<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1433">
  <Title>Rhetorical structure in dialog*</Title>
  <Section position="3" start_page="0" end_page="247" type="intro">
    <SectionTitle>
2 Previous work
</SectionTitle>
    <Paragraph position="0"> We are engaged in tim construction and inlplemenration of a theory of content-planning for complex, mixed-initiative task-oriented dialogs based on corpus analysis, for use in dialog systems such as the TRIPS system (Allen et al., 2000) 1 . Our basic premise is that a conversational agent should be able to produce whatever a human can produce in similar discourse situations, and that if we can explain why a human produced a particular contribution, &amp;quot; This work w~ supported by ONR research grant N0001495-l-1088, U.S. Air Force/Rome Labs research contract no. F30602-95-1-0025, NSF research grant no. IRI-9623665 and Columbia University/NSF research grant no. OPG: 1307. We would like to thank the armuymous reviewers and l)r. Jason Eisner for their helpful comments on earlier drafts of this paper.</Paragraph>
    <Paragraph position="1"> IWe are using the Monroe corpus (Stent, 2000), with reference t.o the TRAINS corpus (Heeman and Allen, 1995) and the HCRC Mapta~sk corpus (Anderson et al., 1991).</Paragraph>
    <Paragraph position="2">  we can program a conversational agent to produce something similar. Therefore, in examining our dialogs the question we must answer is &amp;quot;Why did this speaker produce this?&amp;quot;.</Paragraph>
    <Paragraph position="3"> RST is a descriptive theory of hierarchical structure in discourse that identifies functional relationships between discourse parts based on the intentions behind their production (Mann and Thompson, 1987). It has been used in content planning systems for text (effectively text monolog) (e.g. (Cawsey, 1993), (How, 1993), (Moore and Paris, 1993)). It has not yet been used much in content planning for spoken dialog.</Paragraph>
    <Paragraph position="4"> Because the dialogs we are examining are taskoriented, they are hierarchically structured and so provide a natural place to use RST. In fact, in order to uncover the full structure behind discourse contributions, it is necessary for us to use a model of rhetorical structure. Certain dialog contributions are explained by the speaker's rhetorical goals, rather than by task goals. In example 1, utterance 3 is justification for utterance 1 but does not directly contribute to completing the task.</Paragraph>
    <Paragraph position="5"> Example 1 A 1 They can't fix that power line at five ninety and East B 2 \Veil it A 3 Because you got to fix the tree first The details of how to apply RST to spoken dialog are unclear. If we mark rhetorical structure only within individual turns (as has generally been the case in annotations of text dialog, e.g. (Moser et al., 1996),(Cawsey, 1993)), we miss the structure in contributions like example 1 or example 2. There is also tile question of how to handle dialog-specific behaviors: grounding utterances and back-channels (utterances that maintain the comnmnication), and al)andoned or interrupted utterances.</Paragraph>
    <Paragraph position="6">  In our first attempt to annotate, we removed abandoned utterances, back-channels, and simple acknowledgments such as &amp;quot;Okay&amp;quot;. We used utterances as minimal units; utterances were segmented using prosodic and syntactic cues and speaker changes (see 3.2.2). We did occasionally split an utterance into two units if it consisted of two phrases or clauses separated by a cue word such as &amp;quot;because&amp;quot;. Two annotators, working separately, marked one complete dialog using Michael O'Donnell's RST annotation tool (1997). They used the set of relations in (Mann and Thompson, 1987), and some additional relations specific to dialog or to our domain. Examples of the additional relations are given in figure 1. When we compared the results, the tree structures obtained were similar, but the relation labels were very different, and in neither case was the entire dialog covered. Also, the annotators found structure not covered by the relations given. As a result, we stopped the annotation project and started developing an annotation scheme that would retain rhetorical relations while dealing with the difficulties we had encountered. The rest of this paper describes this new annotation scheme. An example of the type of analysis we are looking for appears in figure 3.</Paragraph>
  </Section>
class="xml-element"></Paper>