<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-1046">
  <Title>On Representing the Temporal Structure of a Natural Language Text*</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Discourse Relations
</SectionTitle>
    <Paragraph position="0"> It is clear that the idea about tile order of events introduced by sequences of sentences that wc have sketched in the introduction is an oversilnpfification.</Paragraph>
    <Paragraph position="1"> Iu the folh)wing we will restrict ourselves to the tenses prevailing in narrativcs, silnple past and past perfect, and to only some of the text pllenomena related to these tenses (as we have done ill tile LILOG ilnl)lelnentation with respect to tile German counterparts of these tenses).</Paragraph>
    <Paragraph position="2"> }3csi(les tile disl:ourse relations used ill CXalnple (1), i.e. tl~shback and continuation, there are at least two nlore, which we call elabo~vltlon and backgTound.</Paragraph>
    <Paragraph position="3"> Conlpare the two fi)\]lowing exalllples:  (2) Mouday Hans went to Par~s. Tuesday he met Gabt.</Paragraph>
    <Paragraph position="4"> (3) Monday Hans went to Paris. At the border he had some trouble.</Paragraph>
    <Paragraph position="5"> (2) dcscril)es a continuation. This can lie deduced from the characteristics of the location times.</Paragraph>
    <Paragraph position="6"> 2 8 9 PRec. oY COLING-92, NANTES, AUG. 23-28, 1992  In contrast to that, (3) describes an elaboration. We call the relation between a new event e2 and its reference event el an elaboration, if e2 describes et on a nmre fine-grained level (which gives rise to the temporal condition of inclusion (C_) between the new event and tile reference event within the representation of the text).</Paragraph>
    <Paragraph position="7"> In (3) the relation of elaboration can be deduced only if there is some input from the knowledge base equipping the resolution component with facts about journeys to foreign countries. Notice that, here, the temporal inclusion is due to the hommnorphic local inclusion of tile border into the path of the travel. This illustrates that an intelligent temporal resolutimt component cannot do without having access to local reasoning.</Paragraph>
    <Paragraph position="8"> Now compare the following pair of examples, taken from Dowty (1986) and from Hinrichs (1986) respectively:  (4) John entered the president's office. The clock on the wall ticked loudly.</Paragraph>
    <Paragraph position="9"> (5) Jameson entered the room, shut the door carefully and switched off the hght. It was pitch-dark around hint because the venetian blinds were closed.</Paragraph>
    <Paragraph position="10"> hi both (4) and (5) the last sentence introduces not  all event proper, but a state. As widely accepted, the normal case of relating a new state e2 to an existing reference event el seems to be that of inelusiml (e, C_ e2). l However, as exaalples like (5) make clear, tills default can be overwritten. Here ton, it seems that backgrmmd knowledge gives the decisive indication about when the default has to be upset.</Paragraph>
    <Paragraph position="11"> We call the discourse relation of (4) background and that of (5) causally introduced background. For cases like (5) we introduce tile new state e~ a.s following the actual reference event el (el -&lt; e2) and as including a dummy reference time t (t C e:) with respect to which e2 introduces a backgrouud, ht the filrther course of processing the text, t has to be instautiated by a new event which is understood as following el.</Paragraph>
    <Paragraph position="12"> The T-list makes use of these discourse relations a-s structuring elements. For instance, processing the samlfle text (6), which integrates tile text phenomena reflected by the examples (1) - (4), should result in a representation which has as index the T-list (7): {6) John left Paris by the Porte de 5 ~ Cloud late in the evening (el). He went to Frankfort (e2).</Paragraph>
    <Paragraph position="13"> At the border he was stopped (e3). It was pretty dark (eJ. Only a .few lights were burning (es). He had to open hz.~ boot (es). In Paris he had bought five cartons of e:garettes (e7) and had hidden them in the car (es). The customs officers missed them (eg}.</Paragraph>
    <Paragraph position="14"> IThis parallels, for instance, the insights of Partee(1984), H inrichs( 1986l, Kamp and Rohrer(1985) and others. Early in the morning John entered a bar near the mess tower in Frankfort (elo), ordered a beer (el1) and lighted a cigarette (el2). He was happy (eta).</Paragraph>
    <Paragraph position="16"> Ill (7) the structure of (6) is stored in the following way: el, c2, sis, el 1, el2 (the leaving of Paris, the going to Frankfort, the entering tile bar, the ordering of beer and the lighting the cigarette) on tile one hand and ea,eo, e0 (tile stoppiug, the opening of the boot, the failure of the officers) on the other form continuation chains (tied up in the picture by -- ). The second of these chains elaborates an event of the first, e2, tile transition to Prankfort (el/\). e4 and e~ serve as background to ea, cla is the background of e12 (bg/\).</Paragraph>
    <Paragraph position="17"> Finally, the continuation chain c7 - ea is a flashback seen from the perspective es.</Paragraph>
