SITUATION SEMANTICS AND MACHINE TRANSLATION.
C.J. Rupp CCL, UMIST P.O. Box 88 Manchester M60 1QD
Introduction Situation Semantics: The Basics. Situation Semantics is one of the most Situation Semantics is an informational recent and controversial theories in approach to formal semantics. The formal semantics. Machine Translation philosophical basis of this theory is laid (MT) is a highly complex application out in Barwise and Perry's Situations domain, in which research is expensive and Attitudes (1983) (henceforth B&P). of both time and resources. On the Most of the structure of the theory can surface, the space for interaction be seen as arising out of three basic between these two fields would seem concepts. fairly limited, and in practice the the application of formal semantics in MT Attunement: has been very limited, a notable ... an organism must be attuned to exception being the Rosetta project similarities between situations, what (Landsbergen 1982, 1987). The abstract we have called uniformities, and to translation problem however remains relationships that obtain between and any application must be based on these uniformities... (B&P plO) some formalisation of the problem. Constraints: The purpose of this paper is demonstrate that the enriched ... systematic relations of a special theoretical vocabulary of Situation sort between different types of Semantics offers a more intuitive situation .... These systematic characterisation of the translation constraints are what allow one process, than was .possible using more situation to contain information traditional semanuc theories. This about another. Attunement to these demonstration will take the form of a constraints is what allows an agent to formalisation of the most commonly used pick up information from one method for MT in terms of Situation situation about another. (B&P p94) Semantic constructs. In this respect this Partiality: paper follows on from a previous paper (Johnson, Rosner & Rupp 1988), in Situation types are partial. They which MT was presented as a testing don't say everything there is to say ground for semantic representation about everyone or even everything languages. This paper will turn the issue about the individuals appearing in the around and consider what the theory of situation type. (B&P p9) Situation Semantics has to offer to an The other main features of the theory MT application. The abstract description can be seen as arising out of the of the MT system to be considered will interaction of these three concepts. The therefore remain the same. combination of attunement with The paper consists of a basic constraints, when applied to the problem introduction to the machinery of of linguistic meaning, leads to the Situation Semantics, an examination of relation theory of meaning. This states the problem of translation, a formal that language users are attuned to a description of a transfer-based MT particular type of constraint which system and some examples of the kind relates the situation in which an of lexical transfer one would expect to utterance is made with the situation it is define in such a system. about. Put more formally: a sentence
- 308 - relates an utterance situation, u, and a original notion. Speaker connections are described situation, s. therefore the set of culturally specific constraints to which the users of a u[ ls particular language are attuned in order The notion of efficiency arises out of to permit them to assign meaning to the interaction of this relation theory of occurrences of its expressions. meaning with the notion of partiality. Before considering some more recent Natural language expressions only carry developments associated with Situation a certain amount of information and so Semantics, it will be useful to sketch only partially determine the range of some of the distinctions between this appropriate utterance and described theory and more traditional semantic situations. They can therefore be said to theories, such as Montague Semantics, be efficient in that they can be used for with particular reference to the various purposes. The notion of implications that these may have for an efficiency implies a clear distinction MT methodology, as in Landsbergen's between meaning and interpretation. It Rosetta (Landsbergen 1982, 1987). is only possible to arrive at a full Partiality is the most obvious interpretation by anchoring the utterance characteristic of Situation Semantics, situation and, as a consequence, the when compared to traditional possible described situation to actual situations. world semantics of the Montagovian The sentence itself carries only meaning. variety. In traditional theories truth This theory is sufficiently fine-grained conditions take priority over content. to permit further distinctions within the The interpretation of a sentence is the utterance situation, which contains two set of possible worlds in which it would differing types of information: the be true. Each such world is total and discourse situation and speaker therefore fully determines the answer to connections. The discourse situation is any possible question that could be that part of the utterance situation asked about it. Some sentences will be concerned with the external facts of the necessarily true and be assigned the set discourse, such as the identity of the of all possible worlds as an speaker and hearer, the temporal and interpretation making them spatial location of the conversation and indistinguishable from one another. perhaps even information about the Others, including all sentences with a mental states of the speaker and hearer. necessary truth as a constituent part, The discourse situation must be form sets of logically equivalent anchored before an interpretation can be sentences each receiving the same determined. Speaker connections are interpretation. This results in a situation concerned with the linguistic attunement where attempts to generate a sentence that must be shared by the speaker and from its interpretation might result in a hearer for effective communication. The sentence with completely different meaning relation can