<?xml version="1.0" standalone="yes"?>
<Paper uid="W91-0222">
  <Title>A Two-Level Knowledge Representation for Machine Translation: Lexical Semantics and Tense/Aspect</Title>
  <Section position="3" start_page="0" end_page="258" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Recently, there has been much discussion in the literature regarding the interaction of tense and aspect with lexical-semantics (see, for example, \[Bennett et hi., 1990\], \[Hinrichs, 1988\], \[Maybury, 1990\], \[Moens and Steedman, 1988\], \[Nakhimovsky, 1988\], \[Passonneau, 1988\], \[Pustejovsky, 1988, 1989\], and \[Tenny, 1989\]). Among those who have studied the problem of tense and/or aspect, many have taken the work of \[Reichenbach, 1947\] as a starting point (see, for example, \[Brent, 1988\], \[Moens and Steedman, 1988\], \[Hornstein, 1990\], and \[Passonneau, 1988\]), while others have built on the work of \[Allen, 1983, 1984\] (see, for example, \[Vilain et hi., 1990\], and \[Williams, 1990\]), and still others have based their investigation on a combination of Allen and Reichenbach's work (see, for example, \[Yip, 1985\]). Among those who have studied lexical-semantic representations, many have taken the work of \[Jackendoff, 1983, 1990\] as a starting point (see, for example, \[Brent, 1988\], \[Dorr, 1989, 1990a, 1990b, 1990c\], \[Levin and Rappaport, 1985\], and \[Siskind, 1989\]), while others have built on the work of \[Dowty, 1979\] and \[Vendler, 1967\] (see, for example, \[Bennett et hi., 1990\], \[Moens and Steedman, 1988\], \[Nakhimovsky, 1988\], and \[Passonneau, 1988\]), and still others have followed \[Mourelatos, 1981\] and \[Comrie, 1976\] (see, for example, \[Bach, 1986\] and \[Pustejovksy, 1989\]).</Paragraph>
    <Paragraph position="1"> This paper proposes a two-level model that integrates tense and aspect information, based on theories by both Hornstein (in the spirit of l:teichenbach) and Allen, with lexical-semantic information based on an extended version of Jackendoff's theory that includes a verb classification system proposed by Dowty and Vendler. The model is intended to be extensible to realms outside of the temporal domain (e.g., the spatial domain).</Paragraph>
    <Paragraph position="2">  The integration of tense and aspect with lexical-semantics is especially critical in machine translation because of the lexical selection process during generation: there is often a number of lexical connective and tense/aspect possibilities that may be produced from a lexical semantic representation, which, as defined in the model presented here, is largely underspecified. The use of tense and aspect information constrains the choice of target-language terms, which, in turn, limits the possibilities for the generation of tense and aspect; thus, there is a two-way communication channel between the two processes: lexical selection and tense/aspect selection.</Paragraph>
    <Paragraph position="3"> The following section defines the dividing line between non-lexical knowledge (i.e., the tense and aspect of the surface sentence) and lexical knowledge (i. e., the lexical tokens that make up the surface sentence), and then discusses how these two types of knowledge are integrated in a two-level knowledge representation model for a machine translation system.</Paragraph>
    <Paragraph position="4"> Section 3 addresses the issue of cross-linguistic applicability of these two knowledge types, and section 4 discusses possible extensions of temporal knowledge to the spatial domain.</Paragraph>
    <Paragraph position="5"> 2 Interaction of Tense and Aspect with Lexical Semantics null The hypothesis proposed by \[Tenny, 1987\] is that the mapping between cognitive structure and syntactic structure is governed by aspectual properties. The implication is that lexical-semantic knowledge exists at a level that does not include aspectual or temporal information (though these two types of knowledge may depend on each other in some way). This is the view that is adopted here: it is assumed that lexical semantic knowledge consists of such notions as predicate-argument structure, well-formedness conditions on predicate-argument structures, and procedures for lexical selection of surface-sentence tokens; all other types of knowledge must be represented at some other level.</Paragraph>
    <Paragraph position="6"> The lexical-semantic representation that is adopted as the interlingua for the UNITRAN machine translation system \[Dorr, 1989, 1990a, 1990b, 1990c\] is an extended version of lezical conceptual structure (henceforth, LCS) (see \[Jackendoff, 1983, 1990\]). This representation is the basis for the lexical-semantic level that is included in the knowledge representation (KR) component (see figure l(a)). The second level that is included in this component is the tense and aspect structure.</Paragraph>
    <Paragraph position="7"> In addition to the KR component, there is also a syntactic representation (SR) component (see figure l(b)) that is used for manipulating the syntactic structure of a sentence. Together, the KR and SR operate bidirectionally in order to analyze the source-language sentence and synthesize the target-language sentence. We will omit the discussion of the  SR component of UNITRAN (see, for example, \[Doff, 1987\]) and will concern ourselves only with the KR component for the purposes of this paper.</Paragraph>
