<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2103">
  <Title>Discourse Generation Using Utility-Trained Coherence Models</Title>
  <Section position="3" start_page="0" end_page="803" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Various theories of discourse coherence (Mann and Thompson, 1988; Grosz et al., 1995) have been applied successfully in discourse analysis (Marcu, 2000; Forbes et al., 2001) and discourse generation (Scott and de Souza, 1990; Kibble and Power, 2004). Most of these efforts, however, have limited applicability. Those that use manually written rules model only the most visible discourse constraints (e.g., the discourse connective &amp;quot;although&amp;quot; marks a CONCESSION relation), while being oblivious to fine-grained lexical indicators. And the methods that utilize manually annotated corpora (Carlson et al., 2003; Karamanis et al., 2004) and supervised learning algorithms have high costs associated with the annotation procedure, and cannot be easily adapted to different domains and genres.</Paragraph>
    <Paragraph position="1"> In contrast, more recent research has focused on stochastic approaches that model discourse coherence at the local lexical (Lapata, 2003) and global levels (Barzilay and Lee, 2004), while preserving regularities recognized by classic discourse theories (Barzilay and Lapata, 2005). These stochastic coherence models use simple, non-hierarchical representations of discourse, and can be trained with minimal human intervention, using large collections of existing human-authored documents.</Paragraph>
    <Paragraph position="2"> These models are attractive due to their increased scalability and portability. As each of these stochastic models captures different aspects of coherence, an important question is whether we can combine them in a model capable of exploiting all coherence indicators.</Paragraph>
    <Paragraph position="3"> A frequently used testbed for coherence models is the discourse ordering problem, which occurs often in text generation, complex question answering, and multi-document summarization: given a0 discourse units, what is the most coherent ordering of them (Marcu, 1996; Lapata, 2003; Barzilay and Lee, 2004; Barzilay and Lapata, 2005)? Because the problem is NP-complete (Althaus et al., 2005), it is critical how coherence model evaluation is intertwined with search: if the search for the best ordering is greedy and has many errors, one is not able to properly evaluate whether a model is better than another. If the search is exhaustive, the ordering procedure may take too long to be useful.</Paragraph>
    <Paragraph position="4"> In this paper, we propose an Aa1 search algorithm for the discourse ordering problem that comes with strong theoretical guarantees. For a wide range of practical problems (discourse ordering of up to 15 units), the algorithm finds an optimal solution in reasonable time (on the order of seconds). A beam search version of the algorithm enables one to find good, approximate solutions for very large reordering tasks. These algorithms enable us not only to compare head-to-head, for the first time, a set of coherence models, but also to combine these models so as to benefit from their complementary strengths. The model com- null bination is accomplished using statistically well-founded utility training procedures which automatically optimize the contributions of the individual models on a development corpus. We empirically show that utility-based models of discourse coherence outperform each of the individual coherence models considered.</Paragraph>
    <Paragraph position="5"> In the following section, we describe previously-proposed and new coherence models.</Paragraph>
    <Paragraph position="6"> Then, we present our search algorithms and the input representation they use. Finally, we show evaluation results and discuss their implications.</Paragraph>
  </Section>
class="xml-element"></Paper>