A mapping of EuroWordNet Top Ontology into Upper Cyc Ontology is presented. The mapping is expressed in terms of a CycL microtheory encoding of the EuroWordNet Top Ontology, because it is impossible to be made just by means of equivalence and subsumption relations. However we provide also a simplified relational view that is sufficient for many purposes.
The mapping will facilitate a better understanding of those ontologies. It could also be used as a tool for the linking of the actual lexical items in the wordnets for the EWN-covered languages to the knowledge represented in Cyc KB.

Introduction

In this paper we will be concerned with mapping the EuroWordNet Top Ontology (see [Vossen 98]) to the Upper Cyc Ontology (see [Cycorp 97]). A direct and in the same time formally correct mapping via  equivalence and subsumption relations is impossible because of the complexity of the Upper Cyc Ontology. This is why the mapping is expressed in terms of a CycL (the knowledge representation language of Cyc) microtheory encoding of the EuroWordNet Top Ontology. We also provide, however, a simplified relational view of the mapping that is sufficient for many purposes. The purpose of the mapping is manifold: (1) to ensure linking between the concepts in the two ontologies; (2) to be used as a tool for the linking of the actual lexical items in several languages to the knowledge represented in Cyc KB; (3) to provide a more detailed semantic context information for the lexical items in EuroWordNet.

The Two Ontologies

A set of base concepts (BC) is developed in order to unify the conceptual knowledge represented among the different wordnets. The base concepts were selected from the resources available in each language according to their importance defined by two criteria: (1) the number of relations connected to this concept and (2) its position in the hierarchy of concepts.  The set "language independent" base concepts in the ILI is produced by merging (in a complex way) the sets for each language. Additionally, the set of BC is grouped in coherent clusters by means of the Top Ontology (EwnTO).  It comprises 64 concepts defining the fundamental semantic classes. Because of the choice of the BC it is expected that all the words could also be classified under the semantic features of the EwnTO via their relations to the BC.

The Upper Cyc Ontology (UCO) (see [Cycorp 97]) is the publicly available part of the Cyc® Knowledge Base and is devoted to the representation of language independent encyclopaedic knowledge. The Upper Cyc Ontology consists of about 3000 constants. Many of the constants are defined as unary predicates and could be viewed as concepts. Others denote relations, logical operators and so forth. The name of each Cyc® constant begins with the string #$. The set of constants in UCO is hierarchically organized by means of two structuring relations:

• #$isa - a relation connecting an instance with a collection. Formally, (#$isa X Y) means exactly that X is element of the set Y.
• #$genls - subsumption relation connecting one collection with another, more general one. Formally, (#$genls X Y) means exactly that X is a subset of Y.

The important point about UCO is that it contains (according to its developers) enough concepts to structurally classify properly each new concept. So it should be a good candidate for an upper ontology.

Representation of EwnTO in CycL

;;; #$EuroWordnetMt
(#$genlMt #$EuroWordnetMt #$BaseKB)
(#$comment #$EuroWordNetMt "#$EuroWordNetMt microtheory ...")

;;; #$EuroWordnetTCType
(#$ist #$EuroWordnetMt (#$isa #$EuroWordnetTCType #$PredicateCategory))
(#$ist #$EuroWordnetMt (#$comment #$EuroWordnetTCType "Collection of the EuroWordnet top-concepts. Each TC is a Cyc predicate."))

The highest concept in EwnTO (Top) is stated to be the same as the peak of the hierarchy of UCO:

;;; #$TopTC
(#$ist #$EuroWordnetMt (#$equals #$TopTC #$Thing))
(#$ist #$EuroWordnetMt (#$isa #$TopTC #$EuroWordnetTCType))

In the following assertions we will leave implicit the part which explicitly states that the assertions are in #$EuroWordnetMt microtheory namely (#$ist #$EuroWordnetMt ASSERTION). Below are some of the definitions of the auxiliary predicates used in the mapping:

;;; #$exactType
(#$genlPreds #$exactType #$isa)
(#$genls #$MappingPredicate #$TaxonomicSlotForCollections)
(#$equivalent (#$exactType ?COL ?TYPE)
                       (#$and (#$equivalent (#$isa ?X ?TYPE) (#$genls ?X ?COL))
                              (#$isa ?COL ?TYPE))

#$exactType is a specification of  the #$isa predicate which relates two collection such that the second collection ?TYPE is a collection of collections and it contains all sub-collections of ?COL (including ?COL itself) and only them. Using this predicate we can state as equivalent a part of the collection hierarchy and a collection of collections, i.e. concept and concept type.

;;; #$specificType
(#$isa #$specificType #$DefaultMonotonicPredicate)
(#$isa #$specificType #$TaxonomicSlotForCollections)
(#$equivalent (#$specificType ?COL ?TYPE)
                       (#$implies (#$isa ?X ?TYPE) (#$genls ?X ?COL))

Sometimes we need to state that a given collection contains only some collections from the sub-hierarchy. For this purpose we use the predicate #$specificType. In the above assertions #$equivalent is a logical operator easily definable in CycL.
 