    <Paragraph position="8"> The translation example shown here illustrates the fact that the English sentence John went to the store when Mary arrived can be translated in two ways in Spanish.</Paragraph>
    <Paragraph position="9"> This example is addressed further throughout this paper. The remainder of this section defines the dividing line between non-lexical knowledge (i.e., tense and aspect) and lexical knowledge (i.e., properties of predicates and their arguments), and discusses how these two types of knowledge might be integrated in a two-level knowledge representation model for a machine translation system.</Paragraph>
    <Section position="1" start_page="251" end_page="254" type="sub_section">
      <SectionTitle>
2.1 Tense and Aspect Structure
</SectionTitle>
      <Paragraph position="0"> The information required for the realization of tense and aspect is considered to be outside of the scope of the lexical knowledge. Aspect is taken to have two components, one that distinguishes between states and events, 1 and the other that defines the perspective (i.e., simple, progressive, and perfective). (See, for example, \[Dowty, 1979\] and \[Vendler, 1967\].) Tense, on the other hand, is taken to be the external time relationship between a given situation and others. (See, for example, \[Bennett et al., 1990\]).</Paragraph>
      <Paragraph position="1"> In the example of figure 1, the source- and target-language sentences consist of two event structures, each of which is associated with its own tense and aspect structure. In the case of go (John went to the store), the event is associated with the Reichenbachian Basic Tense Structure (BTS) E,K_S, which indicates that the event is in the past. 2 The aspect of this clause is &amp;quot;simple&amp;quot; (as opposed to progressive or perfective). In the case of arrive (Mary arrived), the event is associated with the same Reichenbachian temporal representation (E,R S) and aspect (simple), since it too is in the simple past tense. As for relating these two events, the approach adopted here is based on a neo-Reichenbachian framework proposed by \[Hornstein, 1990\] in which the basic tense structures are organized into a complex tense structure (CTS) as follows: the first event (i.e., the matrix clause) is written over the BTS of the second event (i.e., the adjunct clause) and the S and R points are then associated. 3 The entire temporal/aspectual structure for this example would be specified as follows:  theoretical entities: S (the moment of speech), R (a reference point), and E (the moment of the event). The key idea is that certain linear orderings of the three time points get grammaticalized into six basic tenses in English. The corresponding Basic Tense Structures are:  The S, R, and E points may be separated by a line (in which case, the leftmost point is interpreted as temporally earlier than the other) or by a comma (in which case, the points are interpreted as contemporaneous).</Paragraph>
      <Paragraph position="3"> Both tense and aspect are considered to be non-lexical in that they are determined by factors relating not to the lexical-semantic structure or particular lexical tokens of the surface sentence, but to the temporal/aspectual features of the context surrounding the event coupled with certain linguistically motivated constraints on the tense structure of the sentence. In particular, it has been persuasively argued by \[Hornstein, 1990\] that all sentences containing a matrix and adjunct clause are subject to a linguistic (syntactic) constraint on tense structure regardless of the lexical tokens included in the sentence. For example, Hornstein's linguistic Constraint on Derived Tense Structures (CDTS) requires that the association of S and R points not involve crossover in a complex tense structure: (2) * John went to the store when Mary arrives.</Paragraph>
      <Paragraph position="5"> Here, the association of R2 and R1 violates the CDTS, thus ruling out the sentence.</Paragraph>
      <Paragraph position="6"> Note that this linguistic constraint is a syntactic restriction on the manipulation of tense structures, not on the temporal interpretation of tensed sentences. Thus, the constraint holds regardless of the lexical token that is chosen as the connective between the {- } before after as soon as while two events.</Paragraph>
      <Paragraph position="7">  (3) * John went to the store Mary arrives.</Paragraph>
      <Paragraph position="8"> * The connecting word that relates the two events must be selected independently of the  temporal/aspectual structure associated with the sentence since the order of E1 and E2 in a given CTS does not necessarily correspond to the order imposed by the interpretation of the connective. For example, the CTS for John went to the store before Mary had arrived is identical to the CTS for John went to the store after Mary had arrived, even though El is placed linearly after E2 in both cases:</Paragraph>
