<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1426">
  <Title>Can text structure be incompatible with rhetorical structure?</Title>
  <Section position="3" start_page="0" end_page="194" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Text planning (or more broadly, document planning) can be divided into two stages. In the first stage, material is selected, perhaps from a knowledge base, and organized rhetorically. In the second stage, the rhetorical structure is realized by a text structure (or document structure), through which the material is distributed among sentences, paragraphs, vertical lists, and perhaps even diagrams. The RAtS (1999) proposal for a standard NLG architecture distinguishes tile outputs of these two phases by the data types l:l.hetRep (rhetorical representation) and DocRep (document representation).</Paragraph>
    <Paragraph position="1"> We focus in this paper on the second stage of text planning -- the passage from RhetRep to DocRep. NLG researchers have addressed this issue in various ways, but everyone assumes some kind of structural compatibility between rhetorical structure and text structure.</Paragraph>
    <Paragraph position="2"> The most popular discourse framework in NLG is R ST (Mann a.nd Thompson. 1988). which makes the crucial distinction between nucleus, which is the most important part of a message, and satellite, which is the peripheral part of the message. Scott and Souza (1990) provide guidelines for the realisation of RST trees into a coherent text. One of them is to avoid dangling sentences, that is, to avoid putting &amp;quot;information that is only weakly relevant to the message&amp;quot; in a separate sentence because it will feel as if it has been introduced as an afterthought or as introducing a new topic which is then abruptly abandoned, disrupting the comprehensibility of the text. As an example, the authors provide the attributive satellite of an elaboration relation. null Marcu (1996), in order to build a valid text plan from a set of rhetorical assertions, uses the &amp;quot;nuclearity principle&amp;quot;, that is the observation in Mann and Thompson's framework that &amp;quot;whenever two large text spans are connected through a rhetorical relation, that rhetorical relation holds also between the most important parts of the constituent spans&amp;quot;. Therefore, the resulting text plans are valid in the sense that they are isomorphic with one of the rhetorical structures that can be built from the rhetorical assertions using this nuclearity principle.</Paragraph>
    <Paragraph position="3"> Our aim in this paper is to formulate more precisely a notion of structural compatibility which is necessary in order to describe the realisation of a RhetRep into various DocReps, and then .to discuss some examples (mostly taken from the domain of patient information leaflets) of apparently acceptable texts in which this no.tion of compatibility is violated..:To discuss this issue clearly, an assmnption must be made about the kinds of information represented by rhetorical and text structure; we outline in section 2 the common assumption that these representations are trees, labelled respectively with rhetorical and textual categories, the rhetorical structure being unordered and the text struc- null ture ordered. Section 3 then defines a notion of.structural compatibility that:is weaker than isomorphism; section 4 shows that we can find plausible counterexamples even to this weaker formulation, and discusses why these passages occur. Section 5 discusses some implications for NLG, and finally, section 6 raises further important issues.</Paragraph>
  </Section>
class="xml-element"></Paper>