    <Paragraph position="18"> We define tile accessibility of all event as reference event in such a way that, having processed the whole text, the only accessible event will be elz, whereas, in the case of having processed the text till the integration of es, the accessible events will be es, es and e2 (with decreasing prominettce). In this situation es is the so called actual reference event. When integrating a new sentence, one will try to expand the node represented by the actual referent first. But, finally, what level has to be chosen and what kind of anchoring has to be stored in the next step is decided by the system rules. So, for instance, in the case of (6), by purely linguistic reasons, ea cannot be related to tile actual reference event e8, because there is a tense switch from past perfect to simple past between es and es which indicates the termination of the flashback. This results in tile re-initialization of the level ttle flashback started from as the level providing the new actual referent. In the case of (6) this is cs which is a member of an elaboration. Actual elaborations may be terminated by knowledge indicating that expanding the structure at the actual level is not plausible or even not possible. In this case the system will try to relate the new event with respect to tile (accessible) events of higher levels of the elaboration hierarchy. Since, with respect to (6), eg is compatible with e2 the system has no reason to terminate the level marked by e0. We think that the data confirm the concept of a hierarchy of reference events as used in the system. The exact defiuition of the accessible reference events of a T-list resides in the definition of the specific access-functions for T-lists that we sketch ACTE.S DE COLING-92, NANTES, 23-28 AOt~r 1992 2 9 0 PRec. OF COLING-92, NANTES, AUG. 23-28, 1992 in section 5. 2 In fact, ill tile LILOG implementation, tile items of the index, i.e. of the. T-list, do not consist of the pure events. They are terms with different slots filled hy the relevant temporal information such ms tile tense form which introduces the corresponding event, the ,xspect, the actual speech tittle etc. Skiplfing such technical details, the simplified syntax of the T-list is the following:</Paragraph>
    <Paragraph position="20"> where E(D-List): list item with E: the discourse referent for the event, and D-List: tile list of the items depending on the</Paragraph>
    <Paragraph position="22"> Here, of course, embeddings which are given by the dimensions by, fb, elin turn stand for background, flashback and elaboration. T'-List is a T-list where each D-list is the empty list. We use T'-List for the items ill bg-lists, because we do Ilot think that a background give rise to further embeddings. The neighborhood of T-list items of the same level signifies continuation, except the case where such a list is fl'onted by bg. In this case, in the corresponding DtLS the elements of the list include the same event and, therefore, overlap each other. Continuation is reflected in the reversed order, because we use T-list as a stack. For (6), according to (7), we write (8). Here mid in the following we abbreviate E(\[\]) by E:</Paragraph>
    <Paragraph position="24"> Of course, there are other discourse relations ~ those mentioned. Especially, if one tries to analyse the structure of texts on a more fine-grained level than the purely temporal one, the relations described are not sufficient, others like explanation, contrast etc. are needed, s For lack of space, here, we can nothing say about these. The relations described are those 2Anchoring new evetlts to non-accessible reference events clearly is possible, provided there is additional information to motivate this clmice, such as, for instance, definite descriptions using event nomlnalizations in order to establish ccreferentiallty, But, at present, such possibilities are not implemented. null</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Partly for other purposes discourse or rhetorical relations
</SectionTitle>
    <Paragraph position="0"> are tlsed, for instance~ in Grosz and Sidner(1985}, Thompson and Mann (1987}.</Paragraph>
    <Paragraph position="1"> dealt with ill tile LILOG system.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Background Knowledge
</SectionTitle>
    <Paragraph position="0"> The LILOG knowledge base provides us with a sort hierarchy for discourse referents which is related with respect to events to tile classification given in Vendler(1967). This sort hierarchy is used as one of several information packages within the temporal resolution. With respect to the hierarchy we mainly use calendar knowledge and incompatibilty-statements about event types. For instance, all event introduced with the location time 1987 (or with subintervals of 1987) cannot overlap with an event with location time 1988. Exactly the same is true when the new event has an event type which is incompatible with the type of the reference event. A SWIMMING-event is incompatible with a WRITING-event if tile relevant therustic roles are the same, i.e., here, the agent-role. Of course, when fornlulating such incompatibilities one tries to make use of the sort hierarchy. One will define the incompatibility for very general sorts (if possible) so that it is inherited by subsorts.</Paragraph>
    <Paragraph position="1"> Knowledge of this kind is hard information so to speak. In addition to this, the system uses default knowledge about event types which is similar to what is cldled event f~nmes (cf. Miusky (1975), Schank and Abelson (1977), Bartsch (1987) and others). Typical (simplified) examples of the two kinds of knowledge are (9) and (10):</Paragraph>
