File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/89/e89-1006_metho.xml
Size: 20,833 bytes
Last Modified: 2025-10-06 14:12:13
<?xml version="1.0" standalone="yes"?> <Paper uid="E89-1006"> <Title>TENSES AS ANAPHORA*</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Representing events in DKT </SectionTitle> <Paragraph position="0"> In the framework of Discourse Representation Theory (DRT) (Kamp(1981)) a Discourse Representation Structure (DRS) is a pair < U, K > consisting of a set U of discourse referents (DRFs) and a set K of conditions. Discourse Referents are assumed to be sorted according to the following sort system (for our purposes only the temporal branch is relevant): all, a temporal, t non-temporal, x dynamic state, s event, e process, p We shall use the indicated letters for DRFs of the corresponding sort 2. Conditions take the following forms:</Paragraph> <Paragraph position="2"> P(al,..., an), where P is an n-ary predicate symbol and the ai are discourse referents. Conditions of this form are also called atomic. For 2-place temporal relations we will also use infix notation.</Paragraph> <Paragraph position="4"> t: DRS, where t is a temporal DRF Thus DRT uses a variant of the Davidsonian method of talking about events: they are treated as a kind of objects. But DRT deviates from the Davidsonian representation in that instead of using additional argument places in the predicates 2For simplicity we will use the term event not just for the corresponding subsort of dynamic but also for dynamic or temporal objects in general. For the same reasaa we do not make a clear distinction between &quot;aspect&quot; and &quot;Aktionsarten&quot;. The intended meaning should be clear from the context. The classification is related to the one given in VendieT(1967).</Paragraph> <Paragraph position="5"> an event discourse referent is put before a DttS representing the nature of the event. This allows for a recursive definition of aspects and to account for aspectual change. So the process p of z's travelling would be written in DRT as and the event of x's travelling to Paris would be written as p x paris e': p: I travel(x) 1 goal(p, paris) DRSs containing conditions as these are embedded into models by mapping the temporal DRFs onto objects in a domain of temporal objects. A proper embedding is realized if the value of the event-DRF fulfills conditions represented by the sub-DRS which the event introduces with respect to an interpretation function defined among other things on predicates such as travel in the example above. An advantage of this representation on the one hand is that, following the usual definition of accessibility of DRSs from a DRS used in DRT, restrictions on the accessibility of DRFs as possible antecedents for anaphoras can be expressed, though for our purposes this will play only a subordinate role. On the other hand different modes of existence can be discriminated for the objects DRFs stand for, depending on the position of the sub-DRSs, where these DRFs are introduced, in the DRS of the whole text. In the case of reported speech for instance, the real existence of a reported episode is not necessary. Extensions of the original modeltheory ofDRT (cf. Kamp(1981)) which point in this direction are given e.g. in (Asher(19S6), Reyle(1985), Eberle(1988b)).</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 Basic Assumptions of the </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Semantics of Tenses and Adverbials </SectionTitle> <Paragraph position="0"> The basic tenses of French narrative texts are imparfait and pass~ simple. The interaction of these tense forms is often described by two pairs of opposition: On the one hand the pass~ simple is used to describe actions of the narrative, &quot;the course of events&quot;, whereas the imparfait serves to paint the -44background of the story. On the other hand the imparfait can be used to describe events in progress as viewed from &quot;inside&quot;, whereas the pass~ simple presents the event as a punctual entity. In order to reflect these dichotomies we require that an event introduced by pass~ simple serves as new reference time which must not start before the old one, and that the imparfait introduces a state which includes temporally the existing reference time. In this case no new reference time Is created.</Paragraph> <Paragraph position="1"> In the case of pass~ simple we do not require that the new reference time has to follow completely the old one in order to deal correctly with discourse situations as 'elaboration' or 'contrast' and others. To discriminate such textual functions an elaborate inference component is needed, which at present is not available. In addition there are cases where this inference component would need information about the proceeding of the whole story. This cannot be made available at this stage of processing.</Paragraph> <Paragraph position="2"> Thus an underspecified relation not-before is necessary which can be defined, as other relations like subset (needed for the imparfait), out of the relations overlap and before, the only basic temporal relations used in the system s .