therefore be content or the required sentence restated in terms of a discourse conjoined with a potentially infinite set situation, d, speaker connections, c, and of necessar)', truths. Hence B&P's described situation, s. argument m favour of partial interpretations which contain only a d,c[~]s fixed amount of information. This is also The notion of speaker connections one of the reasons why MT systems assumed in this paper differs slightly based on Montague Semantics have from that used in B&P, which was been predominantly concerned with concerned primarily with determining the derivation rather than representing the reference of certain clearly referential interpretation of sentences. phrases, such as proper names and The relation theory of meaning also definite descriptions. In this paper it is represents a much greater balance assumed that most words are in some between context and content than more sense referential although their referents traditional theories, where context is may be complex partial objects; this usually limited to the determination of a would seem a natural extension of the few indexical terms. Although it has not
- 309- yet been adequately explored, the comment (Barwise & Perry 1985, p143). contextual side of the meaning relation Unification-based approaches to does implicitly contain the possibility of Situation Semantics generally require representing aspects of the informational the inclusion of a level of abstract structure of texts, which is of essential representation which only partially importance in producing representations determines the range of possible for languages such as Japanese or interpretations. This can be seen as the German, where informational structure meaning carried by the sentence under a directly affects syntax. It is not possible range of interpretations, but it will be to treat such languages in any depth less ambiguous that the original even with derivational techniques, when sentence as different syntactic analyses only truth conditional information can be will give rise to different representations. recorded. It is possible to state the meaning Finally traditional semantic theories relation imposed by such a assume a static and total interpretation representation or situation schema in function, which assigns denotations to the same terms as were used for the lexical items. This poses two distinct sentence. problems when considering translation. Firstly in the case of words with more d,c[sit.~]s than one sense it is not obvious how to The relation between a sentence, decide which denotation to choose. and its representation sit.qb will be given What is required is a more dynamic by a grammar G, which maps strings of a mechanism which permits the preferred language L to members of a class of reading to vary according to the context. representations R, (which will be in the Secondly there is the implicit form of Directed Acyclic Graphs). In assumption that the range of possible order to reflect the semantic relation denotations is common to both between these two objects it will be languages concerned, and if we reject necessary to define two auxiliary this assumption we are faced with the "interpretation functions" which metaphysical problem of constructing determine the set of possible appropriate denotations for each interpretations so that language out of an unknown set of primitives, with no philosophical InL(¢) = {
- 310- therefore not in the minds of the the text, but also the range of language users but in the communicative uniformities that it is possible to refer to. interaction between them. This makes This association of linguistic forms with this level of representation the perfect uniformities in the real world is provided medium for the study of translational by speaker connections and these will equivalences. be the domain over which translation must operate. Speaker connections do Translation Equivalence within however cover certain text internal a Situational Framework. forms of reference, such as anaphoric binding; these should also remain This section is concerned with the two predominantly impervious to translation. essential problems of any approach to It is mainly those connections involved MT: the nature and extent of the in reference into the described situation information that must be preserved, and that must be altered. While this domain the nature of the alteration which must only represents a very small part of the be effected. Following on from the situational formalisation of the meaning previous section it would seem that a relation it still represents a vast area of partial representation which carried the potential variation. content of the text ought to supply sufficient information to be preserved. Transfer-based MT. This would represent the meaning of the text while leaving ambiguities of The problem space for MT is interpretation underspecified. This would traditionally viewed as being triangular effectively freeze the described in shape (Vauquois 1979). In this model situation, leaving the context side of the the problem of translating between texts meaning relation as the only domain for is reduced to that of a transfer mapping translation operations. between abstract representation of those texts. It is usually assumed that A text places fewer constraints on its there is a direct relationship between context than a conversation, because the the complexity of the transfer operations author and reader know a lot less about and the level of abstraction of the each other than do the corresponding representations; some of the issues speaker and hearer. It follows from this involved in this trade-off are discussed that much of what a text does have to in Krauwer & des Tombe (1984). The say about its context will remain limit case is where the representations constant under translation. If an author are sufficiently abstract that transfer assumes his reader to have specialised becomes vacuous: this is exemplified by knowledge of a particular subject domain the interlingual approach adopted in then this requirement should not be Rosetta (Landsbergen 1982). Increased affected by