Mapping the highest levels

Four concepts are defined one step below 1stOrderEntity concept  - Origin, Form, Composition, Function, which determine the main aspects of each time and space thing and correspond to the Qualia structure presented in [Pustejovsky 95]. In our view these concepts are more like attributes of existing objects than object classes or orthogonal dimensions of explanation. We encode them as classes of predicates rather than predicates. The next layer under these four concepts defines the concepts representing the actual values for these attributes. Here we give an example of the encoding of one of these concepts in  #$EuroWordNetMt:

;;; #$1stOrderEntityTC
(#$equals #$1stOrderEntityTC #$SomethingExisting)
(#$isa #$1stOrderEntityTC #$EuroWordnetTCType)

;;; #$1stOrderEntityTCType
(#$isa #$1stOrderEntityTCType #$PredicateCategory)
(#$genls #$1stOrderEntityTCType #$EuroWordnetTCType)
(#$exactType #$1stOrderEntityTC #$1stOrderEntityTCType)

;;; #$OriginTCType
(#$genls #$OriginTCType #$1stOrderEntityTCType)

;;; #$NaturalTC
(#$equals #$NaturalTC #$NaturalTangibleStuff)
(#$isa #$NaturalTC #$OriginTCType)

This example demonstrates the most simple mapping - the case there exists a Cyc constant with the same meaning as the EWN Top Concept.

The hierarchy below 2ndOrderEntiry is similar. The first level defines two dimensions (qualia) for the characteristics of a situation: SituationTypes and SituationComponents. The former divides  situations in dynamic and static while the later defines clustering of the situations according to the presence of a specific aspect in the description of the situation content. Thus we can follow the above pattern in the definition of these concepts also.

;;; #$SituationTypeTCType
(#$genls #$SituationTypeTCType #$2ndOrderEntityTCType)

;;;  #$SituationComponentTCType
(#$genls #$SituationComponentTCType #$2ndOrderEntityTCType)

Below these we often have more complicated mapping. For example, we need to define MentalTC as follows

;;; #$MentalTC
(#$equals #$MentalTC (#$UnionFn #$MentalEvent #$MentalAttribute))

Here an EWN top concept is represented as a disjunction of two UCO constants because UCO doesn't contain a common concept for dynamic and static situations which also to account for the mental aspect of a situation.

A more interesting case of classification is the mapping of the EWN's PartTC. In UCO there are some constants devoted to distiguishing some special kinds of parts as parts of organisms, parts of buildings and others but obviously there is no general definition part. Therefore all we can point out are examples of parts and we can say that in principle only in individual things can constitute a part of something. Here is the mapping definition:

;;; #$PartTC
(#$genls #$PartTC #$Individual)
(#$genls #$OrganismPart #$PartTC)
(#$genls #$CellPart #$PartTC)
(#$genls #$PartOfBuilding #$PartTC)
(#$isa #$PartTC #$CompositionTCType)

This definition reflects another characteristic of EwnTO which we will discuss in detail below. As a part of the Composition dimensions (or qualia) PartTC could be applied to many entities below 1stOrderEntity, but its value is not significant for many of the concepts: "It is not the case that all persons will be classified as Parts because they may be part of group." ([Vossen 98]) The notion of "intentional" significance is important in the EWN classification of word meanings but it is very hard to  represent them on a general level in Cyc. We decided to leave this concept underspecified.

A harder problem than the described above is the mapping of the top concept TimeTC. It is defined in [Vossen 98] as: "Situations in which duration or time plays a significant role; Static yesterday, day, pass, long, period, Dynamic e.g. begin, end, last, continue." This gives us a hint at to how to constrain the concept from above by the disjunction of two UCO constants (#$Event or #$StaticSituation) both of which have a temporal aspect. However this is still too general to cover the EwnTO meaning of TimeTC which doesn't include a lot of events and states which are classified under #$Event or #$StaticSituation.  The only appropriate concept in UCO is #$TemporalRelation but it is more specific and the mapping is problematic because it is not a specialization of #$SituationType.
 

Handling the differences in representation and structure

Mismatching taxonomic structure

Some of the problems arise because of differences between the formal definitions in the UCO and the EWN Top Ontology (with respect to the base taxonomic relations) for concepts which glosses are matching. In those cases the mapping was made according to the glosses.

