File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/88/j88-2005_metho.xml
Size: 66,168 bytes
Last Modified: 2025-10-06 14:12:11
<?xml version="1.0" standalone="yes"?> <Paper uid="J88-2005"> <Title>A COMPUTATIONAL MODEL OF THE SEMANTICS OF TENSE AND ASPECT</Title> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> < O </SectionTitle> <Paragraph position="0"> Situation type: process Kinesis: active Boundedness: unbounded Progressive aspect has often been compared to lexical stativity.l~ Here the commonality among sentences like (9) and (10) is captured by associating the feature of unboundedness both with stative lexical items and with progressive aspect. The temporal structures of states and unbounded processes are thus identical with respect to boundedness. However, the distinction between the kinesis of (9) and (10) is retained by distinguishing active from stative intervals.</Paragraph> <Paragraph position="1"> In (11) the interval associated with the alarm sounding is unspecified for boundedness, meaning that the clock time may occur within the interval for which the alarm sounded, or at its onset or termination.</Paragraph> <Paragraph position="2"> 11. The alarm sounded at 0800.</Paragraph> <Paragraph position="3"> Situation type: process Kinesis: active Boundedness: unspecified In (10), where the verb is progressive, the clock time is interpreted as falling within the unbounded interval of sounding, but in (11), where the verb is not progressive, the clock time can be interpreted as falling at the inception of the process or as roughly locating the entire process. ~2 Nonprogressive forms of process verbs exhibit a wide variation in the interpretation of what part of the temporal structure is located by tense. The influencing factors seem to be pragmatic in nature, rather than semantic. The solution taken here is to characterize the event time of such predications as having an unspecified relation to the active interval associated with the denoted process, represented graphically above by the dashed line around the event time. Transition Events. A transition event is a complex situation consisting of a process which culminates in a new state or process. The new state or process comes into being as a result of the initial process. Since states have no kinesis, they cannot culminate in new situations. The temporal structure of a transition event is thus an active interval followed by--and bounded by-a new active or stative interval.~3 That there are these three distinct components of transition events can be illustrated by the following sentences in which the time adverbials modify one of the three temporally distinct parts of the predicated event.</Paragraph> <Paragraph position="4"> 12. It took 5 minutes for the pump to seize.</Paragraph> <Paragraph position="5"> 13. The pump seized precisely at 14:04:01.</Paragraph> <Paragraph position="6"> 14. The pump was seized for 2 hours. 14 The duration 5 minutes in (12) above applies to the interval of time during which the pump was in the process of seizing. The clock time in (13) corresponds to the moment when the pump is said to have made a transition to the new state of being seized. Finally, the measure phrase in (14) corresponds to the interval associated with the new state.</Paragraph> <Paragraph position="7"> Following Dowty 1986, Vendler's (1967) two classes of achievements and accomplishments are collapsed here into the single class of transition events, and for Computational Linguistics, Volume 14, Number 2, June 1988 47 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect much the same reasons. That is, achievements differ from accomplishments in being typically of shorter duration and in not entailing a sequence of subevents, but they nevertheless do in fact have some duration (Dowty 1986:43). Even so-called punctual events (e.g., They arrived at the station; She recognized her long-lost friend) can be talked about as if they had duration (Talmy 1985, Jackendoff 1987), apparently depending on the granularity of time involved. It is my belief that handling granularity depends on appropriate interaction with a relatively rich model of the world and of the current discourse, but would not require new units of time; depending on the level of detail required, moments could be exploded into intervals, or intervals collapsed into moments. For these reasons, punctual events are not treated here as a separate class.</Paragraph> <Paragraph position="8"> With verbs in Vendler's class of achievements, the same participant generally participates in both the initial process and the resulting situation, as in (15): 15. The engine failed at 0800.</Paragraph> <Paragraph position="9"> > Situation type: transition event Kinesis: active Boundedness: bounded Here, the engine participates in some process (failing), which culminates in a new state (e.g., being inoperative). In each case, however, there are two temporally distinct intervals, as shown in the diagram above, one bounded by the other.</Paragraph> <Paragraph position="10"> Causative verbs typically denote accomplishments involving subevents in which the action of one participant results in a change in another participant, as in (16): 16. The pump sheared the drive shaft.</Paragraph> <Paragraph position="11"> Here, a process in which the pump participated (shearing) is asserted to have caused a change in the drive shaft (being sheared). The consequence of the different argument structures of (15) and (16) on the event representation is discussed in the next section. The boundary between the two intervals associated with a transition event, the transition bound, is defined as a transitional moment between the initial active interval and the ensuing active or stative interval associated with the new situation. An important role played by the transition bound is that it is the temporal component of transition events that locates them with respect to other times. For example, (15) asserts that the moment of transition to the new situation coincides with 0800. In contrast with examples 9-11, the status of the engine prior to 0800 is asserted to be different from its status at 0800 and afterwards. The components of temporal structureJproposed here are intended to provide a basis for deriving what is said about the relative ordering of situations and their durations, rather than to correspond to physical reality. Thus a transition bound is a convenient abstraction for representing how transition events are perceived and talked about. Since a transition event is one which results in a new situation, there is in theory a point in time before which the new situation does not exist and subsequent to which the new situation does exist. This point, however, is a theoretical construct not intended to correspond to an empirically determined time. It corresponds exactly to the kind of boundary between intervals involved in Allen's (1983, 1984) meets relation.</Paragraph> <Paragraph position="12"> The intervals for which situations hold are closely linked with the semantic decompositions of the lexical items used in referring to them. This allows PUNDIT to represent precisely what kinds of situations entities participate in and when. The decompositions include not only N-ary relational predicates among the verb's arguments (Passonneau 1986), but also the aspectual operators for processes and events proposed in Dowty 1979. The main clauses for examples 9, 10, 15, and 16 are given below as examples 17-20.</Paragraph> <Paragraph position="13"> 17. The pressure was low.</Paragraph> <Paragraph position="14"> Decomposition: low(patient(\[pressure 1 \])) 18. The alarm was sounding.</Paragraph> <Paragraph position="15"> Decomposition: do(sound(actor(\[alarm 1 \]))) 19. The engine failed.</Paragraph> <Paragraph position="16"> Decomposition: become(inoperative(patient(\[engine 1 \])) 20. The pump sheared the drive shaft.</Paragraph> <Paragraph position="17"> Decomposition: cause(agent(\[pump 1 \]),become(sheared(patient (\[shaftl\])))) In (17), the semantic predicate low is associated with the predication be low, and is predicated over the entity referred to by the subject noun phrase, the pressure. 