<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0203">
  <Title>Using a probabilistic model of discourse relations to investigate word order variation</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1 Introduction: Non-canonical main
</SectionTitle>
    <Paragraph position="0"> clause word order in English Users of natural languages have many ways to encode the same propositional content within a single clause. In English, besides the &amp;quot;canonical&amp;quot; word order, (1), options for realizing a proposition like GROW(MYRA,EGGPLANTS), include topicalization, left-dislocation, it-clefts, and wh-clefts, shown in (2-5), respectively.</Paragraph>
    <Paragraph position="1">  (1) Myra grows eggplants.</Paragraph>
    <Paragraph position="2"> (2) Eggplants, Myra grows.</Paragraph>
    <Paragraph position="3"> (3) Eggplants, Myra grows them.</Paragraph>
    <Paragraph position="4"> (4) It's eggplants that Myra grows.</Paragraph>
    <Paragraph position="5"> (5) What Myra grows are eggplants.</Paragraph>
    <Paragraph position="6">  Corpus-based research has shown that these forms are appropriate only under certain discourse conditions (Prince, 1978; Birner and Ward, 1998); among others. These include the membership of referents in a salient set of entities (left-dislocations and topicalizations) or the salience of particular propositions (topicalizations and clefts). For example, in (6), the topicalization is felicitous because there is a salient set KINDS OF VEGETABLES and a salient open proposition, that Myra stands in some relation X with an element of that set.</Paragraph>
    <Paragraph position="7"> (6) Myra likes most vegetables, but eggplants she adores.</Paragraph>
    <Paragraph position="9"> The discourse conditions licensing the use of these non-canonical syntactic forms are necessary conditions. When they do not hold, native speakers judge the use of the form infelicitous. They are not, however, sufficient conditions for use because salient sets and open propositions are ubiquitous in any discourse context, but these non-canonical forms are rare. Each type alone makes up &lt; 1% of utterances, across a variety of genres (Creswell, 2003).</Paragraph>
    <Paragraph position="10"> In addition to their information structure functions, one additional communicative goal these word orders fulfill is providing information about how an utterance is related to other discourse segments (Creswell, 2003). Native speaker intuitions about the appropriateness of non-canonicals in particular contexts provide anecdotal evidence (i.e.</Paragraph>
    <Paragraph position="11"> based on listing individual examples) for this discourse function. To provide broader support for this claim, however, we need to be able to generalize across many tokens.</Paragraph>
    <Paragraph position="12"> Ideally, a corpus annotated with discourse relations would be used to measure the correlations between the presence of non-canonical word order and particular discourse relations. However, explicit annotation of discourse relations is a difficult task, and one heavily dependent on the specific theory from which the set of discourse relations is chosen. Instead, this paper describes how a set of more easilyannotated features can be used to create a simplified approximation of the discourse context surrounding non-canonical (or canonical control) utterances. These features are then used as the independent variables in a statistical model which provides evidence for claims about how speakers use non-canonical word order to communicate information about discourse relations.</Paragraph>
    <Paragraph position="13"> The remainder of the paper is organized as follows: Section 2 describes how some non-canonical word orders in English contribute to the establishment of certain discourse relations. Section 3 describes how these relations can be approximated with a probabilistic model composed of more easily annotated features of the discourse context. Section 4 presents results and discussion of using such a model to measure the correlations between discourse relations and word order. Section 5 concludes and suggests improvements and applications of the model.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Additional meaning of non-canonical
</SectionTitle>
    <Paragraph position="0"> syntax: discourse relations The meaning of a multi-utterance text is composed not only of the meaning of each individual utterance but also of the relations holding between the utterances. These relations have syntactic aspects, such that single utterances can be grouped together and combined into segments recursively and are often modeled as a hierarchical tree structure (Grosz and Sidner, 1986; Webber et al., 1999). Discourse relations may also have a semantic or meaning component; this property, when treated in the literature, is often referred to as coherence, subject matter, or rhetorical relations (Kehler, 2002; Halliday, 1985; Mann and Thompson, 1988).</Paragraph>
    <Paragraph position="1"> The use of an utterance with non-canonical word order helps hearers make inferences about both the syntactic and semantic properties of discourse relations between the utterance and the rest of the discourse. For both aspects of discourse relations, it is the fact that the non-canonical order marks part of the utterance's information as salient or discourse-old that assists these inferences.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 Syntax of discourse relations
</SectionTitle>
