<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0613">
  <Title>Probabilistic Head-Driven Parsing for Discourse Structure</Title>
  <Section position="3" start_page="0" end_page="96" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Achieving a model of discourse interpretation that is both robust and deep is a major challenge. Consider the dialogue in Figure 1 (the sentence numbers are from the Redwoods treebank (Oepen et al., 2002)).</Paragraph>
    <Paragraph position="1"> A robust and deep interpretation of it should resolve the anaphoric temporal description in utterance 154 to the twenty sixth of July in the afternoon. It should identify that time and before 3pm on the twentyseventh as potential times to meet, while ruling out July thirtieth to August third. It should gracefully handle incomplete or ungrammatical utterances like 152 and recognise that utterances 151 and 152 have no overall effect on the time and place to meet.</Paragraph>
    <Paragraph position="2"> According to Hobbs et al. (1993) and Asher and Lascarides (2003), a discourse structure consisting of hierarchical rhetorical connections between utterances is vital for providing a uni ed model of a wide 149 PAM: maybe we can get together, and, discuss, the planning, say, two hours, in the next, couple weeks,  150 PAM: let me know what your schedule is like.</Paragraph>
    <Paragraph position="3"> 151 CAE: okay, let me see.</Paragraph>
    <Paragraph position="4"> 152 CAE: twenty, 153 CAE: actually, July twenty sixth and twenty seventh looks good, 154 CAE: the twenty sixth afternoon, 155 CAE: or the twenty seventh, before three p.m., geez. 156 CAE: I am out of town the thirtieth through the, 157 CAE: the third, I am in San Francisco.</Paragraph>
    <Paragraph position="5">  range of anaphoric and intentional discourse phenomena, contributing to the interpretations of pronouns, temporal expressions, presuppositions and ellipses (among others), as well as in uencing communicative goals. This suggests that a robust model of discourse structure could complement current robust interpretation systems, which tend to focus on only one aspect of the semantically ambiguous material, such as pronouns (e.g., Strcurrency1ube and Mcurrency1uller (2003)), de nite descriptions (e.g., Vieira and Poesio (2000)), or temporal expressions (e.g., Wiebe et al. (1998)). This specialization makes it hard to assess how they would perform in the context of a more comprehensive set of interpretation tasks.</Paragraph>
    <Paragraph position="6"> To date, most methods for constructing discourse structures are not robust. They typically rely on grammatical input and use symbolic methods which inevitably lack coverage. One exception is Marcu's work (Marcu, 1997, 1999) (see also Soricut and Marcu (2003) for constructing discourse structures for individual sentences). Marcu (1999) uses a decision-tree learner and shallow syntactic features  to create classi ers for discourse segmentation and for identifying rhetorical relations. Together, these amount to a model of discourse parsing. However, the results are trees of Rhetorical Structure Theory (RST) (Mann and Thompson, 1986), and the classi ers rely on well-formedness constraints on RST trees which are too restrictive (Moore and Pollack, 1992). Furthermore, RST does not offer an account of how compositional semantics gets augmented, nor does it model anaphora. It is also designed for monologue rather than dialogue, so it does not offer a precise semantics of questions or non-sentential utterances which convey propositional content (e.g., 154 and 155 in Figure 1). Another main approach to robust dialogue processing has been statistical models for identifying dialogue acts (e.g., Stolcke et al. (2000)). However, dialogue acts are properties of utterances rather than hierarchically arranged relations between them, so they do not provide a basis for resolving semantic underspeci cation generated by the grammar (Asher and Lascarides, 2003).</Paragraph>
    <Paragraph position="7"> Here, we present the rst probabilistic approach to parsing the discourse structure of dialogue.</Paragraph>
    <Paragraph position="8"> We use dialogues from Redwoods' appointment scheduling domain and adapt head-driven generative parsing strategies from sentential parsing (e.g., Collins (2003)) for discourse parsing. The discourse structures we build conform to Segmented Discourse Representation Theory (SDRT) (Asher and Lascarides, 2003). SDRT provides a precise dynamic semantic interpretation for its discourse structures which augments the conventional semantic representations that are built by most grammars. We thus view the task of learning a model of SDRT-style discourse structures as one step towards achieving the goal of robust and precise semantic interpretations.</Paragraph>
    <Paragraph position="9"> We describe SDRT in the context of our domain in Section 2. Section 3 discusses how we encode and annotate discourse structures as headed trees for our domain. Section 4 provides background on probabilistic head-driven parsing models, and Section 5 describes how we adapt the approach for discourse and gives four models for discourse parsing. We report results in Section 6, which show the importance of dialogue-based features on performance.</Paragraph>
    <Paragraph position="10"> Our best model performs far better than a baseline that uses the most frequent rhetorical relations and right-branching segmentation.</Paragraph>
    <Paragraph position="12"/>
  </Section>
class="xml-element"></Paper>