<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1020">
  <Title>Representing discourse coherence: A corpus-based analysis</Title>
  <Section position="2" start_page="0" end_page="1" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> An important component of natural language discourse understanding and production is having a representation of discourse structure. A coherently structured discourse here is assumed to be a collection of sentences that are in some relation to each other. This paper aims to present a set of discourse structure relations that are easy to code, and to develop criteria for an appropriate data structure for representing these relations.</Paragraph>
    <Paragraph position="1"> Discourse structure relations here refer to informational relations that hold between sentences or other non-overlapping segments in a discourse monologue. That is, discourse structure relations reflect how the meaning conveyed by one discourse segment relates to the meaning conveyed by another discourse segment (cf. Hobbs, 1985; Marcu, 2000; Webber et al., 1999).</Paragraph>
    <Paragraph position="2"> Accounts of discourse structure vary greatly with respect to how many discourse relations they assume, ranging from two (Grosz &amp; Sidner, 1986) to over 400 different coherence relations, reported in Hovy and Maier (1995). However, Hovy and Maier (1995) argue that taxonomies with more relations represent subtypes of taxonomies with fewer relations. This means that different taxonomies can be compatible with each other.</Paragraph>
    <Paragraph position="3"> We describe an account with a small number of relations in order to achieve more generalizable representations of discourse structures; however, the number is not so small that informational structures that we are interested in are obscured.</Paragraph>
    <Paragraph position="4"> The next section will describe in detail the set of coherence relations we use, which are mostly based on Hobbs (1985). Additionally, we try to make as few a priori theoretical assumptions about representational data structures as possible. These assumptions will be outlined in the next section.</Paragraph>
    <Paragraph position="5"> Importantly, however, we do not assume a tree data structure to represent discourse coherence structures. In fact, a major goal of this paper is to show that trees do not seem adequate to represent discourse structures.</Paragraph>
    <Paragraph position="6"> 2 Collecting a database of texts annotated with coherence relations This section describes (1) how we define discourse segments, (2) which coherence relations we used to connect the discourse segments, and (3) how the annotation procedure worked.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 Discourse segments
</SectionTitle>
      <Paragraph position="0"> Discourse segments can be defined as non-overlapping spans of prosodic units (Hirschberg &amp; Nakatani, 1996), intentional units (Grosz &amp; Sidner, 1986), phrasal units (Lascarides &amp; Asher, 1993), or sentences (Hobbs, 1985). We adopted a sentence unit-based definition of discourse segments.</Paragraph>
      <Paragraph position="1"> However, we also assume that contentful coordinating and subordinating conjunctions (cf.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Coherence relations
</SectionTitle>
      <Paragraph position="0"> We assume a set of coherence relations that is similar to that of Hobbs (1985) and Kehler (2002).</Paragraph>
      <Paragraph position="1"> Table 1 shows the coherence relations we assume, along with contentful conjunctions that can signal the coherence relation.</Paragraph>
      <Paragraph position="2"> cause-effect because violated expectation although; but condition if...then; as long as similarity (and) similarly contrast but; however elaboration also, furthermore attribution ...said, according to...</Paragraph>
      <Paragraph position="3"> temporal sequence before; afterwards  The same relation, illustrated by (9), is an epiphenomenon of assuming contiguous distinct elements of text. (a) is the first segment and (c) is the second segment of what is actually one single discourse segment, separated by the intervening discourse segment (b), which is in an attribution relation with (a) (and therefore also with (c), since  [according to some analysts,] b [is expected to improve by early next year.] c Cause-effect, violated expectation, condition, elaboration, temporal sequence, and attribution are asymmetrical or directed relations, whereas similarity, contrast, temporal sequence, and same are symmetrical or undirected relations (Mann &amp; Thompson, 1988; Marcu, 2000). The directions of asymmetrical or directed relations are as follows: cause effect for cause-effect; cause absent effect for violated expectation; condition consequence for condition; elaborating elaborated for elaboration, and source attributed for attribution.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.3 Coding procedure
</SectionTitle>
