An Adaptive Multi-Agent System for Ontology Co-evolution

Souad Benomrane

, Zied Sellami

, Mounir Ben Ayed

and Adel M. Alimi

REGIM-Lab.: REsearch Groups in Intelligent Machines, University of Sfax, ENIS, BP 1173, Sfax, 3038, Tunisia

LIP6 - Equipe ACASA, 4 Place Jussieu, Boite Courrier 169, 75252 Paris Cedex 05, France

Keywords:

Dynamic Ontology, Co-Evolution, Adaptive Multi-Agent System.

Abstract:

A dynamic ontology evolution reﬂects the ontology adaptation, to a set of changes and their propagation to the

other dependent components, to ensure its consistency. This process needs a frequent involvment of the user

(ontologist), which is a complex and time consuming task. As a solution, in this paper we present an extension

of an ontology evolution tool called DYNAMO MAS based on an adaptive multi-agent system (AMAS). We

improve agents by adding new behaviour to adapt to ontologist actions in order to improve the proposals

already made and to propose others.

1 INTRODUCTION

The number of ontologies that are developed and used

for various applications are in increase. The major

problem faced in ontologies is their change or their

evolution. Ontologies can change for various rea-

sons, for example when user needs change or when

the knowledge of the domain evolve. Sometimes, the

original representation needs some conceptualization

adjustment.

Ontologies must be modiﬁed regularly while

maintaining their coherence, to adapt to the evolution

of ﬁeld to which they relate. Thus, (Stojanovic, 2004)

deﬁnes ontology evolution as a timely adaptation to

the arisen changes and the consistent propagation of

these changes to dependent artefacts.

In the litterature, several approaches have been

proposed to manage the process of ontology evolu-

tion, (Stojanovic, 2004) (Klein, 2004) (Luong, 2007)

(Djedidi and Aufaure, 2010) (Tissaoui et al., 2011)

(Sellami et al., 2012). All these proposals are based

on the approach of Stojanovic (Stojanovic, 2004),

which is the ﬁrst that proposed a process of ontology

evolution. It contains six steps: (1) change capturing,

(2) representation phase, (3) semantics of change, (4)

the implementation of change, (5) change propaga-

tion and (6) the validation phase. Throughout these

steps, the ontologist was involved in every stage. It

is the ontologist who chooses to apply the changes in

the ontology, placing the new elements (concepts, re-

lations), etc.

To reduce this frequent involvement of ontologist,

(Ottens et al., 2009), (Sellami et al., 2012) propose

an approach and a system based on an adaptive multi-

agent system (AMAS) called DYNAMO MAS which

automates three steps of stojanovic process (change

capturing, representation phase, implementation of

change). The role of ontologist in ontology evolution

is fundamental. Thus, we must take into considera-

tion its interventions. (Sellami et al., 2012) developed

a system which allows to answer these several interac-

tions with ontologist and one of its components is DY-

NAMO MAS which includes the ontology that will

be proposed to the ontologist. Its inputs are candidate

terms and lexical relations obtained from text analy-

sis and an OWL ontology. It provides as output an

ontology expressed in OWL. The ontologist gives his

reaction by approving, moving, or rejecting the terms

and concepts proposed by the AMAS.

Our contribution in this work is an extension ofthe

approach of DYNAMO MAS (Sellami et al., 2012) by

adding to the agents new behaviour to adapt and learn

from ontologist feedback. His reaction is considered

as perturbation by the DYNAMO MAS. As a feed-

back, it reorganizes the agents and produces another

ontology proposition. This interactive process is re-

peated until a satisfactory state of the ontology is ob-

tained. The originality of our system is to ensure the

consistency of an ontology without the frequent in-

volvement of the ontologist. It is an entire automatic

adaptation instead of (Tissaoui et al., 2011) approach

where the ontologist is involved in choosing the rele-

vant scenario to apply the list of changes. In our ap-

proach, the ontologist only acts to give his desire for

216

Benomrane S., Sellami Z., Ben Ayed M. and Alimi A..

An Adaptive Multi-Agent System for Ontology Co-evolution.

DOI: 10.5220/0005257602160221

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 216-221

ISBN: 978-989-758-073-4

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

a proposal.