Lets take as an example MoneyRepresentationTC  that is subsumed in EWN by RepresentationTC. They are mapped into UCO as follows:
    (#$equals #$RepresentationTC #$InformationBearingObject)
    (#$equals #$MoneyRepresentationTC #$TenderObject)
However it is NOT true in UCO that
    (#$genls #$TenderObject #$InformationBearingObject)
The reason for this is the different structuring of the conceptualizations used by the creators of both ontologies. Thus #$Currency (a specialization of #$TenderObject) is an #$InformationBearingObject (IBO) but doesn't covers some of the meanings of MoneyRepresentationsTC, for example "shares". The closest UCO concept for the latter one is #$Stock, but it is a specialization of #$SalesAgreement and it is not declared as a kind of  #$TenderObject. The reason for this is that #$Stock covers only the abstract aspect of the stock without its material (paper) media which is represented by #$StockCertificate. Unfortunately the later one is not declared to be a specialization #$TenderObject. The mismatch can be partially corrected if  #$TenderObject could be classified in UCO as a kind of IBO. The last assertion would be correct because each of  #$TenderObject instances could play this role. However the correctness and relevance of the last assertion is arguable. Finally, the comparison would be easier if #$StockCertificate be classified as a #$TenderObject, but the last is not obvious for some kinds of #$Stock.

This was a typical example in which we mapped MoneyRepresentation to #$TenderObject following the matching glosses. Our motivation for this decision was that it could be expected that the knowledge enterers using the ontologies (especially those with linguistic background) are more likely to also give preference to the meaning stated in the gloss. It is also a fact that the formal meaning encoded by the taxonomic relations is just a small fraction of what is described in the gloss.

Concept Types in UCO rather than Concepts

Some UCO constants stand for concept types (collections of collections) rather than for concepts themselves. Such concept types include, for example, the constant #$PositionType. It represents the collection of all concepts (predicates) about occupations (OccupationTC) but it is not a concept itself. The mapping is even harder when such a concept type covers just part of the sub-concepts of a top concept. Such is the case for example between #$SocialAttaributeType and SocialTC. The former covers only the static situations clustered under SocialTC.

The mapping between Time and  #$TemporalRelation (continuing the above example explanation) is even more interesting because the later constant does not represents a concept - it is a concept type or in other words meta-concept. That means that the concepts related with it (like #$after) are not its specializations (sub-concepts). They are its instances that in Cyc will be expressed via (#$isa #$after #$TemporalRelation) rather then using one of the subsumtion predicates (#$genls, #$genlPreds, etc.). In UCO there are many other cases like this - meaning that is conceptualized as a concept type rather than as a concept.

With respect to the above phenomena we need the following relations between a EWN top concepts and UCO constants to be used in the mapping:

Missing subsumtion relations in UCO

There are subsumption relations that are not precise in UCO. For example Occupation could be directly mapped to #$IntendedFunction. On the other hand #$PositionType (that has no relation with #$IntendedFunction) is still relevant to Occupation because all of its instances are specializations of Occupation. In this case we included #$PositionType as a constant that is not "the Mapping" but still relevant to Occupation. This case is additionally complicated by the fact that #$PositionType itself is a type of concept rather then concept. The mapping is:

    (#$equals #$OccupationTC #$IntendedFunction)
    (#$specificType #$OccupationTC #$PositionType)
 

Conclusion

The mapping defined here shows the level of compatibility between the EWN top ontology and UCO. It became clear that there are important aspects that could  not be properly covered in UCO. We  should say that this mapping could suffer in quality because of two main reasons: (1) the complexity of UCO, that the authors can not pretend to fully understand; (2) underspecification of some of the EWN top concepts, especially some of those below 1stOrderEntity. As a result it could be expected that some of the mappings could be improved.

The mapping we have constructed will be used in two future works. First, it will be extended with the base concepts of EuroWordNet and then used  for the creation of a Bulgarian lexical knowledge base connected to EuroWordNet base concepts and thus to UCO. Second, we envisage a use of the mapping for the analyses of text on the idea of lexical chains (see [Hirst and St-Onge 1998]) which will be used to determine the right ontology chunks assigned to the text along the lines of [Kiryakov and Simov 1999].
 

REFERENCES

[Cycorp 97] Cyc® Public Ontology.

[Fellbaum 98] Fellbaum, Christiane (editor), WORDNET: an electronic lexical database. MIT Press, 1998.

[Hirst and St-Onge 1998] Graeme Hirst and David St-onge, Lexical Chains as Representations of Context for the Detection and Correction of Malapropisms. In [Fellbaum 98]. MIT Press, 1998.

[Kiryakov and Simov 1999] Atanas Kiryakov and Kiril Simov. Ontologically Supported Semantic Matching. In the proceedings of NoDaLiDa'99 conference. Trondheim, Norway, 1999.

[Pustejovsky 95] James Pustejovsky. Generative Lexicon. The MIT Press. Cambridge, MA, 1995.

[Vossen 98] Vossen Piek (ed.), EuroWordNet General Document. Version 3, Final, July 19, 1999
 
 

APPENDIX: Relational mapping between EWN Top Concepts and Upper Cyc Ontology constants