      <Paragraph position="0"> One substructure of a coherent discourse structure is its attentional structure, which can be modeled as a stack of focus spaces (Grosz and Sidner, 1986). Each segment in the discourse tree has a corresponding focus space containing the currently salient discourse entities. When a segment begins, its focus space is pushed onto the stack on top of any other incomplete segments' spaces. When the segment ends, the focus space is popped off the stack.</Paragraph>
      <Paragraph position="1"> When an utterance continues in the same segment, the focus stack is unchanged.</Paragraph>
      <Paragraph position="2"> Non-canonical utterances instruct hearers about where to attach segments to the discourse tree. Because of the necessary conditions that license the use of a non-canonical, in most cases the open proposition or set is part of a focus space pushed onto the stack previously. So, the non-canonical form evokes the old proposition or set and thus reactivates the salience of that focus space. Reactivating the salience of the focus space in turn activates the salience of the discourse segment. As a result, the hearer infers that the new segment associated with the non-canonical utterance should be attached at the same level as this reactivated discourse segment, i.e. at a non-terminal node on the tree's right frontier. Any intervening segments should be closed off, and their focus spaces should be popped off the stack.</Paragraph>
      <Paragraph position="3"> To illustrate, in (7) the use of the it-cleft occurs after an intervening discussion of a separate topic.</Paragraph>
      <Paragraph position="4"> It-clefts are used to indicate that an existential closure of an open proposition is presupposed, here [?]t.YOU GOT TO MICHIGAN STATE AT TIME t. This presupposed material allows speaker B to mark the question as related to the prior discussion. In a tree structure of this discourse, the cleft corresponds to an instruction to &amp;quot;pop&amp;quot; back to a higher level in the tree when attaching the utterance, where speaker G's career at Michigan State was under discussion.</Paragraph>
      <Paragraph position="5"> The canonical version in (8) is an abrupt and infelicitous continuation of the discourse, as if B is unaware of the previous discussion of G's arrival at</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Michigan State.1
</SectionTitle>
      <Paragraph position="0"> (7) G: So for two years, I served as a project officer for grants and contracts in health economics that that agency was funding. I decided to go to academia after that and taught at Michigan State in economics and community medicine. One thing I should mention is that for my last three months in government, I had been detailed to work on the  (8) In what year did you get to Michigan State?</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Semantics of discourse relations
</SectionTitle>
      <Paragraph position="0"> The contribution of non-canonical utterances to the inference of semantic aspects of discourse relations is also related to the fact that these word orders mark (part of) an utterance's content as discourse-old or presupposed. Non-canonical word order is 1Varying the placement of the primary prosodic stress may improve the version in (8); see Delin (1995) and Creswell (2003) for comparison of the discourse function of prosody and syntax.</Paragraph>
      <Paragraph position="1"> used to indicate relations of RESEMBLANCE rather than CONTIGUITY.</Paragraph>
      <Paragraph position="2"> A CONTIGUITY relation is the basic relation found in narratives. According to Labov (1997), utterances in a prototypical narrative describe in the order they took place a sequence of causally-related events which lead up to a MOST REPORTABLE EVENT. Kehler (2002), following Hobbs (1990), says the events should be centered around a system of entities, and each event should correspond to a change of state for that system. To infer a CONTIGUITY relation between two utterances, the hearer must infer that their eventualities correspond to a change of state for that system.</Paragraph>
      <Paragraph position="3"> Inferring a RESEMBLANCE relation between two utterances depends on a very different type of information. To establish RESEMBLANCE, the hearer must identify a common relation R that relates the propositional content of two utterances and also the number and identity of their arguments (Kehler, 2002). Resemblance relations include PARAL-LEL, CONTRAST, EXEMPLIFICATION, GENERAL-IZATION, EXCEPTION, and ELABORATION.</Paragraph>
      <Paragraph position="4"> Non-canonicals are useful in resemblance relations because 1) the presence of 'old' material in a non-canonical helps overrule the default coherence relation of CONTIGUITY by making that interpretation less likely, and 2) the use of old material and a structured proposition assists the hearer in identifying a common relation and corresponding arguments needed to establish RESEMBLANCE.</Paragraph>
      <Paragraph position="5"> This is illustrated in (9). The use of a left-dislocation tells the hearer that the referent of a lot of the doctors is in a salient set. By identifying that set as {PROFESSIONAL PEOPLE}, the hearer can realize that the information being added about a lot of the doctors is going to be in an EXEMPLIFICATION relation with the earlier statement that professional people in general began to think of themselves as disabled.</Paragraph>
