<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1018">
  <Title>Modeling Local Coherence: An Entity-based Approach</Title>
  <Section position="3" start_page="141" end_page="141" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Local coherence has been extensively studied within the modeling framework put forward by Centering Theory (Grosz et al., 1995; Walker et al., 1998; Strube and Hahn, 1999; Poesio et al., 2004; Kibble and Power, 2004). One of the main assumptions underlying Centering is that a text segment which foregrounds a single entity is perceived to be more coherent than a segment in which multiple entities are discussed. The theory formalizes this intuition by introducing constraints on the distribution of discourse entities in coherent text. These constraints are formulated in terms of focus, the most salient entity in a discourse segment, and transition of focus between adjacent sentences. The theory also establishes constraints on the linguistic realization of focus, suggesting that it is more likely to appear in prominent syntactic positions (such as subject or object), and to be referred to with anaphoric expressions.</Paragraph>
    <Paragraph position="1"> A great deal of research has attempted to translate principles of Centering Theory into a robust coherence metric (Miltsakaki and Kukich, 2000; Hasler, 2004; Karamanis et al., 2004). Such a translation is challenging in several respects: one has to specify the &amp;quot;free parameters&amp;quot; of the system (Poesio et al., 2004) and to determine ways of combining the effects of various constraints. A common methodology that has emerged in this research is to develop and evaluate coherence metrics on manually annotated corpora. For instance, Miltsakaki and Kukich (2000) annotate a corpus of student essays with transition information, and show that the distribution of transitions correlates with human grades. Karamanis et al. (2004) use a similar methodology to compare coherence metrics with respect to their usefulness for text planning in generation.</Paragraph>
    <Paragraph position="2"> The present work differs from these approaches in two key respects. First, our method does not require manual annotation of input texts. We do not aim to produce complete centering annotations; instead, our inference procedure is based on a discourse representation that preserves essential entity transition information, and can be computed automatically from raw text. Second, we learn patterns of entity distribution from a corpus, without attempting to directly implement or refine Centering constraints. null</Paragraph>
  </Section>
class="xml-element"></Paper>