File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/04/w04-2309_metho.xml
Size: 14,260 bytes
Last Modified: 2025-10-06 14:09:24
<?xml version="1.0" standalone="yes"?> <Paper uid="W04-2309"> <Title>But What Do They Mean? An Exploration Into the Range of Cross-Turn Expectations Denied by &quot;But&quot;</Title> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 A Methodology for Distinguishing the </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Underlying Expectations </SectionTitle> <Paragraph position="0"> We outline a novel methodology for distinguishing features involved in these relations using linguistic substitution tests involving the cue phrase taxonomy presented in Knott's thesis (1996). Knott presents a taxonomy of cue phrases distinguished as feature-theoretic constructs rather than markers of one or more of a set of rhetorical relations as postulated in RST. Rather than finding data to describe a conceptualised theory of rhetorical relations, he uses data containing cue phrases to drive the creation of his taxonomy of cue phrases, which reveals psycholinguistic features involved in conveying or interpreting meaning, i.e., the data drives his theory of linguistic production. We enquire into the nature of the relations underlying these denied cross-turn expectations using the following methodology: 1. Take original &quot;but&quot; example and determine expectation being denied via algorithm in (Thomas and Matheson, 2003).</Paragraph> <Paragraph position="1"> 2. Invert example so that the consequent of the expectation is asserted rather than denied in B's turn, (i.e., omitting the &quot;but&quot;). (So the dialogue conveys that the expectation in Step 1 succeeds.) 3. Determine what sort of expectation this inverted pair of turns seems closest to, given Knott's taxonomy. Determine whether the cues conveying this relation are substitutable in this inverted dialogue: (a) Test all the high-level categories in the taxonomy and see which ones work by substitution tests involving cues belonging to those categories. Then determine whether the category chosen captures the nature of the expectation (i.e., intuitively, following annotator's judgment).</Paragraph> <Paragraph position="2"> (b) If so, i. test whether hyponyms1 of these high-level cues work in the inverted dialogue. The most specific hyponyms that work indicate the maximally specific set of features that pertain to the relation underlying the expectation.</Paragraph> <Paragraph position="3"> ii. Now confirm that these cues that work in the inverted dialogue do not also work in the original (denied) dialogue. Those cues that work in the inverted example but not in the denied (original) dialogue are indicative of the nature of the underlying relation that's denied in DofE.</Paragraph> <Paragraph position="4"> iii. Look up the feature-value definitions for the maximally specific cues that work in this inverted (not denied) case. Comparing these to the feature-value definition of hyponyms of &quot;but&quot; that deny the same expectation will reveal which feature-values are denied/inverted in the denied case.</Paragraph> <Paragraph position="5"> Comparing the intersection of feature-values for hyponyms that work in the inverted case to the intersection of feature-values for hyponyms that work in denied case shows precisely which features are being denied.</Paragraph> <Paragraph position="6"> (c) If Knott's taxonomy does not provide a category that works for the inverted dialogue, i. check whether any of Knott's categories fit the original denied expectation by testing which cues are substitutable in the orig- null inal example; a good place to start is with hyponyms of &quot;but&quot;. ii. For cues that work, check their hyponyms to determine the maximally specific set of features that apply to the relation between turns. Note that this only specifies the relation underlying the denied expectation and does not shed light on the original (not denied) expectation.</Paragraph> <Paragraph position="7"> iii. If no cues besides &quot;but&quot; work in the original dialogue, then &quot;but&quot; must be the maximally specific cue that works, and we cannot determine more precisely the nature of the denied expectation, so assume that the turns are related by simple contingency/co-occurrence.</Paragraph> <Paragraph position="8"> 1Hyponyms inherit the features of their parent (higher-level) cues in the directed acyclic graph structure of the taxonomy. So all hypernyms (higher-level parents) should also be substitutable in the given case. Hypernyms are far less specific and therefore less precise.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Modal Status Actual Actual 3.1 An example </SectionTitle> <Paragraph position="0"> So to determine how A and B might be related (in B's perspective) for Ex.1, we find that the following cues work in the inverted dialogue below with the expectation asserted rather than denied: (4) Example 4.</Paragraph> <Paragraph position="1"> A: Greta was beautiful.</Paragraph> <Paragraph position="2"> B: (Yes) a0 indeeda1 she a0 evena1 married.</Paragraph> <Paragraph position="3"> The asserted expectation works with Knott's &quot;additional information&quot; category of cues, and &quot;even&quot; and &quot;indeed&quot; are the most specific of these cues which work. In the original denied example below, (with &quot;*&quot; indicating unacceptable cues): (5) Example 5.</Paragraph> <Paragraph position="4"> A: Greta was beautiful.</Paragraph> <Paragraph position="5"> B: a0 However/even so/in spite of this/all the same/despite this/nevertheless/then again/*indeeda1 she never married. two of the most specific of these negative polarity cues which work are &quot;despite this&quot; and &quot;then again&quot;, which differ from &quot;indeed&quot; and &quot;even&quot; in polarity (the former are negative, the latter positive) and focus of polarity (the former are anchor-based, the latter, counterpart-based). Also, the cues which work in the inverted case do not work in the denied case. Furthermore, these negative polarity cues are defined for some values which are undefined for these additional information cues, and the two pairs also share some feature-values in common. But the features that are defined for both and differ are the ones which are the most informative; they specify which features are being denied in the DofE case, and which ones asserted in the inverted case. So here we find that DofE involves denying polarity and focus of polarity in the underlying expectation.</Paragraph> </Section> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Modelling Issues </SectionTitle> <Paragraph position="0"> Although this methodology requires human judgment to assess the results of the substitution tests, it is a first step towards distinguishing underlying relations in DofE.