<?xml version="1.0" standalone="yes"?>
<Paper uid="J03-4002">
  <Title>c(c) 2003 Association for Computational Linguistics Anaphora and Discourse Structure</Title>
  <Section position="2" start_page="0" end_page="550" type="abstr">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> It is a truism that a text means more than the sum of its component sentences. One source of additional meaning are relations taken to hold between adjacent sentences &amp;quot;syntactically&amp;quot; connected within a larger discourse structure. It has been very difficult, however, to say what discourse relations there are, either theoretically (Mann and Thompson 1988; Kehler 2002; Asher and Lascarides 2003) or empirically (Knott 1996).</Paragraph>
    <Paragraph position="1"> Knott's empirical attempt to identify and characterize cue phrases as evidence for discourse relations illustrates some of the difficulties. Knott used the following theory-neutral test to identify cue phrases: For a potential cue phrase ph in naturally occurring text, consider in isolation the clause in which it appears. If the clause appears incomplete without an adjacent left context, whereas it appears complete if ph is removed, then ph is a cue phrase. Knott's test produced a nonexhaustive list of about two hundred different phrases from 226 pages of text. He then attempted to characterize the discourse relation(s) conveyed by each phrase by identifying when (always, sometimes, never) one phrase could substitute for another in a way that preserved meaning. He showed how these substitution patterns could be a consequence of a set of semantic features and their values. Roughly speaking, one cue phrase could always substitute for another if it had the same set of features and values, sometimes do so if it was less specific than the other in terms of its feature values, and never do so if their values conflicted for one or more features.</Paragraph>
    <Paragraph position="2">  [?] School of Informatics, University of Edinburgh, 2 Buccleuch Place, Edinburgh, EH8 9LW, UK. E-mail: bonnie@inf.ed.ac.uk.</Paragraph>
    <Paragraph position="3"> + Department of Computer Science, Rutgers Universtiy, 110 Frelinghuysen Road, Piscataway, NJ 08854-8019. E-mail: mdstone@cs.rutgers.edu.</Paragraph>
    <Paragraph position="4"> ++ Department of Computer &amp; Information Science, University of Pennsylvania, 200 South 33rd Street, Philadelphia, PA 19104-6389. E-mail: joshi@linc.cis.upenn.edu.</Paragraph>
    <Paragraph position="5"> SS Department of Computer Science, University of Otago, P.O. Box 56, DUNEDIN 9015, New Zealand.  Computational Linguistics Volume 29, Number 4 By assuming that cue phrases contribute meaning in a uniform way, Knott was led to a set of surprisingly complex directed acyclic graphs relating cue phrases in terms of features and their values, each graph loosely corresponding to some family of discourse relations. But what if the relational meaning conveyed by cue phrases could in fact interact with discourse meaning in multiple ways? Then Knott's substitution patterns among cue phrases may have reflected these complex interactions, as well as the meanings of individual cue phrases themselves.</Paragraph>
    <Paragraph position="6"> This article argues that cue phrases do depend on another mechanism for conveying extrasentential meaning--specifically, anaphora. One early hint that adverbial cue phrases (called here discourse connectives) might be anaphoric can be found in an ACL workshop paper in which Janyce Wiebe (1993) used the following example to question the adequacy of tree structures for discourse:  (1) a. The car was finally coming toward him.</Paragraph>
    <Paragraph position="7"> b. He [Chee] finished his diagnostic tests, c. feeling relief.</Paragraph>
    <Paragraph position="8">  d. But then the car started to turn right.</Paragraph>
    <Paragraph position="9"> The problem Wiebe noted was that the discourse connectives but and then appear to link clause (1d) to two different things: then to clause (1b) in a sequence relation (i.e., the car's starting to turn right being the next relevant event after Chee's finishing his tests) and but to a grouping of clauses (1a) and (1c) (i.e., reporting a contrast between, on the one hand, Chee's attitude toward the car coming toward him and his feeling of relief and, on the other hand, his seeing the car turning right). (Wiebe doesn't give a name to the relation she posits between (1d) and the grouping of (1a) and (1c), but it appears to be some form of contrast.) If these relations are taken to be the basis for discourse structure, some possible discourse structures for this example are given in Figure 1. Such structures might seem advantageous in allowing the semantics of the example to be computed directly by compositional rules and defeasible inference. However, both structures are directed acyclic graphs (DAGs), with acyclicity the only constraint on what nodes can be connected. Viewed syntactically, arbitrary DAGs are completely unconstrained systems. They substantially complicate interpretive rules for discourse, in order for those rules to account for the relative scope of unrelated operators and the contribution of syntactic nodes with arbitrarily many parents.</Paragraph>
    <Paragraph position="10">  We are not committed to trees as the limiting case of discourse structure. For example, we agree, by and large, with the analysis that Bateman (1999) gives of (2) (vi) The first to do that were the German jewellers, (vii) in particular Klaus Burie. (viii) And Morris followed very quickly after, (ix) using a lacquetry technique to make the brooch, (x) and using acrylics, (xi) and exploring the use of colour, (xii) and colour is another thing that was new at that time.</Paragraph>
