<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1019">
  <Title>Modelling the substitutability of discourse connectives</Title>
  <Section position="3" start_page="0" end_page="149" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Discourse coherence relations contribute to the meaning of texts, by specifying the relationships between semantic objects such as events and propositions. They also assist in the interpretation of anaphora, verb phrase ellipsis and lexical ambiguities (Hobbs, 1985; Kehler, 2002; Asher and Lascarides, 2003). Coherence relations can be implicit, or they can be signalled explicitly through the use of discourse connectives, e.g. because, even though.</Paragraph>
    <Paragraph position="1"> For a machine to interpret a text, it is important that it recognises coherence relations, and so as explicit markers discourse connectives are of great assistance (Marcu, 2000). When discourse connectives are not present, the task is more difficult.</Paragraph>
    <Paragraph position="2"> For such cases, unsupervised approaches have been developed for predicting relations, by using sentences containing discourse connectives as training data (Marcu and Echihabi, 2002; Lapata and Lascarides, 2004). However the nature of the relationship between the coherence relations signalled by discourse connectives and their empirical distributions has to date been poorly understood. In particular, one might wonder whether connectives with similar meanings also have similar distributions.</Paragraph>
    <Paragraph position="3"> Concerning natural language generation, texts are easier for humans to understand if they are coherently structured. Addressing this, a body of research has considered the problems of generating appropriate discourse connectives (for example (Moser and Moore, 1995; Grote and Stede, 1998)). One such problem involves choosing which connective to generate, as the mapping between connectives and relations is not one-to-one, but rather many-to-many.</Paragraph>
    <Paragraph position="4"> Siddharthan (2003) considers the task of paraphrasing a text while preserving its rhetorical relations. Clauses conjoined by but, or and when are separated to form distinct orthographic sentences, and these conjunctions are replaced by the discourse adverbials however, otherwise and then, respectively.</Paragraph>
    <Paragraph position="5"> The idea underlying Siddharthan's work is that one connective can be substituted for another while preserving the meaning of a text. Knott (1996) studies the substitutability of discourse connectives, and proposes that substitutability can motivate theories of discourse coherence. Knott uses an empirical methodology to determine the substitutability of pairs of connectives. However this methodology is manually intensive, and Knott derives relationships for only about 18% of pairs of connectives. It would thus be useful if substitutability could be predicted automatically.</Paragraph>
    <Paragraph position="6">  This paper proposes that substitutability can be predicted through statistical analysis of the contexts in which connectives appear. Similar methods have been developed for predicting the similarity of nouns and verbs on the basis of their distributional similarity, and many distributional similarity functions have been proposed for these tasks (Lee, 1999). However substitutability is a more complex notion than similarity, and we propose a novel variance-based function for assisting in this task.</Paragraph>
    <Paragraph position="7"> This paper constitutes a first step towards predicting substitutability of cnonectives automatically. We demonstrate that the substitutability of connectives has significant effects on both distributional similarity and the new variance-based function. We then attempt to predict substitutability of connectives using a simplified task that factors out the prior likelihood of being substitutable.</Paragraph>
  </Section>
class="xml-element"></Paper>