A COMPOUND STRATEGY FOR ONTOLOGIES COMBINING

Dominique Renaud, Cecilia Zanni-Merk and François Rousselot

LGeCo, INSA Strasbourg, 24 boulevard de la Victoire, Strasbourg, France

Keywords: Expert systems, Ontology engineering, Intelligent multi-agent systems, Ontology sharing and reuse,

Ontology matching and alignment.

Abstract: This article reports our work on a strategy for results aggregation coming from a multi agent system. Each

set of results is related to a specific ontology. The application is the organisation analysis of Small and

Medium Enterprises (SME). In this context, different Knowledge Bases (KB) are used. Depending on their

origin, the different KB may be close, complementary and sometimes contradictory. The proposed approach

uses a strategy based on two key ideas. The first one is general and aims at selecting a combining method of

ontologies and the second one is focused on the selection and combining of sub-parts of ontologies. The

combination of these two strategies should improve the understanding of the results produced by the multi

agent system.

1 INTRODUCTION

In recent years, software agents (MAS) have become

a well studied and frequently applied technical

implementation for distributed systems. In this work,

a MAS is used as a multi-experts eco-system for

analysis and diagnostic of Small and Medium

Enterprises (SME) organization. The targeted

system analyzes the management activities of an

enterprise and provides suggestions to help address

those areas in which it is less successful. To

introduce the notion of multiple point of views, each

agent is associated to a particular knowledge base

(KB) and ontology. This situation implies the

production of many pieces of results related to a

limited topic. To be well understood by an external

user, all the produced results must be aggregated by

topic. The aggregation of results raises some issues

about combining multiple ontologies.

Section 2 describes the context of these works

and the following sections present the problem of

aggregation of results associated with related

ontologies (section 3). A first draft solution is

proposed in the form of an assembled ontology in

section 4. Finally, section 5 presents our conclusions

and perspectives of future work.

2 CONTEXT

MAEOS is a project on the modelling of the support

to the organizational and strategic development of

SMEs.

The main objective of MAEOS is to improve the

efficiency and performance of business advice to

SMEs. To do this, the main part of this work is

devoted to the modelling of knowledge coming from

management sciences.

Unlike the current trends, which are to create a

homogeneous KB covering the domain of a

problem, our choice is different. It is to keep to a

maximum the plurality of each KB with their field of

interest, constraints and richness. The interest and

the difficulty of this project are to combine a large

variety of sources and origins of knowledge around

SME topics.

The targeted knowledge is separated into two

kinds of expertise. On the one hand, the theoretical

knowledge in the area of change in SMEs

(organization, strategy,...) are used as core models

and on the other hand, expert knowledge

accumulated during practice is used as

complementary knowledge.

The main outputs of this project are a set of

methods and software tools for analysis and

diagnosis of SMEs. The software tools must be able

to evolve according to the state of the art on SMEs

and, in particular, their administrative or legal

environments. In addition, they must also be able to

reflect the richness and contradictions inherent to the

models coming from management sciences.

To achieve these objectives, a multidisciplinary

team was created. Three main research areas are

represented: artificial intelligence, software

engineering and management sciences. This work

involves, therefore, two major points consisting of

200

Renaud D., Zanni-Merk C. and Rousselot F. (2009).

A COMPOUND STRATEGY FOR ONTOLOGIES COMBINING.

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development, pages 200-205

DOI: 10.5220/0002299802000205

 SciTePress

the building of KBs related to SMEs and the

implementation of an expert system based on

software agents. This part of the project uses three

technologies: the ontologies are described with

OWL-DL (OWL-DL, 2004), OMV is used for meta-

data (OMV, 2007) and the MAS is programmed in

Java.

Knowledge bases, in progress, are designed to

cover a significant portion of aspects relating to

organization and managerial behaviours of SMEs.

An ontological study was conducted with the

aim of providing the theoretical foundations

necessary for the development. Several ontologies

have been studied. Our main sources were the

ontology MASON (Lemaignan, 2006), TOVE (Fox,

1992, 1998) and ENTERPRISE (Uschold, 1998).

