This Seminar has been set up with the particular purpose in mind to provide the generation community with an opportunity to focus on problematic aspects of natural language generation that are under-reported in papers designed to be accepted into conferences, and to produce results requiring face-to-face activity of a significant portion of the commu— nity. The aim hereby was to strenghten the theoretical foundation of natural language generation by identifying reasons why a systematic approach is so difficult in this area, and by stating principles for this research field. As it has turned out, success in these directions has been partially achieved, the principles identified are yet tentative or formulated on an admittedly abstract level. Nevertheless, we believe to have done an important step in the right direction.

Unlike with most scientific meetings, the form of this Seminar was rather informal, which is quite in accordance with the overall goals. A few introductory talks have been prescheduled to prepare the ground for subsequent discussions. Some additional talks have been given according to actual course of development. The major part of the time available has been devoted to tasks to be addressed by relatively small working groups in parallel although, as it turned out, a lot of time was also spent in general discussions after talks and in connection with working group reports.

Four working groups have been formed altogether. Two of them have addressed issues of lexicalization, including architectural concerns, the other two working on text planning issues. Some comparably simple examples taken from working systems have been prepared for the Seminar to provide adequate material for concrete actions. These examples were taken from the areas of text (report generation) and dialog, augmented by a specialized environment (hypertext). Heavily use was made of these examples during group work, which is also reflected by the reports produced.

In the course of the Seminar, some reasons have turned out to be primarily responsible for the difficulty to address generation in a systematic way-(they partially overlap in scope):

• Apart from a subtask commonly viewed as realization, it is almost impossible to structure the whole generation process in a principled way that is meaningful for a variety of different types of applications.
• The initial representations from which a generation process potentially starts differ significantly in type of data (e.g., visual data or tables) and depth of representation.
• The requirements on the functionality of a generator may vary significantly according to the scope and kind of task
• The perspective adopted by the motivation underlying the generation task (be it psychological, engineering or theoretical purposes) may cause crucial differences in the approach taken (some consider this distinction a source for a separation of the whole field).

Nevertheless, we think that some tentative, abstract principles have been identified, which should be taken into account in the design of a generator:

• clearly describe the task to accomplish by performing corpus anayses or comparable procedures, in order to identify the functionality of the system to build
• Design a model representation in such a way that the conceptual distinctions it emails are all needed and sufficient to get the envisioned behavior
• Select and adapt methods that exploit the conceptual distinctions made to produce the results according to these distinctions; unfortunately, classical divisions into types of application like dialog system, report generation, etc. are too course grained to provide a suitable basis for selecting methods.

Creative Commons BY 3.0 DE