FINDING TRANSLATION EQUIVALENTS: AN APPLICATION OF GRAMMATICAL METAPHOR
John A. Bateman* USC/Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292 U.S.A. (e-mail: bateman@isi, edu)
April 4, 1990
Abstract linguistic structures generated by monolingual gen- erators for distinct languages for a given abstract se- In this paper I describe how a significant class of cases mantic specification to be structurally similar: there- that would involve (possibly complex) structural trans- fore, translations that involve very diverse structures fer in nmchine translation can be handled avoiding trans- are readily obtainable if they are semantically moti- fer. This is achieved by applying a semantic organization vated. Monolingual generation components of the developed for monolingual text generation that is suffi- type described are also independently motivated by ciently abstract to remain invariant, within theoretically language processing tasks that do not involve MT specifiable limits, across different languages. The further and so are ill any case required. Providing for their application of a mechanism motivated from within mono- usage within the MT context also is therefore doubly lingual text generation, 'grammatical metaphor', then beneficial. allows candidate appropriate translations to be isolated. The particular mechanism developed within text The incorporation of these essentially monolingual mech- anisms within the machine translation process promises generation that I will apply to MT problems here to significantly improve translational capabilities; exam- is that of grammatical raelaphor (Ilalliday, 1985); grammatical metaphor was originally developed ples of this are presented for English and German. within the tradition of Systemic-Functional Linguis- tics and is now beginning to be applied within 1 Introduction the PENMAN text generation system (Mann and Matthiessen, 1985; The Penman Project, 1989). Historically there has rather little interaction be- Systemic-functional linguistics posits sets of map- tween work in text generation and machine trans- pings between semantic information to be expressed lation (MT) -- even though language generation and grammatical features. One class of mappings is needs to be an integral component of any cornplete termed 'congruent', in that it offers an unmarked re- MT system. Current text generation systems are, alization of a semantic concept type -- e.g., that however, achieving results which can be beneficially a processual semantic entity is realized as a ver- applied in the MT context. In this paper, I de- bal constituent in the grammar -- while a fur- scribe one such area of possible interaction between ther class of mappings is 'noneongrnent', in that it. mechanisms developed for monolingual text gener- enables marked correspondences between semantics ation and the requirements of MT. In particular, and grammar --e.g., in nominalizations where a the increasing concern that text generation theories processual ,;emantic entity is realized as a nominal show for higher levels of semantics and its realization constituent in the grammar. Below I show how the in linguistic form makes it possible to move away set of linguistic structures related by grammatical from lower-level, 'structural' transfer between lan- metaphor form a useful equivalence class for MT: i.e., guages. In Iket, some of the semantic specifications when seeking an appropriate translation tbr some now being uncovered within text generation arc suf- sentence, one will often be found in the set of tar- ficiently abstract as to capture significant informa- get language sentences formed by generating from tional invariances across languages. Sophisticated the corresponding abstract semantic specification in- monolingual generation components are able to gen- voldng grammatical metaphor for variations in the erate appropriate linguistic structures from such ab- structures generated. stract informational specifications. This avoids, in To show how this works in more detail, I first de- many cases, problems of (possibly complex) struc- scribe the level of abstract semantic information that tural transfer. There is no reason a priori for the is currently used within the PENMAN system -- this we term the upper model- and the user interface to *Much of the work reported here was carried out with the financial support of the Institut fi',. Angcwandte hfformations- the text generation system that it. supports -- which forschung at the University of the Saarland, S~mrbri'mken', ad- is called SPL. Second, 1 provide more concr¢~te exam- ditional support was provided by AFOS[{ contract, F49620-87- pies of two distinct kinds of grammatical metaphor C-0005, and in part by DAI{PA contract MDAg03-87-C-641. that we can now implement compntationally. And 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 finally, I go on to show their applica~:.ion to MT using for significant contributions. examples from English and German.