Each knowledge base which we are building is

divided into three parts: an ontology, a collection of

best practices with facts and/or rules, and meta-data.

Later in the project, other KBs will be added to the

existing ones. All of them will build sets of SMEs

modular models.

The expert system is intended to use the different

KBs. It will be an ecosystem of reactive agents

(Figure 1). At present, the ecosystem is written in

Java and is not complete. Indeed, no direct

communication exists between agents. All

exchanges are made through a common bag. An

agent is associated with a particular KB. Therefore,

all agents are characterized by a knowledge field

through an ontology, a collection of facts and/or

rules and a set of meta-data.

Model

Profil



Query

Bag

Use

IsBoundto

gents

Knowledge

Bases

OMV

OWL‐DL

Facts/Rules

Enrich

Figure 1: The multi-agents system.

Each agent picks information up in the common

bag. It accomplishes its deduction tasks. At the end,

it adds the results to the bag. an agent is triggered

when there are pieces of information in the common

bag matching some of its characteristics. The

process is considered as finished when the agents

have nothing new to add to the common bag.

3 AGGREGATION OF RESULTS

BASED ON ONTOLOGIES

In the context of this ecosystem, our research

activities focus on the post-processing of results.

The main issue is related to the aggregation of

results coming from related domain ontologies.

Indeed, the KBs used in this project are limited to

SMEs management. Depending on their origin, their

contents may be close, complementary and

sometimes contradictory.

The current trend is to create a homogeneous

ontology covering the domain of a problem. The

choice of this project is different. It is to keep to a

maximum the plurality and the growth potential of

each KB over the process of analysis and diagnosis.

The objectives are to address the constraints, on the

one hand, related to the expression of the richness

and contradictions inherent to the models and, on the

other hand, related to the evolution of a MAS and its

KB.

Our answer is intended to create local ontologies

to the problems. Each set of results from the MAS is

supported by an ontology of its own. The main

approach is to push the ontology combining as far as

possible in the process. This is to keep our goals of

multiple point of views.

3.1 Aggregating Results

The aggregation of results based on ontologies

requires more than a mere correspondence between

terms or parts of models. This is because the

production of results is carried out with several facts

bases and/or rules and because each of these bases is

related to a specific ontology. Each ontology that is

used contains its own taxonomy, roles and axioms

and is built with an intention and a point of view.

This results aggregation must ensure a coherent

semantics. Therefore, it is, at best, the integration of

several ontologies into a new one covering all the

results. Finally, once aggregated, the results must

also be consistent with the facts and rules

implemented by the software agents.

3.2 Combining Ontologies

There are many tools and works on ontologies

combining (Klein, 2001), (Choi, 2006), (Bruijn,

2006), (Flouris, 2007). Four classes of methods are

applicable to the MAEOS problem:

A COMPOUND STRATEGY FOR ONTOLOGIES COMBINING

201

 “Merging” ontologies implies the creation of a

new one by linking up existing ones. Each

important concept is selected, based on its

relevance. There are particular cases, such as

the “integration” that is to complement or

build an ontology with smaller specific

ontologies or the “inheritance” method that

uses the notion of a “is_a” relationship to

merge several ontologies from the most

general to the most specific concept.

 “Mapping” is about building a translation

model, via a bijection or not, among

ontologies. A special case called “refinement”

appears when the atomic concepts of a first

ontology have their equivalent in the non-

atomic concepts of a second one.

 “Alignment” corresponds to a partial or

complete translation of concepts and/or the

possible addition of relations in order to create

a new ontology from several ones. The

operation is called “unification” if the

alignment is done on all the concepts of an

ontology over another.

 “Mediation” is quite a different class. It uses

the idea of negotiation to find the best

building compromise for an ontology from

several ones.

All these methods cannot always be applied in a

systematic and / or automated way. As highlighted

by (Noy, 1999), the intervention of an expert may be

required.

Some methods, such as the alignment is better

suited to a fusion where different ontologies are

complementary or have different semantic levels. It

is necessary to know the criteria for selecting a

combining method as well as the limitations of these

methods.

The combining of several ontologies implies, at

least, the presence of common or relative conceptual

entities in them.

