<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-2108">
  <Title>Preventing False Temporal Implicatures: Interactive Defaults for Text Generation*</Title>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
Ac'II~,S DE COLlNG-92, NANWI..'S, 23-28 AOt~-r 1992 7 2 1 Pnoe. or: COLING-92, NnrctT.s, Auo. 23-28, 1992
</SectionTitle>
    <Paragraph position="0"> syntactic markers indicate this causal structure. By contrast, in (1) the reader realises what's going on only at tile last sentence. The discourse structure is therefore unclear until the whole text is heard, for the narrative requires a common topic which is only stated at the end.</Paragraph>
    <Paragraph position="1"> So (2)'s a better discourse than (1); but we would never generate it, if textual order had to mirror eventual order. If a generation system were permitted to generate (2), however, a price must be paid. The proper interpretation of (2) relies on the recruitment of certain causal information, left implicit by the utterance. The generator thus bas some responsibility for ensuring that the interpreter accomplishes the required inferences. A formal model of impticature must be folded into the generation process, so that the appropriate reasoning can proceed.</Paragraph>
    <Paragraph position="2"> States |nteract with causal information Ill La.scarides and Oberlander \[1992\], we considered in detail the following pair of examples:  (3) Max opened the door. The room was pitch dark.</Paragraph>
    <Paragraph position="3"> (4) Max switched off the light. The room was pitch  dark.</Paragraph>
    <Paragraph position="4"> Now, no-one would want to say that (3) involved a room becoming pitch dark immediately after a door was opened. Rather, most accounts (such as those based in or around DRT, such as ttinrichs \[1986\]) will take the state of darkness to overlap tile event of door-opening. That's how one might say states are dealt with in a narrative: events move things along; states leave them where they are. But if we have a piece of causal information to hand, things axe rather different. In (4), it seems that the state doesn't overlap the previously mentioned event.</Paragraph>
    <Paragraph position="5"> If one wishes to preserve the assumption about the role of states in narrative, it would have to be weakened to the constraint that states either leave things where they are, or move them along. This is not a very convincing move. An alternative is to formalise the role of the additional causal knowledge. Informally, the basis for the distinct interpretations of (3) and (4) is that the interpretation of (4) is informed by a causal preference which is lacking in the case of (3): if there is a switching off of the light and a room's being dark that are connected by a causal, part/whole or overlap relation, then normally one infers that the former caused the latter. This knowledge is defeasible, of course. In generation, such knowledge will constrain the space of adequate utterances; if H lacks the defeasible causal knowledge that switching off lights cause darkness, then (4) won't be adequate for H, who will interpret (4) in the same way as (3), contrary to S's intentions. Given this, S must contain a defeasible reasoning component to compute over such knowledge.</Paragraph>
    <Paragraph position="6"> The important point for now is that even if we describe things in the order in which they are assumed to happen, this doesn't necessarily make the candidate utterance a good one. if the speaker and the hearer possess differing world knowledge, there may be problems in retrieving the correct causal-temporal structure.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Two Methods of Generating with
Defeasible Knowledge
Generation by Defeasible Reasoning
</SectionTitle>
      <Paragraph position="0"> There is a very general way in which we might view interpretation and generation in terms of defensible reasoning. Consider the process of discourse interpretation as one of KB extension. The K8 contains an utterance-interpretation, and a set of knowledge resources; the latter may include general knowledge of the world, knowledge of linguistic facts, knowledge about tire discourse so far, and about the speaker's knowledge state. We then try to extend the KB so as to include the discourse interpretation. Consider now the process of generation; it too can be thought of as KB extension. Tills time, the KB contains a temporal-causal structure, and a set of knowledge resources, perhaps identical to that used in interpretation. We now try to extend the KB so as to include the realization of a linguistic structure's semantic features (with predicates, arguments, connectives, orderings), where these features ensure that the final linguistic string describes the causal structure in the KB. This view might be described as generation by defeasible reasoning.</Paragraph>