    <Paragraph position="11"> 1 A reviewer has suggested an alternative analysis of (1) in which clause (1a) is elaborated by clause (1b), which is in turn elaborated by (1c), and clause (1d) stands in both a sequence relation and a contrast relation to the segment as a whole. Although this might address Wiebe's problem, the result is still a DAG, and such a fix will not address the additional examples we present in section 2, in which a purely structural account still requires DAGs with crossing arcs.</Paragraph>
    <Paragraph position="12">  Webber et al. Anaphora and Discourse Structure  Possible discourse structure for example (1). Each root and internal node is labeled by the type of relation that Wiebe takes to hold between the daughters of that node. (i) uses an n-ary branching sequence relation, whereas in (ii), sequence is binary branching.</Paragraph>
    <Paragraph position="13">  Simple multiparent structure.</Paragraph>
    <Paragraph position="14"> in which clause (ix) stands in a manner relation with clause (viii), which in turn stands in a succession (i.e., sequence) relation with clause (vi). This is illustrated in Figure 2, which shows a DAG (rather than a tree), but without crossing dependencies.</Paragraph>
    <Paragraph position="15"> So it is the cost of moving to arbitrary DAGs for discourse structure that we feel is too great to be taken lightly. This is what has led us to look for another explanation for these and other examples of apparent complex and crossing dependencies in discourse. The position we argue for in this article, is that whereas adjacency and explicit conjunction (coordinating conjunctions such as and, or, so, and but; subordinating conjunctions such as although, whereas, and when) imply discourse relations between (the interpretation of) adjacent or conjoined discourse units, discourse adverbials such as then, otherwise, nevertheless, and instead are anaphors, signaling a relation between the interpretation of their matrix clause and an entity in or derived from the discourse context. This position has four advantages: 1. Understanding discourse adverbials as anaphors recognizes their behavioral similarity to the pronouns and definite noun phrases (NPs) that are the bread and butter of previous work on anaphora. This is discussed in section 2.</Paragraph>
    <Paragraph position="16"> 2. By understanding and exploring the full range of phenomena for which an anaphoric account is appropriate, we can better characterize anaphors and devise more accurate algorithms for resolving them. This is explored in section 3.</Paragraph>
    <Paragraph position="17"> 3. Any theory of discourse must still provide an account of how a sequence of adjacent discourse units (clauses, sentences, and the larger units that they can comprise) means more than just the sum of its component  Computational Linguistics Volume 29, Number 4 units. This is a goal that researchers have been pursuing for some time, using both compositional rules and defeasible inference to determine these additional aspects of meaning (Asher and Lascarides 1999; Gardent 1997; Hobbs et al. 1993; Kehler 2002; Polanyi and van den Berg 1996; Scha and Polanyi 1988; Schilder 1997a, 1997b; van den Berg 1996) If that portion of discourse semantics that can be handled by mechanisms already needed for resolving other forms of anaphora and deixis is factored out, there is less need to stretch and possibly distort compositional rules and defeasible inference to handle everything.</Paragraph>
    <Paragraph position="18">  Moreover, recognizing the possibility of two separate relations (one derived anaphorically and one associated with adjacency and/or a structural connective) admits additional richness to discourse semantics.</Paragraph>
    <Paragraph position="19"> Both points are discussed further in section 4.</Paragraph>
    <Paragraph position="20"> 4. Understanding discourse adverbials as anaphors allows us to see more clearly how a lexicalized approach to the computation of clausal syntax and semantics extends naturally to the computation of discourse syntax and semantics, providing a single syntactic and semantic matrix with which to associate speaker intentions and other aspects of pragmatics (section 5.) The account we provide here is meant to be compatible with current approaches  to discourse semantics such as DRT (Kamp and Reyle 1993; van Eijck and Kamp 1997), dynamic semantics (Stokhof and Groenendijk 1999), and even SDRT (Asher 1993; Asher and Lascarides 2003), understood as a representational scheme rather than an interpretive mechanism. It is also meant to be compatible with more detailed analyses of the meaning and use of individual discourse adverbials, such as Jayes and Rossari (1998a, 1998b) and Traugott, (1995, 1997). It provides what we believe to be a more coherent account of how discourse meaning is computed, rather than an alternative account of what that meaning is or what speaker intentions it is being used to achieve.</Paragraph>
    <Paragraph position="21"> 2. Discourse Adverbials as Anaphors 2.1 Discourse Adverbials Do Not Behave like Structural Connectives  We take the building blocks of the most basic level of discourse structure to be explicit structural connectives between adjacent discourse units (i.e., coordinating and subordinating conjunctions and &amp;quot;paired&amp;quot; conjunctions such as not only ...but also, and on the one hand ...on the other (hand) and inferred relations between adjacent discourse units (in the absense of an explicit structural connective). Here, adjacency is what triggers the inference. Consider the following example: (3) You shouldn't trust John. He never returns what he borrows.</Paragraph>