translation. This type of abstraction can however lead to the loss information is external to the text and of relevant information and implies therefore would not appear in the recourse to a culturally independent set representation of the content and so of primitives. The adoption of a would not be affected by translation. The situational framework for an MT system only major alteration required in the places interesting constraints on the context of the text is that the reader and method to be employed, because both author are considered to be users of the the abstract representational level and target language rather than the source the nature of the transfer mapping are language. This will not greatly affect the determined by the theoretical external facts of the interaction so the framework. Interestingly, this turns out discourse situation can remain constant. to be the kind of transfer-based method It will however drastically affect the most commonly advocated within linguistic attunements that the author syntactically oriented approaches to MT. and reader must share in order to communicate. These are culturally Within the current model, with conditioned and affect not only the way situation schemata functioning as the that words may be used to refer to the representational level in a transfer- uniformities that make up the content of based MT system, the abstraction from text to representation would be that
~,.-v - 311 - defined by the grammar relation, G, transfer mapping to a lexical level above, except that two versions of this smacks of regression towards primitive relation are now required. The parsing word-for-word translation, but with the relation would be given by a source assistance of recent developments language grammar, GsougcE. within unification-based grammar formalisms nothing could be further from Gsot~c~ = { <~,sit.~> I~ ~ Ls, the truth. There are two features shared sit.~ E Rs, by UCG (Zeevat et al. 1987), HPSG InRs(sit.~) c_ InLs(~) } (Pollard & Sag 1987) and recent versions of LFG (Fenstad et al. 1987, Generation would, similarly, use a Halvorsen 1987) which make the target language grammar, GTARG~T implementation of such lexical transfer possible. The fin'st is the combination of GTARGET = { <~',sit.¢~'> I ~' ~ Lt, syntactic, semantic and even sit.~' ~ Rt, phonological information expressed in InRt(sit.¢~') ~ InLt(¢~') } the same form at all levels of the grammar. This allows for the incremental The transfer relation can then be evaluation of constraints across these defined as a translation relation across various domains. The second is the representations, TR, expressed in terms concentration of information in the of: the two representations, sit.~ and lexicon, including information concerning sit.¢~', a constant described situation, s, the combinatory behaviour of individual and for the purposes of this model a lexical items. These two principles, constant discourse situation, d. The known as constraint propagation and lexicalism, should make it possible to actual mapping, K, will be defined across define lexical transfer relations in terms the two sets of speaker connections, c of the representations associated with and c'. individual words of the language, TR = {
- 312 - considering relations between attribute- Some examples may demonstrate how value graphs is the kind of lexical rule this technique might be applied to some found in PATR type environments of the less trivial equivalences between (Shieber 1984, Karttunen 1986). These representations. The often quoted are essentially relations between graphs equivalence between the Dutch sentence and are used to treat such phenomena Ik zwem graag. as passivisation. A similar mechanism could be used to implement lexical I swim willingly. transfer relations. There would however with the representation be one major change in the formulation of and the English such rules, namely the fact that the representations to be related belong to I like to swim. different grammars and so are involves a difference in the syntactic associated with different syntactic category used to express essentially the structures. This would affect the way same uniformity. This would be reflected that the root of the graph was associated with the lexical item and the in the structure of the semantic way that information about the representations assigned to these surrounding context was passed on. In a sentences. lexical rule information to be preserved can simply be equated, but here a translational equivalence is required. Ind ~ype sit-'] There are a number of ways in which this correspondence between elements of different domains might be treated. Cond el I- FTY e re-q" These include the kind of structural correspondence used for relating syntax and semantics in recent work on LFG (Halvorsen 1987, Kaplan & Halvorsen 1988, Kaplan, Netter, Wedekind & Zaenen 1989) and also bilingual lexical _coo./ entries as in Beaven & Whitelock (1988). The UD formalisation given here will assume a slightly more flexible version of the latter approach, in that not only is the requirement to associate entries of different lexicons recognised, but also the need to be free of the immediate syntactic structure. Rel c(graag) Before commenting on a UD implementation of such lexical relations it is necessary to point out that the UD environment does not support lexical (D rules of the form mentioned above. There is instead a more generalised notion of c(zwemmen)U relational abstraction over the Pol 1 representational domain. This permits the relational characterisation of most of the phenomena usually treated by lexical Fig. 1 - A Situation Schema for the rules, but not the interpretation under Dutch sentence: Ik zwemgraag. which such rules carry out non- monotonic operations on existing structures. Relational abstractions also The Dutch representation (Fig.l) shows permit lexical relations to be broken graag as a relation over pairs of down into more specific generalisations relational objects where the English allowing for a more modular treatment of (Fig.2) represents like as a relation such phenomena. between an entity and a situation.
