File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/91/p91-1008_metho.xml
Size: 21,155 bytes
Last Modified: 2025-10-06 14:12:48
<?xml version="1.0" standalone="yes"?> <Paper uid="P91-1008"> <Title>Open Explana~ lanatidegn Closed Open Narration Explanation/// ~xplanation Closed ~ Open Narration</Title> <Section position="4" start_page="0" end_page="55" type="metho"> <SectionTitle> Social Research Council. 1 Extensive classifications of discourse relations are of- </SectionTitle> <Paragraph position="0"> fered in (Polanyi, 1985), (Scha & Polanyi, 1988) and (Thompson & Mann, 1987).</Paragraph> <Paragraph position="1"> of the event described in a, i.e. fl's event is part of the preparatory phase of a's: 2 (2) The council built the bridge. The architect drew up the plans.</Paragraph> <Paragraph position="2"> Explanation(a, fl): For example the event described in clause fl caused the event described in clause a: (3) Max fell. John pushed him.</Paragraph> <Paragraph position="3"> Background(a, fl): For example the state described in fl is the 'backdrop' or circumstances under which the event in a occurred (so the event and state temporally overlap): (4) Max opened the door. The room was pitch dark.</Paragraph> <Paragraph position="4"> Result(a, fl): The event described in a caused the event or state described in fl: (5) Max switched off the light. The room was pitch dark.</Paragraph> <Paragraph position="5"> We assume that more than one discourse relation can hold between two sentences; the sickness in (6) describes the circumstances when Max took the aspirin (hence the sentences are related by Background) and also explains why he took the aspirin (hence the sentences are related by Explanation as well).</Paragraph> <Paragraph position="6"> logical forms. They indicate, therefore, that the constraints on the use of the above discourse relations cannot rely solely on the logical forms of the sentences concerned.</Paragraph> <Paragraph position="7"> No theory at present is able to explain the distinct temporal structures of all the above texts. Webber (1988) observes that Kamp & Rohrer (1983), Partee (1984), Hinrichs (1986) and Dowty (1986) don't account for the backwards movement of time in (2) and (3). Webber (1988) can account for the backwards movement of time in (2), but her theory is unable to predict that mismatching the descriptive order of events and their temporal order is allowed in some cases (e.g. (2) and (3)) but not in others (e.g. (1), which would be misleading if the situation being described were one where the greeting happened before Max stood up).</Paragraph> <Paragraph position="8"> Our aim is to characterise the circumstances under which each of the above discourse relations hold, and to explain why texts can invoke different temporal relations in spite of their similar syntax.</Paragraph> <Paragraph position="9"> Dahlgren (1988) represents the difference between (1) and (3) in terms of probabilistic laws describing world knowledge (WK) and linguistic knowledge (LK). Our approach to the problem is generally sympathetic to hers. But Dahlgren's account lacks an underlying theory of inference. Furthermore, it's not clear how a logical consequence relation could be defined upon Dahlgren's representation scheme because the probabilistic laws that need to interact in certain specific ways are not logically related. Unlike Dahlgren (1988), we will supply an inference regime that drives the interpretation of text.</Paragraph> <Paragraph position="10"> The properties required of an inference mechanism for inferring the causal structure underlying text is discussed in (Lascarides & Oberlander, 1991). The work presented here builds on this in two ways; first by supplying the required notion of inference, and second by accounting for discourse structure as well as temporal structure.</Paragraph> <Section position="1" start_page="55" end_page="55" type="sub_section"> <SectionTitle> Temporal Relations and Defeasible Reasoning </SectionTitle> <Paragraph position="0"> Let us consider texts (1) and (3) on an intuitive level. There is a difference in the relation that typically holds between the events being described. Intuitively, world knowledge (WK) ineludes a causal 'law' gained from perception and experience that relates falling and pushing: 3 * Causal Law 3 Connected events el where x falls and e2 where y pushes z are normally such that e2 causes el.</Paragraph> <Paragraph position="1"> There is no similar law for standing up and greeting. The above law is a de feasible law. Our claim is that it forms the basis for the distinction between (1) and (3), and that defeasible reasoning underlies the temporal interpretation of text.</Paragraph> <Paragraph position="2"> First consider text (1). Intuitively, if there is no temporM information at all gained from WK or syntactic markers (apart from the simple past tense which is the only temporal 'expression' we consider here), then the descriptive order of events provides the only vital clue as to their temporal order, and one assumes that descriptive order matches temporal order. This principle is a re-statement of Grice's (1975) maxim of Manner, where it is suggested that text should be orderly, and it is also motivated by the fact that the author typically describes events in the order in which the protagonist perceives them (cf. Dowty (1986)). This maxim of interpretation can be captured by the following two laws:</Paragraph> </Section> </Section> <Section position="5" start_page="55" end_page="55" type="metho"> <SectionTitle> * Narration </SectionTitle> <Paragraph position="0"> Unless there's information to the contrary, clauses a and j3 that are discourse-related are such that Narration(a, ~) holds.