      <Paragraph position="10"> Thus, it is assumed that the knowledge required to determine the temporal/aspectual structure associated with a sentence exists at a level that is independent from the lexical-semantic knowledge required to select the appropriate lexical items for the surface sentence. null The claim that there is a separation of temporal/aspectual knowledge from lexical-semantic knowledge is further strengthened if one considers that a given lexical-semantic representation may correspond to more than one tense and aspect structure, depending  on the context of the linguistic utterance. For example, suppose we are given the fact that John went to the store before Mary arrived; this fact can be represented as a single lexical-semantic structure, but it may be associated with a number of tense and aspect structures, each of which corresponds to a different surface-sentence utterance:  (5) John went to the store before Mary arrived.</Paragraph>
      <Paragraph position="11"> (6) John went to the store before Mary had arrived.</Paragraph>
      <Paragraph position="12"> (7) John had gone to the store before Mary arrived.</Paragraph>
      <Paragraph position="13"> (8) John had gone to the store before Mary had arrived.</Paragraph>
      <Paragraph position="15"> All of these surface realizations are perfectly valid given that they are consistent with our temporal knowledge, i.e., that the &amp;quot;going to the store&amp;quot; event occurs before the &amp;quot;arriving&amp;quot; event. Thus, the lexical-semantic structure for these two events does not constrain the choice of potential tense and aspect structures with which it may be associated. This provides further evidence that temporal/aspectual knowledge exists at a level that is independent from lexical-semantic knowledge.</Paragraph>
      <Paragraph position="16"> Note that Hornstein's neo-Reichenbachian theory crucially relies on an asymmetry between the matrix and adjunct clauses. Thus, there is an important distinction between \[Hornstein, 1990\], in which the asymmetrical property is fundamental to the theory, and \[Yip, 1985\], in which the asymmetrical property is entirely abandoned. I suggest that Hornstein's intuition is the correct one given that we cannot arbitrarily interchange the matrix and adjunct clauses. For example, Yip's theory predicts that we should be able to replace &amp;quot;El after E2&amp;quot; with &amp;quot;E2 before El,&amp;quot; which is not always the case:  (9) (i) John will go to the store after Mary has arrived.</Paragraph>
      <Paragraph position="17"> (ii) * Mary has arrived before John will go to the store.</Paragraph>
      <Paragraph position="18"> (10) (i) John will go to the store after Mary arrives.</Paragraph>
      <Paragraph position="19"> (ii) * Mary arrives before John will go to the store.</Paragraph>
      <Paragraph position="20">  Given this asymmetrical property, it would not be possible to randomly select a matrix/adjunct order and an appropriate temporal connective for a surface sentence solely on the basis of lexical information. What is needed is the temporal relation between the two events and the constraints on their combination before it is possible to derive the surface structure of the sentence. In addition, the aspectual information must be determined before the two events can be combined since this information will be necessary for retrieving the tense structure (e.g., simple vs. progressive) and selecting the lexical connective (e.g., when vs. while). We will return to the problem of the interaction between tense/aspect and lexical semantics in section 2.3, but first we will turn to a brief description of the lexical-semantic representation that is used as the interlingua for the machine translation system.</Paragraph>
    </Section>
    <Section position="2" start_page="254" end_page="255" type="sub_section">
      <SectionTitle>
2.2 Lexical-Semantic Structure
</SectionTitle>
      <Paragraph position="0"> Lexical-semantic structure exists at a level of knowledge representation that is distinct from that of tense and aspect in that it encodes information about predicates and their arguments, plus the potential realization possibilities in a given language. In terms of the representation proposed by \[Jackendoff, 1983, 1990\], the lexical-semantic structures for the two events of figure 1 would be the following:  (11) (i) \[~.,:, GOLo= (\[Th,=s John\], \[Po.l,io. TOLoc (\[Th,oS John\], \[L .... io, Store\])\])\] (ii) \[s.,., GOLoC/ (\[T~,.s Mary\], \[,o.i,o. TOLoC/ (\[T~,., Mary\], \[~oC/.,o. e\])\])\]'  Although temporal connectives are not discussed in the theory proposed by \[Jackendoff, 1983, 1990\], it is assumed that these two structures are related by means of a lexical-semantic token corresponding to the temporal relation between the two events. As it turns out, this Jackendoff-style lexical-semantic representation is sufficient for the purposes of producing an interlingual representation for machine translation. (See, for example, \[Dorr, 1990b, 1990c\].) However, a richer lexical-semantic representation is necessary in order to accommodate a framework in which predicates of different temporal/aspectual categories are readily distinguished. Although the lexical-semantic representation provided by Jackendoff distinguishes between events and states, this distinction alone is not sufficient for choosing among similar predicates that occur in different temporal/aspectual categories. In particular, events can be further subdivided into more specific types so that non-atomic events (i.e., events that are allowed to have an extended interpretation) such as destroy can be distinguished from atomic events (i. e., events that never have an extended interpretation) such as obliterate. This distinction is not captured in Jackendoff's framework, but it is a crucial distinction given that these two similar words cannot be interchangeably used in all temporal/aspectual contexts: John destroyed the house { for an hour. } (12) (i) until Jack arrived.</Paragraph>