i 13 / //
2 The Upper Model and SPL inalienable-possessions of that actor. 1 The position- ing of the general types and relationships within the The PENMAN tipper model organizes the 'proposi- upper model provides much of the information that tional type' meanings that need to be expressed in a grammar needs for constraining possible surface text; it provides a general semantic taxonomy of realizations. 2 classes of experiences and objects. This classifica- tion can also be seen as an inheritance hierarchy that organizes concepts according to how they may be ex- 3 Grammatical Metaphors pressed. For example, the inheritance of certain roles defines the types of participants that processes may A powerful property of the relationship between the have and the types of qualities that may be ascribed upper model and the grammar is the existence of to particular objects, while class-subclass relations grammatical metaphor. Grammatical metaphor oc- capture generalizations about possible grammatical curs when meanings are realized through 'nontypical' and lexical realizations of concepts. Significantly, selections of grammatical features; such realizations this orientation towards supporting grammatical re- can be described systematically and bring their own alization renders the upper model independent of distinctive contributions to the meaning expressed particular domains -- the semantic taxonomy offers (for a general typology, see Ravelli, 1985). Two types an organization that is required for any domain if it is of grammatical metaphor which we are beginning to to support natural language generation. Rather fine be able to control computationally are rankshifling distinctions are drawn by the current upper model, and complexity metaphors. Figure 1 shows the first which contains approximately 200 concepts; details grammatical decisions that must be made when be- are provided in Bateman, Kasper, Moore, and Whit- ginning generation of a grammatical unit according ney (1990). to a systemic-functional grammar such as that used within PENMAN. For each meaning to be expressed, The upper model supports specifications of sen- tence 'meaning' that may serve as input to a text grammatical decisions must first be made as to the generation system (Kasper, 1989a), or as output rank of the grammatical unit that will be used -- from an analysis component (Kasper, 1989b). Such i.e., clauses, groups (nominal, verbal, adverbial and specifications, which are expressed in the PENMAN prepositional), words, and morphemes -- and the Sentence: Plan Language (SPL), abstract beyond complexity of that grammatical unit -- i.e., whether many syntactic variations; they capture basic mean- the unit is a single unit or some combination of simi- ings -- defined in terms of the upper model -- that lar units (e.g., conjunction). Traditionally, the map- may be given a variety of linguistic realizations de- ping between meanings to be expressed and these pending on other, specifiable, criteria. One conse- grammatical choices has been rather inflexible and quence of the abstractness of this representation is limited to congruenl realizations; i.e., by and large, a that many sentences taken from distinct languages process-type meaning (activity, state, etc.) would be that would require complex structural transfer for realized by a selection of a simple clause, an object- translation simply share a common SPL representa- type meaning by a selection of nominal, etc. Gram- tion, thus requiring no transfer at all. However, it is tactical metaphor makes explicit the fact that this important to note that this is motivated by a com- relationship between meaning and grammatical form monality in linguistic function when a sufficiently ab- is considerably more flexible and many noncongruent stract standpoint is taken rather than on any claims realizations are possible. of universality; further details on the theoretical sta- Consider, for example, the SPL specification for tus of the shared representations are presented in example sentence (2) The discharge of electricity re- Bateman (1989). We will see examples of the avoid- suited in a breakdown of the system. ance of structural transfer in Section 5. ((cO / cause-effect :domain discharge A simple example of an SPL specification using :range breakdown) the upper model is shown below. This shows the (discharge / directed-action SPL specification for the sentence: (1) Mary cut her :actee (electricity / substance)) (breakdown / nondirected-action finger. :actor (system / object)))
(cut / directed-action When the grammatical alternatives of the grammar :actor (mary / person) are considered, rather than automatically selecting :actee (finger / body-part :inalienable-possession-inverse 1 Definitions of most of these upper model terms are given in Bateman, Kasper, Moore and Whitney (1990). ]: have ex- mary) ) tended the actual classification given there b~, t'ncluding in- verse relations where necessary and by increasing the level of detail of description for the purposes of discussion. In addi- The labels cut, mary, and finger are instan- tion, as with all the SPL examples given in this paper, I have tial variables which may provide lexical infor- made the simplification of removing specifications of all non- upper model related information since these are not relevant mation; each of these has to be assigned to for the present discussion. a type (directed-action, person, body-part) 2 The additional information necessary to produce a unique drawn from the generalized meanings defined in utterance is factored in SFL into three functional components: the logical, textual, and interpersonal ~metafunctions'. The the upper model. Relationships between entities information in the SPL is drawn from a fourth component: are specified by means of roles (:actor, :actee, the experiential metafunction. Only by combining meanings from all four components are sufficient constraints provided :inalienable-possession-inverse), which are for a unique utterance (Matthiessen, 1987:258). In Sections also defined in the upper model. We could, there- 3-5, we see examples of variations in the constraints of the logical metafimction, and in the Conclusion I return briefly fore, gloss the meaning of this expression as a, to the textual metafunction. Interpersonal constraints have dircclcd-action of an actor which affects one of the been assumed constant throughout.