Different criteria can be applied to identify the

similarities between two conceptual entities

(Maedche, 2002):

 The similarity of terms;

 The similarity of properties;

 The similarity of the entities subsuming or

being subsumed.

In real situations, several penalizing cases may

appear at different levels. Disparities in the

definitions may not only arise at the conceptual,

terminological or taxonomy level but also at the

syntactic level. Between two close ontologies, it is

common to have the same term with different

meanings or several terms referencing the same

concept. Depending on the ontology author’s

viewpoint, several definitions may relate to the same

concept. Mismatches among ontologies are

numerous. They are summarized in (Klein, 2001),

(Visser, 1997) and (Hameed, 2004) with a series of

examples (Figure 2).

These differences affect the implementation of

the combining methods. The most extreme case

happens when disjoint ontologies are considered and

makes impossible the application of any combining

method. In the case of close ontologies, a choice

cannot be made if the degree of similarity among

several terms is equivalent (Colomb, 2007).

Ontology

Mismatches

Conceptualisation

Mism atch

Explication

Mismatch

Class

Mismatch

Relation

Mismatch

Categ o rization

Mismatch

Aggregation-level

Mismatch

Structure

Mismatch

Attribute-Assignem ent

Mismatch

Attribute-Type

Mismatch

Concept

Mismatch

Concept & Definiens

Mismatch

Definiens

Mismatch

Term

Mismatch

Concept & Term

Mismatch

Term & Definiens

Mismatch

Figure 2: Ontology mismatches taxonomy.

Next, even if connections are established among

conceptual entities, there is no guarantee that they

will be bijections. Conflicts at semantic level may

also appear. Finally, the difference of granularity

between ontologies can result in the elimination or

aggregation of some entities. It should be noted that

the number of mismatches cases increases when

ontologies are larger. Different ways should be

studied in order to minimise these mismatches.

4 THE AGGREGATION

STRATEGY

4.1 A Compound Strategy

For this project, two solutions are combined to

reduce the incidence of mismatch in ontologies

combining: selecting the combining method and

using small size ontologies.

4.2 Method Selection

In general, choosing a combining method for

ontologies is a critical issue. It becomes even more

problematic if these combinations have to be

KEOD 2009 - International Conference on Knowledge Engineering and Ontology Development

202

performed in an automated way. The use of multiple

ontologies may be reduced to alignment of parts or a

full ontology merging.

A partial alignment may be sufficient in the case

of the use of ontologies on complementary

knowledge domains or of different semantic levels.

Merging is more appropriate if the contents of the

ontologies overlap. Finally, in the case where a

choice can not be operated in an automated way,

there is the mapping solution with a previous work,

or the use of mediation. This last method remains

complex to implement and is now set aside.

The criteria based on the works of (Flouris,

2007) and (Colomb, 2007) allow the following cases

(Figure 3):

 Merging is used when the implemented

ontologies are complementary. For two

ontologies A and B , merging both of them

implies that the two ontologies are treated as

sub-parts of a more comprehensive ontology.

In other words, ontologies have common parts

and distinct ones. The easiest situation is the

merging by “inheritance” when the most

general concepts of one ontology correspond

to more specific concepts of the other.

 Alignment can be achieved when ontologies are

close and do not correspond to the merging

situation.

 Mapping is used in preparation of merging

among multiple ontologies. It is used when the

ontologies do not seem to have clear common

concepts. Although this technique is very

reliable, it requires some previous work.

1 Does mapping exist ?

3 Does overlap exist (modularity) ?

5 Is alignment applicable?

Begin

End

Use fusion

Use Alignement

Use mapping 2

Yes

Figure 3: Combining method algorithm.

Finally, in the context of this research work, if

no choice can be made, ontologies are not combined.

4.3 The Use of Reduced Ontologies

It is not always possible to be in the best situation

for combining ontologies. The size of the ontologies

has an important influence on the possibilities of

combining: big ontologies are more complex to

combine. Indeed, the cases of mismatch are much

more frequent if ontologies are important.

Use of small complementary ontologies can

facilitate the construction of a more comprehensive

one. At best, a solution can be to use modular

ontologies or at least, that are possible to be split.