      <Paragraph position="6"> (9) During the Depression an awful lot of people began to think of themselves as disabled, especially professional people, who depended on clients whose business was on a cash basis-there was no credit, this was a universe without credit cards. A lot of the doctors, they were doing an awful lot of charity work. They couldn't support themselves. They'd have a little heart attack. They'd have disability insurance. They went on the insurance company rolls. A lot of doctors had disability insurance and a lot of others too. A lot of the insurance companies stopped underwriting disability insurance. They couldn't afford it.</Paragraph>
      <Paragraph position="8"> between utterances (U) by examining lexical discourse cues (M) and relations between entities (e) (10) A lot of the doctors were doing an awful lot of charity work.</Paragraph>
      <Paragraph position="9">  Without the left-dislocation, identifying the inclusion relationship between the set of professional people and doctors is quite difficult. The preferred interpretation of the canonical version in (10) is only that the doctors were doing charity work for professional people who had no credit cards. The left-dislocation supports the additional inference that the exemplification described above holds too.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Probabilistic model of discourse
</SectionTitle>
    <Paragraph position="0"> relations and non-canonical syntax To provide evidence beyond individual examples for the phenomena in Section 2, we need to measure the correlation between discourse relations and syntactic form, but annotating discourse relations directly is problematic. Annotation of hierarchical discourse structure is difficult and subjective although efforts have been made (Creswell et al., 2002; Marcu et al., 1999). Even annotating linear segmentation is challenging, particularly in the vicinity of segment boundaries (Passonneau and Litman, 1997). Annotation of the semantics of discourse relations requires a predetermined set of relation types, on which theories vary widely, making theory-neutral generalizations about the role of non-canonical syntax impossible.</Paragraph>
    <Paragraph position="1"> This project attempts to overcome these difficulties by indirectly deriving discourse relations by mapping from their known correlates to the use of certain non-canonical forms. The correlates used here are referential relations across utterance boundaries and the presence and type of lexical discourse markers or cue words. These features are annotated with respect to a three-utterance window centered on a target utterance Ui, shown schematically in Figure 1.</Paragraph>
    <Paragraph position="2"> These referential and lexical features build on the work of Passonneau and Litman (1997), who use them in discourse segmentation. Their use here is extended to also derive information about the se-</Paragraph>
    <Paragraph position="4"> dependent variables mantic and syntactic properties of the relations between utterances.</Paragraph>
    <Paragraph position="5"> As illustrated in Figure 2, discourse relations (e.g. R1 ) influence observable patterns of referential relations (e.g. x) and discourse markers (e.g. m). We want to test whether discourse relations also influence the use of certain sentence types. However, the discourse relations themselves are not observable directly. To measure their correlation with sentence-level syntax, we will only look at correlations of referential patterns and discourse markers with syntactic form. In the logistic regression analysis performed here, syntactic form is the dependent variable; referential relations and lexical cues are the independent variables.</Paragraph>
    <Paragraph position="6"> This analysis only measures the direct influence of the independent variables on the dependent variable, and does not model the existence of a mediating set of (unobserved) discourse relations, the result being that it is unable to capture correlations among the independent variables. This inherent inadequacy of the model will be discussed further below. Despite this inadequacy, a logistic regression analysis is used because it is a mathematically convenient and well-understood way to model which features of the independent variables are significant in predicting the occurrence of each syntactic form, while taking into account the rare prior probabilities of the non-canonical syntactic forms.</Paragraph>
    <Paragraph position="7"> In order to decide whether the featural models provide evidence to support the claims about discourse relations and syntactic forms, we first need to make clear our assumptions about how referential relations and lexical markers correlate with discourse relations. Based on those assumptions, testable predictions can then be made about how referential relations and lexical markers should correlate with syntactic forms.</Paragraph>
    <Paragraph position="8"> Ref Utterances share center of attention; Cp of  The lexical discourse marker feature is annotated for the target Ui and its preceding (Ui[?]1) and following (Ui+1) utterances and has five values: and, but, so, other or none. The predictions about the correlations between these lexical features and syntactic forms are based on the assumed correlations between these lexical markers and discourse relations. First, if non-canonicals are indicators of attentional stack pops, they should be more likely at segment boundaries; hence, we expect an increased presence of cue words (Passonneau and Litman, 1997) on non-canonicals compared to canonicals.</Paragraph>
