=Paper=
{{Paper
|id=Vol-2228/paper10
|storemode=property
|title=Ontologies in Category Theory: A Search for Meaningful Morphisms
|pdfUrl=https://ceur-ws.org/Vol-2228/paper10.pdf
|volume=Vol-2228
|authors=Cauã Antunes,Mara Abel
|dblpUrl=https://dblp.org/rec/conf/ontobras/AntunesA18
}}
==Ontologies in Category Theory: A Search for Meaningful Morphisms==
https://ceur-ws.org/Vol-2228/paper10.pdf
Ontologies in Category Theory:
A search for meaningful morphisms
Cauã Antunes, Mara Abel
Instituto de Informática – Universidade Federal do Rio Grande do Sul (UFRGS)
Caixa Postal 15.064 – 91.501-970 – Porto Alegre – RS – Brazil
crantunes@inf.ufrgs.br, marabel@inf.ufrgs.br
Abstract. In order for the entire potential promised by the semantic web to be
achieved, applications must be able to integrate knowledge from multiple
sources. This requires reliable ways of relating disparate ontologies to be
found. Category theory and its morphisms represent a possible solution to this
problem, but careful definition of morphisms is required. This paper analyzes
several category theoretical approaches to ontologies, focusing on the chosen
morphisms and the semantic consequences thereof. Criteria for semantically
evaluating morphisms are then discussed and standing challenges are
identified.
1. Introduction
The Semantic Web was envisioned as a mean to allow computers to access, process and
make inferences over decentralized information [Berners-Lee et al. 2001]. In order to
achieve this, the concepts and the relations between them must be specified formally in
computational objects called ontologies.
Individual ontologies, however, are not enough to provide the full functionalities
conceived by the semantic web, given the distributed nature of the information and
intended applications. These applications must access knowledge spread across multiple
ontologies. In order for this to be possible, semantically sound methods must be found
to integrate information from different ontologies with heterogeneous specifications.
Category theory is a mathematical formalism that focuses on the relations
(referred to as morphisms in related literature) between entities (objects) rather than
emphasizing the entities themselves. Through this emphasis in relations, category theory
abstracts from the representational aspects that hinder the integration of disparate
ontologies. Thus, it provides a formal and sound foundation upon which relations
between ontologies can be studied. Category theory also allows operations over these
relations, as investigated by [Zimmermann et al. 2006], easing the effort necessary to
associate indirectly related ontologies. Analogously, [Seremeti and Kougias 2013]
indicate the composition of morphisms as a powerful tool to easily assimilate a new
ontology into a previously existing network of ontologies.
Nevertheless, this expressiveness power depends on the morphisms used to
express the relations between ontologies. An adequate selection of morphisms must be
made in order to guarantee that the knowledge contained in the ontologies is not
rendered useless by translation failures. This selection can only be done properly if
criteria can be found for the identification of suitable morphisms. In this paper we
discuss a possible set of guidelines to fulfill this role.
The remainder of this paper is organized as follows: Section 2 introduces the
fundamentals of category theory, Section 3 discusses works from the literature that deal
with category theoretical approaches to ontology and their respective morphisms,
Section 4 presents an analysis on what it means for a morphism definition to be good in
a semantic sense and propose criteria for evaluating such definitions, Section 5
discusses challenges that are still faced in the definition of ontology morphisms, and
Section 6 contains a brief conclusion.
2. Category Theory Fundamentals
According to [Adámek et al., 1990], a category is a quadruple A = (O, hom, id, ◦),
consisting of:
1. a class O, whose members are called A-objects,
2. for each pair (A, B) of A-objects, a set hom(A,B), whose members are called A-
morphisms from A to B,
3. for each A-object A, a morphism idA: A→A, called the A-identity on A,
4. a composition law associating each A-morphism f : A→B and each A-morphism
g : B→C to an A-morphism g ◦ f : A→C, called the composite of f and g,
subject to the following conditions:
a) composition is associative, i.e., for morphisms f: A→B, g: B→C and h:C→D,
the equation h ◦ (g ◦ f) = (h ◦ g) ◦ f holds,
b) A-identities act as identities with respect to composition, i.e., for any morphism
f: A→B we have idB ◦ f = f = f ◦ idA,
c) the sets hom(A,B) are pairwise disjoint.