      <Paragraph position="1"> Modulo more minor differences, these notions are close to the ideas of interpretation as abduction (Hobbs et al \[1988\]) and generation as abduction (ltobbs et al \[1990:26-28\]), where we take abduction, in the former case for instance, to be a process returning a temporal-causal structure which can explain the utterance in context. Correspondences between a defensible deduction approach and an abductive approach have been established by Konolige \[1991\]; he shows that the two are nearly equivalent, tire consistency-based approach being slightly more powerful \[1991:15-16\], once closure axioms are added to the background theory. Lascarides &amp; Oberlander \[1992b\] discuss ill detail how such a generation process produces temporally adequate utterances.</Paragraph>
      <Paragraph position="2"> Interactive defaults Here, we turn to another, less powerful but simpler, method of applying defensible reasoning: the Interactive Defaults (ID) strategy introduced by Joshi, Webber and Weischedel \[1984a, 1984b, 1986\]. Rather than considering the defeasible process a.s applying directly to the KS's causal network, we instead consider its role as constraining or debugging candidate linearised utterances, generated by some otimr process; here we will remain neutral on the nature of that originating process.</Paragraph>
      <Paragraph position="3"> A speaker S and a hearer H interact through a dialogue; a writer S and a reader tl interact through a text. Joshi et al argue that it is inevitable that both S and H infer more from utterances than is explicitly contained within them. Taking Griee's \[1975\] Maxim of Quality seriously, they argue that since both .5' and H know this is going to happen, it is incumbent upon S to take into account the implicatures II is likely to make on the basis of a candidate utterance. If S detects that something S believes to be false will be among H's implicatures, S must block that inference somehow. The basic way to block it is for S to use ACRES DE COLING-92, NANrF~, 23-28 AotYr 1992 7 2 2 PROC. OF COLING-92, NANTES, AUG. 23-28. 1992 a different utterance; one which S does not believe will mislead H.</Paragraph>
      <Paragraph position="4"> In terms of defeasible reasoning, the point is that S must use it to calculate the consequences of the candidate utterance; if the process allows the derivation of something S believes to be false, the utterance should not be used in its current form. Joshi et al illustrate with tile following example; given the KB in (5), and the question in (6), they want the process to show why the answer in (7b) is preferred to that in (7a):  (5) Sam is an associate professor; most associate professors are tenured; Sam is not tenured.</Paragraph>
      <Paragraph position="5"> (6) ls Sam an associate professor? (7) a. Yes.</Paragraph>
      <Paragraph position="6"> b. Yes, but he is not tenured.</Paragraph>
      <Paragraph position="7">  We wish to elaborate this interactive defaults strategy 0D), and consider in greater formal detail the defeasible reasoning al)out causal-temporal strnctures that S and H are assumed by S to indulge itl; and to consider which candidate utterances arc eliminated on this basis.</Paragraph>
      <Paragraph position="8">  ID requires a theory of implicatnrc in terms of defaults, and an underlying logical notion of nonrnonotonic or defensible inference. We also require a formal eharacterisation of the properties an adequate candidate utterance must possess; we define these below in terms of temporal coherence, reliability and precision. Fnrthermore, we assume a model of discourse structure is required. For certain discourse relations, such as Narration and Explanation, are implicated from candidate utterances (cf. texts (1) and (2)), and these impose certain temporal relations on tile events described. We turn to this latter issue first.</Paragraph>
      <Paragraph position="9">  The basic model in which we embed ID assumes that candidate discourses possess hierarchical structure, with units linked by discourse relations modelled after those proposed by Hobbs \[1985\]. Lascarides &amp; Asher \[1991\] use Narration, Explanation, llaekground, Result and Elaboration. They provide a logical theory for determining the discourse relations between sentences in a text, and the temporal relations between the events they describe. The logic used is the nonmonotonic logic Common Sense Entailment (CE) proposed by Asher &amp; Morreau \[1991\]. Implieatures are calculated via default rules. For example, they motivate the following rules as manifestations of Gricean-style pragmatic maxims and world knowledge, where the clauses a and/3 appear in that order in the text. Informally:  event described in fl caused that described in C/v, then normally Ezplanalion(e~,\[t) holds.</Paragraph>