15 The time component recognizes this structure as a stative predication because it contains no aspectual operators.</Paragraph> <Paragraph position="18"> The decomposition for (18) consists of a basic semantic predicate, sound, its single argument, and the aspectual operator do, indicating that its argument is in the class of process predicates; the actor role designates the active participant.</Paragraph> <Paragraph position="19"> The decompositions of transition-event verbs contain the aspectual operator become, whose argument is a predicate indicating the type of situation resulting from the event. With inceptive verbs, as in (19), the actor of the initial process is also the patient or theme of the resulting situation, although this dual role is not represented explicitly in the decomposition. If a distinct actor causes the new situation, the verb falls into the class of causatives and the actor of the initial process is conventionally called an agent, as in (20). Other decompositional analyses (Dowty 1979, Foley 1984) conventionally represent the initial process of transition-event verbs by associating an activity predicate (e.g., do) with the actor or agent of the initial process (e.g., cause(do(agent()), 48 Computational Linguistics, Volume 14, Number 2, June 1988 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect become(inoperative(patient())))). The decompositions in (19) and (20) can be considered abbreviated versions of these more explicit predicate/argument structures. The become operator of transition-event verbs thus provides a crucial piece of information used when deriving representations of transition events. Given a reference to a specific transition event that has already taken place, the temporal component deduces the existence of the new situation that has come into being by looking at the predicate embedded beneath the become operator. This is described more fully in Section 4.2.3. As will be shown in Section 4, PUNDIT represents actual situations as predicates identifying the situation type as a state, process, or event. In order to familiarize the reader with the representation schema without needless repetition of detail, a single example of a situation representation is given below for (17).</Paragraph> <Paragraph position="20"> 17. The pressure was low.</Paragraph> <Paragraph position="21"> state(\[lowl\], low(patient(\[pressurel\])), period(\[lowl\])) Each situation representation has three arguments: a unique identifier of the situation, its semantic decomposition, and its time argument, in this case, the interval (or period) over which the predicate holds. The same pointer (e.g., \[lowl\]) is used to identify both a specific situation and its time argument because the actual time for which a situation holds is what uniquely identifies it. The participants in a situation help distinguish it from other similar situations, but while the same entities can participate in other situations, time never recurs.</Paragraph> <Paragraph position="22"> Having introduced the distinct situation types and the temporal structures that distinguish them, the next steps are to show how they are computed and how they permit a simple computation of temporal location. This will be done in Section 4. Since the preceding discussions also introduced the representation of lexical aspect and the relevance of the verbal categories, it is now possible to clearly summarize the input which the temporal analysis component receives.</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 INPUT TO THE TEMPORAL COMPONENT </SectionTitle> <Paragraph position="0"> PUNDIT's time component performs its analysis after the sentence has been parsed and recursively after the semantic decomposition of each predicating element in the sentence has been created (Palmer 1986). Although this paper focuses on the temporal analysis of certain kinds of tensed verbs, the basic algorithm described here has been extended to handle other cases as well.</Paragraph> <Paragraph position="1"> Describing the full input to the temporal component provides an opportunity to mention some of them.</Paragraph> <Paragraph position="2"> The input to the time component for each tensed clause includes not only the surface verb and its tense and aspect markings, but also the decomposition produced by analyzing the verb and its arguments (cf.</Paragraph> <Paragraph position="3"> Section 2.2.2.). The input to the time component is thus a list of the following form: \[\[Tense, Perfect, Progressive\], Verb, Decomposition, {Context}\] Each element of the list will be described in turn.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.1 VERBAL CATEGORIES </SectionTitle> <Paragraph position="0"> The first element in the input list is itself a list indicating the form of the verb, i.e., its grammatical inflection.</Paragraph> <Paragraph position="1"> \[\[Tense, Perfect, Progressive\], Verb, Decomposition, {Context}\] The tense parameter is either past or present.16 If the verb is in the progressive or perfect, the corresponding parameter appears while absence of either in the input sentence is reflected in its absence from the list.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.2 THREE ORDERS OF VERBS </SectionTitle> <Paragraph position="0"> The next two elements in the input to the time component are the surface verb and its decomposition. Lexical aspect is encoded in the decomposition as described in Section 2.2.2 for the cases where it is relevant. However, it is a more fundamental classification pertaining to the verb which helps determine the cases where aspect is relevant.</Paragraph> <Paragraph position="1"> \[\[Tense, Perfect, Progressive\], Verb, Decomposition, {Context}\] Since this information is only for treating more complex cases than are described in this paper, the following discussion is intended only to indicate that the model has been extended to cover verbs whose semantic structure contains temporal information of a different order than the inherent temporal structure of an actual situation. After a brief description of three temporal orders of verbs, the discussion will return to explication of the input required for implementing the basic model.</Paragraph> <Paragraph position="2"> In addition to the aspectual distinction among state, process, and transition-event verbs, there are other distinctions related to temporal semantics. A particularly significant one is among what I call first-, second-, and third-order verbs, by analogy with the distinction among first-, second-, and third-order logics. A first-order verb is one whose arguments are concrete entities, e.g., humans, machines, and other physical objects. A second-order verb takes as its arguments states, processes, and events, but does not in and of itself refer to a situation. Rather, its semantic content is primarily temporal or aspectual (e.g., occur, follow). Third-order verbs refer to complex situations (e.g., result, cause) whose participants are themselves situations. The aspectual distinctions among verbs referring to states, processes, and transition events are only relevant to first-order verbs.</Paragraph> <Paragraph position="3"> Second-order verbs can be identified by the impossibility of temporal modification of a situation referred to by the verb, independent of the situation(s) referred to by the verb's argument(s) (Newmeyer 1975), as can Computational Linguistics, Volume 14, Number 2, June 1988 49 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Order Examples Definition </SectionTitle> <Paragraph position="0"> First &quot;fail&quot;, &quot;operate&quot; verbs that denote situations and whose arguments are not propositional' Second &quot;occur&quot;, &quot;follow&quot; verbs that provide temporal information about their propositional arguments Third &quot;result&quot;, &quot;cause&quot; verbs that denote situations but which also provide temporal information about their propositional arguments by contrasting examples 21 and 22 with (23) and The failure occurred on Tuesday.</Paragraph> <Paragraph position="1"> The failure was discovered on Tuesday.