      <Paragraph position="0"> In order to code the coherence relations of a text, annotators used a procedure consisting of three steps. In Step One, a text is segmented into discourse segments as described above. In Step Two, adjacent discourse segments that are topically related are grouped together. For example, if a text discusses inventions in information technology, there could be groups of a few discourse segments each talking about inventions by specific companies. There might also be subgroups of several discourse segments each talking about specific inventions at specific companies. Thus, marking groups determines a partially hierarchical structure for the text. In Step Three, coherence relations are determined between discourse segments and groups of discourse segments. Each previously unconnected (group of) discourse segment(s) is tested to see if it connects to any of the (groups of) discourse segments in the already existing representation of discourse structure.</Paragraph>
      <Paragraph position="1"> In order to help determine the coherence relation between (groups of) discourse segments, the (groups of) discourse segments under consideration are connected with a contentful conjunction like the ones shown in Table 1. If using a contentful conjunction to connect (groups of) discourse segments results in an acceptable passage, this is used as evidence that the coherence relation corresponding to the contentful conjunction holds between the (groups of) discourse segments under consideration.</Paragraph>
    </Section>
    <Section position="4" start_page="0" end_page="1" type="sub_section">
      <SectionTitle>
2.4 Statistics on annotated database
</SectionTitle>
      <Paragraph position="0"> In order to evaluate hypotheses about appropriate data structures for representing coherence structures, we annotated 135 texts, from the Wall Street Journal 1987-1989 and the AP Newswire 1989 (Harman &amp; Liberman, 1993), with the coherence relations described above. For the 135 texts, the mean number of words was 545 (min.: 161; max.: 1409; median: 529), the mean number of discourse segments was 61 (min.: 6; max.: 143; median: 60).</Paragraph>
      <Paragraph position="1"> Each text was independently annotated by two annotators. In order to determine inter-annotator agreement for the database of annotated texts, we computed kappa statistics (Carletta, 1996). For all annotations of the 135 texts, the agreement was 88.45%, per chance agreement was 24.86%, and kappa was 84.63%. Annotator agreement did not differ by text length (kh  Most accounts of discourse coherence assume tree structures to represent coherence relations between discourse segments in a text (Carlson et al., 2002; Corston-Oliver, 1998; Lascarides &amp; Asher, 1993; Longacre, 1983; Grosz &amp; Sidner, 1986; Mann &amp; Thompson, 1988; Marcu, 2000; Polanyi, 1988; van Dijk &amp; Kintsch, 1983; Walker, 1998; Webber et al., 1999). Other accounts assume less constrained graphs (Hobbs, 1985).</Paragraph>
      <Paragraph position="2"> The proponents of tree structures argue that trees are easier to formalize and derive than less constrained graphs (Marcu, 2000). We tested whether coherence structures of naturally occurring texts can be represented by trees, i.e. if these structures are free of crossed dependencies or nodes with multiple parents. However, we found a large number of both crossed dependencies as well as nodes with multiple parents in the coherence structures of naturally occurring texts. Therefore we argue for less constrained graphs as an appropriate data structure for representing coherence, where an ordered array of nodes represents discourse segments and labeled directed arcs represent the coherence relations that hold between these discourse segments.</Paragraph>
      <Paragraph position="3">  The following two sections will give examples of coherence structures with crossed dependencies and nodes with multiple parents. The section after that will present statistical results from our database of 135 coherence-annotated texts.</Paragraph>
    </Section>
    <Section position="5" start_page="1" end_page="1" type="sub_section">
      <SectionTitle>
3.1 Crossed dependencies
</SectionTitle>
      <Paragraph position="0"> Crossed dependencies are rampant and occur in many different forms in the coherence structures of naturally occurring texts. Here we will give some examples. Consider the text passage in (10).</Paragraph>
      <Paragraph position="1">  Other accounts also acknowledge examples that cannot be represented in tree structures (Webber et al., 1999). In order to maintain trees, these accounts distinguish non-anaphoric coherence structures, represented in a tree, and anaphoric coherence structures, which are not subject to tree constraints. However, e.g., Haliday &amp; Hasan (1976) stress the importance of anaphoric links as a cue for coherence structures. Therefore, by Occam's Razor, we assume a single level of representation for coherence rather than multiple levels.</Paragraph>
      <Paragraph position="2"> Figure 1 represents the coherence relations in (10). The arrowheads of the arcs represent directionality for asymmetrical relations (elaboration) and bidirectionality for symmetrical relations (contrast).</Paragraph>
