<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2003">
  <Title>FINDING TRANSLATION EQUIVALENTS: AN APPLICATION OF GRAMMATICAL METAPHOR</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Historically there has rather little interaction between work in text generation and machine translation (MT) -- even though language generation needs to be an integral component of any cornplete MT system. Current text generation systems are, however, achieving results which can be beneficially applied in the MT context. In this paper, I describe one such area of possible interaction between mechanisms developed for monolingual text generation and the requirements of MT. In particular, the increasing concern that text generation theories show for higher levels of semantics and its realization in linguistic form makes it possible to move away from lower-level, 'structural' transfer between languages. In Iket, some of the semantic specifications now being uncovered within text generation arc sufficiently abstract as to capture significant informational invariances across languages. Sophisticated monolingual generation components are able to generate appropriate linguistic structures from such abstract informational specifications. This avoids, in many cases, problems of (possibly complex) structural transfer. There is no reason a priori for the *Much of the work reported here was carried out with the financial support of the Institut fi',. Angcwandte hfformationsforschung at the University of the Saarland, S~mrbri'mken', additional support was provided by AFOS\[{ contract, F49620-87-C-0005, and in part by DAI{PA contract MDAg03-87-C-641. The opilfions in this report arc solely those of the author. Thanks are due to Erich Steiner, Jgrg Schi'ltz, and C (~cilc Pro'is for significant contributions.</Paragraph>
    <Paragraph position="1"> linguistic structures generated by monolingual generators for distinct languages for a given abstract semantic specification to be structurally similar: therefore, translations that involve very diverse structures are readily obtainable if they are semantically motivated. Monolingual generation components of the type described are also independently motivated by language processing tasks that do not involve MT and so are ill any case required. Providing for their usage within the MT context also is therefore doubly beneficial.</Paragraph>
    <Paragraph position="2"> The particular mechanism developed within text generation that I will apply to MT problems here is that of grammatical raelaphor (Ilalliday, 1985); grammatical metaphor was originally developed within the tradition of Systemic-Functional Linguistics and is now beginning to be applied within the PENMAN text generation system (Mann and Matthiessen, 1985; The Penman Project, 1989).</Paragraph>
    <Paragraph position="3"> Systemic-functional linguistics posits sets of mappings between semantic information to be expressed and grammatical features. One class of mappings is termed 'congruent', in that it offers an unmarked realization of a semantic concept type -- e.g., that a processual semantic entity is realized as a verbal constituent in the grammar -- while a further class of mappings is 'noneongrnent', in that it.</Paragraph>
    <Paragraph position="4"> enables marked correspondences between semantics and grammar --e.g., in nominalizations where a processual ,;emantic entity is realized as a nominal constituent in the grammar. Below I show how the set of linguistic structures related by grammatical metaphor form a useful equivalence class for MT: i.e., when seeking an appropriate translation tbr some sentence, one will often be found in the set of target language sentences formed by generating from the corresponding abstract semantic specification involdng grammatical metaphor for variations in the structures generated.</Paragraph>
    <Paragraph position="5"> To show how this works in more detail, I first describe the level of abstract semantic information that is currently used within the PENMAN system -- this we term the upper model- and the user interface to the text generation system that it. supports -- which is called SPL. Second, 1 provide more concrC/~te exampies of two distinct kinds of grammatical metaphor that we can now implement compntationally. And finally, I go on to show their applica~:.ion to MT using examples from English and German.</Paragraph>
    <Paragraph position="7"/>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 The Upper Model and SPL
</SectionTitle>
    <Paragraph position="0"> The PENMAN tipper model organizes the 'propositional type' meanings that need to be expressed in text; it provides a general semantic taxonomy of classes of experiences and objects. This classification can also be seen as an inheritance hierarchy that organizes concepts according to how they may be expressed. For example, the inheritance of certain roles defines the types of participants that processes may have and the types of qualities that may be ascribed to particular objects, while class-subclass relations capture generalizations about possible grammatical and lexical realizations of concepts. Significantly, this orientation towards supporting grammatical realization renders the upper model independent of particular domains -- the semantic taxonomy offers an organization that is required for any domain if it is to support natural language generation. Rather fine distinctions are drawn by the current upper model, which contains approximately 200 concepts; details are provided in Bateman, Kasper, Moore, and Whitney (1990).</Paragraph>
    <Paragraph position="1"> The upper model supports specifications of sentence 'meaning' that may serve as input to a text generation system (Kasper, 1989a), or as output from an analysis component (Kasper, 1989b). Such specifications, which are expressed in the PENMAN Sentence: Plan Language (SPL), abstract beyond many syntactic variations; they capture basic meanings -- defined in terms of the upper model -- that may be given a variety of linguistic realizations depending on other, specifiable, criteria. One consequence of the abstractness of this representation is that many sentences taken from distinct languages that would require complex structural transfer for translation simply share a common SPL representation, thus requiring no transfer at all. However, it is important to note that this is motivated by a commonality in linguistic function when a sufficiently abstract standpoint is taken rather than on any claims of universality; further details on the theoretical status of the shared representations are presented in Bateman (1989). We will see examples of the avoidance of structural transfer in Section 5.</Paragraph>