</Paragraph> <Paragraph position="2"> *The failure that happened on Monday occurred on Tuesday.</Paragraph> <Paragraph position="3"> The failure that happened on Monday was discovered on Tuesday.</Paragraph> <Paragraph position="4"> Example 22 mentions two distinct situations, a discovery and a failure. In (21), however, the subject of the sentence, the failure, denotes an event, but the verb occur does not denote a separate situation. It provides tense and aspect information for interpreting its argument. In other words, the temporal information in (21) is very similar to that contained in (25): 25. Something failed on Tuesday.</Paragraph> <Paragraph position="5"> A pragmatic difference between the two sentences is that in (21) it is not necessary to mention what failed whereas in (25), the verb fail must have a subject. Other verbs in this class are foUow, precede, continue, happen, and so on. Because these verbs contribute primarily temporal information, they are conventionally referred to as aspectual verbs (Freed 1979, Lamiroy 1987, Newmeyer 1975).</Paragraph> <Paragraph position="6"> It is easy to see that the analysis of aspectual verbs must be implemented somewhat differently from verbs like fail, which directly denote situations. In a sentence like (25), the relevant temporal information is contained in the verb and its tense and aspect marking alone. In contrast, the temporal information in (24) pertaining to the fail event is distributed not only in the verb and its tense and aspect markers, but also in its subject.</Paragraph> <Paragraph position="7"> Temporal analysis of sentences like (24) must be performed not only at the main clause level, but also at the level of embedded propositions. In essence, analysis of aspectual verbs is of a different order. Consequently, verbs like fail are classified here as first-order verbs while the so-called aspectual verbs are classified as second order.</Paragraph> <Paragraph position="8"> PUNDIT's temporal component also handles a third class of verbs, classified as third order. A third-order verb denotes a real-world situation, but its arguments are other situations. Consequently, the verb may contribute temporal information about the arguments as well as about the situation it denotes. The verb result illustrates this type. Sentence 26 asserts the existence of a result situation; the result relationship holds between an instigating situation mentioned in the noun phrase loss of air pressure, and a resulting situation mentioned in the noun phrase failure.</Paragraph> <Paragraph position="9"> 26. Loss of air pressure resulted in failure.</Paragraph> <Paragraph position="10"> Additionally, the meaning of result includes the temporal information that the instigating situation (the loss) precedes the resulting situation (the failure). A full temporal analysis of sentences like (26) requires two steps. The first is to analyze the temporal structure of the situation denoted by the verb. The second is to draw the correct temporal inferences about the verb's propositional arguments. Such verbs combine some of the properties of both first- and second-order verbs and thus constitute a third order of analysis. Classifying a verb as a third-order verb drives the search for temporal inferences associated with its arguments.</Paragraph> <Paragraph position="11"> The classification of these three orders of verbs, summarized in Table 1, is recorded independently of the lexical decompositions used by both the temporal-analysis component and the semantic interpreter. At present, verb-order information is used only by the temporal-analysis component. It essentially selects for the appropriate flow of control through the temporalprocessing procedures. Although PUNDIT recognizes the distinction between first-, second-, and third-order verbs, and processes the relevant temporal information in each case, the remainder of the paper will deal only with the analysis of first-order verbs.</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.3 LEXICAL ASPECT </SectionTitle> <Paragraph position="0"> The third element in the input list is the decomposition structure produced by the semantic analysis of the verb and its arguments.</Paragraph> <Paragraph position="1"> \[\[Tense, Perfect, Progressive\], Verb, Decomposition, {Context}\] The important aspectual features of the decompositions, discussed in Section 2.2.2, can be summarized as follows. If the decomposition of a first-order verb contains a become operator, the verb is in the transition- null event class; otherwise, if it contains a do operator, the verb is in the process class; else, the verb (or other predicate) is stative.</Paragraph> <Paragraph position="2"> 50 Computational Linguistics, Volume 14, Number 2, June 1988 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect</Paragraph> </Section> </Section> <Section position="7" start_page="0" end_page="0" type="metho"> <SectionTitle> 3.4 DISCOURSE CONTEXT </SectionTitle> <Paragraph position="0"> The final element in the input to the temporal component is a data structure representing the current discourse context.</Paragraph> <Paragraph position="1"> \[\[Tense, Perfect, Progressive\], Verb, Decomposition, {Context}\] The first element of this data structure is a list of unanalyzed syntactic constituents. At this stage of processing, PUNDIT has produced a full syntactic analysis of a surface sentence (or sentence fragment), and a semantic decomposition of some predication within the sentence. After the semantic analysis of a clause, the constituent list contains all those syntactic constituents that do not serve as arguments of the verb, e.g., adverbial modifiers of the verb phrase and sentence adjuncts. After the analysis of the main clause of Sentence 27, for example, the constituent list would contain two unanalyzed constituents; the prepositional phrase introduced by during, and the subordinate clause introduced by when.</Paragraph> <Paragraph position="2"> 27. The pump failed during engine start, when oil pressure dropped below 60 psig.</Paragraph> <Paragraph position="3"> This list of constituents is processed after the temporal content of a predication is analyzed in the search for temporal adverbials that modify the predication (cf.</Paragraph> <Paragraph position="4"> Section 5 below). The data structure representing the current discourse context contains temporally relevant information, such as the tense and voice of the main clause. The main-clause tense is used for the analysis of situations mentioned in embedded tenseless constituents, while voice is used in analyzing adjectival passives. null The next section describes an algorithm for interpreting the four pieces of information relevant to actual references to states, processes, and events. It demonstrates how the temporal structure and temporal loca-tion are generated from the verb's grammatical categories of tense, perfect, and progressive, and from its lexical aspect.</Paragraph> </Section> <Section position="8" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 ALGORITHM FOR THE TEMPORAL ANALYSIS OF INFLECTED VERBS </SectionTitle> <Paragraph position="0"> The introductory and discussion sections have undoubtedly reinforced the view that semantic processing of temporal information is a complicated problem, even when the scope of the problem is constrained to the simple cases addressed here. Relevant information is distributed within and across distinct constituents, and their contribution to temporal information can depend upon co-occurring elements. Yet these are in no way insurmountable problems. The fundamental design principles behind my approach to temporal processing have been to carefully separate the analysis into distinct subtasks, to pare down to a minimum the information available to each task, and to provide a simple compositional semantics for each kind of temporal input. In this section, I outline the basic algorithm for the temporal analysis of inflected verbs. This algorithm analyzes the four components of the inflected verb described in the preceding section (lexical aspect, progressive, perfect, tense). The output that is generated can than serve as input for further temporal processing. Section 5 illustrates the integration of this basic algorithm into a more global procedure that successively interprets the main and subordinate clauses of complex sentences where the subordinating conjunction is a temporal adverbial.