    <Paragraph position="3"> (9) regulates the localization of the theme of a movemeat. From this rule we infer the existence of different PLACED-AT-states for the thente within the movement. (10) predicts that normally (D fl)r default) turning a switch when it is lighted results in darkness.</Paragraph>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5 Temporal Resolution
</SectionTitle>
    <Paragraph position="0"> The construction of tile seulantic represeutation for a discourse proceeds in several stages. I shall skip here the steps leading to the DRS of a sentence. But it has to be stressed that tile result of sentence processing is not a completed DRS. There may be variables which have to be instantiated in the process of integrating the sentence-DRS into the text-DRS.</Paragraph>
    <Paragraph position="1"> The sentence-DRS is understood as the input of the aspect-calculus of the prototype which has to contpute the aspect of the events of the DRS. Therefore it uses tile inform~ttion provided by the thematic roles. The starting point is the aspect-entry for the ACTES DE COLING-92, NANTES, 23-28 ^OI3T 1992 2 9 1 PROC. OV COLING-92, NA~CrES, AUO. 23-28, 1992 verb in the lexicon. It can be overwritten by specific role information and morphological information with regard to aspect. For the sake of simplicity in the following we assume that there are just two contrasting aspects corresponding to the main sorts of the upper structure of the event sort hierarchy, the heterogeneous events proper (ev) and the honmgeneous states (st). The aspect calculus used is based on ideas fi'om Kritka (1987)2So, for Peter wrote a book we will get the aspect value ev, whereas for Peter wrote books we will get st. For the example (6) we get that e4, es, el3 are states, the rest will be analyzed as events proper.</Paragraph>
    <Paragraph position="2"> After tile computation of the aspect we come up with an indexed DRS its descrihed iu the second and third section with the value of the aspect feature instantiated. null Within the process of anchoring we use several access-functions with respect to the T-list. Among other things these are act-tel(T-list) which gives us the most topicalized item in the list, the actual referent, el-embedder( T-list, E), fb-embedder( T-list,E) which givo us the points elaborations and flashbacks containing E start from. It has to be stressed that reference events have to be events proper. They caunot be states. The actual referent is the first element of the stack, provided that this element does not contMn fl~hback- or elaboration-embcddings. If it does we start the recursivc call of the search procedure with respect to the first of these embedded lists, tense{El, aspect(El, and tp(E) give us the valnes of tile tense and aspect features (sp, ppf, ev, st: simple past, past perfect, event and state) and the temporal perspective from which E is seen.</Paragraph>
    <Paragraph position="3"> In order to decide the anchoring of new events, the rules of the system make use of three filters: (F1) tile tense/aspect filter which decides whether a particular discourse relation is possible with respect to tense and aspect, (F2) the &amp;quot;consistency&amp;quot; filter, which decides whether tile choice of a particular discourse relation is logically compatible with the information of the preceding text against the background of tile information from the inference component (this consists mainly of testing sort subsumptions ill the context of teml)oral incompatibilities between calendar units and event sorts), and (F3) the evidence filter which tries to support the relation tested by means of rules like (9) and (10).</Paragraph>
    <Paragraph position="4"> To illustrate the system we restrict ourselves to some simple eases (compare figure 1). We suppose that the new sentence introduces just one event anti we only sketch simplified versions of tile rules for continuation, 1.), for elaboration, 2.), for flashback, 4 For a detailed description of the calculus refer to Eberle {1991}. 3.), for causally introduced background, 4.), for background, 5.), for termination of elaborations, 6.), and for the termination of flashbacks, 7.). We skip rules which deal with specific subcases of these relations.</Paragraph>
    <Paragraph position="5"> For instance, we skip rules which deal with dummy referents t introduced by the causal background and rules which allow for more precise temporal localizations of tile new events. We use rewrite rules which depict the amalgamation of a sentence DRS with a text DRS.</Paragraph>
    <Paragraph position="6"> In rule 1.) of figure 1, (*) reflects the particular case where e,.ef is ntember of all elaboration. If it is not, tile corresponding condition in 1(,+1 is omitted.</Paragraph>
    <Paragraph position="7"> Ill tile cases 1.) - 3.) (F1) requires that es is heterogeneous (aspect(es) = ev). Tile tense forms of reference- and new event have to be identical</Paragraph>
    <Paragraph position="9"> arrangement allows also for continuation and elaboration on flashback levels and for iteration of flashbacks.</Paragraph>
    <Paragraph position="10"> For 4.) and 5.) tile tense forms must be identical. In contrast to 1.) - 3.), it is required that aspect(es) = st.</Paragraph>