</Paragraph> <Paragraph position="3"> The plusqueparfait can be understood as perfective state giving background information with respect to the actual reference time of the story, (Jean avait dgj~ mange'), or as introducing or continuing a flashback.</Paragraph> <Paragraph position="4"> The conditionnel is understood as a counterpart to the plusqueparfait describing an anticipation with respect to a perspective point in the past. We think that pass~ compos~ in (written) narrative texts should be treated as analogue to the pass4 simple with respect to pure temporal relations 4 .</Paragraph> <Paragraph position="5"> Temporal adverbials provide a location time for events in relation to the temporal structure of the preceding text. They can differ from each other by their characterization of the location time and their anaphoric behavior. Deictic adverbials like demain, ia semaine derni~re for instance create location times of different temporal extension with different ordering conditions regarding the evaluation time (after and before), but they are similar in that the evaluation time must be the speech tem cf. Kamp/Rohrer(1985), Eberle(1988c).</Paragraph> <Paragraph position="6"> time, whereas in case of le lendemain, la semaine prdcddante the temporal relations and extensions are equivalent to the deictic analogues but the time of evaluation has to be a past reference time or perspective point. Frame-time adverbials like ce jourla can be distinguished from punctual time adverbials like fi trois heures and from adverbials, like puis, which simply state a temporal relation between the event to be introduced and a temporal antecedent. Some adverbials, like puis and ensuite, do not restrict the nature of the antecedent, it is just a reference time. But maintenant e.g. requires that the evaluation time is a perspective point of the text. The resolution component has to take into account such phenomena.</Paragraph> </Section> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Aspects of Implementation </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.1 Architecture </SectionTitle> <Paragraph position="0"> The construction of the semantic representation for a discourse proceeds in several stages: each sentence is parsed using a Lexical Functional Grammar (LFG) (Kaplan/Bresnan(1982), Eisele/DSrre(19S6)) which analyzes sentences into functional structures (f-structures), augmented by indices to indicate the linear order of words in the input string. The f-structure serves as input for the construction of a proto-semantic representation (cf. Reyle(1985), Reinhardt/Kasper(1987)). The last stage consists in integrating this representation into a semantic representation for the discourse, mainly by doing the necessary resolutions for anaphoric expressions. Accordingly, the system consists of three major modules: LFG-Parser: I f-structure Composer: I</Paragraph> <Paragraph position="2"> The whole system is implemented in PRO-LOG. Here mainly the Composer and the Resolver will be discussed with respect to the treatment -45of the tenses in these modules. The inference machine and knowledge base are at present not implemented. The proto-semantic representation for a sentence built up by the Composer differs from the semantic representation proper in that it not just contains the semantic information available from the sentence but also morpho-syntactic information from the f-structure needed to constrain the resolution process. Thereby, this information is passed to the Resolver which separates semantic and syntactic information and uses it. What sort of morpho-syntactic information is passed will be discussed later.</Paragraph> <Paragraph position="3"> The concept of resolution here is broader than the usual one which comprises mainly determining the reference of anaphoric expressions like pronouns. We use the term as covering all kinds of context dependency beyond the single sentence level where something in the sentence has to be related to some entity in the preceding discourse. The term temporal resolution will be used to refer to the process of determining the temporal structure of the events the discourse is about.</Paragraph> <Paragraph position="4"> The Resolver is intended to implement good heuristics based on purely linguistic knowledge.</Paragraph> <Paragraph position="5"> The evaluation of the readings produced should be left to the inference machine which also can access non-linguistic world knowledge.