    <Paragraph position="9"> Predictions about the type of cue words are based on the survey of lexical cue meanings from Hirschberg and Litman (1994). Because and is an indicator of segment-continuation and the relation CONTIGUITY, we expect decreased incidence on non-canonicals. However, we expect greater incidence on Ui+1 when Ui is non-canonical because Ui should be used to start a new segment. The presence of but indicates a CONTRAST relation. Non-canonicals should have a greater likelihood of being in contrast with either of the utterances surrounding them,2 so we expect a greater incidence of but on both Ui and Ui+1 for non-canonicals than for canonicals. The presence of so can indicate RE-STATEMENT or RESULT, so so should appear more often on Ui for wh-clefts, which are often used in ELABORATION relations.</Paragraph>
    <Paragraph position="10"> The referential features are four-valued and annotated with respect to pairs of utterances, (Ui[?]1, Ui) and (Ui, Ui+1). The values here, described in Table 1, form an implicational scale from strongest to weakest connections, and the utterance pair is labeled with the strongest relation that holds.</Paragraph>
    <Paragraph position="11"> In general, the more semantic content two utterances share, the more likely they are to be related. Referential connections are the measure of shared content used here. Discourse relations vary in their likelihood to be associated with certain values of 2See Creswell (2003) for examples and discussion.</Paragraph>
    <Paragraph position="12"> Ref, shown in Table 2. For example, an utterance immediately following a discourse pop, should be unlikely to share a center with its immediately preceding utterance and be highly likely to share no references at all. Two utterances in a RESEMBLANCE relation (other than ELABORATION) are likely to have inferential connections without coreferential connections. Note that for nearly all of these patterns, the correlation between a referential feature value and the syntax or semantics of a discourse relation is not absolute but only more or less likely. Using a probabilistic model, however, allows for patterns of relative likelihood in the data.</Paragraph>
    <Paragraph position="13"> Based on the assumptions in Table 2, we can now make predictions about expected correlations between the referential features and utterances with non-canonical word orders. These predictions are based primarily on how we expect non-canonical utterances to compare with canonical utterances.</Paragraph>
    <Paragraph position="14"> However, when we test them on our data, we will also compare each type of non-canonical utterance with the others.</Paragraph>
    <Paragraph position="15"> * Non-canonicals should be more likely than canonicals to follow a POP and begin a new segment. They should have weaker referential ties to the preceding utterance. They should have a higher incidence of having no referential ties to Ui[?]1, and a lower incidence of having no referential ties to Ui+1.</Paragraph>
    <Paragraph position="16"> * Non-canonicals should be less likely than canonicals to have a NARRATIVE relation with either Ui[?]1 or Ui+1. This situation predicts that with respect to both of the utterances surrounding a non-canonical utterance, these utterances will be less likely to share the same center of attention than when Ui is canonical.</Paragraph>
    <Paragraph position="17"> * Non-canonicals should be more likely than canonicals to be in RESEMBLANCE relations with Ui[?]1 and/or Ui+1. So, a greater likelihood of reference to inferentially-related entities and smaller likelihood of reference to coreferential entities or shared centers in both the preceding and following utterance is expected.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Results and discussion
</SectionTitle>
    <Paragraph position="0"> To test the predictions about non-canonicals and discourse relations, a corpus of 799 utterances with non-canonical word order were extracted from 58 transcribed interviews from the Social Security Administration Oral History Archives (SSA), a corpus with [?]750,000 words and 44,000 sentences. In addition to the four types of non-canonicals, 200 randomly-selected controls with canonical word order were also included. Table 4 lists the breakdown  by syntactic type. The two lexical and three referential features described in the previous section were annotated for each of the 999 utterances.</Paragraph>
    <Paragraph position="1"> Logistic regression models for binary comparisons between each of the five sentence types were then created. For 9 of 10 comparisons, at least one of the five features were found to be significant.3 Table 4 lists all features found to be significant for each of the ten comparisons, i.e. features that individually have a significant effect in improving the likelihood of a model when compared to a model that uses no features to predict the distribution of the two classes.4 For comparisons with multiple features significant at the five-percent level, the p-value of the model fit in comparison with a fully saturated model is listed in the fourth column of Table 4.</Paragraph>
    <Paragraph position="2"> In order to understand the most likely context in which a form to appears, we need to examine the weights assigned to each feature value by the regression analysis. The detailed feature weights in the best model are listed in Table 5.</Paragraph>
    <Paragraph position="3"> Table 6 summarizes the general conclusions we can draw from these weights about the most favorable discourse contexts for each of the four types of non-canonicals. For considerations of space, we discuss in detail only one of the four types here, whclefts. Wh-clefts are particularly relevant with respect to the insights they provide into the inherent limitations of our model of discourse relations.</Paragraph>