14 2 ~ CLA U~;F-..., N NGMINAL.~ These are all options provided by the current PEN- MAN grarrmlar. Example sentence (2) is then a com- ~.~ GROU!~~ V~RBAL'~ bination of variation (V-C2) at clause rank and fur- ~'-'~ PREPOSITIONAI~. ther similar variations at nominal rank.
4 The Application to MT
We can also characterize certain kinds of variation observable across languages in translation equiva- lents in exactly the terms provided by rankshifting and complexity grammatical metaphors. For exam- ple, a German translation equivalent to example sen- tence (1) would be: (3) Mary schnitt sich in den Finger. (1) and (3) present problems for transfer Figure 1 The rank and complexity systems of the because the number of participants and surface sen- grammar fence structures differ. This requires that the pro- cesses involved be assigned to different classes and structural transfer rules are then necessary to relate the translatkmal possibilities. However, the state of some 'topmost' node for immediate assignment to affairs in (1) and (3) can be described in terms of clause rank, the mechanism of rankshifting gram-- the same upper model categories regardless of the matical metaphor interposes a decision procedure particular language conside~vd, and this permits the which selects an input term most appropriate for generation of (3) from the same SPL specification as realization as a clause. Similar decisions are made (1). concerning each rank in the grammar and the com- Each monolingual grammar and lexicon will al- plexity at each rank. This gives rise to the following ready directly control the congruent assignment of possibilities for expression (focusing for the moment just on the expression of the cause-effect relationship terms in the SPL to ranks of the grammar, although and approximately preserving the textual organiza- this assignment will differ across languages just as tion): it
3 15 add to the complexity of the realization mechanism, 5.3 Realizations of modality since the grammatical metaphor mechanism which supports it is already required for monolingual gener- Consider the following translation pair (Schlitz, ation; once a monolingual grammar for any language 1989), which is again problematic in a transfer-based is in place, these realizational possibilities for MT fl'amework that relies on representations less ab- follow also. Grammatical metaphor thus provides stract than that of the upper model and SPL because a mechanism by which a syntactically neutral SPL the structures are very different: specification licenses a set of metaphorically related (5) John is likely to implement the algorithm realizations in any language for which an appropri- A ) John implementiert wahrseheinlich den ate grammar has been constructed. 3 A contextually lgorithmus suitable translation is then often found in the set of sentences related by grammatical metaphor that a (5) and (6) both share a common SPL representa- target language supports. The kind of translational tion: mapping this provides is shown in Figure 2. (I implement~ / directed,action ) :actor John 5 Further examples of the ..ctoe Igor t m--T utility of the mechanism for [ MT / probable)) As before, several differential realizations of this 5.1 Verb and prepositional phrase SPL can be generated by allowing the grammatical combinations metaphor mechanisms to order the consumption of terms as heads at differing ranks in the grammar. A likely German translational equivalent for example For example, the orderings [D A B C], [A D B C] and sentence (2) above is: 4 (4) Wegen der elektrischeu [A B C D] support clause structures with the follow- Enlladung brach das System zusammcn. Again, the ing dependency organizations respectively, showing structures of the two sentences are quite different realizations of the statement of probability at clause -- a simple process in (2) seems to have been de- complex rank, verbal group rank, and circumstantial composed into a prepositional phrase in (4) -- and adjunct rank: this would require complex structural transfer. How- ever, the SPL provided for sentence (2) already en- (i) [1 It is likely] tails both realizations: the clause structure of (4) is [z that [3 John] implements identical to variation type (V-C4) of the metaphor [, tile algorithm]] setfi All that needs to be altered is the lexical infor- (ii) [, [, John] [2 [3 is likely to] mation pointed at by the instantial variables. implement] [s the algorithm]] (iii) [, [2 John] will [4 probably] implement [s the algorithm]] [x [2 John] implementiert [4 wahrscheinlich] 5.2 Nominalizations [3 den algorithmus]] The principles of grammatical metaphor also ap- ply across languages within nominal phrases. For What would previously have required rather com- example, in the translation pair of sentences (2) plex transfer mechanisms, therefore, is here an auto- and (4), we can also see a difference in the struc- matic consequence of tile realizational mechanisms. ture of the corresponding nominal phrases: discharge of electricity and elektrische Entladung. In terms of the possible realization variants for this nominal 5.4 Across the text/grammar phrase: the English has selected (V-N3), the Ger- boundary man (V-N5). The same holds for the other process As a final example of the generality of the mecha- in the pair concerning the system breakdown. Here nisms described here (for further examples, see Bate- again, therefore, the metaphor set contains appropri- man, 1989), it is also possible to consider dependency ate translations whose structures are rather different. relations defined above the clause complex as possible 3 PENMAN provides an appropriate grammar for English, targets. The 'highest' rank assignment for the cause- and a similarly appropriate grammar for German is currently effect relationship suggested so far has been that of under construction wathJn the KOMET text generation project (Steiner et al., 1989). Bateman and Li (1988) reports on clause complex; this is the highest rank available in earlier experiments with similar grammars for Japanese and the grammar. However, in current work in the PEN- Clfinese. MAN group on Rhetorical Structure Theory (RST: 4 Note that I am not claiming that these sentences provide the only translations of each other. The argument rests upon Mann and Thompson, 1987), there is the possibility these being possible desired translation equivalents in some of extending this view to consider the dependency particular circumstances. relationships defined across sentences in texts. If the 3 Although a similar range of possibilities (V-C1 - V-C4) are theoretically available for the realization of the SPL in following translation pair were desirable, for exam- Germma, not all possible variations are equally acceptable in ple, then we see precisely this kind of variation mul- all contexts. For example, the realization of the cause-effect relation at clause rank calls for a process to realize that re- tilingually: lation: a plausible selection on the basis of similarity with the English sentence might appear to be the verb 'resultieren' (7) After you had explained your views, I -- however, this is very marked. The realization at circum- could see them much better. stantial adjunct group rank (in German with, e.g., a wegen- (8) Du erkl~irtest deine Position, Danach adjunct) is often preferable. verstand ich sie viel besser.
16 4 MggAMIORIC~LLY- MRTAM IORIC,%LI.Y. RELATED Ill~LATIr l) 8~T
att.~ o*~lnx ,S t TI ¢lat~ eoml~lu -.... SPL (p0 / pmc~t~ / : role, : rol~ a : )
S, /
L~i~ La~c~eo~rn~u~r Lz L:
Figure 2: Mapping between sets related by grammatical metaphor
Focusing on the temporal/causal relationship ex- ® a particular class of mapping relations between pressed between the two situations described, both this abstract semantic organization and lin- sentences can receive the combined SPL and RST guistic form- these mappings are motivated specification: by the notion of 'grammatical metaphor'.
