Summarizing is a vital information processing task. We have proposed a Dagstuhl seminar on automatic summarizing, now,for the following reasons:

• there has been a rapid growth of interest and activity in discourse and text processing in general and summarizing in particular
• there has been useful progress in computational text interpretation and generation, for example through the Message Understanding Conference
• there is a chance to integrate recent and pertinent results from related disciplines, from cognitive science to document retrieval, and to rethink system design in the light of their findings
• there is a manifest need to broaden the approaches used in existing summarizing systems, and to adopt new ones, to obtain systems able to handle texts that vary in length, topic type, and form
• there is a clear requirement, for intelligent communication in man-machine interaction, for summaries tailored to specific user needs.

More powerful summarizing systems than those developed so far are clearly needed. Today's information systems are not able to summarize in an intelligent way, deriving significant information from their text sources in order to provide the user with material of an appropriate scale, depth and orientation, in a coherent text form. The few systems built so far have been limited in both approach and implementation. They have been based either on shallow, statistically-oriented approaches to identifying salient source content, or on deeper analysis but only within a prespecified topic framework. They have thus been either only weak and uncertain in their ability to capture key information, or effective only within a very narrow application context and not readily extensible.

But summarizing is not just a primary task for text handling systems. It is also called for as a subtask in many other information management contexts, for instance in interactive consultation or instruction, so building a flexible summarizing capability into systems for these purposes will enhance their performance. At the same time, since summarizing depends on discourse interpretation, transformation and generation, it is a crucial test of discourse theory. Developing adequate theories of discourse structure and processing for summarizing will thus benefit all areas of language processing.

The Dagstuhl seminar we report on is built on the premise that summarizing is a complex, knowledge-based task to which many different language, text, and world knowledge resources jointly apply. This view of summarizing as a composite task implies an interdisciplinary seminar. bringing together both those who are directly engaged with summarizing and those contributing to parts or aspects of the whole.

We therefore invited researchers working on automatic summarizing, practitioners with experience of abstracting and information retrieval, researchers engaged with text retrieval, computational linguists working on methods of text analysis and production, members of the NLP community attacking specific tasks like message processing and data extractions, or concerned with user interfaces, linguists studying pragmatics and discourse, and psychologists and cognitive scientists especially concerned with text and discourse processing. All of these have an important contribution to make, both to the development of an intermediate summarizing technology operating on text, and to longer term research on summarizing for which full text understanding is required.

Creative Commons BY 3.0 DE