- 313 - definition is framed solely in terms of the s,t--1 semantic representations and the direct connection between the two representations is made by embedding Cond Rel c(like) one under the lexical leaf of the other. This method of representing the correspondence between actual lexical Arg 1 c(/) entries is highly experimental and has not yet been applied to any of the larger grammars developed within the UD formalism. It does however avoid one of the more basic problems with the UD implementation of situation schemata: the fact that while it is relatively easy to _ t on' I m assert the existence of a piece of Pol 1 representation, it is not possible to -- Rel c(swim) ensure that this representation be associated with an actual piece of syntactic structure. It is relatively easy Arg 1"~ c(/)--'] to emulate projection mechanisms such as the a and @ mappings of Halvorsen & Kaplan (1988) by the use of attributes Pol 1 within a larger representation, but it is not currently possible to reproduce the Fig.2 - A Situation Schema for the corresponding inverse mappings. English sentence: I like to swim. The relation defined above could be described using conventional LFG The relation between the semantic notations and projections, including a 't representations of the two words can be mapping for translation between expressed by the abstraction LTR semantic representations. This would (Lexical Transfer) as follows: however require a slight alteration in the representation language to permit only LTR(Dutch,English) one condition on each object. The
- 314 - (o* ind type) =(1;o * ind type) (o* ind type) = sit I ,,e (o* cond tel ind type) = rel (o* cond tel cond rel) = graag o.d Re, Indp,,e re ]l ('¢o* cond rel) = like x(o* arg 1) = (zo* arg 1) '¢(o* cond rel cond arg 2) = / "clJ/ (xo * cond arg 2 cond rel) (xo* cond arg 2 ind type) = sit L~ond ~J (o* cond rel ind type) = rel m (o* cond rel cond arg 1) = Rel e(courir) (o* cond rel ind val) (xo* cond arg 2 cond arg 1) = (xo* cond arg 1) Arg This also expresses the translational equivalence of the two words purely in e(descendre~_ terms of their semantic representations, E but this formalism does in principle Pol 1 m permit the definition of inverse m projections so that o'lxo* would be Arg 1 c(i/) the c-structure node associated with the word like. In order to take advantage of 2 this device it is however necessary to "-~drType e~- sacrifice the increased expressive power of a representation language defined in UD and the highly modular treatment that the use of relational abstractions provides. While both of these _.Cond I I l_. formalisms permit the specification of the required lexical transfer relations / without the need to route all information 1_ ue)] through the source language syntax, it Rei c(r would seem more appropriate to explore the range of appropriate transfer relations within the UD formalism. A more complex example of this kind of lexical transfer relation might be taken Pol 1 from comparing the use of verbs of motion in English and French. In French Fig.3 - A Situation Schema for the the verb often describes uniformities associated with motion and its direction French sentence: and any specification of manner might I1 descend la rue en courant. have to rely on an adverbial modifier as in In English however verbs are usually I1 descend la rue en courant. more concerned with the manner of the He go-down the road runningly motion and the direction is left to a The representation associated with prepositional phrase, as in the this sentence is given in Fig.3. corresponding He runs down the road. which receives the representation in Fig.4.
