<?xml version="1.0" standalone="yes"?>
<Paper uid="P87-1019">
  <Title>NOMINALIZATIONS IN PUNDIT</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
NOMINALIZATIONS IN PUNDIT
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ABSTRACT
</SectionTitle>
    <Paragraph position="0"> This paper describes the treatment of nominalizations in the PUNDIT text processing system.</Paragraph>
    <Paragraph position="1"> A single semantic definition is used for both nominalizations and the verbs to which they are related, with the same semantic roles, decompositions, and selectional restrictions on the semantic roles. However, because syntactically nominalizations are noun phrases, the processing which produces the semantic representation is different in several respects from that used for clauses. (1) The rules relating the syntactic positions of the constituents to the roles that they can fill are different. (2) The fact that nominailzations are untensed while clauses normally are tensed means that an alternative treatment of time is required for nomlnalizations. (3) Because none of the arguments of a nominallzation is syntactically obllgatory, some differences in the control of the filling of roles are required, in particular, roles can be filled as part of reference resolution for the nominalization. The differences in processing are captured by allowing the semantic interpreter to operate in two different modes, one for clauses, and one for nominalizations. Because many nomlnalizations are noun-noun compounds, this approach also addresses this problem, by suggesting a way of dealing with one relatively tractable subset of noun-noun compounds.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="131" type="metho">
    <SectionTitle>
1Formerly SDC-A Burroughs Company.
1. Introduction
</SectionTitle>
    <Paragraph position="0"> In this paper we will discuss the analysis of nominalizations in the PUNDIT text processing system. 2 Syntactically, nomlnalizations are noun phrases, as in examples (I)-(7).</Paragraph>
    <Paragraph position="1">  (1) An inspection of lube oil filter revealed metal particles.</Paragraph>
    <Paragraph position="2"> (2) Lou of lube oll preuure occurred during operation.</Paragraph>
    <Paragraph position="3"> (3) SAC received hifh ueafe.</Paragraph>
    <Paragraph position="4"> (4) In~eeti#ation revealed adequate lube oil. (5) Request replacement of SAC..</Paragraph>
    <Paragraph position="5"> (6) Erosion of impellor blade tip is evident. (7) Unit has low output air pressure, resulting  in ale*# gae turbine atarte.</Paragraph>
    <Paragraph position="6"> Semantically, however, nominaliTatlons resemble clauses, with a predlcate/argument structure like that of the related verb. Our treatment attempts to capture these resemblances in such a way that very little machinery is needed to analyze nominalizations other than that already in place for other noun phrases and clauses.</Paragraph>
    <Paragraph position="7"> There are two types of differences between the treatment of nomlnalizatlons and that of clauses. There are those based on linqui~tle differences, related to (1) the mapping between syntactic arguments and semantic roles, which is I The research described in this paper was supported in part by DARPA under contract N000014-85-C-0012, administered by the Office of Naval Research. APPROVED FOR  different in nomlnalisations and clauses, and (2) tense, which nomlnallsations lack. There are also differences in control; in particular, control of the filling of semantic roles and control of reference resolution. All of these issues will be discussed in detail below.</Paragraph>
  </Section>
  <Section position="4" start_page="131" end_page="131" type="metho">
    <SectionTitle>
2. Clause analysis
</SectionTitle>
    <Paragraph position="0"> The semantic processing to be described in this paper is part of the PUNDIT s system for processing natural language messages. The PUNDIT system is a highly modular system, written in Prolog, consisting of distinct syntactic, semantic and discourse components. ~-lirschman1985\], and~-lirschman1986\], describe the semantic components of PUNDIT, while ~)ah11986, Palmer1988, Passonneau1986\], describe the semantic and pragmatic components.</Paragraph>
    <Paragraph position="1"> The semantic domain from which these examples are taken is that of reports of failures of the starting air compressors, or sac's, used in starting gas turbines on Navy ships.</Paragraph>
