<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1428">
  <Title>EXEMPLARS: A Practical, Extensible Framework For Dynamic Text Generation</Title>
  <Section position="3" start_page="0" end_page="267" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> In &amp;quot;NLG vs. Templates,&amp;quot; Reiter \[Reiter 95\] points out that while template-based text generation tools and techniques suffer from many well-known drawbacks, they do nevertheless enjoy numerous practical advantages over most tools and techniques developed in the NLG community in many circumstances.</Paragraph>
    <Paragraph position="1"> These advantages include, among others, efficiency, simplified system architectures~ -full control over Output, and much reduced demands on knowledge .acquisition and representation. This leads Reiter to suggest that, from a practical perspective, one should use NLG techniques in hybrid systems that mix template-style and more sophisticated techniques; moreover, to facilitate adoption, NLG technologies should be developed so that they Can be used without &amp;quot;getting in the way.&amp;quot; In line with this thinking, we have. been developing EXEMPLARS, an object-oriented, rule-based framework for dynamic text generation, with an emphasis on ease-of-use, programmatic extensibility and run-time efficiency. Exemplars \[Rambow et al. 98\] are schenm-like text planning rules that are so called because they are meant to capture an exemplary way of achieving a communicative goal in a given communicative context, as determined by the system designer. Each exemplar contains a specification of the designer' s intended method for achieving the Communicative goal. In the general case envisioned in Rainbow et al., these specifications can be given at the level of intentional-rhetorical, conceptual, lexicosyntactic, or formatting/hypertext structures. The present framework currently supports specifications only at the level of formatting/hypertext structures -- using any SGML-based representation, such as HTML -- or RealPro abstract syntactic structures \[Lavoie &amp; Rambow 97\]. A more complete range of specifications is instead Supported in PRESENTOR \[Lavoie &amp; Rambow 98\], a parallel implementation of the general approach with a complementary emphasis; while PRESENTOR emphasizes representation, we 266 . J~ have instead emphasized extensibility and classification-based planning. In future work, we plan to merge the best of the two implementations.</Paragraph>
    <Paragraph position="2"> In comparison to existing hybrid systems (e.g. \[Reiter et al. 95\]; \[Milosavljevic et al. 96\]; \[Knott et al. 96\]), we believe the present framework offers the following novel features: * Extensible classification-based text planning mechanism: The text planner's rule selection mechanism involves a decision tree-style traversal of the exemplar specialization hierarchy, where the applicability conditions associated with each exemplar in the hierarchy are successively evaluated in order to find the most specific exemplar for the current context. Viewed in this way, the rule selection mechanism naturally forms the *basis of an efficient, deterministic approach to text planning, where communicative actions are classified in context and then recursively executed, much as in \[Reiter &amp; Mellish 92\]. In contrast to Reiter and Mellish's approach, however, we emphasize extensibility, supporting inter alia discourse-sensitive conditions.</Paragraph>
    <Paragraph position="3"> * Java-based definition language: Exemplars are defined using a superset of Java, and then compiled down to pure Java. This approach makes it possible to (i) reuse basic Java constructs as well as *Java's inheritance mechanism, exceptions, threads, etc., (ii) directly and efficiently integrate with other application objects, and (iii) take advantage of advanced Java-based system architectures.</Paragraph>
    <Paragraph position="4"> * Advanced HTML/SGML support: With exemplars, the designer can bootstrap the authoring process using existing HTML or (normalized) SGML, then annotate the specification to produce dynamic content. Moreover, in contrast to other HTML template approaches (e.g. that provided with JavaSoft's Java Web Server \[Sun 98\]), we allow the designer to generate HTML in a truly hierarchical fashion.</Paragraph>
    <Paragraph position="5"> To date we have developed three systems 1 with the framework at CoGenTex --*namely the Project Reporter \[CoGenTex 97\], CogentHelp \[Caldwell &amp; White 97\] and EMMA systems \[McCullough et al.</Paragraph>
    <Paragraph position="6"> 98\] -- and are currently engaged in using it to develop a natural language query tool for a large data warehousing company. The framework has benefited substantially from feedback received during its use with these projects.</Paragraph>
    <Paragraph position="7"> The rest of this paper is organized as follows. In Section 2, we describe how the EXEMPLARS framework can be used to dynamically generate HTML using objects from an application object model. In Section 3, we focus on the role of specialization and extensibility in managing textual variation, In Section 4, we compare our classification-based approach to text planning to that of \[Reiter &amp; Mellish 92\], as well as to systemic and schema-based approaches, plus HTML template approaches taken Outside the NLG community. In Section 5, we conclude with a discussion Of the types of generation systems for which we consider the framework to be appropriate.</Paragraph>
    <Paragraph position="8"> 'Project Reporter is currently in the pre-beta release stage of development; CogentHelp and EMMA are operational prototypes.</Paragraph>
  </Section>
class="xml-element"></Paper>