The semantic representation of spatial configurations: 
a conceptual motivation for generation in Machine Translation 
Cornelia Zelinsky- Wibbelt 
IAI-EUROTRA-D 
Martin- Luther-Str. 14 
D-6600 SaarbrOcken 3 
E-mail: cor@iai.uni-sb.de 
Abstract 
This paper deals with the automatic translatiou of prepositions, 
which are highly polysemous. Moreover, the same real situation 
is often expressed by different prepositions in different 
languages. We proceed from the hypothesis that different usage 
patterns are due to different conceptualizations of the same real 
situation. Following cognitive principles of spatial 
conceptualization, we design, a semantic interpretation process 
for spatial relations in which our translation system uses 
semantic features derived from a semantic sort hierarchy. Thus 
we can differentiate subtle distinctions between spatially 
significant configurations. 
1. Introduction 
This p.qper deals with a general phenomenon of (machine) 
translation. The same real situation is often expressed differently 
in different languages. This is especially true for situations 
which are expressed by prepositions. We hold that different 
usage patterns resulting from this fact are due to different 
conceptualizations of the same real situation. The motivation 
for this phenomenon is given by the theory of Cognitive 
Grammar: Proceeding from the hypothesis of basic cognitive 
domains (cf. Langacker 1988.pp.54) it is possible to define all 
and only the properties which nmy become salient in spatial 
conceptualizations of entities and tht,s distinguish sufficiently 
one spatial conceptualization from another. Basic cognitive 
domains refer to basic cognitive capacities as, e,g., the ability to 
conceptualize space and time. They are basic, because in any 
hierarchy schematizing conceptual complexity they constitute 
the lowest level and thereby the range of the conceptual 
complexity. They are the most general cognitive capacities and 
are fixed in certain sonsomotorie as well as logical schemata, and 
as we have said, in some languages of tim world they are 
morphologically manifested. This is in line with ontologies based 
on prototype semantics, as they have been developed by 
Dahlgren (1986,1989), Hobbs (1987), Moens (1989)), Miller & 
Johnson-Laird (1976) for word sense disambiguation in 
different NLP systems. A hierarchy of semantic sorts is defined 
over basic cognitive domains, which is used for lexical 
representation, thus facilitating transfer (cf. Zelinsky-Wibbelt 
1988, 1989). We design a semantic interpretation process for 
spatial relations in which our translation system uses semantic 
features derived from the semantic sort hierarchy. The 
implementation of this interpretation process is based on 
assumptions about the cognitive process of configuring spatially 
significant entities by operating on conceptual representations of 
word meanings. The EUROTRA-D CAT2 system shows that 
with the definition of word meanings with respect to cognitive 
domains, we can interpret subtle distinctions between spatially 
significant configurations in an econmnic and elegant way. 
2. Schematlzatlon as the spatial organization of entities 
The main point we want to illustrate with the translation of 
prepositions is how certain circumstances and purposes of an 
utterance instantiate intportant principles of conceptualization. 
We will consider particularly pragmatic factors resulting from 
the specific environment of the language and the situation of 
utterance; we will investigate how these pragmatic factors 
determine the relevance, salience and typicality of the entities 
constituting the conceived situation. Moreover, we consider 
these moaning constituting factors which depend on the 
culture-specific environment of the language user to be 
conceptual motivations of meaning in the broadest sense, namely 
in the sense that abstract situations, which may not be perceived 
sensorically are conceived in terms of concrete, sensorically 
perceivable situations. This perceptually driven 
conceptualization of abstract scenes, by which metaphors are 
created and interpreted, cannot, however, be considered in this 
paper (We deal with this phenomenon in Zelinsky-Wibbelt 
1989b and Zelinsky-Wibbelt forthcoming). 
Prepositions are called relational expressions because they 
express how the conceptualizer profiles the relation between two 
participants: between the moving of moveable lrajeetor, usually 
referred to by the NP mentioned first, and the more stationary 
backgrouttded landmark, usually referred to by the second NP. 
In this assymetrical partitioning of the real scene, the relation 
between trajector and landmark is profiled in that the trajector 
is located with respect to tbe landmark. The asymmetric relation 
between trajector and landmark becomes obvious when we try 