    <Paragraph position="2"> The goal of semantic analysis is to produce a representation of the information conveyed by the sentence, both implicit and explicit. This involves 1) mapping the syntactic realization onto an underlying predicate argument representation, e.g., assigning referents of particular syntactic consltuents to predicate arguments, and 2) mak\]Jig implicit argument fillers expllclt. We are using an algorithm for semantic interpretation based on predicate decomposition that integrates the performance of these tasks. The integration is driven by the goal of filling in the predicate arguments of the decomposition.~almer1986\].</Paragraph>
    <Paragraph position="3"> In order to produce a semantic representation of a clause, its verb is first decomposed into a semantic predicate representation appropriate for the domain. The arguments of the predicates constitute the SEMANTIC ROLES of the verb, which are slml\]ar to cases 4 For example, fall decomposes into become inoperatlve, with patient as its  only semantic role. Semantic roles can be filled either by a syntactic constituent or by reference PUBLIC RELEASE, DISTRIBUTION UNLIMITED.</Paragraph>
    <Paragraph position="4"> s PUNDIT UNDderstands and Integrates Text 4 In this domain the semantic roles include: agent, In null stigator, experiencer, Instrument, theme, Ioeatlon, actor, patient, source, reference_pt and goal. There are domain specific criteria for selecting a range of semantic roles. The criteria which we have used are described resolution from default or contextual information. We have categorized the semantic roles into three classes, based on how they are filled Semantic roles such as theme, actor and patient are syntactically OBLIGATORY, and must be filled by surface constituents. Semantic roles are categorized as semantically ESSENTIAL when they must be filled even if there is no syntactic constituent avaUahle, s In this case they can be filled pragmatically, making use of reference resolution, as explained below. The default categorization is NON-ESSENTIAL, which does not require that the role be filled. The algorithm in Figure 1 produces a semantic representation using this information.</Paragraph>
    <Paragraph position="5"> Each step in the algorithm will be illustrated at least once in the next section using the following (typical) CASREPS text.</Paragraph>
    <Paragraph position="6"> ~a(c) failed.</Paragraph>
    <Paragraph position="7"> Pump sheared.</Paragraph>
    <Paragraph position="8"> Ineestifatiort reeealed metal eontamlnation in filter.</Paragraph>
    <Section position="1" start_page="131" end_page="131" type="sub_section">
      <SectionTitle>
2.1. A Simple Example
</SectionTitle>
      <Paragraph position="0"> DECOMPOSE VERB - The first example uses the fall decomposition for Sac failed: fall &lt;beeomeP (inoperatlveP (patlent(P))).</Paragraph>
      <Paragraph position="1"> It indicates that the entity filling the OBLIGATORY patient role has or will become inoperative. null FOR patient ROLE -</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="131" end_page="132" type="metho">
    <SectionTitle>
PROPOSE SYNTACTIC CONSTITUENT
FILLER - A mapping rule indicates that the syn-
</SectionTitle>
    <Paragraph position="0"> tactic subject is a likely filler for any patient role. The mapping rules make use of intuitions about syntactic cues for indicating semantic roles first embodied in the notion of case ~lllmore1968,Palmer1981\]. The mapping rules can take advantage of general syntactic cues like &amp;quot;SUBJECT goes to PATIENT&amp;quot; while still indicating particular context sensitivities. (See ~almer1985\] for details.) in{Paseonneau198611 s We are in the process of defining criteria for categorizing a role as ~SSeNTIAL. It is clearly very domain dependent.</Paragraph>
  </Section>
  <Section position="6" start_page="132" end_page="132" type="metho">
    <SectionTitle>
CALL REFERENCE RESOLUTION - See is the
</SectionTitle>
    <Paragraph position="0"> subject of ma(c) failed, and is suggested by the mapping rule as a 1Lkely filler of the patient role.</Paragraph>
    <Paragraph position="1"> At this point the semantic interpreter asks noun phrase analysis to provide a unique referent for the noun phrase subject. Since no sac, have been mentioned previously, a new name is created: sael.</Paragraph>
  </Section>
  <Section position="7" start_page="132" end_page="132" type="metho">
    <SectionTitle>
TEST SELECTION RESTRICTIONS - In addi-
</SectionTitle>