      <Paragraph position="10"> * Axiom on Explanatloxt If Ezplanation(c~,\[t) holds, then event el described by c~ does not occur bcfore event e2 described by/3.</Paragraph>
      <Paragraph position="11"> * Causal Law If clauses c~ and fl are discourse-related, and (~ describes the event c I of x fidling and fl the event e2 of y pushing x, then normally c2 causes el.</Paragraph>
      <Paragraph position="12"> * Causes Precede Effects  If event e~ eanses el~ t\]lell c I doesn't occur bcfore e2.</Paragraph>
      <Paragraph position="13"> The rules for Narration and l&amp;quot;xplanation constitute defe~iblc tingnistic knowledge, and the Axioms on them, indefeasible linguistic knowledge. Thc Causal Law is a mixture defea-sible linguistic knowledge and worhl knowledge: given that tim clauses are diseourse-rclated somehow, the events they describe must he commetcd in a causal, part/wholc or overlap relation; here, given the events in question, they must staud illa causal relation~ if things are norreal. That Causes Precede the.it Etfcets is in(lethtmible world knowledge. These rules arc used under the cE inference regime to infer the discourse structures ofcandidate texts. Two i)atterns of inference are particularly relevant: Defensible Modus Ponens (birds normally fly, Twecty is a bird; sn Tweety flies); and the Penguin Principle (all penguins are birds, birds normally fly, penguius normally don't fly, q'weety is a penguin; so Tweety doesn't fly).</Paragraph>
      <Paragraph position="14"> For example, in thc absence of information to the contrary, the only one of the rules whose antecedent is satisfied in interpreting text (8) is Narration.</Paragraph>
      <Paragraph position="15"> (8) Max stood up. John greeted hinl.</Paragraph>
      <Paragraph position="16"> Other things being equal, wc infer via Defeasible Modus Ponens that the Narration relation holds between (8)'s clauses, thus yielding, assuming logical omniscience, an interprctation where the descriptive order of events matches their temporal order. On the other band, in interl)reting text (9), in the absence of further information, two defanlt laws haw~ their antecedents satisfied: Narration and the Causal Law.</Paragraph>
      <Paragraph position="17"> (9) Max fell. John pusbed him.</Paragraph>
      <Paragraph position="18"> The consequents of these default laws cannot both hold in a consistent Ks. By the Penguin Principle, the law with the more specific antecedent wins: the Causal Law, because its antecedent logically entails that of Narration. \]\[lence, (9) is interpreted a.s a ca.se where the pushing caused the falling. In turn, this entails that the antecedent to Explanation is verified; and whilst conflicting with Narration, it's more specific, and hence its consequent--Ezplanation-follows by the Penguin Principle. Compare this with (8): similar logical forms, bnt different discourse structures, and different temporal structures: IThe formal details of how the logic oE models these AOrES DE COLING-92, NANTES, 23-28 Aour \] 992 7 2 3 I)ROC. OF COLING-92. NANTES, AtJ(J. 23-28, 1992</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Temporal Constraints
</SectionTitle>
      <Paragraph position="0"> So against this background, what are tile properties we require of cmldidate utterances? We concentrate on those constraints that are central to temporal import. Following Bach \[1986\], we take 'eventualities' to cover both events and states. We define lemporal coherence, temporal reliabilily and lerapetal precision--the notions that will characterise the adequacy of an utterauce--iu terms of a set C of relations between eventualities. This set intuitively describes when two eventualities are connected. The relations ill C are: causation, the part/whole relation, 2 temporal overlap, and the immediately precedes relation (where 'et immediately precedes e2 ' means that el and e 2 stand ill a causal or part/whole relation that is compatible with el tcm porally preceding e2). s The definitions are:</Paragraph>
    </Section>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
* Temporal Coherence
</SectionTitle>
    <Paragraph position="0"> A text is temporally coherent if the reader can infer that at least one of tile relations in C holds between the eventualities described m tile sentences.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
* Temporal Reliability
</SectionTitle>
    <Paragraph position="0"> A text is temporally reliable if one of the rclations in C which the reader infers to hold does in fact hold between tile eventualities described in the sentences.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
* Temporal Precision
</SectionTitle>