to turn the relationship around as in the following examples: 
Tile cat is on tile mat./~'The mat is t~nder the cat. 
Industry is situated on tile Rhine./'The Rhine is situated 
on industry. 
Assuming that language does not express reality but bow we 
conceive of reality, tile semantic distinctions made by language 
with respect to our spatial environment do not necessarily agree 
with the entity's real spatial extension, but with its conceptual 
sehematization. Schematization is the fundamental principle 
underlying the linguistic expressions of spatial configurations 
(cf. Talmy 1983:225). It is the selection of those properties 
which become salient with the conceptualization of a scene, 
while the non-salient properties do not participate in this 
process. This means that in addition to prepositions being highly 
polysemons, most entities are Ioxically vague with respect to 
their possible spatial properties which they may realize in the 
respective configurations. 
The process of schematization is led by the following related 
principles which iustantiate certain spatial properties: 
- By the salience principle we prefer to associate an entity 
with a certain shape of one of its parts in a given 
relation. For instance in the conceptualization of the 
sentence "lhe children are riding o,r the bus" "bus" does 
not refer to the whole entity, which is 
three-dlmensional, but only to its two-dimensional floor. 
This becomes salient by the typical relation assumed 
between the entities, which implies the localization of 
the trajector within the space occupied by the landmark 
(place=(tr=part of Im)). This is an example of 
IDEALIZATION by which we focus on the salient 
dimension(s) of an entity and abstract from the 
non-salient dimensions, in this case the vertical 
dimension. The INTRINSIC ORIENTATION of "bus", which is 
FRONTAL, iS also abstracted away from in this example; 
within this relation it is salient that the bus has an 
interior bottom which functions as a SURFACE. 
1 299 
The relevance principle implies that dependent on the 
communicative goals we can choose one of disjoint 
prepositions for a specific spatial configuration. The 
communicative goals depend on the speaker's viewpoint, 
which in turn depends on the situation of utterance. The 
following example might illustrate this principle: Imagine 
a scene in which Mary is inside the building of a 
supermarket. If the speaker is far away from the scene 
he would designate Mary's location by saying "Mary is at 
the supermarket', thus expressing that his idealization of 
the three-dimensional extension of the supermarket to a 
point: by using the preposition "at" he asserts ttmt Mary's 
position coincides with that of the supermarket. If the 
speaker, however, is himsetf on the premises of the 
supermarket, he would designate Mary's location by 
saying "Mary is in the supermarket" thus referring to the 
three-dimensional extension of the building of the 
supermarket functioning as an ENCLOSURE. Thus 
different utterance situations result in disjoint 
conceptualizations and hence different expressions of the 
same real situation. 
The tolerance principle controls the pragmatic conditions 
under which expressions chosen by the speaker are 
adequate. This principle may for instance control 
whether idealization of trajector and landmark to a 
point, as in the above given example, is adequate with 
respect to the specific position of the speaker. The 
tolerance principle also controls the specific range of 
PLASTICITY of a relation. PLASTICITY is the general 
possibility of stretching the boundaries of a spatial 
schematization type with respect to the range of possible 
scenes conforming to it. Thus "in front of" may be used 
even if the located entity is not exactly in front of but 
also beside another entity (of. Herskovits 1988), as 
indicated in the following schema, where by focal 
adjustment we may view X to be located in front of Y 
within the given orientation: 
Front 
X 
Back 
X is located in/rent el Y 
Discourse situation 
The discourse situation may be 
provided by the verbal predicate 
of the sentence. 
Tile typicality principle implies the designation of a 
spatial configuration in dependance of typical relations 
existing between the entities. In our example "The 
children are riding on the bus*, the motion verb implies 
the discourse situation which instantiates the typical 
relation to be that the children are located inside of the 
bus and not on top of it. 
The typicality principle also makes possible the 
interpretation of an entity's INTRINSIC or TYPICAL 
orientation as the default case, if information to the 
contrary is lacking. For instance, human bodies, 
churches and other buildings have an intrinsic frontal 
orientation and a prominent vertical axis. 
Other properties by which we can distinguish different spatial 
