File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/concl/85/e85-1016_concl.xml
Size: 14,848 bytes
Last Modified: 2025-10-06 13:56:03
<?xml version="1.0" standalone="yes"?> <Paper uid="E85-1016"> <Title>PARAHETRIZED ABSTRACT OBJECTS FOR LINGUISTIC INFORMATION PROCESSING</Title> <Section position="6" start_page="110" end_page="113" type="concl"> <SectionTitle> TEMPORAL OBJECTS </SectionTitle> <Paragraph position="0"> Ue assume that temporal information in a text can be represented by proceedin~ in three steps of increasin~ difficulty: - Tense in main clauses.</Paragraph> <Paragraph position="1"> - Tense in subordinate clauses. - Tense in texts.</Paragraph> <Paragraph position="2"> This hierarchy follows that Of Ejerhed and Janlert ( 1981).</Paragraph> <Paragraph position="3"> To summarize the discussion of the previous paragraph, the followina elements of temporal information have to be abstracted: - temporal deixis (enunciative vs. aoristic situations) . Followin~ Comrie (1976), we use the term &quot;situation&quot; as a generic term coverin~ states, events or processes .</Paragraph> <Paragraph position="4"> - inception and termination of a situation.</Paragraph> <Paragraph position="5"> - information relative to the completion of the situation. - local inferences on situations. - mass/count properties of situations. In our system, these elements are reconstructed from the followin~ linguistic data: - tenses in the finite forms of verbs. temporal specifiers (temporal adverbials).</Paragraph> <Paragraph position="6"> -semantic types of situations (computed from the semantic type of the verbs & la Vendler, and the syntactic structure of the proposition). At this point, most existin~ systems of representation make a choice, since a notion of duration has to be included in the model as well. Either one conceives of the basic elements as points, and the notion of an interval has to be introduced; or an interval is a basic element, and a second relation (of overlapplna or inclusion ) is introduced. In fact, in most existina models of time, the basic elements of time are conceived as elements or subsets of (a subset oPS) the real line (or some ramified structure built from it).</Paragraph> <Paragraph position="7"> The choice of either &quot;points&quot; or &quot;intervals&quot; as basic elements leads to definite advantages and particular difficulties. The point model is basically simpler, but in some way harder to justify semantically. However, as shown by Kamp (1979) and Van Benthem ( 1980 ), both points of views are essentially equivalent.</Paragraph> <Paragraph position="8"> Our claim in the matter follows the ~eneral philosophy of abstraction: Instead oPS the nature of the basic elements, consider their intended properties and combination rules for buildin~ derived elements. This combinatorial point of view is implicit, for example, in Allen's model (Allen 83), where a set of &quot;intervals&quot; is abstractly characterized by the relations holdin~ between its elements. It can be shown (Bestou~eff and Li~ozat ,1984) that any set theoritic model of Allen's axioms is equivalent to (a subset) of the intervals (that is, couples of points) on a totally ordered set.</Paragraph> <Paragraph position="9"> In our model,the basic elements are typed boundaries, with a (partial) order defined on them. As shown in (Bestougeff and Li~ozat, 1984) an alternative way of considerin~ the same abstraction would be in terms of &quot;intervals&quot;, where an interval is a couple (bl, b2) of boundaries with bl <b2 . In other words, the term &quot;interval&quot; has only the notions of a be~innina and an end associated with it, and it is immaterial whether one or the other terminology is used. No topolowical p~operties are implied , only combinatorial properties (in terms of the types of boundaries ) are retained in the abstraction. Boundary types are introduced in the model in order to represent aspectual properties of the data.</Paragraph> <Paragraph position="10"> As an example, consider again sentences (1) to (12).</Paragraph> <Paragraph position="11"> The state be ill in (i) holds upon an interval whose left and right boundaries are &quot;closing&quot; and &quot;opening&quot;, respectively. This is a general situation for states in an enunciative setting. The event John repair my ca.._~ in (5), conversely, holds upon an interval with resp. &quot;opening&quot; and &quot;closina&quot; left and right boundaries. Consequently, the adjacent resultin~ state &quot;my car is repaired&quot; holds on an interval with a &quot;closing&quot; left boundary, as a state should. The combination in (5) of a verb of accomplishment with a &quot;closing&quot; risht boundary insures that such a resulting state does indeed exist. This is to be contrasted with the situation in (6). There, the right boundary is an &quot;opening one&quot;, which prevents the inference of a completed action from bein8 made* It seems that the introduction of such typed boundaries is enough to capture the intuition behind the use of topological intervals in systems representing tense and time. The approach chosen here prevents the overloading of the objects with unnecessary or undesirable properties, as is the case when a concrete model like the real line is adopted.