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.2 Temporal Features in the F- Structure </SectionTitle> <Paragraph position="0"> In the Kamp/Rohrer system the tenses are analyzed by means of four features which have temporal and aspectual meaning: tions in discourse. Plusqueparfait for instance has a reading as flashback, and a perfective state reading with the temporal perspective &quot;speech-time&quot; or actually reached &quot;reference time&quot;. Imparfait can have the perspective in the past at the reference time or at the speech time. Following Kamp/Rohrer(1985), the meanings of the main narrative tenses can be described in the following way: biguities syntactically we use the Kamp/Rohrer categories in a slightly different way to get unique descriptions of the tenses. It is completely left to the resolver to account for these ambiguities. Since we exclude the TP-feature we need the additional TENSE-value conditionnel. To mark tenses in indirect discourse the transposed-feature is added: feature value</Paragraph> <Paragraph position="2"> Since we do not discuss embedded clauses in this paper, in the following the transposed-feature is skipped. The tenses are analyzed by these means as shown in Table 1.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.3 Tenses and Temporal Adjuncts </SectionTitle> <Paragraph position="0"> in the Composer Here we will discuss what sort of information the Composer adds to the Proto-DRS when it encoun- null ters a tense feature or temporal adjunct or subclause. It consists basically of two kinds: 1. DRS-conditions, which do not depend on the resolution process 2. Interface structures for the Resolver, called occurrence information and represented as a 6place occ-term.</Paragraph> <Paragraph position="1"> The occurrence information is used to transmit morpho-syntactic information from the parser to the resolver. For the tenses this occurrence information has the form oec(DRF, Pos, TF, Tense, tense, Pointer) - 46 imparfait: perf-- -, prog - +, tense ---- past pass~ simple: perf--- -, prog = -, tense ---- past pass~ compos~: perf -- ~-, prog -- -, tense -- present plusqueparfait: perf -- -I-, prog ---- _, tense -- past</Paragraph> <Paragraph position="3"> with the following slots: DttF the temporal discourse referent, which the resolution process has to locate, usually introduced by the verb Pos verb position in surface structure, TF tf(Perf, Prog, Tr), the temporal features term. The Perf-slot marks the analytic tenses, Prog serves to distinguish for instance imparfait and passd simple, Tr stands for the value of the lransposed feature.</Paragraph> <Paragraph position="4"> Tense past I present I future I conditionnel, values of the tense feature, tense marker, indicating that the occ-term stems from a tense feature. It is also the trigger for temporal resolution.</Paragraph> <Paragraph position="5"> Pointer indicates the occurrence of the tense in main or embedded clauses. The clauses are indexed in a unique way (by natural numbers). The pointer is a pair consisting of the index of the superordinate clause and the index of the clause itself. Main clauses point to themselves. Such an indication is necessary for the treatment of embedded sentences. The pointer encodes a simplified tree structure for the sentence and allows moving around in the tree. Temporal adjuncts and subclauses also provide occurrence information marked in a special way. They contribute an occ-term of the following kind: occ(DRF, Pos,_ ,Rel, sel- lime, Pointer) DRF here represents the time introduced by the adverbial Pos its position in the surface structure Rel the temporal relation introduced. For instance trois jours avant introduces before.</Paragraph> <Paragraph position="6"> set-time indicates a special resolution mode for temporal adjuncts and indicates that this information was contributed by a temporal adjunct. In the resolution process the marker will cause the DRFs of the tense markers to resolve to DRF.</Paragraph> <Paragraph position="7"> The information shown in table 2 is introduced by the Composer for the tenses (the &quot;&quot; represent initially empty slots which get filled in in the process of combining the meanings).</Paragraph> <Paragraph position="8"> It will be noticed that the tenses do not introduce new conditions into the DRS since the temporal relations cannot be determined without respect to the nature of the temporal 'antecedent' and therefore have to be generated in the Resolver.</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.4 Temporal Resolution </SectionTitle> <Paragraph position="0"> For temporal resolution the Resolver uses a stack of a system of times consisting of quintuples of the form: - 47 1. reference time (usually the last event) 2. temporal perspective point 3. temporal location time (usually identical to the reference time) 4. speech time (at present kept constant for the whole discourse) 5. last resolved tense (with its occurence information) null Every resolution process generates such a quintuple which gets stacked. If the temporal perspective point is changed (plusqueparfalt and conditionnel), a substack is created and used till the original perspective point is restored.</Paragraph> <Paragraph position="1"> The resolver removes the occurrence information for the tenses and temporal adjuncts from the proto-DRS. The tenses get resolved according to the rules discussed below. The presence of temporal adjuncts changes the flow of resolution as it requires that the temporal DRF introduced by the adjunct has to be resolved in accordance with the DRF introduced by the tense.