page 1  (8 pages)
2to next section

Towards the Configuration of

Generation Systems:

Some Initial Ideas

Stephan Busemann?

Deutsches Forschungszentrum fur Kuntstliche Intelligenz

Stuhlsatzenhausweg 3

D-66123 Saarbrucken, Germany

[email protected]

1 Introduction

Many systems for NL generation that serve as NL front end to some application system (e.g. a database, or an expert system) are designed in such a way that general, linguistic parts are not carefully distinguished from application-dependent parts. Often decisions made in the course of generation implicitly rely on properties of the application system, or they are based on knowledge that combines domain-specific aspects with more general ones. For instance, a generator used in dialogues about appointment scheduling might deterministically generate the speech act Request if it encounters the domain concept arrange since, in this application situation, arrange always represents the intention of arranging a meeting.

If such domain dependencies are placed in a generation front end, it is difficult, or even impossible, to transport the system to other applications. There is an obvious need for reusable generation components. What is called for is a better way of modularization. The problem is well known, and partial solutions have been suggested. For instance, transportable surface generators do exist and are used for various application classes. Beyond the surface-oriented levels, however, there is as yet no overall design strategy for generators that would help in achieving better modularity and reusability.

There is an obvious reason for this deficiency. The large variety of possible generation tasks, including discourse generation, dialog contributions and machine translation1, obviously cannot be dealt with by one single system. Moreover, the kind of input given to a generator by an application system is in no way standardized. For instance, focus

?This work was supported by the German Ministry for Research and Technology (FKZ ITW 9002 0). 1We exclude the problems of generating spoken output in this paper.