</Paragraph> <Paragraph position="1"> We address how this information might be modelled in the PTT (Poesio and Traum, 1998) model of dialogue by adapting the Information State (IS) (Matheson et al., 2000) in order to facilitate more responsive generation from the system upon hearing the DofE.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.1 Utilising Knott's Feature Definitions </SectionTitle> <Paragraph position="0"> Knott argues that his data-driven definition of relations is compatible with the view that relations are planning operators with preconditions and effects, where the relations' preconditions are defined via the speaker's intentions and applicability conditions specified for what the speaker wants to convey, and the effects are simply the intended effects the conveyed relation has on the hearer.</Paragraph> <Paragraph position="1"> More practically speaking, the features are defined in terms of variable bindings and relationships which describe the relations concisely. For example, the polarity feature describes whether the defeasible rule a2 a0a4a3 holds, based on whether A=P and C=Q (positive) or A=P and C is inconsistent with Q (negative), where A and C are the propositional contents of the two respective related clauses. To address how polarity might be determined in a dialogue situation, if a speaker believes a0a5a3 , then this is in her Private Beliefs field in the IS. If her turn is mapped onto Q, and the prior turn is mapped successfully onto P by matching first-order logic representations of the material in the two turns, then if her turn maps onto Q, we can assume positive polarity; if her turn maps onto a negated Q, then we assume negative polarity.</Paragraph> <Paragraph position="2"> While mapping speakers' turns onto the variables which define Knott's features might be difficult, we can automate some of the feature assignment to update the IS by maintaining an exhaustive (i.e. complete) static table of cue-phrase definitions2. This way, once the most specific cue-phrases that work in the inverted and denied expectations are determined, we can automatically assign feature-value-pair bundles to these dialogues which describe the underlying relation being denied. Then comparing the feature-values for the maximally specific cues for both the asserted and denied cases (as we saw in the previous section), we can determine precisely which features are being denied in a given DofE, and the IS can be updated with this information, so that in the next turn of the dialogue, the speaker can compare these feature-value assignments to his own (in his private beliefs) and respond accordingly with a highly specific response to the DofE which targets precisely where he disagrees or agrees.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.2 Information State Modification </SectionTitle> <Paragraph position="0"> We propose, given information about the nature of the underlying relation via the feature-value differences involved in the DofE as well as broader information about the category(ies) to which a cue-phrase belongs in Knott's taxonomy, to include this in the a6a8a7a10a9a12a11 Conversational Act as follows for Ex.3: 2Knott provides a partial table of cue-phrase definitions like this in Appendix D.1 of his thesis.</Paragraph> <Paragraph position="2"> a0 with the more specific temporal ordering relation as the link between A and B's turns; the last field includes specific features being denied.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.3 Responding to DofE Appropriately </SectionTitle> <Paragraph position="0"> Upon hearing B's DofE, A must then respond appropriately. If A also infers the nature of B's denied expectation, this can lead to much more responsive generation.</Paragraph> <Paragraph position="1"> (Thomas and Matheson, 2003) address how interpreting DofE in the IS can facilitate better generation. We argue that their algorithm cannot predict the correct expectations in cases involving noncontingency related expectations (i.e., cases unlike Ex.1). E.g., in Ex.3, their algorithm would predict that B has the expectation that &quot;having a child in '43 a0 not married in '47&quot;, since according to their original formulation of defeasible rules, B's turn is negated to form the consequent, so depending on A's beliefs he would respond accordingly. E.g., If A disagrees with both B's assertion and inferred expectation, then neither must be in his beliefs, and he might respond: &quot;She didn't marry in '47, and anyway just because she had a child in '43 doesn't mean she should be married in '47.&quot; (I.e., A does not understand that B sees the events as temporally ordered.) With our added information about the nature of this expectation, namely that it involves temporal ordering, we can improve upon Thomas and Matheson's scheme by predicting the following more appropriate responses.</Paragraph> <Paragraph position="2"> Notice that given this added information about the relation underlying the DofE, denying the DofE now means denying the underlying relation licensing the expectation.</Paragraph> <Paragraph position="3"> This means that A can be much more relevant when generating a response: 1. If A disagrees with both B's assertion and inferred expec null tation, then neither must be in his beliefs, and he might respond: &quot;She married before '43, and anyway lots of people back then had children before marrying.&quot; 2. If A only agrees with B's assertion, then this assertion must be in his private beliefs, and he might respond: &quot;Yeah, but lots of people back then had children before marrying.&quot; 3. If A only agrees with B's expectation, then this must be in his private beliefs, and he might say: &quot;But she married before '43.&quot; 4. If A agrees with both B's assertion and expectation, he might say (minimally): &quot;Yes, that's odd.&quot; Notice that these responses indicate that A has understood B's temporal ordering that underlies these events and is the source of the denied expectation, and this allows B to correct possible misunderstandings. E.g., if B realises that A thinks she believes that people need to marry before having children, and this is an incorrect inference on A's part, she can indicate this by responding, e.g., to the situation in which A disagrees with both B's inferred expectation and assertion, B: &quot;OK, but I don't think that people had to marry before having children.&quot; B needs to recognise specifically that A failed to interpret her temporal ordering expectation in order to correct A's misassumption. In cases in which specific features are the precise source of the DofE, if the hearer of the DofE can recognise that the wrong polarity is being attributed to his utterance, he (A) might indicate this misassumption by saying &quot;but not marrying is common among beautiful people&quot; (Ex.1).</Paragraph> </Section> </Section> class="xml-element"></Paper>