The adopted strategy consists in selecting only the

necessary concepts for the aggregation of the results.

The objective is to only keep the necessary

knowledge for the interpretation of the results

supplied by the MAS. This is to facilitate the

combining of the ontologies.

The decomposition of ontologies in sub-

ontologies seems to be an attractive possibility.

However, it supposes several assumptions:

 There are ontologies that are modular or

decomposable into partitions

 There exist coherent sub-ontologies

 The number of extracted concepts is sufficient

for the combining of the sub-ontologies

And for our system:

 The results produced by the MAS are relative to

close concepts

 All the ontologies graphs are single “is-a” trees.

It is evident that these assumptions cannot apply

to every combining of ontologies. The context

defined by all the produced results is important. This

context helps collecting close sub-parts of ontologies

around a particular subject.

4.4 Selection of Sub-parts of an

Ontology

The basic idea is to extract, from a set of ontologies,

the smallest consistent sub-ontologies with a

maximum coverage of the concepts used by the

results. For that purpose, an algorithm based on the

properties of partitioning and modularity of graphs is

used.

The algorithm considers an ontology as a

semantic network. It treats the network as a directed

graph that has nodes as concepts and edges as roles

with their properties and their constraints.

It is clear that some characteristics are taken into

account in the handling of these graphs. The

semantic networks are graphs containing a tree

A COMPOUND STRATEGY FOR ONTOLOGIES COMBINING

203

structure related to the taxonomy of the described

subject, relationships related to their constraints

(transitive, symmetrical…) and to their properties

(mandatory, attributes…). Within this framework,

the search for a partition in the graph of a semantic

network respects certain criteria.

These criteria aim at selecting the concepts

intervening in the interpretation of the results and at

only preserving a coherent sub-graph. They are

expressed as:

 A sub-graph must contain all the necessary

concepts to link every part of the result.

 A sub-graph can be extracted if and only if it is

connected to the rest of the graph by incident

edges.

 A sub-graph must preserve the hierarchy

formed by the “is_a” relations.

 Each node must keep its concept definition.

These facts lead to the algorithm in Figure 4.

The first step selects all concepts used by the

results. To maintain the consistency, steps two and

three extend the selection to sibling concepts.

The second step completes the previous selection

with the "is-a" relation tree. The third one adds the

shortest mandatory paths between the selected

concepts.

The main loop, from step four to step eleven,

aims at selecting complementary concepts in relation

with the current selection.

Finally, the secondary loop allows enumerating

and choosing the necessary relations and neighbour

concepts.

The algorithm stops when no concept or no

relationship can be selected.

Yet, this strategy has some limitations. On the

one hand, the extracted sub-ontology can represent

all the ontology in particular cases:

 If the graph is connected or strongly connected.

The high number of edges among nodes

requires the extraction of a bigger sub-graph.

 If the selected concepts belong to a clique

located at the bottom of the “is_a” relations

tree.

 If the useful concepts are distributed in a too

homogeneous way in the graph. The paths

making possible to go from a selected node to

another are then more important.

Begin

((Incidential relation) & (cardinality > 0))

| ((selected concept) & (cardinality > 0)) ?

Yes

2 Select is-a subtrees

3 Select shortest path between concepts

1 Select mandatory concepts related to results

For each local relation

7 Select relation

(unselected concept) ?

9 Select concept

12 Extract sub-ontology

Last relation ?

Last concept ?

Yes

End

For each concept

Figure 4: Sub-ontology selection algorithm.

On the other hand, the selected sub-ontology

does not necessarily contain all the concepts

required to be combined with another ontology. In

that case, this sub-ontology must be completed.

This extraction algorithm allows selecting

partitions of ontologies. It is integrated in the main

aggregation algorithm.

4.5 The Main Algorithm

The strategy of results aggregation aims at

producing concise knowledge and at facilitating its

interpretation. In the MAS ecosystem of MAEOS,

the closeness of the contents of the KBs generates

many similar results. For that purpose, it is

necessary to be able to combine the whole contents

of the bag into groups of homogeneous results.