This article is organised as follows. Firstly, we

present an overview of DYNAMO MAS and show

our motivation to extend it. Secondly, we detail our

extension in DYNAMO MAS. Finally, we conclude

and plan some perspectives.

2 DYNAMO SYSTEM OVERVIEW

2.1 Ontology: A Self-organizing

Multi-Agent System

DYNAMO system is an Adaptive Multi-Agent

System (AMAS) based tool, supporting the co-

construction and the evolution of a Terminological

and Ontological Resources (TOR) to index docu-

ments and to allow semantic information retrieval. A

TOR (called ontology in the rest of the paper) con-

tains a set of domain concepts and a set of associated

terms. Every term denotes at least one concept. DY-

NAMO system has a corpus of documents and even

an ontology in OWL as input and generates an ontol-

ogy as output.

The DYNAMO system includes three compo-

nents: a corpus analyser, a MAS and a proposal man-

ager (Fig. 1).

Figure 1: The general architecture of DYNAMO system

(Sellami et al., 2013).

• The corpus analyzer: The goal of the corpus an-

alyzer is to prepare inputs for the MAS. It runs

the YaTeA term extractor (Sellami et al., 2013), a

lexical relation generator, and a term and lexical

relations selector. It outputs (Ti, Rel, Tj) triplets

where Ti and Tj are candidate terms and/or terms

(if the term belongs to the ontology) and Rel is a

lexical relation label;

• The MAS: it includes the ontology that will be

proposed to the ontologist. Its inputs are the

triplets provided by the corpus analyzer and an

OWL ontology. It provides as output an OWL on-

tology.

Every agent has a relevance value which is the

maximum conﬁdence score of all its relations with

its neighbors. During its life cycle, the agent com-

putes its relevance value (Sellami et al., 2013). If

it exceeds the proposal threshold (between 1 and

10), the agent proposes itself to be part of the on-

tology by sending a request to the proposal man-

ager;

• The proposal manager: it enables the ontologist to

both visualize the ontology and the MAS propos-

als and to interact with the MAS.

2.2 Agent Behaviour

DYNAMO MAS consists of two types of agents: (i)

term agents represent candidates for the terminolog-

ical part of the ontology and (ii) the concept agents

represent candidates for the conceptual part of the on-

tology. The relations of the ontology as well as the

lexical relations are part of an agent knowledge. A set

of interactions between these agents enabling them to

ﬁnd the right position in the MAS organization.

Each agent acts in the MAS according to two types of

behaviours:

• Nominal behaviour: it represents the basic algo-

rithm allowing term agents and concept agents to

move throw the MAS to retrieve their appropriate

position (Sellami et al., 2011);

• cooperative behaviour: it represents the coopera-

tive agents aptitude to resolve some exception sit-

uations called Non Cooperative Situations (NCS).

Thanks to this behaviour, the AMAS ensures that

it will converge for a solution. Thus, this coop-

erative process will put an end to a set of infor-

mations exchange to ﬁnally reach a stabilization

(Sellami et al., 2011).

Obtained results (Sellami and Camps, 2012) show

that the DYNAMO MAS is a very useful tool to the

construction and the evolving of an ontology. How-

ever, these results present two key issues that should

be solved so that the DYNAMO performs better:

(i) The term and concept agents are missing additional

adaptive behaviour to react to ontologist actions. To

ﬁx it, we are supposed to enhance the behaviour of the

AnAdaptiveMulti-AgentSystemforOntologyCo-evolution

217

agents (nominal and cooperative) already developed

and implemented in DYNAMO MAS (Sellami et al.,

2013). We are adding these agents, adaptive skills, to

detect its uselessness, to avoid the useless and wrong

proposals and to propose others.

(ii) The time taken to automatically evolve an on-

tology is longer than necessary to manual mainte-

nance. This is due to the response time during val-

idation, deletion or modiﬁcation of a proposal using

PROMPTDIFF tool. We believe that optimizing the

graphical interface between user and system, the time

of the interaction with the ontologist will be consider-

ably shortened.

The next section presents our proposed adaptive be-

haviour (an extension of DYNAMO MAS algorithm

(Sellami et al., 2011)).