    <Paragraph position="0"> A text is temporally precise if whenever the reader infers that one of a proper subset of the relations in C holds between the eventualities described in the sentences, then she is also able to infer which.</Paragraph>
    <Paragraph position="1"> A text is temporally incoherent if the natural interpretation of the text is such that there are no inferrable relations between the events. A text is temporally unreliable if tim natural interpretation of the text is such that the inferred relations between tile events differ from their actual relations in the world.</Paragraph>
    <Paragraph position="2"> In addition, a text is temporally imprecise, or as we shall say, ambiguous, if the natural interpretation of tile text is such that the reader knows that one of a proper subset of relations in C holds between the eventualities, but the reader can't infer which of this proper subset holds.</Paragraph>
    <Paragraph position="3"> It follows from the above definitions that a text call be coherent but unreliable. On the other hand, there may be no questiou about reliability simply because we cannot establish a temporal or causal relation between the two eventualities. At any rate, a generated utterance is adequate only if it is temporally coherent, reliable and precise. We intend to apply tile ID strategy to eliminate candidate utterances that are inadequate in this sense.</Paragraph>
    <Paragraph position="4"> interpretations, and those of (3) versus (4), are given in Lascarides &amp; Asher \[1991\]. Note that although double applications of the Penguin Principle, as in (9), are not valid in genera\], they show that for the particular case considered here, o~ validates the double application.</Paragraph>
    <Paragraph position="5"> 2We think of 'el is part of e2 ~ in terms of Moens and Steedman's \[1988\] event terminology, as 'el is part of the preparatory phase or consequent phase of e2'.</Paragraph>
    <Paragraph position="6"> aWe a~SUllle that an eYeut el precedes an event e2 if el's culmination occurs before e2's. So there are part/whole relations between el and e~ that are compatible with el temporally preceding e2.</Paragraph>
    <Paragraph position="7"> Applying the ID strategy Before applying ID with temporal constraints, we must consider the possible relations between the knowledge of speaker S and that which speaker S has about hearer H's knowledge state. Notice, incidentally, that Joshi et al explicitly adopt the view that \]D is for debugging candidate utterances. In principle, their framework, however, is more general.</Paragraph>
    <Paragraph position="8"> Although the idea of debugging is intuitive, we shall sometimes talk in terms of constraining the space of possible utterances, rather than of debugging specific utterances. The definitions of temporal constraints are relevant either way.</Paragraph>
    <Paragraph position="9"> Relative KBs Let B(S) be S's beliefs about the KS, linguistic  knowledge (LK) and world knowledge (WK). Let B+(H) be S's beliefs about what H believes about ttle KB, LK and WK. And let B-(H) be S's beliefs about what H doesn't know about the KIL LK and WK (so B+(H) and B-(H) are mutually exclusive).</Paragraph>
    <Paragraph position="10"> Problems concerning reliability and precision arise when B(S) and B+(H) are different, and when S's knowledge of what H believes is partial (i.e. for some  p, p C/ 13+(H) and p q B-(II)). Suppcme that S's goal is to convey the content of a proposition corn tained in his KB, say q. Suppose also that a WFF p is relevant to generating a particular utterance describing q. Then there are several possible relations  S thinks H is mistaken in believing p: p ff 11(S) and p e B+(H) Of course, the cases where both S and H both believe p (p E B(S) and p e B+(II)) and where neither do (p q B(S) and p C B-(H)) are unproblematic, and so glossed over here. We look at each of the above cases in turn, considering tile extent to which tile definitions of reliability, coherence and precision hell) us eonstraiu the utterance space (or alternately, debug candidate utterances).</Paragraph>
    <Paragraph position="11"> Case 1: ~'C/ knows more about p than H We now examine the problems concerning reliability that arise when p E B(S) and p E B-(H). There are two possibilities: either p represents defeasible knowledge of tile lmlguage or the world, or p is some fact in the Km We investigate these in turn.</Paragraph>
    <Paragraph position="12"> p is defeaslble knowledge Let p be a dcfeasible law that represents knowledge that S has and which S knows H lacks. ~lb illustrate, take the case where p is the causal preference introduced earlier: ACrEs DE COL1NG-92, NANTES, 23-28 ho~r 1992 7 2 4 PROC. OF COL1NG-92, NANTES, AUG. 23-28, 1992</Paragraph>
  </Section>
class="xml-element"></Paper>