The proposed algorithm is separated in four steps

(Figure 5).

KEOD 2009 - International Conference on Knowledge Engineering and Ontology Development

204

Sort results by Ontology

Aggregate results by Ontology

Publish to Common Bag

Matches between Ontologies?

Select required concepts

Merge Ontologies

order of relatedness

Begin

End

Yes

Figure 5: The main algorithm.

The first one aggregates results related to the

same ontology. The following one selects the sub-

ontology relative to the results. The third one

combines sub-ontologies and verifies their validity.

And finally, the results are to be aggregated

according to the new ontologies.

5 CONCLUSIONS

In this article, we presented an approach for results

aggregation coming from multiple ontologies. This

approach aims at solving the many limitations

resulting from the use of ontologies whose contents

are closely related.

The suggested strategy is articulated around two

key points: the choice of the combining method and

the partitioning of ontologies.

The first tests carried out showed the interest of

the approach by sub-ontologies. However, the

applied strategies are only efficient on close

ontologies with a simple “is_a” relationship tree

graph and that are slightly connected or modular.

Our next works will be to improve and expand

the selection of sub-ontologies. Indeed, our initial

investigations only apply to simple “is_a”

hierarchies. To be more robust and versatile, the

algorithm must be used on more complex

ontologies.

REFERENCES

de Bruijn J., M. Ehrig, C. Feier, F. Martíns-Recuerda, F.

Scharffe, and M. Weiten, 2006,Ontology Mediation,

Merging, and Aligning,In

Choi N., I.-Y. Song, and H. Han, 2006, A survey on

ontology mapping, In SIGMOD Rec., vol. 35.

Colomb, R. M. and Ahmad, M. N., 2007. Merging

ontologies requires interlocking institutional worlds,

In Appl. Ontol. 2, 1.

Flouris G., D. Manakanat, H. Kondylakis, D. Plexousakis,

and G. Antoniou, 2007,Ontology change:

classification and survey, In The Knowledge

Engineering Review.

Hameed, A., Preece, A., Sleeman, D., 2004, Ontology

reconciliation. In: Steffen, S., Studer, In R. (eds.)

Handbook on Ontologies, Springer, Berlin (2004)

Klein M., 2001,Combining and relating ontologies: an

analysis of problems and solutions, In

IJCAI'01,Workshop on Ontologies and Information

Sharing, G. A. Perez, M. Gruninger, H.

Stuckenschmidt, and M. Uschold, Eds.

Maedche A., Motik B., Silva N. & Volz R., 2002,

MAFRA: a mapping framework for distributed

ontologies, in EKAW’2002, Proceedings of the

International, Springer LNAI 2473.

Noy N. & Musen M. A.,1999, SMART : automated

support for ontology merging and alignment, in

KAW’1999, Proceedings of the Workshop on

Knowledge Acquisition, Modeling and Management.

OMV, 2007, http://omv.ontoware.org/

OWL-DL, 2004, http://www.w3.org/TR/owl-guide/

Visser, P.R.S., Jones, D.M., Bench-Capon, T.J.M., Shave,

1997, M.J.R.: An analysis of ontology mismatches:

heterogeneity versus interoperability. In: AAAI 1997

Spring Symposium on Ontological Engineering,

Stanford.

Lemaignan, S., Siadat, A., Dantan, J.-Y., et Semenenko,

A., (2006) MASON: A proposal for an ontology of

manufacturing domain. Distributed Intelligent

Systems, Collective Intelligence and Its Applications.

DIS 2006. IEEE Workshop 15-16, 195–200.

Fox, M., (1992) The TOVE project: A commonsense

model of the enterprise, industrial and engineering

applications of artificial intelligence and expert

systems. Lecture Notes in Artificial Intelligence (604),

25–34.

Fox, M., et Grüninger, M., (1998) Enterprise modeling. AI

Magazine, 109–121.

Uschold, M., Moralee, M. K. anf S., et Zorgios, Y., (1998)

The enterprise ontology. The Knowledge Engineering

Review 13, 31–89.

A COMPOUND STRATEGY FOR ONTOLOGIES COMBINING

205