A diagram in a category A is a functor (a structure preserving morphism between
categories) D: I→A. A source for a diagram is a pair (A,(fi)i∈ I I), consisting of an object A
and a family of morphisms fi : A→Ai with domain A, indexed by I. A cone (also called a
natural source) is a source such that for each I-morphism d:i→j the triangle
A→ Di →Dj ←A commutes, i.e., d ◦ fi = fj. A limit is a terminal cone, i.e., it is a cone (A,
(fi)i∈ I I) such that for every other cone (Aj,(fi)ji∈ I I) there is a unique morphism gj : Aj →A in
order that the resulting diagram commutes. A limit for a diagram with two objects and
no morphisms (other than identities) is called a product; a limit for a diagram of the
form B→A←C is called a pullback; and a limit for a diagram of the form A ⇉ B is called
an equalizer.
In category theory, the concept of duality plays an important role. A dual for a
categorical construct is obtained by reversing the domain and codomain of its
morphisms. Thus, for a category A = (O, hom, id, ◦), the dual category of A is the
category Aop = (O, homop, id, ◦op), where homop(A,B) = hom(B,A) and f ◦op g = g ◦ f.
Similarly, the concepts previously introduced also present their own dual constructs.
The duals of sources, cones, limits, products, pullbacks and equalizers are called,
respectively, sinks, cocones, colimits, coproducts, pushouts and coequalizers. These
categorical structures have different interpretations (with similar properties) over
distinct domains.
Figure 1 shows a product (the object A and associated morphisms) and a cone
(A’ and associated morphisms) over a diagram containing only objects A1 and A2 (1), a
coproduct and a cocone over the same diagram (2), a pullback and a cone over the
diagram A1 → A0 ← A2 (3), a pushout and a cocone over the diagram
A1 ← A0 → A2 (4), and an equalizer and a cone over the diagram A1 ⇉ A2 (5).
(1) A’ (2) A1 A2
A
A
A1 A2 A’
(3) A’ (4) A0
A1 A2
A A
A1 A2
A0 A’
(5) A’
A A1 A2
Figure 1. Categorical constructs: product (1), coproduct (2), pullback (3),
pushout (4), and equalizer (5).
3. Related Works
In one of the earliest works to define a category of ontologies with its morphisms,
[Bench-Capon and Malcom 1999] base their specifications on morphisms outlined by
[Goguen and Meseguer 1992] between order-sorted theories, i.e., theories that classify
the objects in the universe in different sorts (types) that may be related by sub-sorting.
These morphisms are monotonic in relation to the ordered sorts, preserving the sub-sort
relations. It should be noted that this definition of morphism does not account for any
non-hierarchical relation in the ontologies. In the same work relations between two
ontologies O1 and O2 are defined as structures composed of a third ontology O and
morphisms xi : O→Oi for i = 1,2. This structure is a cone for the diagram containing
only O1 and O2, therefore a limit for the same diagram (a product) would represent the
maximum possible alignment between the two ontologies.
[Zimmerman et al. 2006] expand on these relations, renaming them as V-
alignments due to their shape and defining the operations of ontology merging (a
pushout over the alignment), alignment composition, union and intersection. The V-
alignment is found lacking the expressive power needed for complex alignments, such
as between ontologies that contain concepts related not by equivalence but through
subsumption relations. Two solutions are then proposed, the first using W-alignments,
which are similar to V-alignments with the addition of a bridge ontology, and the
second with the use of morphisms capable of expressing relations other than
equivalence between concepts, namely strict inclusion, strict containment, disjointness
and overlapping with partial disjointness. A composition table for such relations is also
provided. Similar morphisms were later used by [Euzenat 2008]. V- and W-alignments
were also used in posterior works by [Wang et al. 2008] and [Codescu et al. 2014].