configurations are 
the BOUNDARY CONDITIONS of an entity, including 
whether it is COUNT or MASS, but they may also be 
related to the SHAPE properties; that is, it may be of 
importance whether an ENCLOSURE is BOUNDED like a 
suitcase, PARTIALLY BOUNDED llke a bowl, or UNBOUNDED 
like an area. Moreover, the BOUNDARy CONDITIONI\] imply 
whether the entity is temporally UNBOUNDED like a 
STATE or an ACTIVITY or temporally BOUNDED like an 
ACHIEVEMENT or an ACCOMPLISHMENT (cf. Vendler 1967 
and Dowty 1979 for this classification; for the 
explication of the spatio-temporal analogy cf. e.g. Talmy 
1983:pp.255). 
The GRANULARITY of an entity. This refers to an entity's 
subdivision, which may be conceptualized with a more 
finegrained or more coarsegrained resotution. 
The process of schematization results in the asymmetric relation 
between trajector and landmark. This means that prepositions 
are two-place predicates (which is an old assumption of formal 
semantics; of. also Hawkins 1985.61). The relational concept - 
the trajector's spatial disposition - is designated by the 
preposition. 
Let us now organize the relations developing between trajeetor 
and landmark with respect to how they condition each other and 
how they result from the relevance, salience, and typicality of 
the entities constituting the conceived situation. This will be the 
precondition for implementing these relations in the form of 
rules in our translation system. In figure 1 we represent those 
relations which we assume to determine obligatorily the process 
of coneeptuallzing the spatial configuration of entities. We start 
with the relational concept which is part of the interpretation of 
the source language and which keeps constant during translation. 
Universal relation 
/ 
An interllngually constant PLACE relation is pre- \[ 
supposed that implies the global space in which 
the trajector is localized with respect to the 
landmark. 
~ an__aa laodmar____~k 
urse situation instantiates 
I eertaln spatial properties of trnjeetor land 
landmark. 
\] As a consequence of the spatial properties of I 
I trajector and landmark the language specific I 
I relation is conceptualized. \[ 
300 
Figure I Conceptualizing the spatial configuration of entities 
2 
Discourse situation 
The verb provides the discourse 
situation by predicating a motion. 
Landmark (the bus) 
As the sentence is about MOTION, with tile 
PLACE of the trajector within tile space _ - 
of the landmark, tile bus has the normal,- 
typic'd function of a LAROI'I VEIIIf3L}~; 
together with this the SUR',~'ACE of the 
floor become salient. 
Universal relation 
"I'he relation PLACE=(tr=part of Ira} 
is presupposed. 
Trajector (the children) 
Depending on the situation of MOTION, with the 
trajeetor within the space of the landmark, and tile 
normal VEIIICLI~ function of the bus being salient, 
the TRANSPORTABILITY of the children is tile 
typical property ill this relation. 
V 
Language specific relation (o~t) 
The salient part of the landmark being the 
SURFACE of its floor and tile typical property 
of the trajector being its TRANSPORTABII, ITY, 
the relation of support is conceptualized. 
Figure 2 lnstantiation of the spatial properties of children and bus 
Let us now exemplify how these properties and relations get 
irrstantiat~,d with an English speaker's conceptualization of our 
example sentence "The children are ridhtg Oil tire bl*s" on tile 
basis of the German source sentence, in figure 2 we give an 
overview of this. 
In the configuration of the designated spatial scetle "the bus" 
establishes tile landmark. Its spatial conceptualization is the 
condition for the relational concept designated by "o~t": 
Depending on the action of "riding" and tile TYPICAL FUNCTION 
Of "bus" being that of a LAI~O~ VElllCLI~, the T~'LANSPORTAnlL1TY 
becomes the typical property of the trajcctor, which is realized 
by the "chiidrepff' (see rule 9 below): This view of the entities 
excludes all other schematizations from being possible (e.g., that 
in which the children are on top of the bus). An additional 
condition for this schernatization is the information about the 
PLACE relation, which keeps constant during translation and 
which implies tile fact that the trajector is located within the 
space occupied by the landmark. In this configuration the 
SURFACE of the floor of the bus becomes salient, because it is 
relatively large, thus instantiating the relation of SUPPORT to be 
conceptualized between "children" and "bu.s" {see rule 11 below). 
A concel:~tualization where the surroundings become more salient 
has to be expressed by the preposition "hd', which designates the 
relation of INCLUSION as for example in "the customer in tire 
taxi". 
3. The process of schematization 
within a unification based environment 
Our implementation is done in the CAT2 system (cf. Sharp 