</Paragraph> <Paragraph position="12"> In terms of implemented objects, this corresponds to the definition of abstract intervals from typed boundaries and predicate information (the latter can be empty).</Paragraph> <Paragraph position="13"> AS an example , the explicit definition of an interval is as follows : abstype intv =boundary # pred# boundary with make_intv (11,12,13)= abs_intv (11,12,13) and left I =fst (rep intv i) and right 1 = snd (snd (rep_intv i)) and 8etp 1 =fst (snd (rep_Intv l)) and purl (b,i)= if fst (rep_intv i) =U then abs_intv (b,fst(snd(rep_intv)), snd(snd (rep_intv i))) else i and putr(b,i)= if snd(snd(rep_intv i)) = U then abs_intv(fst(rep_intv i), fst(snd(rep_intv i)),b) else i and show_intv 1 =rep_intv i;; The signature of this object is the set of typed operators: make_in#v=-: (boundary # pred# boundary) -> intv le~t=-: intv -> boundary right=-: intv -> boundary getp =-: intv -> pred purl =-: (boundary#in#v) -> intv purr =-: (boundary # intv ) -> intv show intv =-: intv ->(intv + (intv # nseq)) It seems intuitively satisfying to consider the stretch of time involved in a simple clause as totally ordered. The local inferences operate on this restricted scope.</Paragraph> <Paragraph position="14"> To abstract this phenomenon we introduce interval sequences with constraints on the boundary types of adjacent intervals: absrectype nseq= intv + in# # nseq force=-: intv -> nseq ncons=-:(intv#nseq) -> nseq make tnseq=-:(intv # nseq) -> nseq show nseq=-: nseq -> (intv + (intv # nseq)) The central object in the model corresponds to a simple clause. It is called a polytyped stt-ing (or PTS). It is obtained from an interval sequence by addin~ the information about temporal where the functional type &quot;pred-> pts&quot; denotes the set of functions which build PTS's from predicate information.</Paragraph> <Paragraph position="15"> The predicate information is given through the &quot; rules&quot; where &quot;status&quot; is the information relative to the enunciative vs. aoristic status; &quot;tense&quot; denotes the morphological tense of the clause, &quot;vendler&quot; , the Vendler class (i.e. state, activity, accomplishement or achievement ) computed from classe(s) assigned to verbs in the dictionary and the syntactical configurations ; finally &quot;adverbial&quot; corresponds to information attached to the time adverbials.</Paragraph> <Paragraph position="16"> Up to this point , we have described the fundamentals of the system of representation. Of course, the actual construction of the representative temporal object for a text in a given language is highly language dependent. For instaxlce, the present tense in French (which is the language we are working on) is not in a simple correspondance with the &quot;corresponding&quot; simple present in English. Consequently , referrin~ to the objects described above, the functions &quot;fi&quot;, and &quot;rules&quot; are quite specific to the structure of the language represented.</Paragraph> <Paragraph position="17"> The temporal relations in longer units of discourse are comparatively much more loosely specified. Consider the following example: (15) Shakespeare is dead. John is ili. And I am not feelin~ well either.</Paragraph> <Paragraph position="18"> Apart from considerations pertainin~ to real-world knowledge, no information is given about the relative order of the beginnings of the three situations considered. So the representation should allow for indeterminacy, either in listin~ all possible alternatives (this would be the case in Allen's model), or in leaving the order unspecified. This more economical solution is chosen here. Compare (15) to the followin@: (16) Mary ~ot pregnant. She married John. (17) Mary married John. She got pregnant. Here, the order of discourse seems pertinent and should be represented. More complex examples in this respect are: (18) John was angry when Mary dropped the V as e .</Paragraph> <Paragraph position="19"> (19) Mary dropped the vaue. John was ansry.</Paragraph> <Paragraph position="20"> (20) John was angry. Mary dropped the vas e.</Paragraph> <Paragraph position="21"> where (19) or alternatively (20) can be a paraphrase of (18).</Paragraph> <Paragraph position="22"> The precedin~ discussion shows that the total ordering at the sentence level cannot in ~eneral be extended to lar~er units in a simple way. The eventual relations between different simple sentences are a result of a computation makin~ use of the temporal structure of those sentences and the order of discourse.