The work of [Cafezeiro and Haeusler 2007] defines morphisms that preserve
relations, both taxonomical (being monotonic in this regard) and otherwise. These
morphisms are understood as mappings between ontologies, and used to build ontology
operations based on categorical constructions: the product and pushout represent, as in
previous works, alignment and merge operations, respectively, while the pullback
denote a search for similarities between two ontologies in the context of a third, broader
ontology. It is also shown how categorical equalizers may be used to hide sensitive
information in an ontology. These operations, with the exception of component hiding,
were used in the work of [Seremeti and Kougias 2013] with the same categorical
interpretations. These morphisms were later expanded by [Cafezeiro et al. 2014] to
include mappings between the axioms of each ontology, where each axiomatic sentence
is translated to its correspondent in the target ontology.
Several works ([Healy and Caudel 2006], [Zimmermann et al. 2006], [Cafezeiro
and Haeusler 2007], [Seremeti and Kougias 2013]) indicate that morphisms are directed
from the less informative ontology to the more informative one. This agrees with the
intuitive notions behind the understanding of categorical product as ontology alignment,
pushout as merge and pullback as similarity search.
Although no morphism definition is provided, [Hitzler et al. 2005] present
possible conditions for suitable morphisms, however without much clarification or
analysis. The presented conditions are:
1. The preservation of class hierarchies,
2. The preservation of types,
3. The taking into account of model-theoretic logical properties, if featured by
the underlying ontology representation language
4. The taking into account of proof-theoretic properties, and
5. The preservation of language classes.
These criteria will be further analyzed in the next section, where we discuss what
makes a good ontology morphism and propose our own set of conditions.
4. What is a Good Morphism?
The main goal that should be achieved by an ontology morphism is relating
meaningfully the concepts present in source and target ontologies. This requires loss of
information not being allowed, or at least severely restricted. A morphism definition
that, for example, allows all concepts in a source ontology to be mapped to a single
concept in the target ontology is not advisable, and also probably not useful at all.
The ability to represent complex relations is also desirable, for it entails greater
expressiveness and semantic power. Without such capabilities, any alignment between
two ontologies with related but not equivalent concepts is doomed to be represented as
some particularly complicated categorical structure, such as the W-alignment. These
intricate structures require greater effort to be operated on: the merge through a W-
alignment, for example, is performed through three successive pushouts with the
requirement of a bridge ontology, while the same operation is performed with a single
pushout over a V-alignment.
From the criteria outlined by [Hitzler et al. 2005] and listed in the previous
section, the condition (1), while satisfied by most ontology morphisms found in the
literature due to monotonicity, is arguably not strong enough. By itself, it does not
prevent multiple concepts of the same hierarchical level in the source ontology from
being projected onto a single concept in the target ontology. This causes information to
be lost when traversing the morphism, contradicting our initial assumptions on what a
morphism should achieve, as well as the commonly found belief that morphisms should
point to the more informative ontology.
Meanwhile, condition (2) needs to be better specified about what is the intended
meaning of “type”. One possible way to define these types is through the use of meta-
properties (that is, properties of properties) such as the ones utilized by [Guarino and
Welty 2009] in their OntoClean methodology, which bases its ontological analysis in
the meta-properties of rigidity, identity, unity and dependence.
Conditions (3) and (5) are dependent on the ontology representation language
used, and thus are not useful guidelines for more general, implementation independent
morphisms. Condition (4) requires a deeper discussion on proof-theoretic properties that
is beyond the scope of this paper. None of these five conditions account for the
representation of complex relations.
In the following subsections, we discuss and propose desirable aspects for good
morphisms, regarding the preservation of information and the capability of representing
complex relations, and present our criteria for the identification of suitable morphisms.
4.1. Information Preservation
As previously discussed, one of the most common expectations on ontology morphisms
is that they are directed from a less informative source ontology towards a more
informative target ontology. This means that a good definition of ontology morphisms
should not allow information to be lost when transitioning between ontologies through
morphisms. This entails the preservation of multiple forms of ontological information.
Bellow, we discuss three of these forms: the preservation of concepts, relations and
meta-properties.