3 THE DYNAMO MAS

EXTENSION

In this paper, we are focusing in the two latest com-

ponents of DYNAMO system (the MAS and the pro-

posal manager). First, the ontologist shows his reac-

tion towards the evolved ontology (validates, reﬁnes,

rejects, adds others terms or concepts ...). These ac-

tions have an additional effects on the ontology ele-

ments (concepts, terms and/or relations). Secondly,

the main part, MAS has to consider the actions of on-

tologist and learn from this feedback and propose a

new ontology (new organization of MAS. This pro-

cess is iterative until we achieve a consistent ontology

(no more suggestions in the ≪proposal manager≫).

3.1 Types of Evolution Changes and

their Consequences on the Ontology

To improve the different suggestions of DYNAMO

MAS, agents behaviour are enhanced to react locally

to a set of changes made by the ontologist (accept,

reject, move, delete, create, split, merge, group... ).

Each change type can generate additive changes, by

adding new elements (concepts, terms and/or rela-

tions) to the ontology without affecting the existing

ones, and substractive changes by deleting some ele-

ments (Stojanovic, 2004).

Ontologist feedback can be a set of reactions towards

the system proposals or some personal suggestions

speciﬁc to him. Two kind of changes can be distin-

guished: elementary changes and composite changes.

Stojanovic (Stojanovic, 2004) deﬁnes an elementary

change as an ontology change that modiﬁes (adds or

removes) only one entity of the ontology model and a

composite change as a change of ontology that can be

decomposed into several elementary changes.

These changes require various ontology evolution

strategies that depend on several criteria. The role of

our AMAS is to determine automatically the appro-

priate strategy by adding to the agents new behaviour

to adapt and learn from ontologist feedback. His re-

action is considered as perturbation by the DYNAMO

MAS. As a feedback, it reorganizes the agents and

produces another ontology proposition. This interac-

tive process is repeated until a satisfactory state of the

ontology is obtained.

Our AMAS is based on the declarative approach of

Stojanovic (Stojanovic, 2004). This approach does

not specify in advance the possible strategies to re-

solve change operations. The ontologist speciﬁes

declaratively his desire (what) and the AMAS exe-

cutes the appropriate evolution strategy (how) to meet

the ontologist needs. The different types of evolu-

tion changes, considered in our approach, are listed

and classiﬁed by elementary and composite changes

in two tables as shown below:

Table 1: Term and Concept agents adaptive behaviour to

ontologist elementary changes.

Term Agent adaptive be-

haviour

Concept agent adaptive be-

haviour

Adaptation to acceptance or

rejection

Adaptation to acceptance

Adaptation to new term cre-

ation

Adaptation to rejection

Adaptation to removal Adaptation to renaming

Adaptation to removal

Adaptation to sub-concept

creation

Table 2: Term and Concept agents adaptive behaviour to

ontologist composite changes.

Term agent adaptive be-

haviour

Concept agent adaptive be-

haviour

Adaptation to moving Adaptation to moving

Adaptation to merging

Adaptation to renaming

Adaptation to split

Adaptation to grouping

3.2 Agents Adaptive Behaviour to

Ontologist Feedback

DYNAMO MAS provides as output an ontology. The

ontologist gives his reaction by applying one of the

different evolution actions enumerated above. Our

goal is to improve the algorithm of term and con-

cept agent by adding new behaviour to adapt and learn

from ontologist feedback. His reaction is considered

as perturbation by the DYNAMO MAS. As a feed-

back, AMAS reorganizes the agents, improves the

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

218

proposals already made and proposes others in order

to generate a new ontology proposition. This interac-

tive process is repeated until a satisfactory state of the

ontology is obtained.

In total, we have 14 agent behaviours classiﬁed ac-

cording to two criteria: Agent type (term or con-

cept) and ontologist evolution action type (elemen-

tary actions or composite actions). In this paper, we

present 4 examples of agent behaviours description

that matches these criteria.