</Paragraph> <Paragraph position="1"> The basic algorithm for the temporal analysis of inflected verbs has a simple tripartite control structure designed to answer three distinct questions: 1. Does the predication denote a specific situation with actual time reference? 2. If so, what is the temporal structure of the situation, i.e., how does it evolve through time and how does it get situated in time? 3. Finally, what is the temporal location of the situation with respect to the time of text production, and what is the temporal vantage point from which the situation is described? Figure 1 illustrates the algorithm's global control structure, with the modules corresponding to each question as well as the relevant input for each module. The first module examines all four temporal parameters described in Sections 3.1 and 3.5 in order to reject certain cases. The second module requires only the two parameters pertaining to the computation of temporal structure. It sends a component of the temporal structure, the event time, to the third module, which locates the event time by analyzing the remaining two temporal parameters, tense and perfect.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.1 MODULE 1: ACTUAL TIME </SectionTitle> <Paragraph position="0"> The first task performed by PUNDIT's temporal component is to identify references to specific situation tokens; that is, instances of situations which have actually occurred. The input is the lexical verb and its grammatical categories. In certain cases, the form of the verb itself can indicate that the predication refers to a type of situation, rather than to a specific token. Thus the screening step described here rejects these cases and otherwise assumes that the predication denotes a specific situation. As pointed out in Section 2.1, the verb itself provides insufficient information in two kinds of cases: those where explicit disconfirming information occurs elsewhere in the sentence (e.g., arguments of the verb, modals, frequency adverbials; cf. examples 2 and 7, repeated below): 2. Tourists flew TWA to Boston.</Paragraph> <Paragraph position="1"> 7. The lube oil pump seized whenever the engine jacked over.</Paragraph> <Paragraph position="2"> and those where pragmatic features of the discourse context affect the interpretation of semantic input (as in a sportscast). While Module 1 currently serves only as a filter, it could be made to generate informative output Computational Linguistics, Volume 14, Number 2, June 1988 51 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect</Paragraph> </Section> </Section> <Section position="9" start_page="0" end_page="0" type="metho"> <SectionTitle> ASPECT PROGRESSIVE PERFECT </SectionTitle> <Paragraph position="0"> for subsequent processing of semantically and pragmatically more complex phenomena.</Paragraph> <Paragraph position="1"> In Section 2.1 it was shown that two classes of inflected verbs generally denote situation types, rather than actual tokens. These are process verbs and transition-event verbs in the simple present tense (i.e., nonprogressive and nonperfect), as exemplified in (28) and (29).</Paragraph> <Paragraph position="2"> 28. Number 2 air compressor operates at reduced capacity. (operate is a process verb.) 29. They replace the air compressor every three years. (replace is a transition event verb.) For the compound tenses, present tense interacts with the progressive and perfect verbal categories. The progressive alters the aspectual properties of nonstative verbs so that they refer to unbounded situations, and unbounded situations--unlike the other temporal structures-can be located in the actual present (cf. Section 4.2.2). With the perfect forms, the situation being referred to is always located in the past, and tense pertains to the situation's reference time rather than its event time (cf. Section 4.2.3). Thus, as shown in Figure 1, all four elements in the temporal data structure are inspected in order to identify the two cases exemplified in (28) and (29).</Paragraph> <Paragraph position="3"> Table 2 summarizes the relation between the inflected verb and actual temporal reference.</Paragraph> <Paragraph position="4"> In the current implementation of PUNDIT, predications that meet the first condition do not receive further temporal analysis.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.2 MODULE 2: COMPUTE TEMPORAL STRUCTURE </SectionTitle> <Paragraph position="0"> Module 2 computes the first type of specific temporal information associated with reference to an actual situation. It generates an explicit representation of the situation's temporal structure. This structure includes one or more time arguments associated with the semantic predicates in the decomposition, and the situation's event time. Each situation type--state, process, transition event--receives an appropriate situation label, time argument(s), and event time. The temporal structure evoked by an inflected verb can be computed entirely on the basis of the values of the two aspectual elements in its input (Lexical Aspect, Progressive), as shown in Table 3, will be described in the following three sections corresponding to the three situation types.</Paragraph> <Paragraph position="1"> Though not shown in the figure or in Table 3, Module 2 also receives another input data structure: the semantic decomposition. The decomposition is analyzed during the processing of transition-event situations in order to associate distinct time arguments with distinct semantic predicates in the decomposition. This procedure is explained in the appropriate section below.</Paragraph> <Paragraph position="2"> As shown in Table 3, if the lexical aspect of the predicate is stative (Aspect = stative), then the progressive parameter is irrelevant for computing temporal structure. Lexical stativity is sufficient to identify the</Paragraph> </Section> </Section> <Section position="10" start_page="0" end_page="0" type="metho"> <SectionTitle> LEXICAL ASPECT PROGRESSIVE PERFECT TENSE ACTION </SectionTitle> <Paragraph position="0"/> </Section> <Section position="11" start_page="0" end_page="52" type="metho"> <SectionTitle> LEXICAL PROGRESSIVE LABEL TIME EVENT ASPECT ARGUMENT TIME (ET) </SectionTitle> <Paragraph position="0"> stative Yes/No State unbounded stative interval includes ET process or transition event Yes Process unbounded active interval includes ET process No Process unspecified active interval has ET transition event No Event transition bound unifies with ET Computational Linguistics, Volume 14, Number 2, June 1988 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect situation as a state whose time argument is an unbounded stative interval.</Paragraph> <Paragraph position="1"> Example 30 gives a simple stative sentence, the relevant input to Module 2, and the final situation representation. Note that (30) illustrates the use of the progressive with a verb in the locative class of statives noted in Dowty 1979, and mentioned in Section 2.2.1.17 30. Metallic particles are clogging the strainer.</Paragraph> <Paragraph position="2"> As soon as the lexical aspect is recognized to be stative, Module 2 generates the state label and period time argument used in creating the representation depicted above. A period time argument in the context of a state representation denotes a stative interval. The situation representation in (30) indicates that a specific state, clogl, holds over the stative interval, period(\[clogl\]); the decomposition in the representation indicates the participants and the relation between them that holds over this interval. By definition, this interval also has an event time associated with it, whose relation to the interval we can determine by its boundedness feature.