((rO / rst-sequence :domain erklaeren The examples of the previous section made the :range verstehen) point that the level of grammatical metaphorical re- (erklaeren / verbal-process :sayer you alization is not preserved during translation, but the :message views) (verstehen / mental-process :senser f. set of possibilities defined by granmmtical metaphor :phenomenon views provides strong candidates for high quality transla- :propel~y-ascription tion equivalents, ttowever, since SPL forms decom- (better / evaluative~qualiZy))) mit from so much of the surface syntax, it is nec- Sentence (8) preserves the realization of the RST essary to control the range of linguistic realizations relation at the text level and so produces two in- that are compatible with the constraints SPL spec- dependent sentences. The next available strategy ifications represent. The types of control that are for realization is within the grammar, at clause rank necessary already constitute major areas of active re- with complexing. This realization is adopted in the search within the PENMAN teXt generation project, English variant (7). Again, therefore, the metaphor- where the principal type of constraint under devel- set provides a powerful link between translationally opment is textual. It is reasonable to assume that equivalent texts, even beyond the boundaries of the textual organization, including the particular pat- grammar. Work on this will continue as our devel- terns of thematic development adopted by a text, opment of RST and text organization continues. is one type of meaning which needs to be preserved in translation. Metaphorically related clauses often have rather different textual organizations and so 6 Conclusmn.o o Future direc- particular choices of clause may be motivated by the need to achieve particular thematic developments. tions A rather more global kind of textual constraint is that created by the 'style' of a text as a whole. Thus, Following from earlier experiments in the integration for example, the translation process should not select of Machine Translation (MT) and text generation as translation equivalents sentences whose stylistic (Bateman, Kasper, Schlitz, and Steiner, 1989), we import were radically divergent. This, however, can have found that the combination of two components only be established on the basis of analyses of what of the Systemic-Functional Linguistic (SFL) model constitutes an appropriate text style for a particular of language (Halliday, 1985; Matthiessen, 1987), such communicative situation. This is being addressed in as are being developed for computational use within text generation in terms of the SFL notion of register the PENMAN text generation system, significantly re- (Bateman and Paris, 1989), and its application to duces tlhe need for structural transfer in machine MT is pursued in Paris (forthcoming). translation (MT) without requiring deep modelling The relevance of textual constraints is also not lim- of specific domains. These two components may be ited to the simple types of grammatical metaphor il- described thus: lustrated in this paper; for example, there are also ® a linguistically motivated organization of gen- 'metaphors of transitivity' which chang(: the process- eral semantic distinctions that are not highly type involved, e.g.: language-specific and that hold across both dif- fering domains and differing languages -- this (9) electricity was discharged : directed- level of organization is realized computation- action ally within the PENMAN system as the upper (10) there was a discharge of electricity: re- model; lationM
5 17 Variations such as these provide significantly dif- [7] Kasper, Robert T. (1989a) A Flexible Interface for ferent thematization options. In certain cases, there- Linking Applications to Penman's Sentence Gen- fore, in order to achieve a particular thematic de- erator. In Proceedings of the DARPA Speech and Natural Language Workshop, Philadelphia, Febru- velopment for a text in a given language, this kind ary 1989. of metaphor may be the only means available for achieving a syntactically thematizeable constituent. [8] Kasper, Robert T. (1989b) 'Unification and Classi- This type of grammaticM metaphor can also be fication: an experiment in information-based pars- ing'. In International Workshop on Parsing Tech- motivated by the need to preserve Aklionsarl (cf. nologies, 28-31 August, 1989, Carnegie-Mellon Uni- Vendler, 1!)67). In the following translation pair, for versity, Pittsburgh, Pennsylvania, USA, ppl-7. example, we need to change the transitivity type of the predicate in order to preserve the accomplish- [9] Mann, William C., and Matthiessen, Christian ment Aktionsart. 