3.2.1 Term Agent Adaptive Behaviour to

Ontologist Elementary Changes:

Acceptance or a Rejection

To adapt to ontologist acceptance or rejection of a

term, we enhance the term agent with the following

behaviours: When a term agent is accepted or rejected

by the ontologist, it informs its neighborhood, not yet

validated, by sending a message. This message al-

lows them to recompute their relevance value. For

example, when the term agent Problem is accepted or

rejected, it informs the term agent Problem Failure

that has been accepted or rejected ❶ (Fig. 2). Then,

Figure 2: Term agent adaptation to an acceptance or a re-

jection.

it recalculates its relevance. If it exceeds the proposal

threshold and it respects the Term Proposal Condition

(TPC), it proposes itself to be part of the ontology.

TPC: a term agent can not be proposed if the concept

that it denotes has not be proposed before it. When

a concept agent is rejected then the term agents, that

have not yet proposed, wait until the concept agent

proposes itself again.

3.2.2 Term Agent Adaptive Behaviour to

Composite Changes: Moving

When a term agent is moved by the ontologist, it in-

forms its neighborhood, not yet validated, by sending

a message. For example, if the term agent (Data Fail-

ure) is moved, then it sends a message to the concept

agent (Data Failure) and the term agent (Problem Fail-

ure).

The term agent (Data Failure) moves to the concept

agent (Failure) to denote it ❶ (Fig. 3). If the concept

agent (Data Failure) is already proposed then it will

be removed automatically.

If the term agents (Data Failure) and (Problem Fail-

ure) are related with a synonymy relation(Data Fail-

ure target)then the term agent (Problem Failure) com-

putes its relevance value. If the moving of (Data Fail-

ure) to (Failure) increases the relevance of the term

agent(Problem Failure) then (Problem Failure) moves

to the concept agent (Failure) ❷.

If there is hyperonymy relation between (Data Fail-

ure) and (Problem Failure), then the term agent (Prob-

lem Failure) sends a request to the concept agent

(Failure) asking for its parent concept agent ❷. The

concept agent (Failure) processes the request and no-

tiﬁes the term agent (Problem Failure) by a mes-

sage containing his parent concept agent (Default)

❸. (Problem Failure) sends a request to (Default) for

establishing denotation link . (Default) accepts and

sends a notiﬁcation to (Problem Failure). Therefore,

(Problem Failure) moves to (Default) to create a de-

notation link with it ❹. If the concept agent (Problem

Failure) is already proposed then it will be removed

automatically.

Figure 3: Term agent adaptation to moving.

3.2.3 Concept Agent Adaptive Behaviour to

Elementary Changes: Removal

When the ontologist proposes to delete a concept

agent, it informs its neighnorhood, not yet validated,

by sending a message. This message allows them to

recompute their relevance value.

For example, if the concept agent (Failute) has to be

removed, it informs its neighnorhood, the term agent

(Failure), its sub-concept agents (Data Failure) and

(Problem Failure) and the concept agent (Exception)

and it proposes a new parent concept agent (Default)

❶ (Fig.4). The concept agents (Data Failure) and

(Problem Failure) and the term agent (Failure) recom-

pute their relevance value. If it exceeds the threshold

AnAdaptiveMulti-AgentSystemforOntologyCo-evolution

219

Figure 4: Concept agent adaptation to removal.

proposal, then, (Data Failure) and (Problem Failure)

send a request for establishing an (is-a) relation with

(Default) and the term agent (Failure) sends a denota-

tion request to (Default) ❷. The latter accepts and no-

tiﬁes them. Therefore, (Data Failure), (Problem Fail-

ure) and the term agent (Failure) move to (Default).

The concept agent (Failure) has no more a term to de-

notate it. Thus, it disappears from the AMAS. Else,

the conecpt agent (Failute) will disappear with their

sub-concept agents (they have no more parent con-

cept agent) from the AMAS. If the relevance score

of theconcept agent (Exception) exceeds the threshold

proposal, then it can propose itself to the ontologist.

3.2.4 Concept Agent Adaptive Behaviour to

Composite Changes: Merging

When the ontologist proposes to merge a concept

agent with a validated one, the AMAS reacts and

adapts to this action by running a set of operations.