    <Paragraph position="0"> tion to the mapping rules that are used to associate syntactic constituents with semantic roles, there are selection restrictions associated with each semantic role. The selection restrictions for fail test whether or not the filler of the patient role is a mechanical device. A sac is a mechanical device so the subject of the sentence mac failed maps straightforwardly onto the patient role, e.g., beeomeP (inoper at|veP (pat|ent (sac1))).</Paragraph>
    <Paragraph position="1"> Since there are no other roles to be filled the algorithm term~-ates successfully at this point and the remaining steps are not applied. The next example illustrates further steps in the algorithm. null</Paragraph>
    <Section position="1" start_page="132" end_page="132" type="sub_section">
      <SectionTitle>
2.2. Unfilled Obligatory Roles
</SectionTitle>
      <Paragraph position="0"> The second utterance in the example, Psmp mheared, illustrates the effect of an unfilled obligatory role.</Paragraph>
    </Section>
  </Section>
  <Section position="8" start_page="132" end_page="132" type="metho">
    <SectionTitle>
DECOMPOSE VERB -
</SectionTitle>
    <Paragraph position="0"> shear, &lt;- eauseP (!nstigator (I),</Paragraph>
    <Paragraph position="2"> Sheer is an example of a verb that can be used either transitively or intransitively. In both cases the patient role is filled by a mechanical device that becomes sheared. If the verb is used transitively, the instigator of the shearinC/, also a mechanical device, is mentioned explicitly, as in, The rotating driee shaft sheared the psmp. If the verb is used intransitively, as in the current example, the instigator is not made explicit; however, the algorithm begins by attempting to fill it in.</Paragraph>
    <Paragraph position="3"> FOR Instigator ROLE - Working from left to right in the verb decomposition, the first role to and relies heavily on what can be assumed from the context. be filled is the instigator role. A mapping rule indicates that the subject of the sentence, psmp, is a likely filler for this role. Reference resolution returns pump1 as the referent of the noun phrase. Since pump is a mechanical device, the selection restriction test passes.</Paragraph>
    <Paragraph position="4"> FOR patient ROLE - There are no syntactic constituents left, so a syntactic constituent cannot be proposed and tested.</Paragraph>
  </Section>
  <Section position="9" start_page="132" end_page="134" type="metho">
    <SectionTitle>
UNFILLED OBLIGATORY ROLES - The
</SectionTitle>
    <Paragraph position="0"> patlent role, a member of the set of obligatory roles, is still unfilled. This causes failure, and the binding of p,*rnpl to the instigator role is undone. The algorithm starts over again, trying to fill the instigator role.</Paragraph>
    <Paragraph position="1"> FOR instigator ROLE- There are no other mapping rules for instigator, and it is nonessential, so Case 4 applies and it is left unfilled, e The algorithm tries again to fill in the patient role.</Paragraph>
    <Paragraph position="2"> FOR patlent ROLE - Two mapping rules can apply to the patient role, one of which suggests the subject, in this case, the pump, as a filler. Reference resolution returns pump1 again, which passes the selection restriction of being a mechanical device. The final representation is: eauseP (instl gator (I), beeomeP (shearedP (patlent (pumpl)))).</Paragraph>
    <Paragraph position="3"> The last sentence in the text, &amp;quot;Inveatlgation re~ealed metal eontaminatlon ~n filter,&amp;quot; is interesting mainly because of the occurrence of two nomlnallzations which are discussed in detail in a separate section.</Paragraph>
    <Section position="1" start_page="132" end_page="133" type="sub_section">
      <SectionTitle>
2.3. Temporal Analysis of Tensed
Clauses
</SectionTitle>
      <Paragraph position="0"> The temporal component determines what kind of situation a predication denotes and what time it is asserted to hold for ~assonneau1988\].</Paragraph>
      <Paragraph position="1"> Its input is the semantic decomposition of the verb and its arguments, tense, an indication of whether the verb was in the perfect or progressive, and a list of unanalyzed constituents which may include temporal adverbials. It generates three kinds of output: an assignment of IIn other domains, the instigator might be an ~SSZN.</Paragraph>
      <Paragraph position="2"> TLU. role and would get filled by pragmatics.</Paragraph>
