<?xml version="1.0" standalone="yes"?>
<Paper uid="W99-0313">
  <Title>A Two-level Approach to Coding Dialogue for Discourse Structure: Activities of the 1998 DRI Working Group on Higher-level Structures*</Title>
  <Section position="2" start_page="0" end_page="101" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> A two-level scheme for coding discourse structure in dialogue has been proposed and undergone initial testing within the DR/effort. In particular, the higher-level structures working group of the third DR/ was charged with the task of creating a coding scheme concerned exclusively with the discourse structure of dialogue. Finding a good starting point The discourse structure working group was chaired by Christine Nakatani (Bell Laboratories, Lucent Technologies) and co-chaired by David Traum (U Maryland). Pre-meeting group participants also included Jean Carletta (U Edinburgh), Jennifer Chu-Carroll (Bell Laboratories, Lucent Technologies), Peter I-Iceman (Oregon Graduate Institute), Juha Hirschberg (AT&amp;T Labs), Masato Ishizaki (JAIST), Diane Litman (AT&amp;T Labs), Owen Rainbow (Cogentex), Jennifer Venditti (Ohio State U), Marilyn Walker (At&amp;T Labs), Gregory Ward (Northwestern U). Participants at the meeting also included Ellen Bard (U Edinburgh), Yasuo Horiuchi (Chiba U), Koichi Hoshida (ATR), Yasuhiro Katagisi (NTT), Kikuo Maekawa (NLRI), Michael Strube (U Pennsylvania), Masafumi Tamato (NTT), Yuki Tateishi (Tokyo U), and Takahiro Wakao (TAO).</Paragraph>
    <Paragraph position="1"> for a consensus coding scheme for discourse structure in dialogue was a non-trivial task. Most discourse structure schemes in fact were geared toward monologue, and most dialogue coding schemes omitted the higher-level structures that were essential to the monologue schemes, or provided only genre or domain-specific higher-level structures.</Paragraph>
    <Paragraph position="2"> Given the limited amount of work in this area, it was impossible to attempt a comprehensive coding scheme for all aspects of discourse structure ill dialogue. Instead, we were guided by an analysis of what choices needed to be made in creating a coding scheme. (Traum, 1998) identifies three dimensions along which discourse structure schemes can be classifted: granularity, content, structuring mechanisms. * Granularity: how much material (time, text, turns, etc.) is covered by the units (minimum, maximum, and average)? Granularity ranges were divided roughly into three categories: Micro - roughly within a single turn Meso - roughly an exchange, IR-unit, &amp;quot;game&amp;quot;, or short &amp;quot;sub-dialogue&amp;quot;, Macro - coherent larger spans, related to over-all dialogue purposes.</Paragraph>
    <Paragraph position="3"> * Content: what is this a structure of(e.g., intentions, accessibility, effects, etc.)? * * Structuring mechanisms: What kinds of units and structuring principles are used (e.g., fiat, set inclusion, hierarchical/CFG structuring, relational)? How many primitive types of units are allowed (one basic unit type, two or three types of units, or several types)? This multi-dimensional space was then used to classify different extant coding schemes as to which aspects they are concerned with.</Paragraph>
    <Paragraph position="4"> Guided by this principled survey of various schemes, we decided on an objective of defining a  pair of coupled schemes at the meso- and macro-levels in order to create a dialogue-oriented scheme for discourse structure analysis. We felt the micro-level of analysis was addressed by the dialogue acts coding effort of DRI, and it seemed most productive to build meso- and macro-levels on top of that, in an independent manner, to see what synergy might arise. It did not seem most fruitful to code the same content at three different levels, or to code three types of content at the macro-level without making any attempt to relate that coding to other schemes in development within the DRI initiative.</Paragraph>
    <Paragraph position="5"> Thus, for our starting point we proposed two original coding schemes within this multi-dimensional space. One scheme which has as content Grounding (Clark and Schaefer, 1989; Traum, 1994), operated at a meso level of granularity, and used non-hierarchical (and possibly discontinuous ) utterance sets as its structuring principle. The second scheme concerned intentional/informational structure (Grosz and Sidner, 1986; Nakatani et al., 1995) as content, operated at a macro level of granularity, and was structured as hierarchical trees (with annotations for capturing discontinuities). In addition, these two schemes were linked by using the resulting structures from meso-level analysis as basic input for macro-level analysis.</Paragraph>
    <Paragraph position="6"> There were several factors motivating the decision to use these particular facets of discourse structure for initial analysis. First, considering intentions, it is clear that aspects of dialogue at all levels of granularity relate to the intentions of the participants. However, not all of these intentional aspects are attuned to well-behaved plan-like structures. One issue is whose intention is under consideration: the speaker, the hearer, or the collaborative &amp;quot;team&amp;quot; of the speaker and hearer together. It is only at the level of grounded content that some sort of joint or shared intentional structure is really applicable. Below this level, one may only properly talk of individual intentions, even though those intentions may be subservient to joint goals (or goals of achieving sharedness). Thus taking grounded units (achieved at the meso-range) as a starting point for the coding of intentional structure is a natural basis for the study of joint intentional structure. Individual intentions at a lower level, especially those relating to communication management rather than task are expected to be captured within the dialogue act level of the DRI coding scheme (Discourse Resource Initiative, 1997; Allen and Core, Draft 1997). Likewise, the phenomena of grounding can occur on multiple levels. However, since macro-level phenomena (such as task summarization) differ from more local feed-back phenomena (including acknowledgments and repairs), restricting the grounding-relating coding to the meso-level allows for a more tractable effort.</Paragraph>
    <Paragraph position="7"> While examining intentional structure at the macro range and grounding structure at a meso range thus had independent motivations, the coding scheme used for this subgroup was designed to test a further novel and previously untested hypothesis that the units of achieving common ground would serve as an appropriate type of basic unit for intentional analysis. Since the phenomena of grounding and intentional task-related structure are somewhat independent, there is reason to believe the structures might not align properly. However, given the utility of having an appropriate meso-level starting point for intentional structure, and lacking any compelling counter-examples, we decided to put the hypothesis to the test in the coding exercises.</Paragraph>
  </Section>
class="xml-element"></Paper>