Temporal Structure of Discourse

In this paper discourse segments are defined and a method for discourse segmentation primarily based on abduction of temporal relations between segments is proposed. This method is precise and computationally feasible and is supported by previous work in the area of temporal anaphora resolution. 1 I n t r o d d c t i o n B. Webber in [22] explains how discourse segmentation contributes to the interpretation of tense. In this paper we discuss how "tense interpretation" contributes to discourse segmentation. Following Webber's operational definition of segments [21] we present the data struco tures necessary for representing discourse segments and an algorithm to perform discourse segmentation. In order to build a discourse segment structure some clear criteria for grouping chunks of text into segments and for grouping these segments into other segments must be available. Our criteriou is based on the maintenance of temporal coherence between segments. It relies on the abduction of temporal relations between segments that necessarily have temporal properties. Abduction of temporal relations between segments is a feasible and precise method for discourse segmentation. This is the leading criterion for segmentation and does not prevent us from using other criteria such as clue words, spatial relations and recognition of state elaboration. Current discourse structure theories use criteria such as rhetorical relations [9, 15, 13], intention recognition [7], narrative discontinuities [18], etc. All of them use a temporal criterion for segmentation embedded ill less manageable *This work has been supported by JNICT, INIC and Gabinet¢ de Filosofia do Conhecimento. criteria such as "the increasing desire of R to perform action N "l . Our discourse segmentation is accomplished in order to enable us to address the following discourse phenomena: • Temporal anaphora the interpretation of tense and aspect depends on the discourse structure [22]. The maintenance of a "Temporal Focus" is suggested by some authors [11, 22, 18, 5] for tense interpretation. Based on their work our main concern is to provide the "temporal focus" for tense interpretation. In our segment structure the temporal focus is the set of visible segments. A pop in the temporal focus structure will imply the closing of one or more segments. • This and that anaphora These pronouns may refer to discourse segments [21]. With our segmentation we provide discourse segments for reference. • Pronominal and definite anaphora the interpretation of pronouns and definite nominals depends on the attentional structure (Grosz and Sidner). The attentional structure depends on the discourse segmentation. With our approach to segmentation the attentional structure can be computed from our discourse structure our visible segments provide visible entities and sequents for these anaphors. • event reference our approach provides a representation for eventualities, the discourse referents and tile attentional structure necessary for solving this kind of anaphora, • temporal coherence is achieved by inferring one of the possible temporal relations between two eventualities [14]. Our segmentation process mainly controlled by abduction of temporal relations between eventualities enables us to check if a text is temporally coherent. Moreover as we propagate temporal constraints through discourse structure the number of evenI I)efinition of the rhetorical relation motivation [15]. ACTES DE COLING-92, NANTES, 23-28 hO'dT 1992 3 3 1 PROC. OF COLING-92, NANTES, AUO. 23-28, 1992 tualit ies tha t mus t be temporal ly related with a new eventuality increases. * temporal reliability is achieved by the existence of a model for the temporal relations inferred. During discourse processing we build a temporal s tructure where all the temporal cons tralnts can be checked. This structure is updated when a new temporal referent or a newly abducted temporal relation is added. Thus temporal reliability is granted. • discourse coherence is difficult to check by using only our discourse structure. It requires more processing, namely the ability to find intentions for the segments. Tense interpretat ion contributes to segmentation by defining the temporal relation between the segment used as reference and the segment tha t represents the tensed sentence to be interpreted. Thus tense interpretation allows the choice of a segment by indicating where to attach the new sentence segment and great ly restricts the possible referents for anaphora resolution. A failure in satisfying s t ructural cons traints results in the choice of another segment referent. The temporal anchoring of eventualities assumes tha t there is some temporal representa t ion for the eventualities. In this paper we use a representation for eventualities close to the event calculus[12], and a graph structure for t ime representation. Updat ing the discourse s tructure will be equivalent to updat ing a temporal da t a base. Discourse referents are existentially quantified variables tha t can be further constrained by the analysis of a new discourse sentence. In the following sections the temporal relations used in this paper are defined, followed by an explanation of our notion of segments, their properties and the algori thm for discourse segmentation. A detailed example will be worked out. Finally a conclusion and a comparison of our work with related work in this area is presented. 2 T e m p o r a l r e l a t i o n s The semantic representation of an eventuality includes a t ime interval where the eventuality mus t be true so tha t the sentence and the discourse can be true. The time interval will be represented by two time points. Temporal relations between two t ime intervals can be expressed by relations between the extremes of the t ime intervals. So instead of using the 13 reintions proposed by Allen [1] we have chosen to use jus t 5 relations. • t , , < t , 2 = t,~s < t,~, this relation is like Allen's relation before or meets. • t , , > t,~ --=ts2s < i,,~ this relation is like Allen's relation after or met-by. • t , , C t.~ = t.