      <Paragraph position="3">  an actual time to the predication, if appropriate; a representation of the type of sRuation denoted by the predication as either a state, a process or a transition event; and finally, a set of predicates about the ordering of the time of the situation with respect to other times explicitly or implicitly mentioned in the same sentence. For the simple sentence, sac /'ailed, the input would consist of the semantic decomposition and a past tense marker:</Paragraph>
    </Section>
    <Section position="2" start_page="133" end_page="134" type="sub_section">
      <SectionTitle>
Deeomposltlons
</SectionTitle>
      <Paragraph position="0"> become (|no per ative (p atlent (is sell ) ))  The output would be a representation of a transitional event, corresponding to the moment of becoming inoperative, and a resulting state in which the sac is inoperative for some period initiating at the moment of transition.</Paragraph>
      <Paragraph position="1"> 8. Nomlnallsatlons Nominallzations are processed very slml\]arly to clauses, but with a few crucial d~erences, both in linguistic information accessed and in the control of the algorithm. The first important linguistic characteristic of the nom;nallzation algorithm is that the same predicate decomposition can be used as is used for the related verb. Secondly, d~erent mapping rules are required since syntactically a nominallsatlon is a noun phrase. For example, where a likely filler for the patient of fail, is the syntactic subject, a llkely filler for the patient of failure is an of pp. Thirdly, nominalisations do not make use of the obligatory classification for semantic roles, since noun phrase modifiers are not syntactically obligatory. In terms of d~rerences in control structure, because nom;nallzations may themselves be anaphorlc, there are two separate role-filling stages in the algorithm instead of just one. The first pass is for filling roles which are explicitly given syntactically; essential roles are left unfilled. If a uominalization is being used anaphorically some of its roles may have been specified or otherwise filled when the event was first described. The anaphorlc reference to the event, the nomina\]izatlon, would automatically inherit all of these role This suggests the hypothesis that OBLIGATORY roles For clause decompositions automatically become BSSeNTL~ roles for nominalization decompositions. This hypothesis seems to hold in the current domain; however, it will have to be tested on other domains. We are indebted to James Allen for this observation.</Paragraph>
      <Paragraph position="2"> fillers, as a by-product of reference resolution.</Paragraph>
      <Paragraph position="3"> After the first pass, the interpreter looks for a referent, which, if found, will unify with the nomlnalisatlon representation, sharing variable bindings. This is a method of filling unfilled roles pragmatically that is not currently available to clause analysis s. However, the first pass was important for filling roles with any explicit syntactic arguments of the nom;nalizatlon before attempting to resolve its reference, since there may be more than one event in the context whkh nominallzation could be specifying. For example, failure of pump and failure of sac can only be distinguished by the filler of the patient role. After reference resolution a second role-filling pass is made, where still unfilled roles may be filled pragmatically with default values in the same way that unfilled verb roles can be filled.</Paragraph>
      <Paragraph position="4"> S.1. Temporal Analysis of Nomlnallzatlons null As with clauses, the temporal analysis of norninallsatlons takes place after the semantic analysis. Also as with clauses, one of the inputs to the temporal analysis of nomlna\]isatlons is the semantic decomposition. The critical d~erence between the two cases is that a nom;nalisation does not occur with tense. PUNDIT compensates by looking for relevant temporal information in the superordinate constituents in which the nominalizatlon is embedded. Currently, PUNDIT processes nomlnalizatlons in three types of condeg texts.</Paragraph>