1988), an extension of the MT prototype (cf. Arnold et al. 1986) 
formerly nsed in EUROTRA. Although differing primarily in 
the implementation, the basic translation philosophy has been 
preserved. The translation procedure is stratificational in that it 
is split up into the translation between several linguistically 
motivated levels, representing constituency, syntactic functions 
and semantic relations. In this paper we are only concerned with 
the semantic level, the Interface Structure (IS), which should 
contain the semantic information required for transfer, analysis 
and synthesis. For a more detailed description of tile current 
CAT2 system and the current 1S conception see Sharp 1988, 
Zelinsky-Wibbelt 1988 and 1989, and Steiner et at. 1988. 
Let us now relate the process of schematizatlon to generating a 
representation by stepwise rule application, where the rules 
include the instantiations of the schematlzatlon principles given 
in section 2 ("st" stands for semantic feature, ";" for disjunction, 
"pred" for predicate, "arg" for argument, and "rood" for 
orodifier). The language-specific semantic representation which 
unifies with tile correct preposition is generated in the 
respective target language component. We illustrate the 
translation of our example sentence "Die Kinder fahren irn Bus" 
into "The chitdretz are riding on the bus". In order to keep the 
representation clear we give the rules in a very simplified 
version, containing only the information relevant in this context, 
namely tile information about the typicality, salience, and 
relevance of basic cognitive domains and domain-specific 
typical functions: 
German lexical rules: 
(1) (pred,{cat=prep,lu=in,place=(tr=part of hn}).\[*\]. 
Feature co-occurrence rules: 
(2) (?,{cat=pP,place=m }).\[(pred,(cat=prep,piace=A)),* \]. 
Transfer rules: 
(3) (mod,{cat=pp,Place=A)) => (mod,{cat=pp,place=A}).\[*\]. 
English lexical rules: 
(4) (pred,{cat=n,lu=bus, 
sf={typieal function=large vehicle}}).\[*\]. 
(5) (pred,{cat=n,lu=ehild,sf={animate=human}}).\[*\]. 
(6) (pred,{cat=prep,lu=on,reteva n t=su ppor t)).\[* \]. 
(7) (pred,(cat=v,iu=rlde, 
s f=(activity=motion }, 
argJ={sf={animate=human}}}).\[*\]. 
3 0 :'~. 3 
Feature co-occurrence I'nles: 
(8) (?,{cat-np,sf=A)).\[(pred,(cat=n.sf~ A }),*\]. 
(9) (?,(cat=s)).\[ 
(pred,{cat=v,sf={aetivity=motion}}), 
(argl ,{cat=np,sf={salient=transportable, 
idealization=three-dimensional))), 
(mod,(cat=pp,plaee={tr=part_of lm}, 
sf of_ np={salient_shape=surface, 
idealization=two-dimensional, 
typical function=large vehicle}}D,*\]. 
(10) (?,(cat=pp,sf_of np=(salient shape=A)D.\[ 
pred, 
argl, 
(arg2,{cat=np,sf=(salient_sbape=A)})l. 
(1 I) (mod,{cat=pp)).\[ 
(pred,(cat=prep,place=(tr=part of Ira, 
relevan t=support))).\[ 
(argl .{cat=np,s f=(typical=transportable})). 
(arg2,(cat=np,sf~(sallent shape=surface}})\]. 
Rule (3) guarantees that the information about Pt,ACE. which is 
preserved during translation, is transferred. Rules (5) to (7) are 
lexieal rules denoting basic cognitive domains, whereas rule (4) 
denotes a typical function of a domain-specific entity. Both 
knowledge types are used in sentence rule (9), which effects that 
in sentences in which the verb inherently predicates a MOTION 
and has a PP-Modifier with an NP argument whose designated 
obiect - the landmark - typically functions as a LARGE VEtIICLE, 
TRANSpOIt.TABILITY is iastantiated as the salient property of the 
first NP argument of the verb, the trajector. What can then be 
instantiated is the idealization of the PP's NP-argument to a 
TWO-DIMENSIONAL SURPACE which is its salient part. Now the 
schematlzation type may be generated by rule (11). This rule 
effects that in a spatial configuration with a typically 
TIt/.,NSPOItTABLE trajector and a landmark which has a surface as 
its salient part, the relevant concept relating both trajector and 
landmark is that of SUPPORT, which unifies with the lexical rule 
(6) for the preposition on. 
The result of the generation process is represented in a 
simplified version in figure 3. 
predicate argtrnentl modifier 
eat=v cat=np cat=pp 
activity=motion index=O place=(tr=part of Im) 
typical=transportable sf of np=(salient shape=surface) 
I 
predicate argumentl 
cat=prep cat=np 
relevant=support type=empty 
index=D 
typical=transportable 
idealization=three-dimensional 
argument2 
cat=n~ 
shape=surface 
Figure 3 Spatial configuration as the result of unification 
While in this example the discourse situation was given 
intrasententially by the action of riding, it will often only be 
given extrasententially. This opens an area for future research, 
which will also comprise interaction with a knowledge base. 
Acknowledgements 
I would like to thank Wiebke Ramm, Randall Sharp, and J6rg 
Sch0tz, as well as many other colleagues whom I cannot mention 
here for helpful discussions. 