4.1.1. Concept Preservation
If an ontology models two concepts as separate entities, this knowledge should not be
muddled by the morphism. That is, two different concepts in the source ontology must
be mapped to two different concepts in the target ontology, i.e., the morphism should be
injective in regard to concepts. Mathematically, given a morphism f : O→O’ and C1, C2
concepts in O,
f(C1) = f(C2) → C1 = C2
4.1.2. Relation Preservation
Similarly, if an ontology models two relations separately, the ontology morphism
should translate the relations from the source ontology to different relations in the target
ontology, even if both relations share domains and codomains. That is, given a
morphism f : O→O’ and two relations R1, R2 in O,
f(R1) = f(R2) → R1 = R2
Additionally, morphisms should not confuse domains nor codomains of
relations. Thus, given a morphism f : O→O’, a relation R relating two concepts C1 and
C2 in O and a relation R’ in O’ such that f(R) = R’,
R(C1, C2) → R’(f(C1), f(C2))
4.1.3. Meta-property Preservation
While concept and relation preservation should cover a great part of the information
contained in ontologies, there are still pieces of ontological knowledge that could be
confused if morphisms are modeled without additional care. This is the case of
ontological arranges that are similar structurally, but carry deeply different semantic
meanings. One example of such structure (shown in figure 2) would be the relations
between a Car and the Factory where it was built and a Student and the University
where he or she studies. While both the Car and the Student depend respectively on the
Factory and on the University, this dependence takes a different form for each of these
cases. The Student’s dependence to the University is a relational one, i.e., the Student
can only exist as such while there exists a University to which he is related. Otherwise,
the Car’s dependence to the Factory is purely historic, meaning that for the Car to exist,
there once must have existed a Factory to build it, but if the Factory ceases to exist in
any posterior moment it makes no difference to the existence of the Car [Guizzardi and
Wagner 2010].
f
University Factory
in
led f in
r ol i lt
en bu
f
Student Car
Figure 2. A morphism f relating concepts Student and Car
Another difference between the two concepts is due to the meta-property of
rigidity: the concept Car is rigid, while Student is not. This means that a Car cannot
cease to be a Car without ceasing to exist, while someone who is a Student can simply
terminate their relation to the University and stop being a Student while keeping its
existence.
Mismatches such as the one presented here can be avoided with the use of
morphisms that preserve meta-properties of concepts.
4.2. Representation of Complex Relations
As previously noted, the limitation of associating concepts only through equivalence
relations results in additional efforts when dealing with ontologies that do not share
equivalent concepts, but contain ones that are otherwise related. [Zimmerman et al.
2006] perceived this issue and proposed a solution through the expansion of the
definition of morphisms to accept also relations of strict inclusion, strict containment,
disjointness and overlapping with partial disjointness.
This solution, nevertheless, is limited only to subsumption relations, which,
despite covering the main taxonomical backbone upon which ontologies are built,
disregards many other possible relations. It would be desirable to define morphisms that
deal with mereological, dependence or even temporal relations, in addition to the
already covered subsumption. For this, a set of relations must be chosen to be
represented by the morphism, and the composition operation needs to be defined
between each of these relations.
4.3. Criteria for Suitable Morphisms
The reasoning presented thus far can be summed up by five rules that should ideally be
satisfied by adequate morphisms:
1. Concepts should be associated injectively;
2. Relations should also be associated injectively;
3. Domains and codomains of relations should be preserved;
4. Meta-properties of concepts should be preserved;
5. Morphisms should be able to represent relations other than equivalence between
concepts.
The inspected morphisms do not fulfill most of these conditions. In particular,
conditions (1), (2) and (4) are not satisfied in any of the reviewed works. Condition (3)
is completely met only in the work of [Cafezeiro and Haeusler 2007] while partially met
in the remainder works, where it is only true for taxonomic relations. Condition (5) is
contemplated only by [Zimmermann et al. 2006] and [Euzenat 2008], but even in these
works the expressiveness of morphisms has been expanded only to subsumption
relations.
5. Future Challenges
Though many works in the area have been published, some challenges still remain to be
faced when defining ontology morphisms. The first and most evident is relative to the
definition of ontology on which the morphisms are based. Many of the works discussed
here ([Bench-Capon and Malcom 1999], [Healy and Caudel 2006], [Cafezeiro and
Haeusler 2007]) provide their own definitions of ontology and build their morphisms
accordingly. These definitions are not necessarily compatible; hence the same is true for
the morphisms. Additionally, these ontology definitions may not be fully reconcilable
with available ontology representation languages.
Another challenge refers to these languages. Apart from the two conditions
presented by [Hitzler et al. 2005] that explicitly deal with ontology representation
languages, the problem surfaces also in relation to meta-property preservation, since it
depends on the support of said meta-properties being provided by the underlying
representation language.