      <Paragraph position="5"> The first context for which a nomlnalisation is temporally processed is when it occurs as the prepositional object of a temporal connective (e.g., before, during, after) and the matrix clause denotes an actual situation. For example, in the sentence sac lube oil pressure decreased belato 60 pslg after engagement, the temporal component processes the main clause as referring to an actual event which happened in the past and which resulted in a new situation. When PUNDIT finds the temporal adverbial phrase after engagement, it assumes that the engagemeat also has actual temporal reference. In such cases, the nomlnalisat|on is processed using the ! Clauses can describe previously mentioned events, as discussed in \[Dahl1987\]. In order to handle cases like these, something analogous to reference resolution for clauses may be required. However a treatment of this has not yet been implemented in PUNDIT.</Paragraph>
      <Paragraph position="6">  meaning of the adverb and the tense of the main clause.</Paragraph>
      <Paragraph position="7"> The second context in which a nominallzation undergoes temporal analysis is where it occurs as the argument to a verb providing temporal information about situations. Such verbs are classified as aspectual. Occur is such a verb, so a sentence like failure occurred would be processed very s~miIarly to a clause with the simple past tense of the related verb, i.e., aomethlng faile&amp; Another type of verb whose nominallzation arguments are temporally processed is a verb which itself denotes an actual situation that is semantically distinct from its arguments. For example, the sentence in,aestlgatlon re~ealed metal C/onfam~natlon i~t oil filter mentions three situations: the situation denoted by the matrix verb reveal, and the two situations denoted by its arguments, ineemt~gatlon and eontamlnatlo~ If the situation denoted by reveal has actual temporal reference, then its arguments are presumed to as well.</Paragraph>
      <Paragraph position="8"> 8.2. Nominallsatlon Mapping Rules We will Use the previous example, ineestlgatlon revealed metal eontamlnatlon in filter, to illustrate the nom~nallsation analysis algorithm. We will describe the eontamlnatlon example first, since all of its roles are filled by syntactic constituents. The dotted llne divides the algorithm in Figure 2 in the Appendix into the parts that are the same (above the line), and the parts that differ (below the llne.)  a noun modifier of contamination, and metall is selected as the filler of the instrument role.</Paragraph>
      <Paragraph position="9"> FOR theme ROLE - The theme of a nominaUnation can be syntactically realized by an of pp or an in pp. The role is filled with fllterl, the referent of/~lPSer.</Paragraph>
      <Paragraph position="10"> At this point the temporal component is called for the nomlnalisation metal eontamlnatlon in oll filter with two inputs: the decomposition structure and the tense of the matrix verb, in this case the simple past. Because this predicate is stative, the representation of the eontamlnatlon situation is a state predicate with the decomposition and a period time argument as well as the unique identifier S, (which will be eventually be instantiated by reference resolution as \[eontaminatel\]): state(S,</Paragraph>
      <Paragraph position="12"> In this context, the past tense indicates that at least one moment within the period of contamination precedes the time at which the report was filed.</Paragraph>
    </Section>
  </Section>
  <Section position="10" start_page="134" end_page="135" type="metho">
    <SectionTitle>
CALL REFERENCE RESOLUTION FOR NOlV\[I-
</SectionTitle>
    <Paragraph position="0"> NALLZATION - There are no previously mentioned (c)ontamlnatlon events, so a new referent, eontamlnatlonl is created. There are no unfilled roles, so the analysis is completed.</Paragraph>
    <Section position="1" start_page="134" end_page="135" type="sub_section">
      <SectionTitle>
8.3. Filllng Essential Roles
</SectionTitle>
      <Paragraph position="0"> The analysis of the other nominallzation, in~emtlgatlon, illustrates how essential roles are filled. The decomposition of investigate has two semantic roles, a NON-ESSENTIAL agent doing the investigation and an OBLIGATORY theme being</Paragraph>
      <Paragraph position="2"> There are no syntactic constituents, so the mapping stage is skipped, and reference resolution is called for the nominallzatlon. There are no previously mentioned investigative events in this example 10, so a new referent, investigat|onl is created. At this point, a second pass is made to attempt to fill any unfilled roles.</Paragraph>