    <Paragraph position="2"> A simple example of an SPL specification using the upper model is shown below. This shows the SPL specification for the sentence: (1) Mary cut her finger.</Paragraph>
    <Paragraph position="3"> (cut / directed-action :actor (mary / person) :actee (finger / body-part :inalienable-possession-inverse mary) ) The labels cut, mary, and finger are instantial variables which may provide lexical information; each of these has to be assigned to a type (directed-action, person, body-part) drawn from the generalized meanings defined in the upper model. Relationships between entities are specified by means of roles (:actor, :actee, :inalienable-possession-inverse), which are also defined in the upper model. We could, therefore, gloss the meaning of this expression as a, dircclcd-action of an actor which affects one of the inalienable-possessions of that actor. 1 The positioning of the general types and relationships within the upper model provides much of the information that a grammar needs for constraining possible surface realizations. 2</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Grammatical Metaphors
</SectionTitle>
    <Paragraph position="0"> A powerful property of the relationship between the upper model and the grammar is the existence of grammatical metaphor. Grammatical metaphor occurs when meanings are realized through 'nontypical' selections of grammatical features; such realizations can be described systematically and bring their own distinctive contributions to the meaning expressed (for a general typology, see Ravelli, 1985). Two types of grammatical metaphor which we are beginning to be able to control computationally are rankshifling and complexity metaphors. Figure 1 shows the first grammatical decisions that must be made when beginning generation of a grammatical unit according to a systemic-functional grammar such as that used within PENMAN. For each meaning to be expressed, grammatical decisions must first be made as to the rank of the grammatical unit that will be used -i.e., clauses, groups (nominal, verbal, adverbial and prepositional), words, and morphemes -- and the complexity of that grammatical unit -- i.e., whether the unit is a single unit or some combination of similar units (e.g., conjunction). Traditionally, the mapping between meanings to be expressed and these grammatical choices has been rather inflexible and limited to congruenl realizations; i.e., by and large, a process-type meaning (activity, state, etc.) would be realized by a selection of a simple clause, an objecttype meaning by a selection of nominal, etc. Gramtactical metaphor makes explicit the fact that this relationship between meaning and grammatical form is considerably more flexible and many noncongruent realizations are possible.</Paragraph>
    <Paragraph position="1"> Consider, for example, the SPL specification for example sentence (2) The discharge of electricity resuited in a breakdown of the system.</Paragraph>
    <Paragraph position="3"> When the grammatical alternatives of the grammar are considered, rather than automatically selecting 1 Definitions of most of these upper model terms are given in Bateman, Kasper, Moore and Whitney (1990). \]: have extended the actual classification given there b~, t'ncluding inverse relations where necessary and by increasing the level of detail of description for the purposes of discussion. In addition, as with all the SPL examples given in this paper, I have made the simplification of removing specifications of all nonupper model related information since these are not relevant for the present discussion.</Paragraph>
    <Paragraph position="4"> 2 The additional information necessary to produce a unique utterance is factored in SFL into three functional components: the logical, textual, and interpersonal ~metafunctions'. The information in the SPL is drawn from a fourth component: the experiential metafunction. Only by combining meanings from all four components are sufficient constraints provided for a unique utterance (Matthiessen, 1987:258). In Sections 3-5, we see examples of variations in the constraints of the logical metafimction, and in the Conclusion I return briefly to the textual metafunction. Interpersonal constraints have been assumed constant throughout.</Paragraph>
    <Paragraph position="5">  some 'topmost' node for immediate assignment to clause rank, the mechanism of rankshifting gram-matical metaphor interposes a decision procedure which selects an input term most appropriate for realization as a clause. Similar decisions are made concerning each rank in the grammar and the complexity at each rank. This gives rise to the following possibilities for expression (focusing for the moment just on the expression of the cause-effect relationship and approximately preserving the textual organization): null  A further range of variation of this kind subsumes the problems of nominatization. This is equally relevant to our current example. For example, the SPL specification of tile discharge situation alone could, simply by altering the rank at which it is to be realize(l, support the following variations: rank realization clause electricity was discharged (V-N1) nominal group electricity being discharged (V-N2) discharge of electricity (V-N3) electricity's discharge (V-N4) We can take this further: the relationship between the proce,~s and the actee of electricity can be realized within the modification systems of the nominal group at an even finer level; this gives (cf. ttalliday, !985:160): (epithet) electrical discharge (V-N5) (classifier} electricity discharge (V-N6) These are all options provided by the current PENMAN grammar. Example sentence (2) is then a combination of variation (V-C2) at clause rank and (V-N3) at nominal rank.</Paragraph>