</Paragraph> <Paragraph position="3"> Stative intervals are assumed to be unbounded unless an endpoint is provided by further processing (e.g., through adverbial modification, inference). For unbounded intervals, the event time is always an arbitrary moment included within the interval. This is represented as a binary predicate of the following form, where the moment time argument is the event time: Event Time = moment(\[clogl\]) such that includes(period(\[clogl\]), moment(\[clogl\])) This predicate and the state representation given above exemplify the output of Module 2 for state situations.</Paragraph> <Paragraph position="4"> The event time generated here is then passed to Module 3 in order to determine its temporal location. We will return to this same example in the discussion of temporal location in Section 4.3.</Paragraph> <Paragraph position="5"> There are three surface forms that denote process situations: nonprogressive process verbs, progressive process verbs, and progressive transition-event verbs.</Paragraph> <Paragraph position="6"> The nonprogressive and progressive cases have distinct temporal structures, due to differences in the relation of the event time to the active interval over which the process holds. Since this is the only difference among the three cases, the similarities in temporal structure will be presented before the event time is discussed.</Paragraph> <Paragraph position="7"> A nonstative predication that either has a process verb or is in the progressive (i.e., the three combinations of nonprogressive process, progressive process, and progressive transition-event) evokes a process representation. Thus the following three example sentences would each be represented with a process label and a period time argument, representing the active interval over which the process holds. Examples 31 and 32 illustrate the two forms of process verbs that evoke process situations; since they receive the same representation, it is shown only once. Example 33 shows the third type of reference to a process, with a progressive transition-event verb.</Paragraph> <Paragraph position="8"> 31. The diesel operated.</Paragraph> <Paragraph position="9"> The process representation for (33) contains the full decomposition for the verb fail with its aspectual operator become. In this context, the become operator does not denote a transition to a new situation, but rather, indicates a process of becoming, which might or might not culminate in such a transition.</Paragraph> <Paragraph position="10"> Referring again to Table 3, we note that the active intervals for both (32) and (33) will be unbounded, in contrast to (31), where the active interval is unspecified for boundedness. The consequence of this difference on the representation of the event time is outlined in the following paragraphs.</Paragraph> <Paragraph position="11"> Unbounded processes. The predicate specifying the relation between the event time of an unbounded process and the period over which the process holds is identical to that for states. That is, the period time argument includes an arbitrary moment, which serves as the situation's event time, as shown below.</Paragraph> <Paragraph position="12"> 32. The diesel was operating.</Paragraph> <Paragraph position="13"> such that includes(period(\[faill\]), moment(\[faill\])) The progressive always implies unboundedness, and in this respect resembles lexical statives. Again, it is important to remember that an unbounded interval can acquire endpoints through further processing (e.g., of temporal adverbials, as in The diesel was operating until the pump failed.).</Paragraph> <Paragraph position="14"> Unspecified processes. For nonprogressive process verbs, the period associated with the predication is unspecified for boundedness (cf. discussion of Example Computational Linguistics, Volume 14, Number 2, June 1988 53 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect</Paragraph> </Section> <Section position="12" start_page="52" end_page="52" type="metho"> <SectionTitle> LEXICAL PROGRESSIVE LABEL TIME EVENT ASPECT ARGUMENT TIME (ET) </SectionTitle> <Paragraph position="0"> transition event No Event transition bound unifies with ET (From) Table 3.</Paragraph> <Paragraph position="1"> 11 in Section 2.2.1). This gives rise to an indeterminate relationship between the event time and the period time argument over which the process holds; i.e., the event time may start, end, or be included within the period. This unspecified relationship is represented by means of a binary has predicate, as shown in example 31.</Paragraph> <Paragraph position="2"> 31. The diesel operated.</Paragraph> <Paragraph position="3"> Event Time = moment(\[operatel\]) such that has(period(\[operatel\]), moment (\[operatel\])) Both event-time predicates given so far (i.e., includes, has) indicate a relation between an arbitrary moment and a single interval over which a state or process holds. There is otherwise nothing distinctive about the moment selected to be the event time of a process or state situation. In contrast, as Table 3 indicates, and as discussed in Section 2.2.1, the event time of a transition event is equated with a distinctive component of its temporal structure, viz., the transition bound between a process that initiates the event and the new situation reached at the culmination of the process.</Paragraph> <Paragraph position="4"> Table 3 shows only one component of the temporal structure of a transition event (the relevant line of the table is repeated below).</Paragraph> <Paragraph position="5"> As noted in Section 2.2.1, a transition event has three temporal components: an initial active interval leading up to a transition, the moment of transition, and the interval associated with the new, resulting situation. In theory, then, one could represent the full temporal structure of a transition-event predication (e.g., The pump failed) as three contiguous states of affairs: an initial process (e.g., failing) leading up to a transitional moment (e.g., becoming inoperative) followed by a new state of affairs (e.g., inoperative). At present, PUNDIT explicitly represents only the latter two components of transition-event predications: the moment (transition bound) associated with an event of becoming, and the period associated with the resulting situation. This representation has been found to be adequate for the current applications. Thus transition events are actually assigned two situation representations: an event representation with a moment time argument, represented with the input decomposition, and a resulting state or process situation with a period time argument, for which a new decomposition is derived from the input decomposition. Example 34 illustrates a typical transition-event sentence, the relevant input for computing temporal structure, and the two situation representations. null 34. The pump failed.</Paragraph> <Paragraph position="6"> The first situation representation corresponds to the transition event itself. Module 2 generates the event label and moment time argument used in creating the type of event representation shown above for nonprogressive transition event verbs. The moment argument of a transition event is the transition bound implying the onset of a new situation. When Module 2 creates an event with a moment argument, it also creates a representation for the implied situation. In Example 34, the new situation is a state. When creating the representation for the situation resulting from a transition event, it is necessary to determine the appropriate situation label, time argument, and semantic decomposition for the new situation. This is where the semantic decomposition for transition events plays a role, as will be described below.</Paragraph> <Paragraph position="7"> All transition-event verbs contain a state or process predicate embedded beneath an instance of the aspectual operator become. The full decomposition represents the type of situation associated with the moment of transition. The portion embedded beneath become is the situation type associated with the new situation. For example, the decomposition passed to the time component for Sentence 34 would be: become(inoperative(patient(\[pump 1 \]))).