the construction of large-scale information systems, we see the following
as two of the most important foundations for future information systems
Linguistically motivated principles enable an information system to design
its information presentation very flexibly. Information is made available
to a user through several filtering and structuring stages. The first filter
is provided by direct manipulation or information retrieval concerning the
information base itself. This permits a user to focus on particular objects
or sets of objects maintained in the knowledge base. The remaining filters
are all under direct control of the information presentation system and
include: selection of further supportive information from the knowledge
base depending on desired text type and user knowledge, the organization
of that knowledge into a rhetorically motivated structure, the division
of this structure into substructures, and the presentation of these substructures
to the user as interrelated texts. Work in progress is being carried out
in close cooperation with the automatic visualization and automatic page
layout groups of the 'Publishing Division'. Some examples of current presentation
capabilities which demonstrate the results of the communicatively effective
combination of graphical and text modes on the basis of our present prototype
can be seen on the web at: http://www.darmstadt.gmd.de/publish/komet/kometpave-pics-96.html.
These examples show screen dumps from a prototype artist biography 'information
provider' capable of presenting information drawn from an object network
containing over half-a-million objects.
- information systems must move to work with representations of information
- information systems must use linguistically motivated principles for
presenting that content in ways that optimally support the information needs
of a user.
The supporting technology under development by KOMET is based on a highly
stratified, communicative-functional model of the linguistic system that
provides generic techniques for the linguistic expression of information,
for accessing domain knowledge, knowledge bases, etc., and for organizing
the presentation of that information rhetorically. Interfacing with domain
knowledge is handled by means of an extensive linguistically and multilingually
motivated ontology called the Generalized Upper Model. The generation technology
as a whole is furthermore text type based, enabling rapid development to
new text types as required by different applications.
A uniform representational medium (systemic networks as defined by systemic-functional
linguistics) is used throughout the different levels of information maintained.
There are also export capabilities to several state of the art computational
typed feature formalisms, which the KOMET system itself does not use for
reasons of efficiency. Example pages such as those that can be seen at the
web address above are produced in real-time.
The current system also includes an extensively multilingual development
environment for the construction and maintenance of large-scale linguistic
resources; this environment, freely available to the research community,
radically speeds up new language resource development. The KOMET text generation
system now produces short texts in English, German, and Dutch. Other language
resources are currently under development (usually in international cooperations)
for French, Italian and Greek, among others. The basic tactical generator
is being used in a number of international generation projects for several
of these languages. The generation capabilities of KOMET have also been
recently extended to include spoken language generation (currently for German).
Due to the crucially empirically based development of linguistic resources
at all levels (eg, large-scale grammatical resources that cover the constructions
that are actually necessary for a given text type, broad characterizations
of text types that are driven by analyses of existing text type phenomena,
etc.), particularly strong tools have also been created for supporting semi-automatic
text analysis. These are now being used both for empirical text analysis
and domain knowledge acquisition.
Current directions of extension include the further division of rhetorically
motivated structures into interlinked substructures that can form the basis
for hypernodes in extended synthetic hyperdocuments. Hyperlinks are guaranteed
to be intelligible and coherent by virtue of the supporting rhetorical organization.
Hypernodes may themselves contain significant substructure which is presented
to the user as a page of information reminiscent of a magazine or encyclopaedia:
ie, combining different texts, graphics, pictures, etc. There are many synergies
to be observed between the general problems of text generation, graphics
generation, and layout; shared processing techniques are therefore also
being investigated across the different modes of information presentation.
More information on KOMET is available on the web at:
Melina Alexa - GMD
Tel: +49 6151 869 809
return to the contents page