    <Paragraph position="6"> These are all options provided by the current PENMAN grarrmlar. Example sentence (2) is then a combination of variation (V-C2) at clause rank and further similar variations at nominal rank.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 The Application to MT
</SectionTitle>
    <Paragraph position="0"> We can also characterize certain kinds of variation observable across languages in translation equivalents in exactly the terms provided by rankshifting and complexity grammatical metaphors. For example, a German translation equivalent to example sentence (1) would be: (3) Mary schnitt sich in den Finger. (1) and (3) present problems for transfer because the number of participants and surface senfence structures differ. This requires that the processes involved be assigned to different classes and structural transfer rules are then necessary to relate the translatkmal possibilities. However, the state of affairs in (1) and (3) can be described in terms of the same upper model categories regardless of the particular language conside~vd, and this permits the generation of (3) from the same SPL specification as (1).</Paragraph>
    <Paragraph position="1"> Each monolingual grammar and lexicon will already directly control the congruent assignment of terms in the SPL to ranks of the grammar, although this assignment will differ across languages just as it &lt;lifters within languages when noncongruent assignments are made by the rankshifting grammatical metaphor mechanism. Thus, for (3), the first concept consumed during generation -- wilich thereby becomes tim 'head' of the highest level structure, i.e.: the process of the clause --- is the directed-action cutling (schneiden); the second concept is the actor relationship between the process and the person Mary --- this is realized by a simple nominative Subject (in the absence of other textual constraints that might effect voice, diathesis, etc.); the third concept consumed is the relationship of inalienable-possesswn between Mary and finger--- this governs the selection of beneficiary and its realization as a reflexive pronoun; the fourth is the actee relationship holding between tile body-part finger and cutting -- this is realized as the prepositional head of a prepositional group; and, finally, tile fifth concept consumed is the concept finger itself --- which gives the dependent nominal within the prepositional group. This contrasts to the case for (1), where assignment of SPL terms to ranks follows the order: cuzting:clause, :actor mary:nominal, :actee finger:nominal, : inalienable-possession-inverse teary: nominal- modifier.</Paragraph>
    <Paragraph position="2"> The grammar of English 'knows' that the :actee / :inalienable-possession combination can be realized as a nominal group (her finger), while the German grammar needs to 'know' that the combination requires both a nominal group (sich) and a location specification (in den Finger). This infof marion is entirely motivated by individual language-internal considerations; both grammars and lexicons are entirely monolingual and independent of each other and of their use for translation.</Paragraph>
    <Paragraph position="3"> An SPL expression can therefore receive a variety of linguistic realizations, even across dis;tinct languages. This extension to muttilingnality does not  add to the complexity of the realization mechanism, since the grammatical metaphor mechanism which supports it is already required for monolingual generation; once a monolingual grammar for any language is in place, these realizational possibilities for MT follow also. Grammatical metaphor thus provides a mechanism by which a syntactically neutral SPL specification licenses a set of metaphorically related realizations in any language for which an appropriate grammar has been constructed. 3 A contextually suitable translation is then often found in the set of sentences related by grammatical metaphor that a target language supports. The kind of translational mapping this provides is shown in Figure 2.</Paragraph>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5 Further examples of the
</SectionTitle>
    <Paragraph position="0"> utility of the mechanism for</Paragraph>
  </Section>
  <Section position="8" start_page="0" end_page="0" type="metho">
    <SectionTitle>
MT
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
5.1 Verb and prepositional phrase
combinations
</SectionTitle>
      <Paragraph position="0"> A likely German translational equivalent for example sentence (2) above is: 4 (4) Wegen der elektrischeu Enlladung brach das System zusammcn. Again, the structures of the two sentences are quite different -- a simple process in (2) seems to have been decomposed into a prepositional phrase in (4) -- and this would require complex structural transfer. However, the SPL provided for sentence (2) already entails both realizations: the clause structure of (4) is identical to variation type (V-C4) of the metaphor setfi All that needs to be altered is the lexical information pointed at by the instantial variables.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
5.2 Nominalizations
</SectionTitle>
      <Paragraph position="0"> The principles of grammatical metaphor also apply across languages within nominal phrases. For example, in the translation pair of sentences (2) and (4), we can also see a difference in the structure of the corresponding nominal phrases: discharge of electricity and elektrische Entladung. In terms of the possible realization variants for this nominal phrase: the English has selected (V-N3), the German (V-N5). The same holds for the other process in the pair concerning the system breakdown. Here again, therefore, the metaphor set contains appropriate translations whose structures are rather different.</Paragraph>
    </Section>
  </Section>
  <Section position="9" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 PENMAN provides an appropriate grammar for English,
</SectionTitle>
    <Paragraph position="0"> and a similarly appropriate grammar for German is currently under construction wathJn the KOMET text generation project (Steiner et al., 1989). Bateman and Li (1988) reports on earlier experiments with similar grammars for Japanese and Clfinese.</Paragraph>
  </Section>
class="xml-element"></Paper>