      <Paragraph position="1"> (ii) * John obliterated the house until Jack arrived.</Paragraph>
      <Paragraph position="2"> Such distinctions are omitted in \[Yip, 1985\], where all events are merged under the single heading dynamic, and a constraint on temporal interpretations is applied uniformly across all verbs in this category. However, as we have seen above in example (12), it cannot be the case that the temporal interpretations of all dynamic verbs adhere to the same constraints given that the temporal/aspectual distributions are not identical.</Paragraph>
      <Paragraph position="3"> A number of lexical-semantic representations have been proposed that more readily accommodate temporal/aspectual distinctions. In particular, \[Dowty, 1979\] and \[Vendler, 1967\] have proposed an aspectually oriented lexical-semantic structure that provides a four-way classification system for verbs: states, activities, achievements, and accomplishments, each of which has a different degree of telicity (i. e., culminated vs. nonculminated),  rationale for this move is that ~11 events have 8ome duration, even in the case of so-called punctual events, depending on the granularity of time involved. (See \[Passonneau, 1988\] for an adaptation of this scheme as implemented in the PUNDIT system.) For the purposes of this discussion, we will maintain the distinction between achievements and accomplishments.</Paragraph>
      <Paragraph position="4">  \[Bach, 1986\] and \[Pustejovksy, 1989\] (following \[Mourelatos, 1981\] and \[Comrie, 1976\]) in which actions are classified into states, processes, and events.</Paragraph>
      <Paragraph position="5"> In light of these observations, the lexicM-semantic structure adopted for UNITRAN is an augmented form of Jackendoff's representation in which events are distinguished from states (as before), but they are further subdivided into activities, achievements, and accomplishments. The subdivision is achieved by means of three features proposed by \[Bennett et al., 1990\] following the framework of \[Moens and Steedman, 1988\] (in the spirit of \[Dowty, 1979\] and \[Vendler, 1967\]): ~dynamic (i.e., events vs. states, as in the Jackendoff framework), :t:telic (i. e., culminative events (transitions) vs. nonculminative events (activities)), and ~atomic (i.e., point events vs. extended events). This featural system is imposed on top of the lexical-semantic framework proposed by Jackendoff. For example, the primitive GO would be annotated with the features \[+d,+t,-a\] for the verb destroy, but \[+d,+t,+a\] for the verb obliterate, thus providing the appropriate distinction for cases such as (12).</Paragraph>
      <Paragraph position="6"> Figure 2 relates the four types of lexical-semantic frameworks outlined above. Note that the system of features proposed by \[Bennett et al., 1990\] and \[Moens and Steedman, 1988\] provide the finest tuning given that five distinct categories of predicates are identified by the feature settings. (This system is essentially equivalent to the Dowty/Vendler proposal, but features are used to distinguish the categories more precisely.) In the next section, we will see how the tense and aspect structure described in section 2.1 and the lexicM-semantic representation described in this section are combined to provide the framework for generating a target-language surface form.</Paragraph>
    </Section>
    <Section position="3" start_page="255" end_page="258" type="sub_section">
      <SectionTitle>
2.3 Combining Tense/Aspect with Lexical Semantics
</SectionTitle>
      <Paragraph position="0"> The tense/aspect component of UNITRAN operates in tandem with the lexical-semantic component in order to provide a surface form that corresponds to the interlingual representation. Tense information is linked with the lexical-semantic representation by means of Allen's temporal relations, e These temporal relations are assumed to be determined from the context in which the source-language sentence is uttered or, perhaps, from some knowledge source such as a database with temporal information.</Paragraph>
      <Paragraph position="1"> For example, if we determine from a knowledge source that event E1 John went to the store and event E2 Mary arrived have both occurred in the past, then the time of the sit is assumed that the reader is familiar with Allen's 13 notational relations: &gt; (after), &lt; (before), = (equal), m (meets), mi (is met by), o (overlaps), oi (is overlapped by), d (during), di (contains), s (starts), si (is started by), f (finishes), and fi (is finished by).</Paragraph>
      <Paragraph position="2">  linguistic utterance S is after the two event times. 7 This means that the only possible BTS's (for both E1 and E2) are: E,RS (past), E S,R (present perfect), and E R S (past perfect). In each of these three cases, the event time E and the speech time S are separated by (at least one) line, thus providing a temporal interpretation in which E occurs before S.</Paragraph>