</Paragraph> </Section> <Section position="6" start_page="55" end_page="67" type="metho"> <SectionTitle> * Axiom for Narration </SectionTitle> <Paragraph position="0"> If Narration(a, fl) holds, and a and fi describe the events el and e2 respectively, then el occurs before e2.</Paragraph> <Paragraph position="1"> Narration is defensible and the Axiom for Narration is indefeasible. The idea that Gricean-style pragmatic maxims should be represented as defeasible rules is suggested in (Joshi, Webber & Weischedel (1984)).</Paragraph> <Paragraph position="2"> The above rules can be defined in MASH--a logic for defensible reasoning described in (Asher & Morrean, 1991). We will demonstrate shortly that an intuitively compelling pattern of defensible inference can then underly the interpretation of (1).</Paragraph> <Paragraph position="3"> MASH supplies a modal semantics for a language with a default or generic quantifier, and a dynamic partial semantics of belief states is built on top of this modal semantics to c~pture intuitively compelling patterns of non-monotonic teasoning. We use a propositional version of MASH here. Defaults are represented as C/ > C/ (read as &quot;C/ then C/, unless there is information to the contrary&quot;). The monotonic component of the theory defines a notion of validity ~ that supports axioms such as ~ \[:3(C/ --* C/) ~ ((X > C/) --~ (X > C/)). The dynamic belief theory supplies the nonmonotonic component, and the corresponding nonmonotonic validity, ~, describes what reasonable entailments follow from the agent's beliefs. supports (at least) the following patterns of common sense reasoning: Defensible Modus Ponens C/>C/,C/ ~ C/ but not C/>C/,C/,-~C/ ~ C/ e.g. Birds fly, Tweety is a bird ~ Tweety flies, but not: Birds fly, Tweety is a bird that doesn't fly ~ Tweety flies.</Paragraph> <Paragraph position="4"> Penguin Principle C/ >C/,C/>CC/>-~,C/ ~-~i but not: C/ > C/,C/ :> (,C/ > -,(,C/ ~ ( e.g. Penguins are birds, Birds fly, Penguins don't fly, Tweety is a Penguin ~ Tweety doesn't fly, and does not ~ Tweety flies.</Paragraph> <Paragraph position="5"> Nixon Diamond not (C/ > C/,I > &quot;C/,C/,( ~ C/ (or --C/)) e.g. There is irresolvable conflict in the following: Quakers are pacifists, Republicans are nonpacifists, Nixon is a Quaker and Republican. We assume a dynamic theory of discourse structure construction in which a discourse structure is built up through the processing of successive clauses in a text. To simplify our exposition, we will assume that the basic constructs of these structures are clauses. 4 Let (4,13) mean that the clause ~ is to be attached to the clause a with a discourse relation, where a is part of the already built up discourse structure. Let me(a) be a term that refers to the main eventuality described by a (e.g. me(Max stood up) is the event of Max standing up). 5 Then Narration and the axiom on Narration are represented in MASH as follows (cl -~ e.~ means &quot;el wholly occurs before e2&quot;): 4The theory should extend naturally to an account where the basic constructs are segments of text; the approach adopted here is explored extensively in Asher (forthcoming).</Paragraph> <Paragraph position="6"> 5me(c~) is formally defined in Lascarides & Asher (1991) in a way that agrees with intuitions.</Paragraph> <Paragraph position="8"> We assume that in interpreting text the reader believes all LK and WK (and therefore believes Narration and its axiom), the laws of logic, and the sentences in the text. The sentences in (1) are represented in a DnT-type framework as follows: 6 (7) \[e1,~1\]\[~1 <now, hold(el,Q),s~andup(rn, el)\] (8) \[~, t~\]\[t2 < now, hold(~2, t2),gr~t(j, m, ~2)\] In words, (7) invokes two discourse referents el and ~1 (which behave like deictic expressions), where el is an event of Max standing up, tl is a point of time earlier than now and et occurs at it. (8) is similar save that the event e2 describes John greeting Max. (7) and (8) place no conditions on the relative temporal order between et and e2. These are derived at a higher level of analysis than sentential semantics by using defensible reasoning.</Paragraph> <Paragraph position="9"> Suppose that the reader also believes that the clauses in text (1) are related by some discourse relation, as they must be for the text to be coherent. Then the reader's beliefs also include (7, 8). The natural interpretation of (1) is derived by calculating the common sense entailments from the reader's belief state. Given the assumptions on this state that we have just described, the antecedent to Narration is verified, and so by Defensible Modus Ponens, Narration(7, 8) is inferred. Since the belief states in MASH support modal closure, this result and the Axiom on Narration entail that the reader believes the main eventuality of (7), namely el, precedes the main eventuality of (8), namely e2. So the intuitive discourse structure and temporal interpretation of (1) is derived by exploiting defeasible knowledge that expresses a Gricean-style pragmatic maxim.