    <Paragraph position="4"> Overall, wh-clefts are favored in contexts where they start a new segment, one with weak connections with the preceding utterance and strong connections with the following utterance. In particular, feature 4, REF(Ui[?]1,Ui), is significant in the 3The comparison of it-clefts and left-dislocations is the exception here. From the lack of significant features in this comparison we can surmise that the it-clefts and left-dislocations are more similar to each other than any of the other forms compared here.</Paragraph>
    <Paragraph position="5"> 4In particular, the measure whose significance is tested is the -2[?](difference in log-likelihoods of the models), which is kh2 distributed, where the number of degrees of freedom is the difference in the total number of feature values between the two models.</Paragraph>
    <Paragraph position="6"> Relation between 3. Shared 2. Coreferring 1. Inferentially-related 0. No shared Uj and Uk center entities entities only reference  comparison of wh-clefts with all other classes. Wh-clefts are much more likely to share no connections at all with Ui[?]1 and less likely to share only inferential connections when compared with any other class. In comparison with everything but leftdislocations, wh-clefts are also less likely to share their center of attention with Ui[?]1.</Paragraph>
    <Paragraph position="7"> In terms of discourse markers, feature 2 and 3 are significant when comparing topicalizations and controls with wh-clefts (although feature 3 is only weakly significant in comparing wh-clefts and controls.) For feature 2, MARKER(Ui), wh-clefts are less likely than either of the other two to appear with and and more likely to appear with so. For feature 3, however, the presence of and on Ui+1 favors wh-clefts over topicalizations and controls.</Paragraph>
    <Paragraph position="8"> The most likely context in which to find wh-clefts then is one with no referential connections to the previous utterance and marked with the discourse adverbial, so. When the Ui+1 begins with and, assumed to be a marker of continuation of the previous content, wh-clefts are also favored. This pattern resembles most closely the descriptions of a preceding discourse POP and a subsequent discourse CON-TINUE or NARRATIVE.</Paragraph>
    <Paragraph position="9"> One use of wh-clefts that is not borne out conclusively in the data is its use in ELABORATION relations, as in (11). Kehler (2002) describes elaborations as a case of RESEMBLANCE where the predicate and its arguments are the same, but described from a different perspective or level of detail. The hearer must infer the identity of the event and en-</Paragraph>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
CONTROL CONTROL CONTROL CONTROL IT-CLEFT IT-CL EFT IT-CLEFT LEFT-DIS. LEFT-DIS. TOPIC.
</SectionTitle>
    <Paragraph position="0"> category) listed first; weights &lt;0.5 favor the second application value. The farther away from 0.5, the stronger the feature value's effect on the distinction between the two classes.</Paragraph>
    <Paragraph position="1"> tities being described in the two segments. If wh-clefts are associated with ELABORATIONS, then we should see an increased incidence of close referential connections with Ui[?]1 and an increased incidence of so, a marker of restatement. In the results, however, we only see evidence for the latter.</Paragraph>
    <Paragraph position="2"> (11) S: How did you develop this Resource-Based Relative Value Scale at this point? H: We basically treated this as a research project because most of us involved realized we had some past failures and we should not over-promise. We should be prepared to face up to the world and say, &amp;quot;We cannot make the theory operational.&amp;quot; So what we did was we continued to accept the theoretical premise, that is the rational and objective price should be based on the cost of the service. Then we asked, &amp;quot;What constitutes the cost of physicians' services and what are the components of physicians' work?&amp;quot; (SSA, hsiao) A possible factor in the absence of evidence here is that wh-clefts are also associated with discourse pops, which increase the likelihood of having no referential connections with the previous utterance.</Paragraph>
    <Paragraph position="3"> The logistic regression model used here aggregates over all possible discourse relations. So, when two discourse relations that give rise to different lexical and referential patterns are both associated with a single sentence type, the patterns of one may obscure the patterns of the other. A more sophisticated statistical model might take into account dependencies between discourse markers and referential patterns and from them posit hidden states which correspond to different discourse relations. Then based upon these hidden states, the model would predict which sentence type would best fit the context. Such a model would be more true to Figure 2.</Paragraph>
    <Paragraph position="4"> Another limitation of the model shown here is that the only indicators in this model of starting a new segment are weak or absent referential relations, presence of a connective like so, and absence of and. These measures will not necessarily distinguish between continuing in the same segment or beginning a new segment which includes recentlymentioned discourse-old entities.</Paragraph>
  </Section>
class="xml-element"></Paper>