A third challenge, and maybe of even greater concern, is related to another
important yet often overlooked part of ontologies, the axioms. For an ontology
morphism definition to be complete, it is necessary for it to describe how the morphisms
deal with these axioms, yet such descriptions are seldom found in the literature.
Finally, there remains an open problem on the representation of complex
relations, that is, the finding of a finite set of relations to be represented by morphisms
and the definition of all possible compositions between them. A compromise must be
reached, since the morphism expressiveness grows with the number of relations that can
be represented, but so does the work needed to investigate the compositions between
them.
6. Conclusion
While many works proposing category theoretical approaches to ontologies can be
found in the literature, the studies on establishing meaningful morphisms have been
scarce and rare. Several works have been presented which defined morphisms for
ontologies, and a discussion was made on criteria for evaluating their semantic
soundness and expressiveness. We have proposed five conditions to be met by suitable
ontology morphisms. None of the morphisms analyzed fulfills satisfactorily these
conditions.
Thus, much work is still needed on the search for meaningful ontology
morphisms, and this work will face challenges related to the chosen definition of
ontology, the underlying ontology representation language, how ontology axioms are to
be dealt with and the composition rules on ontology relations.
References
Adámek, J., Herrlich, H., and Strecker, G. E. (1990). Abtract and concrete categories:
The joy of cats. John Wiley & Sons, Inc., New York.
Bench-Capon, T. and Malcolm, G. (1999). Formalising ontologies and their relations. In
International Conference on Database and Expert Systems Applications, pages 250-
259. Springer.
Berners-Lee, T., Hendler, J., and Lassila, O. (2001) The semantic web. Scientific
American, 284(5):34-43.
Cafezeiro, I. and Haeusler, E. H. (2007). Semantic interoperability via category theory.
In Tutorials, posters, panels and industrial contributors at the 26th international
conference on Conceptual modeling-Volume 83, pages 197-202. Australian
Computer Society, Inc.
Cafezeiro, I., Viterbo, J., Rademaker, A., Haeusler, E. H., and Endler, M. (2014).
Specifying ubiquitous systems through the algebra of contextualized ontologies. The
Knowledge Engineering Review, 29(2):171-185.
Codescu, M., Mossakowski, T., and Kutz, O. (2014). A categorical approach to
ontology alignment. In Proceedings of the 9th International Conference on Ontology
Matching-Volume 1317, pages 1-12.
Euzenat, J. (2008). Algebras of ontology alignment relations. In International Semantic
Web Conference, pages 387-402. Springer.
Goguen, J. A. and Meseguer, J. (1992). Order-sorted algebra I: Equational Deduction
for multiple inheritance, overloading, exceptions and partial operations. Theoretical
Computer Science, 105(2):217-273.
Guarino, N. and Welty, C. A. (2009). An overview of ontoclean. In Handbook on
ontologies, pages 201-220. Springer.
Guizzardi, G. and Wagner, G. (2010). Using the unified foundational ontology (UFO) as
a foundation for general conceptual modeling languages. In Theory and Applications
of Ontology: Computer Applications, pages 175-196. Springer.
Healy, M. J. and Caudell, T. P. (2006). Ontologies and worlds in category theory:
implications for neural systems. Axiomathes, 16(1-2):165-214.
Hitzler, P., Krötzsch, M., Ehrig, M., and Sure, Y. (2005). What is ontology merging? In
American Association for Artificial Intelligence.
Seremeti, L. and Kougias, I. (2013). Networks of aligned ontologies through a category
theoretic approach. Journals of Emerging Trends on Computing and Information
Sciences, 4(10).
Wang, Q., Rong, L., and Yu, K. (2008). A pushouts based knowledge merging method:
Knowledge reorganization in emergency decision-making support. In 4th
International Conference on Wireless Communications, Networking and Mobile
Computing, 2008. WiCOM’08, pages 1-4. IEEE.
Zimmermann, A., Krötzsch, M., Euzenat, J., and Hitzler, P. (2006). Formalizing
ontology alignment and its operations with category theory. Formal Ontology in
Information Systems: Proceedings of the Fourth International Conference.