For example, when the ontologist proposes to merge

the concept agent (Exception) with the validated

conecpt agent (Failure), (Exception) informs its sub-

concepts agents(DB Exception) and (System Excep-

tion) and the term agent (Exception Error) by send-

ing a message containing the new suggested parent

concept agent (Failure) ❶ (Fig. 5). (DB Exception),

(System Exception) send a request message to (Fail-

ure) for establishing an (is-a) relation and (Exception

Error) sends a denotation request ❷. If their new rel-

evance values(with the new parent concept) increase,

then the concept agent (Failure) notiﬁes them by send-

ing a message of acceptation ❸. Therefore, (DB Ex-

ception), (System Exception) and (Exception Error)

move to (Failure).

If (Exception) has no more sub-concepts or term

agents then it disappears automatically from the

AMAS else it will be removed, necessarily, by run-

ning the removal concept operation (detailed in sec-

tion Adaptation to concept removal).

4 CONCLUSION

In this paper we present an automatic ontology evo-

lution tool called DYNAMO MAS. This tool is based

on an adaptive multi-agent system. First, DYNAMO

MAS evolved an ontology in OWL from text. Sec-

ondly, it reacts to ontologist feedback and produces

new ontology evolution proposal. We showed in Sec-

tion 3 the different agents behaviour against elemen-

tary and composite changes. We noticed that the

adaptive skills we added to term and concept agents

allow them to detect the uselessness of some propos-

als, to avoid the useless and wrong ones and to pro-

pose others.

The originality of our work is that we exploit the on-

tologist feedback to improve the future propositions

and reduce his frequent involvement.

To validate the efﬁciency of DYNAMO MAS exten-

sion, we are currently working in three tasks:

• Implementing the adaptive behaviour skills of our

term and concept agents;

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

220

Figure 5: Concept agent adaptation to merging.

• Designing and developinga new DYNAMO MAS

graphical interface between ontologist and MAS

to enhance the time of the interaction and make it

shortened;

• Testing the system with different ontologies in

french and english languages.

ACKNOWLEDGEMENTS

The authors would like to acknowledge the ﬁnancial

support of this work by grants from General Direction

of Scientiﬁc Research (DGRST), Tunisia, under the

ARUB program.

REFERENCES

Djedidi, R. and Aufaure, M.-A. (2010). Onto-evo a l an on-

tology evolution approach guided by pattern modeling

and quality evaluation. In Foundations of Information

and Knowledge Systems, pages 286–305. Springer.

Klein, M. C. A. (2004). Change management for distributed

ontologies.

Luong, P. H. (2007). Gestion de l’´evolution d’un Web

s´emantique d’entreprise. PhD thesis,

Ecole Nationale

Sup´erieure des Mines de Paris.

Ottens, K., Hernandez, N., Gleizes, M.-P., and Aussenac-

Gilles, N. (2009). A multi-agent system for dy-

namic ontologies. Journal of Logic and Computation,

19(5):831–858.

Sellami, Z., Aussenac-Gilles, N., Gleizes, M. P., and

Camps, V. (2012). Dynamo, un outil de construction

et d’´evolution d’ontologies `a partir de textes. Tech-

nique et Science Informatiques, 31(1):97–124.

Sellami, Z. and Camps, V. (2012). Evaluation of a multi-

agent system for the evolving of domain ontologies

from texts. In Advances on Practical Applications

of Agents and Multi-Agent Systems, pages 169–179.

Springer.

Sellami, Z., Camps, V., and Aussenac-Gilles, N. (2013).

Dynamo-mas: a multi-agent system for ontology evo-

lution from text. Journal on Data Semantics, 2(2-

3):145–161.

Sellami, Z., Camps, V., Aussenac-Gilles, N., and Rouge-

maille, S. (2011). Ontology co-construction with

an adaptive multi-agent system: Principles and case-

study. In Knowledge Discovery, Knowlege Engi-

neering and Knowledge Management, pages 237–248.

Springer.

Stojanovic, L. (2004). Methods and tools for ontology evo-

lution.

Tissaoui, A., Aussenac-Gilles, N., Laublet, P., and Her-

nandez, N. (2011). Evonto: ´evolution de ressource

termino-ontologique pour lannotation s´emantique.

Journ´ees Francophones sur les Ontologies (JFO

2011), pages 22–23.

AnAdaptiveMulti-AgentSystemforOntologyCo-evolution

221