      <Paragraph position="3">  (10) Example text (from SAT practicing materials) 0. Schools tried to teach students history of science.</Paragraph>
      <Paragraph position="4"> 1. At the same time they tried to teach them how to think logically and inductively.</Paragraph>
      <Paragraph position="5"> 2. Some success has been reached in the first of these aims.</Paragraph>
      <Paragraph position="6"> 3. However, none at all has been reached in the  The coherence structure for (10) can be derived as follows: there is a contrast relation between 0 and 1; 0 and 1 describe teaching different things to students. There is another contrast relation between 2 and 3; 2 and 3 describe varying degrees of success (some vs. none). 2 provides more details (the degree of success) about the teaching described in 0, so there is an elaboration relation between 2 and 0. Furthermore, in another elaboration relation, 3 provides more details (the degree of success) about the teaching described in 1. In the resultant coherence structure for (10), there is a crossed dependency between {2, 0} and {3, 1}.</Paragraph>
      <Paragraph position="7"> In order to be able to represent the crossed dependency in the coherence structure of (10) in a tree without violating validity assumptions about tree structures, one might consider augmenting a tree with feature propagation (Shieber, 1986) or with a coindexation mechanism (Chomsky, 1973). But the problem is that both the tree structure itself as well as the features and coindexations represent the same kind of information (coherence relations). It is unclear how one could decide which part of a text coherence structure should be represented by the tree structure and which by the augmentation. As pointed out above, coherence structures of naturally occurring texts contain many different kinds of crossed dependencies. This is important because it means that one cannot simply make special provisions to account for list-like structures like the structure of (10) and otherwise assume tree structures. As an example of a non-list-like structure with a crossed dependency (between {3,  0. Susan wanted to buy some tomatoes 1. and she also tried to find some basil 2. because her recipe asked for these ingredients.</Paragraph>
      <Paragraph position="8"> 3. The basil would probably be quite expensive at this time of the year.</Paragraph>
      <Paragraph position="9">  The coherence structure for (11) can be derived as follows: there is a parallel relation between 0 and 1; 0 and 1 both describe shopping for grocery items. There is a cause-effect relation between 2 and 0-1; 2 describes the cause for the shopping described by 0 and 1. Furthermore, there is an elaboration relation between 3 and 1; 3 provides details about the basil in 1.</Paragraph>
      <Paragraph position="10">  (12) from the AP Newswire1989 corpus is an example with a similar structure: (12) Example text (from text ap890109-0012) 0. The flight Sunday took off from Heathrow Airport at 7:52pm 1. and its engine caught fire 10 minutes later, 2. the Department of Transport said.</Paragraph>
      <Paragraph position="11"> 3. The pilot told the control tower he had the  The coherence structure for (12) can be derived as follows: 1 and 0 are in a temporal sequence relation; 0 describes the takeoff that happens before the engine fire described by 1 occurs. 2 and 0-1 are in an attribution relation; 2 mentions the source of what is said in 0-1. 3 and 1 are in an elaboration relation; 3 provides more detail about the engine fire in 1. The resulting coherence structure, shown in Figure 3, contains a crossed dependency between {3, 1} and {2, 0-1}.</Paragraph>
    </Section>
    <Section position="6" start_page="1" end_page="1" type="sub_section">
      <SectionTitle>
3.2 Nodes with multiple parents
</SectionTitle>
      <Paragraph position="0"> In addition to crossed dependencies, many coherence structures of natural texts include nodes with multiple parents. Such nodes cannot be represented in tree structures. For instance, in the coherence structure of (10), nodes 0 and 2 have two parents. Similarly, in the coherence structure of (13) from the AP Newswire 1989, node 1 has  (13) Example text (from text ap890103-0014) 0. &amp;quot;Sure I'll be polite,&amp;quot; 1. promised one BMW driver 2. who gave his name only as Rudolf.</Paragraph>
      <Paragraph position="1"> 3. &amp;quot;As long as the trucks and the timid stay out  The coherence structure for (13) can be derived as follows: 1 states the source of what is stated in 0 and in 3, so there are attribution relations between 1 and 0 and 1 and 3 respectively. 2 and 1 are in an elaboration relation; 2 provides additional detail about the BMW driver in 1. 3 and 0 are in a condition relation; 3 states the BMW driver's condition for being polite, stated in 0; the condition relation is also indicated by the phrase &amp;quot;as long as&amp;quot;.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>