    <Paragraph position="22"> Adjacency leads the hearer to hypothesize that a discourse relation of something like explanation holds between the two clauses. Placing the subordinate conjunction (structural connective) because between the two clauses provides more evidence for this rela2 There is an analogous situation at the sentence level, where the relationship between syntactic structure and compositional semantics is simplified by factoring away intersentential anaphoric relations. Here the factorization is so obvious that one does not even think about any other possibility.  Webber et al. Anaphora and Discourse Structure tion. Our goal in this section is to convince the reader that many discourse adverbials, including then, also, otherwise, nevertheless, and instead, do not behave in this way. Structural connectives and discourse adverbials do have one thing in common: Like verbs, they can both be seen as heading a predicate-argument construction; unlike verbs, their arguments are independent clauses. For example, both the subordinate conjunction after and the adverbial then (in its temporal sense) can be seen as binary predicates (e.g., sequence) whose arguments are clausally derived events, with the earlier event in first position and the succeeding event in second.</Paragraph>
    <Paragraph position="23"> But that is the only thing that discourse adverbials and structural connectives have in common. As we have pointed out in earlier papers (Webber, Knott, and Joshi 2001; Webber et al., 1999a, 1999b), structural connectives have two relevant properties: (1) they admit stretching of predicate-argument dependencies; and (2) they do not admit crossing of those dependencies. This is most obvious in the case of preposed subordinate conjunctions (example (4)) or &amp;quot;paired&amp;quot; coordinate conjunctions (example (5)). With such connectives, the initial predicate signals that its two arguments will follow.  (4) Although John is generous, he is hard to find.</Paragraph>
    <Paragraph position="24"> (5) On the one hand, Fred likes beans. On the other hand, he's allergic to them.  Like verbs, structural connectives allow the distance between the predicate and its arguments to be &amp;quot;stretched&amp;quot; over embedded material, without loss of the dependency between them. For the verb like and an object argument apples, such stretching without loss of dependency is illustrated in example (6b).</Paragraph>
    <Paragraph position="25"> (6) a. Apples John likes.</Paragraph>
    <Paragraph position="26"> b. Apples Bill thinks he heard Fred say John likes.</Paragraph>
    <Paragraph position="27"> That this also happens with structural connectives and their arguments is illustrated in example (7) (in which the first clause of example (4) is elaborated by another preposed subordinate-main clause construction embedded within it) and in example (8) (in which the first conjunct of example (5) is elaborated by another paired-conjunction construction embedded within it). Possible discourse structures for these examples are given in Figure 3.</Paragraph>
    <Paragraph position="28">  (7) a. Although John is very generous-b. if you need some money, c. you only have to ask him for it-d. he's very hard to find.</Paragraph>
    <Paragraph position="29"> (8) a. On the one hand, Fred likes beans.</Paragraph>
    <Paragraph position="30"> b. Not only does he eat them for dinner.</Paragraph>
    <Paragraph position="31"> c. But he also eats them for breakfast and snacks.</Paragraph>
    <Paragraph position="32"> d. On the other hand, he's allergic to them.</Paragraph>
    <Paragraph position="33"> But, as already noted, structural connectives do not admit crossing of predicate-argument dependencies. If we admit crossing dependencies in examples (7) and (8), we get (9) a. Although John is very generous-b. if you need some money- null (Impossible) discourse structures that would have to be associated with (i) Example (9) and (ii) example (10).</Paragraph>
    <Paragraph position="34"> c. he's very hard to find-d. you only have to ask him for it.</Paragraph>
    <Paragraph position="35"> (10) a. On the one hand, Fred likes beans.</Paragraph>
    <Paragraph position="36"> b. Not only does he eat them for dinner.</Paragraph>
    <Paragraph position="37"> c. On the other hand, he's allergic to them.</Paragraph>
    <Paragraph position="38">  d. But he also eats them for breakfast and snacks.</Paragraph>
    <Paragraph position="39"> Possible discourse structures for these (impossible) discourses are given in Figure 4. Even if the reader finds no problem with these crossed versions, they clearly do not mean the same thing as their uncrossed counterparts: In (10), but now appears to link (10d) with (10c), conveying that despite being allergic to beans, Fred eats them for breakfast and snacks. And although this might be inferred from (8), it is certainly not conveyed directly. As a consequence, we stipulate that structural connectives do not admit crossing of their predicate-argument dependencies.</Paragraph>
    <Paragraph position="40">  That is not all. Since we take the basic level of discourse structure to be a consequence of (1) relations associated with explicit structural connectives and (2) relations</Paragraph>
  </Section>
class="xml-element"></Paper>