    <Paragraph position="11"> (F2) should be clear in all cases. For acceptance, the tested solutious have to be consistent.</Paragraph>
    <Paragraph position="12"> With respect to tile cases 1.) - 4.), (F3) accepts the relation tested, if some evidence for this relation can be inferred from the representation of the preceding text on the be~sis of the background knowledge. Ill addition, in tile cases 1.) - 3.) (F3) also accepts the relation, if evidence for the coml)eting relations cannot be entailed. Ill tile case of tet~e(es) = tense(e,.el) = ppf, flashback, elaboration and continuation are pairwise competing relations. Ill the case of tense(es) = tense(e,,~I) = sp, fl,'~shbaek is ruled out as competing relation with respect to elaboration and eontinuatiolL In the case of 5.), background, and 3.) with tense(es) C/ tense(e,.eI) (F3) is eml)ty.</Paragraph>
    <Paragraph position="13"> Evidence for a particular relation X can be inferred, if there exists an event sort P which is more general than the event description from DRSs with respect to tile event sort hierarchy, such that tile knowledge base (KB) together with the representation of the l)receding text (DRS,) predicts an event e auf type P temporally related to e,.~y in a way significant for X, and that KB, DRS, do not entail the corresponding statement with respect to the competing relations of X. Formally, for instance with respect to X=&amp;quot;eontinuation&amp;quot;, X is evident, if for P with Ve (ORS s\[es/e\] ---* P(e)), we can infer that 3e (P(e) A e,~j&amp;quot; -&lt; el, but neither that ~e (P(e)AeCe,.~l)nor3e (P(e)Ae&lt;e,ef).</Paragraph>
    <Paragraph position="14"> Tile whole system is implemented in Prolog. So tile order of the rules is important. Thus, we get a preference relation over the allowed readings. Continuation is preferred to elaboration which is preferred to flashback (ill the case that all alternatives seem possible and that they are supported by (F3)). With respect to 6.) and 7.), we see that local level expansion is preferred to the anchoring of the new event at  T,s =-\[es\] higher levels of the hierarchy. This strategy seems to be validated by the text phenomena.</Paragraph>
    <Paragraph position="15"> For illustration, we briefly sketch the effects of the rules when applied to the sample text (6). Leaving Paris- and Going to Frankfort-events cannot overlap. Therefore, (F2) rules out the alternative &amp;quot;elaboration&amp;quot; for e~ with respect to ca and the remMning competing &amp;quot;continuation&amp;quot; is chosen. Further supl)ort by (F3) is not needed in this c~se. For travels e to h)reign countries wc can infer the existence of PLA CED-AT~states e' fin&amp;quot; the agent or theme of the travelling event and for the borders contained in the path of the travelling such that e ~ C e (compare the KB rule (9)). Since the description of e3, the stop at the bof der. is subsumed by the more general PLACED-AT: sort, (F3) - ia the absence of corresl)onding evidences for the competing relation of continuation - SUl)ports &amp;quot;elaboration&amp;quot;. Since for the states e4,cr,,e13 the alteHlative of &amp;quot;causally introduced background&amp;quot; is not evident, they are anchored by means of the rule 5.).</Paragraph>
    <Paragraph position="16"> In the case of eC/ and e9, the fact that evidence for &amp;quot;elaboration&amp;quot; is missing selects for &amp;quot;contimiation'. ell) which is situated in Fraukfort cannot lie part of the travel. For this reason anchoring el0 with respect to the level of e,.,l, which actually is eg, is not possil)le. Therefore, with rule 6.), we have to climb up the elaboration hierarchy. The anchoring of eT, es, el 1 and el2 should be clear on the basis of what we have said so far.</Paragraph>
    <Paragraph position="17"> In order to arrive at alternative readings of a text we use a specific backtracking routine which can suppress the impact of (F3) for the resolution.</Paragraph>
  </Section>
  <Section position="8" start_page="0" end_page="0" type="metho">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> The described system complltes the event stru(:ture of a discourse. Siuce we think that the semantics of tense and aspect is not sufficient to establish tile temporal dlsfollrse llleallillg, we use backgl'ollnd kllowledge to C/lisambiguate between different readings.</Paragraph>
    <Paragraph position="1"> The system is incomplete at present. The fragment it deals with doesn't allow for relative clauses. In addition the interactioa between NP-resolutlon and temporal resolution is not available. Some temporally relevant discourse relations are not treated. The logic tmderlying the preference relation is only rudimentary sketched. The approach to temporal resolution AcrEs DE COLIN'G-92, NAUrES, 23-28 AOU'T 1992 2 9 3 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 presented here is a condensed version of that suggested in Eberle (1991). The corresponding implementation for the LILOG prototype dates back to 1989. Meanwhile, independent of this approach, similar suggestions have been made (cLAsher and Lascarides (1991)). In the work of Asher and Lascarides the underlying theory of non-monotonic reasoning is MASH (cf. Asher and Morreau (1991)). It is projected to extract a suited proof theory from MASH for the purpose of refining the temporal resolution component suggested, s</Paragraph>
  </Section>
class="xml-element"></Paper>