</Paragraph> <Paragraph position="8"> As shown in (34), this decomposition appears in the representation of the transition event itself. The argument to the become operator is then extracted for use in the new situation representation: inoperative(patient(\[pump 1 \])) The extracted decomposition is inspected to determine its aspectual class, completely analogously to the procedure for determining the aspectual class of the input predicate (cf. Section 3). In this case, the embedded predicate decomposition is stative because it contains no aspectual operators. If it contained the do operator, the new situation would have been a process. 18 In this fashion, the decomposition guides the selection of the 54 Computational Linguistics, Volume 14, Number 2, June 1988 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect situation label and time argument for the situation inferred to result from the transition event.</Paragraph> <Paragraph position="9"> The final piece of temporal structure derived for a transition event is the temporal relation between the moment associated with the transition event (e.g., moment(\[faill\])) and the period associated with the resulting situation (e.g., period(\[fail2\])). The event moment is the onset of the period. Following Allen 1983, this is called a start relationship. By definition, then, every transition bound starts some period. In the case of Example 34, the moment of failure starts the period for which the pump is in an inoperative state.</Paragraph> <Paragraph position="10"> start(moment(\[fail 1\]), period(\[fail2\])) The event time of a transitional event is always identified with the transition bound. Thus for examples like (34), the moment time argument serves as the event time of the transition event. This identity relation is not represented as a predicate, but rather, is handled via unification, as indicated in Table 3.</Paragraph> <Section position="1" start_page="52" end_page="52" type="sub_section"> <SectionTitle> 4.3 MODULE 3: COMPUTE TEMPORAL LOCATION </SectionTitle> <Paragraph position="0"> PUNDIT's temporal component employs a Reichenbachian analysis of tense whereby situations are located in time in terms of three temporal indices: the event time, speech time, and reference time. ~9 It diverges from Reichenbach primarily by distinguishing between the event time and the temporal structure of a situation.</Paragraph> <Paragraph position="1"> While Reichenbach acknowledged that the progressive, for example, pertains to temporal duration, he did not discuss the differences in temporal structure associated with distinct situation types and their interaction with tense. Here, the event time is only a single component of the full temporal structure of a situation. In this section, we will see how this method of defining the event time makes it possible to compute temporal location independently of lexical or grammatical aspect while preserving the distinctive temporal information they contribute to references to actual situations.</Paragraph> <Paragraph position="2"> The tense and perfect parameters specify the sequencing relations among the event time, reference time, and speech time, with each of the four configurations of tense and perfect specifying a distinct ordering, as shown in Figure 1 and repeated below:</Paragraph> <Paragraph position="4"> simple present present perfect simple past past perfect The speech time, or time of text production, is given. It serves as the temporal fulcrum with respect to which the other temporal indices are located. As shown in Table 4, the presence or absence of the perfect indicates whether the event time and reference time are distinct, in which case the event time precedes the reference time, or whether they are identical. Tense is taken to indicate the relation between the reference time and the speech time, following Reichenbach's suggestion: the position of R\[T\] relative to S\[T\] is indicated by the words &quot;past&quot;, &quot;present&quot;, and &quot;future&quot; (Reichenbach 1947).</Paragraph> <Paragraph position="5"> Since we are dealing here with actual time, rather than potential or hypothetical time, there is only past or present. That is, the reference time either precedes or coincides with the speech time.</Paragraph> <Paragraph position="6"> The reference time and the event time are identical to one another for the simple tenses (ET is RT), which has the effect that tense applies to the event time. Thus, for the simple present, the event time and the speech time coincide. Note that a distinction is made here between identity and coincidence of distinct indices. For any speech act or text containing a description of a situation, the speech situation and the described situation are always conceptually and observationally distinct, thus also their respective temporal indices. These indices are therefore represented as distinct times, which, in the present tense, happen to coincide. However, with the simple tenses, there is no reason to create a distinct reference time and a relation saying that it coincides with the event time. Rather, there are two different functions, which, in the case of the simple tenses, are filled by the same temporal index. The function of the reference time is explained more fully below.</Paragraph> <Paragraph position="7"> Webber (this volume) reviews and expands upon the role reference time plays in intersentential temporal reference. Reference time also plays a role in interpreting relational adverbials like now, yesterday, when, and so on. Adverbs like now and yesterday relate the reference time of a predication to an implicit time, viz., the speech time. Relational adverbs like before, after, and when relate the time of the predication they modify to an explicitly mentioned time, i.e., the reference time associated with their syntactic complements. In the absence of the perfect, the reference time is identical with the event time, as in (35) and (36).</Paragraph> <Paragraph position="8"> 35. The pressure is normal now.</Paragraph> <Paragraph position="9"> 36. The pressure was low yesterday.</Paragraph> <Paragraph position="10"> In the perfect tenses, the reference time and event time are distinct. The event time of both the present and past perfect predications in (37) and (38) is past, i.e., the moment of failure is in the past.</Paragraph> <Paragraph position="11"> 37. The pump has now failed, zdeg Computational Linguistics, Volume 14, Number 2, June 1988 55 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect 38. The pump had failed when the gear began to turn.</Paragraph> <Paragraph position="12"> With the present perfect, it is the reference time that is present, as shown in (37) by the admissibility of the adverb now, which also refers to the present. On one reading of (38), the event time, or moment of failure, precedes the reference time, i.e., the time specified by the when clause. The perfect tenses can also be used simply to affirm truth or falsehood,~ thus (38) has another reading in which the perfect does not contribute a distinct reference time, but merely asserts that it is in fact the case that the pump failed when the gear began to turn.</Paragraph> <Paragraph position="13"> The distinct relations of event time to temporal structure corresponding to the three categories of boundedness--unbounded, unspecified, and bounded--correlate with distinctive behavior of the present tense. If the temporal structure associated with a predication is an unbounded interval, the simple present locates some time within the interval coincident with the speech time. Examples 35-38 illustrate the simple present in the context of the four types of predications that hold over unbounded intervals.</Paragraph> <Paragraph position="14"> 35. The pressure is low.</Paragraph> <Paragraph position="15"> In these examples, the predicate is asserted to hold for some interval of unknown duration, which includes the speech time. Since this interval corresponds by definition to actual time, it cannot be known to continue beyond the speech time into the future. However, that it can extend indefinitely into the past is illustrated by (39), where the situations referred to in the first and second conjuncts are assumed to be the same.