</Paragraph> <Paragraph position="23"> This fact is captured as follows: The structure representin~ a text is constructed stepwise. At each step of the construction, the exlstin~ structure provides a context, in which the order of discourse, in particular, is represented. In technical terms, the corresponding object is called &quot;temporal site&quot;. It is composed of a sequence of PTS's together with a set of relations on the boundaries of the constituent PTS's.So the next sentence to be examined, taken in isolation, is represented by a possibly incomplete structure (a polytyped string) with a total order on it, but with possible indeterminacies (for example, in the assignment of time indexes). This new structure is inserted into the old one, (already constructed temporal site ) thereby creatin~ new constraints resulting in the evaluation of some undetermined parameters in both structures.</Paragraph> <Paragraph position="24"> Here a~ain, the precise combination rules are lanKua~e specific, as they depend on the semantic properties of the time relations in the lan~uaae.</Paragraph> <Paragraph position="25"> It is beyond the scope of this paper C/o give the rules used for French.</Paragraph> <Paragraph position="26"> However C/o ~ive an indication of what the construction amounts to , consider the following english sentences: (21) John was in love with Nary; (22) John has built his house; (23) John was building his house when I left for Rome.</Paragraph> <Paragraph position="27"> The analysis of (21) yields: Denoting by &quot;p&quot; the predication: John is in love with Mary, the structure of the representing PTS can be Symbolized by the formula :</Paragraph> <Paragraph position="29"> where the C's and O's denote closin8 and opening boundaries and S and R,points of speech and reference respectively . These are indexed by the order of occurrence of the correspondin~ boundaries. ~ denotes a dummy predication.</Paragraph> <Paragraph position="30"> Consider sentence (22) Here tense : present perfect; status : enunciative (because of the present perfect tense);</Paragraph> <Paragraph position="32"> describes the associated 9TS, where p is John builds his house and reslt(p) is a resultin~ state, obtained by local inference, which expresses : the house is built .</Paragraph> <Paragraph position="33"> Finally, consider sentence (23). The associated temporal site can be This temporal site contains two FTS's, with p = John builds his house and q= I leave for Rome .q is an achievement in Vendler's classification. The additional information concerns the ordering relations between the boundaries of the PTS's, numbered 1 and 2.</Paragraph> <Paragraph position="34"> Ue have been mainly concerned with the representation of what we have termed enunciative situations (as opposed to aoristic ones). This is justified , as such situations play a central role in discourse. Concerning aoristic situations, similar representative s~ructures are used, which are in fact more strictly constrained.</Paragraph> <Paragraph position="35"> Habitual situations (e.g. sentence (2)) involve a particular treatment of the predicative component, but otherwise fit into the general scheme described above. From this point of view, they are no different from factual situations. Dispositional sentences, on the other hand, cannot be discussed without entering the domain of modalitY=. Although this may seem a serious limitation (especially for English, where modality is all pervasive ), we leave it aside in the present consideration of tense and time.</Paragraph> <Paragraph position="36"> The preceding discussion illustrates the use of linguistic inference at three distinct levels: a) At the simple sentence level, building a representation involves a first type of inference, which makes use of morpho-syntactic and lexico-semantic information .</Paragraph> <Paragraph position="37"> b) At the next higher level, as illustrated by examples (15-20), another type of inference is used to specify and build the correspondln~ structures (temporal sites).</Paragraph> <Paragraph position="38"> c) Still another kind of lln~uistic inference should account for the possible derivabilities Or paraphrasings mentioned pr0pos examples (1-12). Its formalization should make it possible to describe this inference which, starting from a ~iven temporal site allows to deduce new sites from it.</Paragraph> <Paragraph position="39"> Whereas the first two types of inference are constitutive of the derivation of temporal structures and are central to our activity, the last type has still to be defined and examined in a systematic way i.e. defined explicitely as derivation rules. In this context, the facilities for self-reference and structural inference in the software environment are of primary relevance.</Paragraph> </Section> class="xml-element"></Paper>