      <Paragraph position="3"> Figure 3 illustrates the combination of the two BTS's into nine possible complex tense structures. (One component of the aspectual representation, simple vs. perfect, is included as well.) The CDTS rules out four of the nine possibilities leaving the following five cases: s  (13) John went to the store when Mary arrived.</Paragraph>
      <Paragraph position="4"> (14) John went to the store when Mary had arrived.</Paragraph>
      <Paragraph position="5"> (15) John has gone to the store when Mary has arrived. \[as in: Typically, John has gone ...\] (16) John had gone to the store when Mary arrived.</Paragraph>
      <Paragraph position="6"> (17) John had gone to the store when Mary had arrived.</Paragraph>
      <Paragraph position="7">  Now that the constraint proposed by Hornstein has pared down the possibilities for the tense combinations, we can further constrain the choices of surface sentences by selecting the most accurate description of the temporal relation between the two events for the connective when. Suppose we have the additional information from the knowledge source that the &amp;quot;going to the store&amp;quot; event occurs before the &amp;quot;arriving&amp;quot; event; in terms of Allen's notation, this would be specified as E1 &lt; E2. The two sentences that guarantee this relation are (16) and (17). We can ensure that only these two realizations are selected for the when connective by using the E1 &lt; E2 relation as an index into a table that associates the aspectual information of the two events with the when connective. The aspectual information in this table includes the featural specifications (i.e., :t:dynamic, +-relic, :t:atomic) for the two lexical-semantic tokens (i.e., GOLoc in both cases) as well as the perspective of the two events (i.e., simple, progressive, or perfective). A portion of such a table is shown in figure 4. From this table we see that there are only two possible aspectual realizations for the two GOLoc events in the current example: since both events are associated with the features \[+d,+t,-a\] and since there are only five legal tense/aspect combinations to choose from, the only admissible aspectual realizations are a perfective matrix clause with a simple adjunct clause, or a perfective matrix clause with a perfective adjunct clause. Thus, sentences (16) and (17) are selected as legal possibilities for the surface sentence.</Paragraph>
      <Paragraph position="8"> Both \[Brent, 1990\] and \[Yip, 1985\] have attempted to compile a table along the lines of the one shown in figure 4. However, in the case of Yip, the perfective aspect is omitted, thus excluding sentences such as (16) and (17) as possible surface sentence forms. In the case of Brent, the table is calculated irrespective of the lexical connective, thus giving rise to spurious temporal assignments such as E1 = E2 for the temporal interpretation of sentences such as John went to the store before Mary arrived. Furthermore, only punctual events are considered; interval events are entirely ignored in Brent's analysis of temporal connectives. Neither Yip nor Brent take telicity or atomicity into account in the construction of the connective table. However, as we will see in section 3, these features are important for providing a cross-linguistically applicable framework for tense and aspect in a machine translation model.</Paragraph>
      <Paragraph position="9"> 7We are assuming that the time of the linguistic utterance S refers to the present time. SAnalogous results would be obtained if we were to switch the matrix and adjunct clauses for this example, although this is not always the case.</Paragraph>
      <Paragraph position="11"> simple, progressive, perfective  Now that we have looked at the constraints on the choice of temporal/aspectual realizations for the matrix and adjunct clauses with respect to the when temporal connective, there is still the question of how the temporal connective is selected in the first place. For example, without any additional information, we are unable to determine, unambiguously, the relation between the clauses of the sentences shown in figure 1: (18) \[s.,.t GOLoc (\[Thins John\], \[po.t,o, TOLoc (\[Vhi.s John\], \[t.oC/,tio, Store\])\])\] 1 \[~,.., GOLoc (\[Th,., Mary\], \[po.i.o. TOLoc (\[Wh,os Mary\], \[co..,,deg. e\])\])\] \] Potential Temporal Relations : &gt;, =, mi, In particular, the connective when may correspond to any of the three relations (i.e., &gt;, =, or mi) depending on the intended temporal interpretation of the associated events.</Paragraph>
      <Paragraph position="12"> The mapping between temporal relations and temporal connectives is a problem in both directions: on the one hand, connectives that are in a one-to-many relation with Allen's temporal relations (e.g., when, which corresponds to &gt;, =, and mi) create a problem for the analysis of the source-language sentence; on the other hand, the temporal relations that are in a one-to-many relation with potential temporal connectives (e.g., the &gt;, which corresponds to after and when) create a problem for the generation of the target-language sentence. The next section addresses how to control this abundance of selection possibilities.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>