</Paragraph> <Paragraph position="10"> But the analysis of (1) is satisfactory only if the same technique of exploiting defeasible rules can be used to obtain the appropriate natural interpretation of (3), which is different from (1) in spite of their similar syntax.</Paragraph> <Paragraph position="11"> eFor the sake of simplicity we ignore the problem of resolving the NP anaphora in (8). The truth definitions of (7) and (8) are llke those given in DRT save that they are evaluated with respect to a possible world index since MASH is modal.</Paragraph> <Paragraph position="12"> (3) a. Max fell.</Paragraph> <Paragraph position="13"> b. John pushed him.</Paragraph> <Paragraph position="14"> As we mentioned before, Causal Law 3 will provide the basis for the distinct interpretations of (1) and (3). The clauses in (3) must be related by a discourse relation for the text to be coherent, and therefore given the meanings of the discourse relations, the events described must be connected somehow. Therefore when considering the domain of interpreting text, one can re-state the above causal law as follows: 7 Causal Law 3 Clauses a and/3 that are discourse-related where a describes an event el of x falling and/3 describes an event e~ of y pushing x are normally such that e2 causes el.</Paragraph> <Paragraph position="15"> The representation of this in MASH is:</Paragraph> <Paragraph position="17"> This represents a mixture of WK and linguistic knowledge (LK), for it asserts that given the sentences are discourse-related somehow, and given the kinds of events that are described by these sentences, the second event described caused the first, if things are normal.</Paragraph> <Paragraph position="18"> The logical forms for (3a) and (3b) are the same as (7) and (8), save that standup and greet are replaced respectively with fall and push.</Paragraph> <Paragraph position="19"> Upon interpreting (3), the reader believes all defeasible wK and LK together with (3a), (3b) and (3a, 3b). Hence the antecedents to two defeasible laws are satisfied: Narration and Causal Law 3.</Paragraph> <Paragraph position="20"> Moreover, the antecedent of Law 3 entails that of Narration, and the laws conflict because of the axiom on Narration and the axiom that causes precede effects:</Paragraph> <Paragraph position="22"> The result is a 'Complex' Penguin Principle: it is complex because the consequents of the two defeasible laws are not ~ and -~ff, but instead the laws conflict in virtue of the above axioms. MASH supports the more complex Penguin Principle: ;'This law may seem very 'specific'. It could potentially be generalised, perhaps by re-stating el as x moving and e2 as y applying a force to x. For the sake of brevity we ignore this generalisation.</Paragraph> <Paragraph position="23"> but not: \[\] (C/ --* C/), C/ > X, C/ > (, o (x 0), n (C/ -. C/ x Therefore there is a defeasible inference that the pushing caused the falling from the premises, as required.</Paragraph> <Paragraph position="24"> The use of the discourse relation Explanation is characterised by the following rule:</Paragraph> <Paragraph position="26"> In words, if a and f~ are discourse-related and the event described in/3 caused the event described in a, then Explanation(a, ~) normally holds. Furthermore, Explanation imposes a certain temporal structure on the events described so that if is a causal explanation of a then fPs event doesn't precede a's:</Paragraph> <Paragraph position="28"> The antecedent to Narration is verified by the reader's beliefs, and given the results of the Complex Penguin Principle above, the antecedent to Explanation is also verified. Moreover, the antecedent to Explanation entails that of Narration, and these laws conflict because of the above axioms. So there is another complex Penguin Principle, from which Explanation(3a, 3b) is inferred.</Paragraph> <Paragraph position="29"> The second application of the Penguin Principle in the above used the results of the first, but in nonmonotonic reasoning one must be wary of dividing theories into 'subtheories' in this way because adding premises to nonmonotonic deductions does not always preserve conclusions, making it necessary to look at the theory as a whole.</Paragraph> <Paragraph position="30"> (Lascarides & Asher, 1991) shows that the predicates involved in the above deduction are sufficiently independent that in MASH one can indeed divide the above into two applications of the Penguin Principle to yield inferences from the theory as a whole. Thus our intuitions about the kind of Certain constraints are imposed on discourse structure: Let R be Explanation or Elaboration; then the current sentence can be discourse related only to the previous sentence a, to a sentence fl such that R(fl, a), or to a sentence 7 such that R(7, fl) and R(~, a). This is a simpler version of the definition for possible attachment sites in Asher (forthcoming). Pictorially, the possible sites for discourse attachment in the example structure below are those marked open: There are structural similarities between our notion of openness and Polanyi's (1985). The above constraints on attachment explain the awkwardness of text (9a-f) because (9c) is not available to (gf) for discourse attachment.