May i speak freely? Between templates and free choice in natural language generation
Year: 1999
Type: article
Abstract: Natural language generation (NLG) technology is currently finding its way into commercial systems. Promising applications are the automatic generation of weather forecasts, information about measurement data, or various kinds of authoring systems that help an author in composing a text. Generally the techniques used in NLG applications and application-oriented NLG systems differ from those utilized in research systems. While the latter typically aim at general, indepth solutions, the former are geared towards solving particular classes of NLG problems. This involves shallow generation such as dealing with canned texts or templates rather than choosing freely from the coverage of complex linguistic grammars. Correspondingly, diffeent systems encode linguistic knowledge at different levels of detail and sophistication. In spite of recent, more unified theoretical accounts of the NLG process, these differences persist. This is, we believe, to a large extent due to different requirements posed by the application tasks. While many theoretical linguistical concepts are difficult to use in practice, the techniques used in application systems often lack theoretical foundation. Their linguistic inadequacy severely restricts the transportability of systems to other tasks and domains. Current work in NLG addresses this problem, e.g., the RAGS project ( http://www.itri.brighton.ac.uk/projects/rags ). Whatever the chosen approach, its adequacy depends on the expected input to, and the desired output of, the NLG system. Input may vary from non-linguistic data to surface-semantic sentence representations. Further generation parameters may be avialable, such as a user model or a discourse model. Output may vary from a single sentence per context to multiple alternatives that allow, or require, the user to choose from. This workshop is, to our knowledge, the first one topicalizing the relation between application tasks and technologies used. It aims at exploring the tension between more general and more specific approaches to NLG, thereby clarifying what NLG technology is suited best for which task. It is intended to be an opportunity to get an overview over existing state-of-the-art technology and its optimal usage. It will be relevant for both developers and users of NLG systems. Exploring conditions for successful NLG applications is a step that should be taken jointly by technology providers and current and potential users of NLG software. The invited speaker, Paul Heisterkamp of DaimlerChrysler AG, will focus on the industrial usage of NLG software, and we appreciate his contribution to this volume. (orig.)
Authors: Becker, T., Busemann, S. (eds.), Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH (DFKI), Kaiserslautern (Germany), Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH (DFKI), Saarbruecken (Germany)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7011336522