- 315 - LTR(English,French) ~nd ~ype sit-]
- 316 - class words that provide most of the Acknowledgements. content of a text. There are however a few domains that are so structured that they are amenable to decomposition. This research was supported by an When languages refer to uniformities in SERC studentship (No. 8632138X). these domains it is usually with I would also like to express my constructs that are systematically gratitude to Rod Johnson for providing incorporated into their morphology or both intellectual and technical support with words from distinctly closed for this research. classes. These domains correspond to areas that have often caused major problems for MT, such as tense, aspect, modality and determination. In some of References. these domains it has already become accepted to appeal to an abstract representation that is essentially Barwise, J. and J. Perry 1983. language independent, as in the work of Situations and Attitudes. Cambridge, van Eynde on tense and aspect (e.g. Mass.: MIT Press. 1988). It is interesting that the Barwise, J and J. Perry 1985. Shifting primitives required for such Situations and Shaken Attitudes. representations correspond to the kind Linguistics and Philosophy Vol 8 No 1, of structural relations required in the 105-161. Dordrecht: Reidel. different object domains of Situation Semantics: locations, relations, Beaven, J.L. and P. Whitelock 1988. situations and entities. It is not possible Machine Translation Using Isomorphic to present any interesting examples of UCGs. Procceedings of COLING 88. Vol the treatment of such phenomena here 1 32-35, Budapest. as there is still much work to be done on Cooper, R. 1985. Aspectual Classes in determining appropriate sets of primitive Situation Semantics. Report No. CSLI- structural relations, though Cooper 85-14 Stanford: CSLI. (1985,1986) presents a basis for the treatment of tense and aspect within Cooper, R. 1986. Tense and Discourse Situation Semantics. Location in Situation Semantics. Linguistics and Philosophy. Vol. 9 No 1, Conclusion. 17-36. Dordrecht: Reidel. This paper has presented a formal Fenstad, J.E., P-K. Halvorsen, T. description of an approach to MT that is Langholrn and J. van Benthem 1987. based on principles drawn from Situation Situations Language and Logic. Semantics, but which utilises the same Dordrecht: Reidel. basic architecture as more syntactically motivated systems. It has also Halvorsen, P-K. 1987. Situation presented some examples of how such Semantics and Semantic Interpretation an approach might be implemented in Constraint-based Grammars. Report within current unification grammar No. CSLI-87-101 Stanford: CSLI formalisms. While this approach has yet Halvorsen, P-K and R.M. Kaplan 1988. to be implemented on any major scale, Projection and Semantic Description in related work at ISSCO, Geneva has Lexical-Functional Grammar. produced grammars for moderately large Proceedings of the International fragments of German and French which Conference on Fifth Generation deliver the kind of representation Computer Systems, FGCS-88. Tokyo. required by such a system. Johnson, R, M. Rosner and C.J. Rupp 1988. Situation Schemata and Linguistic Representation. Presented at the Lugano Workshop on Computational Linguistics and Formal Semantics, September 1988.
- 317 - Johnson, R, and M. Rosner 1989. A Rich Environment for Experimentation with Unification Grammars. Proceedings of the European ACL 1989. Manchester Kaplan, R.M., K. Netter, J. Wedekind and A. Zaenen 1989. Translation by Structural Correspondence. Proceedings of the European ACL 1989. Manchester Karttunen, L. 1986. D-PATR: A Development Environment for Unification-B ased Grammars. Proceedings of Coling 86.74-80. Bonn. Krauwer, S. and L. des Tombe 1984. Transfer in a Multilingual MT System. Proceedings of Coling 84 464-467. Stanford. Landsbergen, J. 1982. Machine Translation Based on Logically Montague Grammars. Proceedings of Coling 82. Landsbergen, J. 1987. Isomorphic Grammars and their Use in the Rosetta Translation System. In King, M. (Ed) Machine Translation Today: the State of the Art. Proceedings of the Third Lugano Tutorial, Lugano, Switzerland, 2-7 April 1984. Edinburgh University Press. Pollard, C. and I.A. Sag 1987. Information-based Syntax and Semantics: Volume 1 Fundamentals. CSLI Lecture Notes Series No. 13 Stanford: CSLI. Shieber, S.M. 1984. The Design of a Computer Language for Linguistic Information. Proceedings of Coling 84 pp 362-366. Stanford. van Eynde, F. 1988. The Analysis of Tense and Aspect in Eurotra. Procceedings of Coling 88 Vol 2 pp 699- 704. Budapest. Vauquois, B. 1979. Aspects of Mechanical Translation in 1979. Conference for Japan IBM Scientific Program (GETA Report). Zeevat, H., E. Klein and J. Calder 1987. Unification Categorial Grammar. In Haddock N, E. Klein and G. Morrill (F_As) Working Papers in Cognitive Science, Vol. 1: Categorial Grammar, Unification Grammar and Parsing. University of Edinburgh: Centre for Cognitive Science.
- 318 -