</Paragraph> <Paragraph position="2"> We will illustrate the effect of some resolution rules reflecting the heuristics of the system by discussing two sample texts.</Paragraph> <Paragraph position="3"> sl Ce jour-l~ il pleuvait tr~s fort.</Paragraph> <Paragraph position="4"> ce jour-ld introduces the location time t for the first state, sl, (pleuvoir). The Composer augments the DRS by the condition day(t) and the Resolver by the condition t C sl. 5 The first times-quintuple consists of a variable for the reference time, (no event is actually mentioned), the perspective point is assumed to be the speech time. The speech time is fixed by &quot;now&quot;, t is the location time and in the last position the occ-term of sl is stored. Since a reference time does not yet exist, the integration of e2 produces a temporal relation with respect to the last location time: e2 C t, i.e. e2 happens within t. A second times-quintuple is put onto the stack with the reference time e2 and the new occterm. The other time coordinates remain constant. SBecause of the definite description ce jour-ld the NP-Resolver has to establish an equation between t and a DRF of the preceeding text which is a day. If there is no such antecedent t has to be accomodated. The latter ease is at present not implemented.</Paragraph> <Paragraph position="5"> e2 serves now as reference time for s3. The new relation e2 C s3 is introduced and the information which stems from the occ-term of s3 and the old tenses is put as a third times quintuple onto the stack. The reference time is not changed. It should be noticed that for new states the &quot;smallest&quot; available location time is used. Normally this is the reference time of the previous quintuple if existent. For it is not necessary that the explicit given loca-tion time, ce jour-lh in the given example, serves as location time for subsequently introduced states as well. e4 is ordered with respect to e2, e5 with respect to e4 by the not-before-relation and the timesstack is updated in the obvious way. All states and events are located obviously before the speech time now. We omit tile full DRS of the example here.</Paragraph> <Paragraph position="6"> According to the different meanings of the plusqueparfait mentioned in section 3 different rules are available to deal with the plusqueparfait of the second sentence. However, especially in cases where a frame time as in the example exists, the possibility to introduce a flashback is preferred. A stack in the stack is created and the new event serves as reference time for subsequent events in the flashback. The last reference time of the higher level is now regarded as temporal perspective for the events occcuring in the flashback. They are localized before that perspective, le lendemain in the third sentence has to be resolved to an existing reference time, i.e. el or e2. Since we are already in a flashback, in processing the plusqueparfait of e3 the continuation of this flashback is preferred. Thus a solution with el as antecedent for le lendernain would lead to a cyclic structure and should be ruled out by the inference component. The correct ordering conditions are given by establishing e2 as antecedent for the time introduced by le iendemain, and t as location time for e3. The perspective time is copied from the stack. Thus we get the conditions: null e2 < t, e3 C t,e3 < el The embedded stack is updated by the new quintuple. null The implemented heuristics require that the tense switch from a plusqueparfait of a flashback to pass~ - 48 simple or pass~ compos~ or the explicit reference to the perspective point, for instance by means of &quot;maintenant&quot;, always lead to a reactivation of the initial level, from which the first flashback started, that is, all substacks are popped. Without maintenant the imparfait of sl could lead to the continuation of the flashback or to the continuation of the main story. It is regarded as the state variant of both categories.</Paragraph> <Paragraph position="7"> The (simpified) DRS of the example above thus looks as follows: In our opinion cases as in the example above cannot be treated without adding new parameters to the Reichenbachian system. At least doing it facilitates the job. Beyond the imparfait/pass~ simple, pass~ compos~-distinction French does not make use of an explicit morphological aspect marking. Therefore, for instance in the case of conditionnel, treated as anticipation of an episode, we use the Aktionsart-characterization of the verb stored in the temporal sort of the DI~F to specify the value of the prog-feature. An episode of states and dynamics then is treated similar to an imparfait-pass~ simple-story transposed by the stored (past) perspective time. By this means we get an interaction of Aktionsarten and tenses.</Paragraph> <Paragraph position="8"> Similar to the case of flashback the tense switch from conditionnel to another past tense form marks the end of the anticipation and the reactivation of the initial level.</Paragraph> </Section> </Section> class="xml-element"></Paper>