      <Paragraph position="3"> I In other domains, the theme can be essential, as in &amp;quot;I heard a noise. Let's investigate.&amp;quot; I0 If the example had been, A sew eaC/iseer isweetlgate&amp; tAe pump. TAe isteetlgntios oeeurre~ just before tAe complete breakdown., a previously mentioned event would have been found, and the agent and theme roles would have inherited the fillers engineer1 and pnmpl from the reference to the previous event.</Paragraph>
      <Paragraph position="4">  FOR agent ROLE - The role is NON-ESSENTIAL, so Case 4 applies, and it is left unfilled.</Paragraph>
      <Paragraph position="5"> FOR theme ROLE - The selection restriction on the theme of an ineestlgation is that it must be a d*msged component or a dauaage causing event. All of the events and entities mentioned so far, the ,ae and the pump, the failsre of the sac and the shcar/ng of the pump satisfy this criteria. In this case, the item in focus, the ,hearing of the pump, would be selected ~)ah11986\]. The final decomposition is: investlgateP (agent(A),theme(shearl))</Paragraph>
    </Section>
  </Section>
  <Section position="11" start_page="135" end_page="135" type="metho">
    <SectionTitle>
4. Other Compounds
</SectionTitle>
    <Paragraph position="0"> In addition to nom~nalisations, PUNDIT deals with three other types of noun-noun compounds. One is the category of nouns with arguments. These include preuure and temperature, for example. They are decomposed and have semantic roles like nominalisations; however, their treatment is different from that of nomlualisations in that they do not undergo time analysis, since they do not describe temporal situations. As an example, the definition of preuure, pressureP (theme(T),loeation(L)), specifies theme and location as roles. The analysis of a noun phrase like sa(c) oil preuure would fill in the loeatlon with the sac and the theme with the oil, resulting in the final representation, pressur eP (theme(oill),loeatlon(sael)).</Paragraph>
    <Paragraph position="1"> The syntactic mapping rules for the roles permit the theme to be filled in by either a noun modifier, such as all in this case, or the object of an o/ prepositional phrase, as in prcuure o/oil. Simllarly, the mapping rules for the location allow it to be filled in by either a noun modifier or by the object of an in prepositional phrase. Because of this flexibility, the noun phrases, sac all presmute, all preuure in sac, and pressure of oi1 in sac, all receive the same analysis.</Paragraph>
    <Paragraph position="2"> The second class of compounds is that of nouns which do not have semantic roles. For these, a set of domain-specific semantic relationships between head nouns and noun modifiers has been developed. These include: area of object, for example, blade tip, materlal-form, such as metal partlclea; and mater|al-objeet, such as metal eyllnder. These relationships are assigned by examining the semantic properties of the nouns. The corresponding prepositional phrases, as in tip o/ blade, particle, o/ metal, and cylinder of metal, have a similar analysis.</Paragraph>
    <Paragraph position="3"> Finally, many noun-noun compounds are handled as idioms, in cases where there is no reason to analyze the semantics of their internal structure. Idioms in the CASREPS domain include ,hip, force, gear *hair, and connecting pin.</Paragraph>
    <Paragraph position="4"> Our decision to treat these as idioms does not imply that we consider them unanalyzable, or noncompositional, but rather that, in this domain, there is no need to analyze them any further.</Paragraph>
  </Section>
  <Section position="12" start_page="135" end_page="135" type="metho">
    <SectionTitle>
5. Previous Computatlonal Treatments
</SectionTitle>
    <Paragraph position="0"> Previous computational treatments of nominalizations differ in two ways from the current approach. In the first place, nominallzations have often been treated simply as one type of noun-noun compound. This viewpoint is adopted by ~inin1980,Leonard1984,Brachman(nuli)\]. Certainly many nomlnalizations contain nominal premodifiers and hence, syntactically, are noun-noun compounds; however, this approach obscures the generalization that prepositional phrase modifiers in non-compound noun phrases often have the same semantic roles with respect to the head noun as noun modifiers. PUNDIT's analysis is aimed at a uniform treatment of the semantic s~ml\]arlty among expressions like repair of enflne, enf~ne repair, and Csomeone) repaired englne rather than the syntactic similarity of engine repair, sir preuure, and metal partleles. Of the analyses mentioned