</Paragraph> <Paragraph position="16"> 39. The pressure is low and has been low.</Paragraph> <Paragraph position="17"> Predications involving process or transition-event verbs in the simple present have already been eliminated by Module 1 on the assumption that sentences like (40) and (41) do not refer to actual time.</Paragraph> <Paragraph position="18"> 40. The pump operates.</Paragraph> <Paragraph position="19"> Lexical aspect: process Progressive: no 41. The pump fails.</Paragraph> <Paragraph position="20"> Lexical aspect: transition event Progressive: no If a predication is not explicitly unbounded, i.e., if it has or may have an endpoint, then the present tense cannot be interpreted as locating the event time in the actual present. An event time located within an unbounded interval corresponds to persistence of the same situation, whereas an event time that may also be an end-point corresponds to a transition. The way in which example:s like (40) and (41) are interpreted can be explained by considering that we cannot announce changes in the world at the exact moment that we perceive them, although in the guise of reportage or sportscasting, we act as though we can.</Paragraph> <Paragraph position="21"> In contrast to the simple present, the simple past can locate the event time of any temporal structure prior to the speech time. What is distinctive about the past tense in the context of the different temporal structures pertains to the temporal structure surrounding the event time. If the temporal structure is an unbounded interval, then the event time is some moment prior to the speech time within a persisting interval, and the same situation extends unchanged forward towards the present and back into the past. Example 42 illustrates the lack of contradiction in asserting the continuation up to the present of the past, unbounded situation mentioned in the first clause.</Paragraph> <Paragraph position="22"> 42. The pump was failing and is still failing.</Paragraph> <Paragraph position="23"> The temporal structure associated with the situation mentioned in the first clause of (43), in the simple past, is an unspecified interval. Here it is unclear whether the two conjuncts refer to the same situation. Since the event time of the first conjunct is represented noncommittally, i.e., it may or may not be an endpoint of the interval, both interpretations are provided for by the representations generated here.</Paragraph> <Paragraph position="24"> 43. The pump operated and is still operating.</Paragraph> <Paragraph position="25"> Finally, the simple past of a predication denoting a transition event definitely locates an endpoint. The event time of (44) is the transitional moment between an initial process of failing and a resulting state of being inoperative.</Paragraph> <Paragraph position="26"> 44. The pump failed and is still failing.</Paragraph> <Paragraph position="27"> The first clause of (44) is represented by PUNDIT to assert the following temporal information: there was a moment of transition at which the pump failed, viz., its event time (moment(\[faill\])); this moment started a period in which the pump was inoperative (start (moment(\[faill\], period(\[fail2\]))); and finally, it preceded the speech time (precedes(moment(\[faill\]), Speech Time)). The second clause cannot refer to the same transition event because a unique transition bound cannot both precede and coincide with the speech time, nor can it both be an endpoint of, and contained within, an interval. Rather, the second clause refers to a distinct situation, either a process that the speaker presumes will eventually result in a new failure, or an iteration of successive failure events. Of these two possibilities for the second clause, PUNDIT currently generates only the former.</Paragraph> <Paragraph position="28"> 56 Computational Linguistics, Volume 14, Number 2, June 1988 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect The perfect tenses have a more complex semantics and pragmatics than the simple tenses. The semantic interpretation given here accounts for the temporal interpretations assigned to the perfect tenses in which the event time and reference time are distinct from one another. There are uses of the perfect that do not have these temporal effects, as pointed out in McCawley 1971, i.e., cases where the event time and reference time would not be distinct. Here we consider only the temporally relevant uses of the perfect, where each perfect tense specifies two temporal relations: in both cases, the event time precedes the reference time; and tense indicates whether the reference time coincides with or precedes the speech time.</Paragraph> <Paragraph position="29"> The following examples illustrate the present and past perfect with a variety of temporal structures. The only difference between these examples and the simple present tenses examined in the preceding section is the relation between the reference time and event time. The relation between temporal structure, event time, and speech time is the same as for the simple past.</Paragraph> <Paragraph position="30"> 45. The engine has been operating. (unbounded process) 46. The engine has operated. (unspecified process) 47. The pump has failed. (transition event) 48. The pressure had been low. (state) 49. The pump had failed. (transition event)</Paragraph> </Section> </Section> <Section position="13" start_page="52" end_page="52" type="metho"> <SectionTitle> 5 INTERPRETING TEMPORAL ADVERBIALS </SectionTitle> <Paragraph position="0"> It is assumed that temporal adverbials can be analyzed in terms of the same components of temporal structure and temporal sequencing constraints that apply to situations. The situation representations developed here provide a foundation for interpreting three distinct types of adverbial modification corresponding to the three features represented in temporal structure, i.e., kinesis, intervals, and moments. Rate adverbs like slowly and rapidly, which modify the manner in which situations evolve through time, modify active intervals and not stative intervals. For an example like (50), no explicit active interval would be represented, thus one would have to be coerced in order to interpret the adverb.</Paragraph> <Paragraph position="1"> 50. The pressure was rapidly low.</Paragraph> <Paragraph position="2"> Examples like (51), on the other hand, provide a motivation for representing the initial active interval of a transition event (cf. Section 4.2.3), since the adverb essentially selects for such an interval.</Paragraph> <Paragraph position="3"> 51. The engine quickly failed.</Paragraph> <Paragraph position="4"> Durational adverbials like for X, where X is a temporal measure phrase, modify any interval, but not their endpoints. Finally, relational adverbs, which specify temporal sequence, modify the reference time of situations. null Adverbials can combine relational and durational elements. In X, where X is a temporal measure phrase, not only specifies a duration, but also relates the endpoint of this duration to some other time, e.g., the time at which the utterance is produced, as in (52).</Paragraph> <Paragraph position="5"> 52. The lights will go off in 10 minutes (e.g., from now).</Paragraph> <Paragraph position="6"> Temporal connectives like before and after can combine with temporal measure phrases to yield complex adverbials specifying both a duration and a relation, as in (53). 53. The engine seized five minutes before the alarm sounded.</Paragraph> <Paragraph position="7"> In this section, we will look briefly at the two types of durational phrases compared in Vendler 1967 in order to demonstrate the advantages of the representations developed here for interpreting them. Then we will look briefly at the algorithm for interpreting complex sentences with subordinate adverbial clauses.