</Paragraph> <Paragraph position="31"> (9) a. Guy experienced a lovely evening last night.</Paragraph> <Paragraph position="32"> b. He had a fantastic meal.</Paragraph> <Paragraph position="33"> c. He ate salmon.</Paragraph> <Paragraph position="34"> d. He devoured lots of cheese.</Paragraph> <Paragraph position="35"> e. He won a dancing competition.</Paragraph> <Paragraph position="36"> f. ?He boned the salmon with great ex- null pertise.</Paragraph> <Paragraph position="37"> According to the constraint on attachment, the only available sentence for attachment if one were to add a sentence to (1) is John greeted him, whereas in (3), both sentences are available. Thus although the sentences in (1) and (3) were assigned similar structural semantics, they have very different discourse structures. The events they flict between defeasible laws whose antecedents axe such that one of them entails the other is resolvable. Thus unwanted irresolvable conflicts can be avoided. describe also have different causal structures. These distinctions have been characterised in terms of defeasible rules representing causal laws and pragmatic maxims. We now use this strategy to analyse the other texts we mentioned above.</Paragraph> <Paragraph position="38"> Elaboration Consider text (2).</Paragraph> <Paragraph position="39"> (2) a. The council built the bridge.</Paragraph> <Paragraph position="40"> b. The architect drew up the plans.</Paragraph> <Paragraph position="41"> We conclude Elaboration(2a, 2b) in a very similar way to example (3), save that we replace cause(me(~), me(a)) in the appropriate defensible rules with prep(me(~), me(a)), which means that rne(~) is part of the preparatory phase of me(a). In Law 2 below, Info(a,~) is a gloss for &quot;the event described in a is the council building the bridge, and the event described in fl is the architect drawing up the plans&quot;, and the law represents the knowledge that drawing plans and building the bridge, if connected, are normally such that the former is in the preparatory phase of the latter:</Paragraph> <Paragraph position="43"> The inference pattern is a Cascaded Penguin Principle again. The two resolvable conflicts are Law 2 and Narration and Elaboration and Narration.</Paragraph> <Paragraph position="44"> Background Intuitively, the clauses in (4) are related by Background. null (4) Max opened the door. The room was pitch dark.</Paragraph> <Paragraph position="45"> The appropriate reader's belief state verifies the antecedent of Narration. In addition, we claim that the following laws hold: describe an event and state and we have no knowledge about how the event and state are connected, gained from WK or syntactic markers like because and therefore, we assume that they temporally overlap. This law can be seen as a manifestation of Grice's Maxim of Relevance as suggested in (Lascarides, 1990): if the start of the state is not indicated by stating what caused it or by introducing an appropriate syntactic marker, then by Grice's Maxim of Relevance the starting point, and is irrelevant to the situation being described. So the start of the state must have occurred before the situation that the text is concerned with occurs. As before, we assume that unless there is information to the contrary, the descriptive order of eventualities marks the order of their discovery. This together with the above assumption about where the state starts entail that unless there's information to the contrary, the state temporally overlaps events or states that were described previously, as asserted in States Overlap.</Paragraph> <Paragraph position="46"> We assume that the logical form of the second clause in (4) entails state(me(~)) by the classification of the predicate dark as stative. So Background is derived from the Cascaded Penguin Principle: the two resolvable conflicts are States Overlap and Narration and Background and Narration. States Overlap and Narration conflict because of the inconsistency of overlap(el,e~) and el -~ e~; Background and Narration conflict because of the axioms for Background and Narration.</Paragraph> <Section position="1" start_page="67" end_page="67" type="sub_section"> <SectionTitle> Result </SectionTitle> <Paragraph position="0"> (5) has similar syntax to (4), and yet unlike (4) the event causes the state and the discourse relation is Result.</Paragraph> <Paragraph position="1"> (5) a. Max switched off the light.</Paragraph> <Paragraph position="2"> b. The room was pitch dark.</Paragraph> <Paragraph position="3"> Let Info(a,fl) be a gloss for &quot;me(a) is Max switching off the light and me(fl) is the room being dark&quot;. So by the stative classification of dark, Info(a, fl) entails state(me(~)). Then Law 5 reflects the knowledge that the room being dark and switching off the light, if connected, are normally such that the event causes the state: 9</Paragraph> <Paragraph position="5"> The use of the discourse relation of Result is characterised by the following:</Paragraph> </Section> </Section> class="xml-element"></Paper>