above, Brachman's analysis seems to be most similar to ours in that it provides an explicit link from the nominalization to the related verb to relate the roles of the noun to those of the verb. The second way in which our approach differs from previous approaches is that PUNDIT's analysis is driven by taking the semantic roles of the predicate and trying to fill them in any way it can. This means that PUNDIT knows when a role is not explicitly present, and consequently can call on the other mechanisms which we have described above to fill it in. Other approaches have tended to start by fitting the explicitly mentioned arguments into the role slots, thus they lack this flexibility.</Paragraph>
  </Section>
  <Section position="13" start_page="135" end_page="136" type="metho">
    <SectionTitle>
6. L|mltat|ons
</SectionTitle>
    <Paragraph position="0"> The current system has two main limitations. First, there is no attempt to build internal structure within a compound. Each nominal modifier is assumed to modify the head noun unless it is part of an idiom. For this reason,  noun phrases like impel\[or blade t~p erosion cannot be handled by our system in its current state because impel\[or b\[a,le tip forms a semantic unit and should be analysed as a a single argument of eroaion. The second problem k related to the first. The system does not now keep track of the relative order of nora|hal modifiers. In this domain, this does not present serious problems, since there are no examples where a different order of modifiers would result in a d~erent analysis. Generally, only one order is acceptable, as in mac oil eo~taminatlon, ~o~\[ both powerful and extenslble, and which will provide a natural basis for further development.</Paragraph>
  </Section>
  <Section position="14" start_page="136" end_page="137" type="metho">
    <SectionTitle>
Acknowledgements
</SectionTitle>
    <Paragraph position="0"> We would like to thank Lynette Hirschman and Bonnie Webber for their helpful commments on this paper.</Paragraph>
    <Paragraph position="1"> 7. Conclus|ons In this paper we have described a treatment of nom~nalisatlons ill which the goal ls to maxim\[se the s~m~\]arities between the processing of nominallsatlons and that of the clauses to whkh they are related. The semantic s~m~\]arltles between nom~nallzatlons and clauses are captured by making the semantic roles, semantk decompositions, and selectional restrictions on the roles the same for nomlna\]isations and their related verbs. As a result, the same semantk representation k constructed for both structures. This s~m;|arity in representation in turn anows reference resolution to find referents for nom;nallsations whkh refer to events previously described in clauses. In addltion, it allows the time component to integrate temporal relationships among events and situations described in clauses with those referred to by non~uaUsations.</Paragraph>
    <Paragraph position="2"> On the other hand, where d~erences between nom~uaUsations and clauses have a clear \]ingulstic motivation, our treatment provides for differences in processing. PUNDIT recognizes that the semantic roles of non~na\]ised verbs are expressed syntactically as modifiers of nouns rather than arguments of clauses by having a d~erent set of syntactic mapping rules. It ls also true in nominallsatlons that there are no syntacticaUy obligatory arguments, so the analysis of a nom;nallsation does not fall when there is an unfilled obligatory role, as is the case with clauses. Finally, the temporal analysis component is able to take into account the fact that nomlnallzatlons are untensed.</Paragraph>
    <Paragraph position="3"> ~rh;le there are many cases not yet covered by our system, in general, we believe this to be an approach to processing nomlnallsatlons which is</Paragraph>
  </Section>
  <Section position="15" start_page="137" end_page="137" type="metho">
    <SectionTitle>
APPENDIX
DECOMPOSE VERB;
FOR EACH SEMANTIC ROLE
CASE I: IF THERE ARE SYNTACTIC CONSTITUENTS -
PROPOSE SYNTACTIC CONSTITUENT FILLER
CALL REFERENCE RESOLUTION
&amp; TEST SELECTIONAL RESTRICTIONS
CASE 2: IF ROLE IS OBLIGATORY AND SYNTACTICALLY UNFILLED -
FAIL
CASE 3: IF ROLE IS ESSENTIAL AND UNFILLED -
CALL REFERENCE RESOLUTION TO HYPOTHESIZE A FILLER
&amp; TEST SELECTIONAL RESTRICTIONS
CASE 4: IF ROLE IS NON-ESSENTIAL AND UNFILLED -
LEAVE UNFILLED
CALL TEMPORAL ANALYSIS ON DECOMPOSITION
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
class="xml-element"></Paper>