References

Arnold, D.J., S. Kranwcr, M. Rosner, I.. des Tombe, G.B. Varile 
1986 "The <C,A>,T Framework in Eurotra: A Theoretically 
Committed Notion for MT." In: t'roceedings o/ the Ilth 
Co,l/ere,ce on Computational Linguistics, Bonn 1986. 

Dahlgren, K, 
1986 "Using Commonsense Knowledge to Disambiguate 
Prepositional Phrase Modifiers." Proceedings o/ the 
AAAI-86 Veil (5th National Con/erence on Arti/icial 
Intelligence. August 11- 15, Philadelphia, PA.589~593. 

Dahlgren, K. and J. McDowell 
1989 "Knowledge Representation for Commonsense 
Reasoning." Computational Linguistics 15(3). 149-170. 

Dowry, D.R. 
1979 Word Meaning and Montague Grammar. The Sema,tics 
o/ Verbs and Times in Ge.erative Semantics and in 
Montague's PTQ, Dordrecht, Reidel. 

Hawkins, B.W. 
1985 The Semantics el EngUsh Spatial Prepositions. L.A.U.T 
Paper 142. 

llerskovits, A. 
1986 Language and Spatial Cognition. Cambridge University 
Press. Cambridge. 

Herskovits, A. 
1988 "Spatial Expression and the Plasticity of Meaning." In: 
Rudzka-Ostyn, B. Tot~ic~ in Cognitive Linguistics. John 
Benjamins Publishing Company. Amsterdam. 

Hobbs, J,R., W. Croft, & T. Davies 
1987 "Commonsense Metaphysics and Lexical Semantics." 
Computational Linguistics t 3.241-250. 

Langacker, R. 
1988 "A View on Linguistic Semantics." In: Rudzka-Ostyn, B. 
Topics in Cognitive Linguistics. John Benjamins 
Publishing Company. Amsterdam. 

Miller, G. A. and P.N. Johnson-Laird 
1976 La,guage and Perception. Cambridge University Press, 
London. 

Moens, M., J. Calder, E. Klein, M. Reape, and 11. Zeevat 
1989 "Expressing Generalizations in Unification Based 
Formalisms." Proceedings of ,.tie ACL. Manchester 
1989.174-181. 

Retz-Schmidt, G. 
1983 "Various Views on Spatial Prepositions." Al--3fagazi.,w 
9.2.95- 105. 

Sharp, R. 
1988 "CAT2 - Implementing a Formalism for Multilingual 
MT. In: Proceedings of the 2nd htternational Conference 
on Theoretical and Methodological Issues in Machine 
Translation o/ Natural La,guages. CMU Carnegie 
Mellon University, Pennsylvania. 

Steiner, E.\[I., U. Eckert, B. Roth and J. Winther-Thielen 
1988 "The development of the EUROTRA-D system of 
Semantic Relations." In: Stelner, E.II., P. Sehmidt and C. 
Zelinsky-WibbelL From Syntax to Semantics. Insights 
from Machine Translation. Pinter Publishers. London. 

Talmy, L. 
19~3 "How Language Structures Space." In: Pick, H.L. & L.P. 
Acredolo. Spatial Orientation. Theory, Research, and 
Application. New York & London. 

Vendler, Z. 
196"/ Linguistics in Philosophy. Cornell Uuiversily Press. 
London. 

Zelinsky-Wibbelt, C. 
1988 "From Cognitive Grammar to the generation of semantic 
interpretation in machine translation." In: Steiner, E.tI., 
P. Schmidt, and C. Zellnsky-Wibbelt, From Syntax to 
Semantics. Insights from Machine Translation. Pinter 
Publishers. London. 

Zelinsky-Wibbelt, C. 
1989a Machine Translation based on Cognitive Linguistics: 
What lexical semantics contributes to the semantic unity 
o/ a sentence. Eurntra-D Working Paper 16. 
IAI-EUROTRA-D. Martin-Luther-Str. 14, D-6600 
Saarbr0cken 3. 

Zelinsky-Wibbelt, C. 
1989b The Spatial Organization of Concrete Entities and their 
Abstract Metaphorical Target Senses. L.A.U.D. Paper A 
257. 

Zelinsky-Wibbelt, C. forthcoming 
1990 "The Boundaries of Universal Semantics: Prepositions as 
a Case for Conventional Meaning." In: Grabski, M. 
Semantics /or Translation between Natural Languages. 
Mouton de Gryuter. Berlin. 