6 M.I.M. (1985) 'A demonstration of the Nigel text generation computer program'. In Benson, J.D. and (11) Wit erarbeiteten uns die Theorie Greaves, W.S. (eds.) Systemic Perspectives on Dis- (12) We worked ourselves into the theory course: Volume 1; Selected Theoretical Papers from the 9th International Systemic Workshop. New Jer- Finally, the lexical entries of certain languages sey, Ablex. pp50-83. may also restrict the types of alternatives that are available. If one language simply does not have a [10] Mann, William C. and Thompson, Sandra (1987) lexical item for realizing a meaning at a given rank, 'Rhetorical Structure Theory: a theory of text or- then another rank will be pursued. ganization'. In Livia Polanyi (ed.) The Structure of Discourse, Norwood, New Jersey: Ablex Publishing Many of the mechanisms that need to be sup- Corporation. ported for these capabilities to be achieved are equally relevant in the generation of texts from sin- [11] Matthiessen, Christian M.I.M. (1987) 'Notes on the gle languages. We expect, therefore, that the results Organization of the environment of a text genera- we are obtaining in a monolingual setting --- par- tion grammar'. In Kempen, G. (ed.)(1987) Natural ticularly those concerned with the creation of texts, Language Generation: Recent Advances in Artificial Intelligence, Psychology, and Linguistics. Martinus rather than sentences -- will generalize well to ap- Nijhoff Publishers, Dordrecht, pp253-278. plication in multilingual environments such as MT. [12] Paris, C6cile. L. (forthcoming) 'User modelling, reg- ister, and machine translation'. To appear as a References EUROTRA-D Working Paper, Institut fiir Ange- wandte Informationsforschung, Saarbriicken. [1] Bateman, John A. (1989) 'Upper Modelling for Ma- chine Translation: a level of abstraction for preserv- [13] The Penman Project (1989) 'The PENMAN docu- ing meaning'. EUROTaA-D Working Paper. Institut mentation: Primer, User Guide, Reference Manual, fiir Angewandte Informationsforschung an der Uni- and Nigel Manual'. USC/ Information Sciences In- versit/~t des Saarlandes, Saarbriicken. stitute. [21 Bateman, John A., Kasper, Robert T., Moore, Jo- [14] Ravelli, Louise J. (1985) 'Metaphor, Mode and hanna D., and Whitney, Richard (1990) 'A general Complexity: an exploration of co-varying pa.tterns'. organization of knowledge for natural language pro- University of Sydney, Department of Linguistics. cessing -- the PENMAN Upper Model'. ISI Research B.A. Honours thesis. Report, USC/Information Sciences Institute, Ma- [15] Schi~tz, Jbrg F.L. (1989) 'Towards a framework for rina del Rey, California. knowledge-based translation'. Proceedings of 13th German Workshop on Artificial Intelligence, 1989, [3] Bateman, John A., Kasper, Robert T., Schlitz, Springer Verlag. Also available as IAI/Eurotra-D JSrg F.L., Steiner, Erich tI. (1989) 'A new view on Working Paper No. 10, Saarbr~ckeu. the process of translation'. Proceedings of the 4th. European Conference of the Association for Com- [16] Steiner, Erich, Bateman, John, Maier, Elisabeth, putational Linguistics, Manchester, England. Also Teich, Elke and Wanner, Leo (1989) 'Of hills to available in the ISI Reprint Series, report RS-89- climb and ships to sink -- generating German 234, USC/Information Sciences Institute, Marina within a functional approach to text generation'. del Rey, California. Arbeitspapiere der GMD, Gesellschaft fiir Mathe- matik und Datenverarbeitung, Integrated Pubhca- [4] Bateman, John A. and Li, Hang (1988) 'The appli- tion and Information Systems Insitute, Darmstadt. cation of systemic-functionM grammar to Japanese and Chinese for use in text generation'. Proceed- [17] Vendler, Z. (1967) Linguistics in Philosophy. Ithaca, ings of the 1988 International Conference on Com- N.Y.: Cornell University Press. puter Processing of Chinese and Oriental Lan- guages, Toronto, Canada, pp443-447. Also available in the ISI Reprint Series, USC/Information Sciences Institute, Marina del Rey, California. [5] Bateman, John A. and Paris, C&ile L. (1.989) 'Phrasing a text in terms a user can understand'. In Proceedings of the International Joint Confer- ence on Artificial Intelligence, IJCAI-89, Detroit, Michigan. [6] Haltiday, M.A.K. (1985) Introduction to Functional Grammar. London: Edward Arnold. 6Thls is a very important capability to provide for the whole area of translation of gemnan and Slavic "Pr~fixverben" (Steiner, p.c.),
18 6