</Paragraph> </Section> <Section position="14" start_page="52" end_page="52" type="metho"> <SectionTitle> 5.1 DURATIONAL ADVERBIALS </SectionTitle> <Paragraph position="0"> Unbounded situations. Predications denoting states and processes have duration, as shown by the interpretation of durational adverbial phrases of the formforX, where X is a time measure, as in (54) and (55): 54. The pressure was low for 10 minutes. (state) 55. The gear was turning for 10 minutes. (unbounded process) However, as noted in preceding discussions, the past tense in reference to states and unbounded processes does not apply to the whole duration. It applies to the moment within the interval designated as the situation's event time. Since in (54) and (55) the event time is past and the speech time is present, the two temporal indices create an explicit temporal extent within which to locate the durational phrases. The for adverbial phrase also evokes an unbounded duration, meaning that the measure phrase does not necessarily encompass the entire duration, as shown by the lack of contradiction in asserting the continuation of the interval up to the present, as in (56) and (57).</Paragraph> <Paragraph position="1"> 56. The pressure was low for 10 minutes and is still low.</Paragraph> <Paragraph position="2"> 57. The gear was turning for 10 minutes and is still turning.</Paragraph> <Paragraph position="3"> The present perfect would allow one to assert something semantically very similar to (56) and (57), but more laconically (e.g., The pressure has been low for 10 minutes.). However, a context in which (56) would be more correct than the corresponding perfect is perfectly possible; it would have.to be a context where the pressure is now low, was low over some interval of 10 minutes' duration, but where this interval is more than 10 minutes prior to the present, and where the pressure has continued to be low up to the present (e.g., A: The alarm should go off if the pressure is low for 10 minutes. B: Well, the pressure was low for 10 minutes and it's still low, but the alarm still hasn't gone off.).</Paragraph> <Paragraph position="4"> The past tense with an unbounded interval evokes a span of time between the past event time and the present speech time within which to situate the measure of time given by a for adverbial. However, there is no Computational Linguistics, Volume 14, Number 2, June 1988 57 Rebecca J. Passonneau A Computational Model of the Semantics of Tense and Aspect such span of time associated with the present tense of an unbounded interval, hence the impossibility of a for measure phrase in examples (58) and (59). 22 58. ? The pressure is low for 10 minutes.</Paragraph> <Paragraph position="5"> 59. ? The gear is turning for 10 minutes.</Paragraph> <Paragraph position="6"> Note that the present perfect, like the simple past, does provide a temporal point prior to the present, thereby creating a span of time for the durational phrase to apply to, as in (60) and (61): 60. The pressure has been low for three hours.</Paragraph> <Paragraph position="7"> 61. The gear has been turning for five minutes.</Paragraph> <Paragraph position="8"> The temporal structures generated for examples like (56)-(59) make it possible to correctly interpret the adverbial phrases they contain. The measure phrases in (56) and (57) can be interpreted not simply because the mentioned situations have duration, but more importantly because of the distinctness of the two temporal indices, event time and speech time. In (58) and (59), where event time and speech time coincide, there is no explicit span of time within which to situate the measure phrase. Cases where there is no explicit component of temporal structure in the situation representation to match up with the temporal structure evoked by a temporal adverbial are probably candidates for the kind of coercion discussed in Moens and Steedman (this volume).</Paragraph> <Paragraph position="9"> The durational adverbial phrases in (62)-(64) not only specify a duration, but also an endpoint (Vendler ~967). Since progressive process predications are unbounded, there is no actual endpoint to be mapped to, hence, under one reading, (62) cannot be interpreted as a situation with an actual time; rather, it seems to refer to an activity that was supposed to take place five minutes from some time previously specified in the discourse context (e.g., paraphrasable as It was to be the case that the gear would turn five minutes from the present).</Paragraph> <Paragraph position="10"> There is another possible reading, paraphrasable as It turned out to be the case that the gear turned five minutes after some previously specified time, as in the context I applied some lubricant to the gear and it was turning in five minutes, which, like (58) and (59) above, may be examples requiring coercion. 23 In contrast, examples 63 and 64 can be interpreted as actual situations whose endpoints coincide with the endpoints of the five-minute duration.</Paragraph> <Paragraph position="11"> 62. The gear was turning in five minutes.</Paragraph> <Paragraph position="12"> 63. The gear turned in five minutes.</Paragraph> <Paragraph position="13"> 64. The engine was repaired in five minutes.</Paragraph> <Paragraph position="14"> The two types of durational adverbials behave differently when modifying the different types of temporal structures in ways that tend to confirm the representations proposed here.</Paragraph> <Section position="1" start_page="52" end_page="52" type="sub_section"> <SectionTitle> 5.2 COMPLEX SENTENCES </SectionTitle> <Paragraph position="0"> The temporal adverbials encountered in the CASREPs domain consisted predominantly of phrases introduced by temporal connectives, e.g., when, before, and after (Smith 1981). The general problem in analyzing the strictly temporal information associated with such connectives is to associate some time evoked by the matrix clause with some time evoked by the complement phrase. In general, connectives are represented as associating the reference time of the matrix clause with the reference time of the compl'ement. The procedure involved in analyzing the temporal relations specified by a beJbre adverb (or other temporal connective) has the six steps illustrated in (65) below.</Paragraph> <Paragraph position="1"> (65) The compressor failed before the pump seized.</Paragraph> <Paragraph position="2"> The compressor failed moment(\[faill\]) before the pump seized moment(\[seize 1 \]) precede(RTl, RT2) Result: precedes(moment(\[faill\]), moment(\[seizel\])) First, the temporal semantics of the main clause is analyzed. One of the outputs of this analysis is the reference time of the main clause, which in this case would be represented as moment(\[faill\]). Then the time component finds the adverbial phrase before the pump seized in the constituent list, which it recognizes as consisting of a temporal connective (before) and a complement. The complement clause is sent to the semantic interpreter (Palmer 1985) and is returned to the time component for temporal analysis. The fourth step, the temporal analysis of the subordinate clause, yields the information that the reference time of the subordinate clause is moment(\[seizel\]). Finally, the time component looks up the predicate structure representing the semantics of the temporal connective. Before is represented as a binary predicatewpreeedes---whose first argument is the reference time of the main clause and whose second argument is the reference time of the complement clause.</Paragraph> <Paragraph position="3"> Currently, relational adverbs like before, after, and when are represented as predicates relating the reference times of the modified and modifying situations. The procedure for handling temporal connectives assumes a priori that the reference times of the syntactically superordinate and subordinate constituents are the required :input. In future work, these and other adverbs will be treated more explicitly as semantic predicates with selectional constraints that guide the search for the appropriate components of temporal structure associated with the referents of the relevant constituents.</Paragraph> </Section> </Section> class="xml-element"></Paper>