~, < t , , , , t , , s < t,~s this relation is like Allen's relation IN. • to~ D t,3 =-t°~+ < t ,~ , , t ,~ < ts~s this relation is like Allen's relation ON. • t°~ c~ t,= = 3t : t C t , , , t C t,2 this relation is like Allen's relation !. These 5 relations are enough for t ranslat ing natural language sentences as it is difficult to express the 13 relations of Allen and their 213 combinations in natural language. F. van Eynde [6] presents the set of relations necessary for the temporal systems he studied (for EEC languages). Our relation set, however, is sma l l e r . As sentence aspect coerces the verb aspectual class to change [16], v. Eynde's overlap relations may be rewritten using the relations < and >. The following examples will demonstra te our use of the temporal relations. For the sake of simplicity only discourse referents introduced by eventualities and time intervals are represented. Eventualities introduced by nominals (as in example I "his key") are discarded. I) John picked off his key(l). He opened the door(2). The eventualities of sentences (1)and (2) are: event(st, pick(john, key1}}, time(eht,~); event(e2, open(john, door1)), time(e2,t°2); the temporal relation is t , , < t+ 2. lI) John bought an umbrella(l). He had lost his umbrella(2). The eventualities of sentences (1) and (2) are: event(el, buy(john, utah1)), t ime(eht° , ) ; event(e2, looseOohu, umb~)) , time(e2,t°~); the temporal relation is ts~ > t , , . III) John bought an umbrella (1). It was raining(2). The eventualities of sentence (1) and (2) are: event(et, buyOohn,umbl)) , time(ea,t,~); event(e2, rain), t ime(e2,t , ,) ; the temporal relation is t , , C t,2. IV) John had a nice meal(l). He ate salmon(Z). The eventualities of sentence (1) and (2) are: event(el, have(john, meal)), t ime(el, t°t); eventCe2, eat(john,salmon)), t ime(e2,t , ,) ; the temporal relation is t , , D t°~. V) It was raining(I). There was a strong wind(2). A ~ DE COLING-92, NANTES, 23-28 ̂ OUT 1992 3 3 2 PROC. OF COLING-92, NANTBS, AUO. 23-28, 1992 Tile eventualities of sentence (x) and (2) are: event(c1, rain), t ime(eht , , ) ; event(e=, wind_stroug), time(e~,t0~); the temporal relation is t, , c~ t°~. V1) John sat down on a chair0}. Mary lied down on a sofa(2). The eventualities of sentence (1) and (2) are: event(e l , sit(john, chair1)), t ime(el , t , , ) ; event(e~, l ie(Mary, sofa)), time(e~,t,~); the temporal relation is tst none to~ as these two eventualities are independent. 3 D i s c o u r s e S e g m e n t s A discourse segment is a discourse object. It is represented by a discourse referent that can be used for later reference. In contrast to other discourse theories, segments ms dynamic structures that help to define context interpretations are considered as real discourse objects. Thus in our approach we use segments as objects with properties that will be defined later. A text is represented by a segment and a segment supplies context information for the semantic interpretation during discourse processing. Next segments will be defined as well as their construction and use in the semantic interpretation. 3.1 K i n d of s e g m e n t s We distinguish two kinds of segments: basic and non-basic ones. A basic segment represents an eventuality plus some features, typically syntactic ones like tense and aspect (tile leaves of fig. 1). A nonbasic segment has one or more subsegments (basic or not) obeying to a set of temporal constraints and a set of features. Every nonbasic segment has a sort depending on the temporal constraints it imposes on its subsegments. Segment features are necessary for discourse reasoning. Some of them may be dropped after a closing but others have to remain until the discourse is completely processed. The features we take into account in this paper are the following: • tense The feature tense is needed for temporal anaphora resolution. • eventuality The semantic representation of an eventuality is important for temporal anaphora resolution, for causal reasoning and other kinds of reasoning that depend on the kind of the eventuality. • eventuality time This is the main issue ill the definition of a segment as the abducted relation between eventuality times determines tile segment structure's behavior. • discourse referents for solving discourse reference. • subsegments an ordered list containing all its subsegments. 3.2 Sor t s o f s e g m e n t s Depending on the abducted temporal relation between eventualities in a discourse, the eventualities are grouped into different sorts of segments. Using the above mentioned five temporal relations seven sorts of segments ~ can he de~ fined, e.g. 1. basic the minimal segment. 2. none this segment does not impose any restriction on tile temporal relation of its subsegments. The discourse of example VI will be represented by this sort of segment. 3. sequence -the subsegments in the list of segments are temporally ordered, e.g. ex I (fig. l.a). 4. fb contains only two subsegments with the first one temporally situated after the second one, e.g. ex H (fig. 1.b). 5. bk has two subsegments with the first one temporally contained in tire second one, e.g.ex III. 6. elab has two subsegments with the first one temporally containing the second one, e.g. ex IV. 7. over every segment in the list of subsegments must temporally intersect a nonempty time interval, e.g. ex V. For each sort of segment it must be defined how to compute its features representing properties from the features of its aubsegments. 3.3 P r o p e r t i e s o f S e g m e n t s Segments that can have a list of subsegments containing an unlimited number of segments are none, sequence sad over. These segments can be augmented during discourse processing. The features of these segments are the following: a none The feature eventuality contains the set of all subsegments' eventualities, while the 2The nantes of these tmgments ar~ abbreviationt of some rhetorical relatlona that impose the marne temporal cormtralnts. There abreviations should not be read as if they meant the same M the rhetorical relations. They jute mean that their sub~egments obey a particular ternporal relation. AcrEs DE COLING-92, NANTES, 23-28 AOt~'r 1992 3 3 3 PROC. OF COLING-92, NANTES, Auo. 23-28, 1992 le~lUenc~ ¢venl(e3,~l(el,e2)) time(e3,t3), t.~lt li,t2q ~me=~. An=t~rf fim~el.t !) tim~e2~t2~ tl<t2 ~m~sv. Av~cd. ~m~e~p. ApffiVcd.