DYNAMIC ONTOLOGY CO-CONSTRUCTION BASED ON

ADAPTIVE MULTI-AGENT TECHNOLOGY

Zied Sellami, Marie-Pierre Gleizes, Nathalie Aussenac-Gilles

IRIT (Institut de Recherche en Informatique de Toulouse), Toulouse University

118 Route de Narbonne, F-31062 Toulouse cedex 9, France

Sylvain Rougemaille

UPETEC (Emergence Technologies for Unsolved Problems), 10 avenue de l’europe, 31520 Ramonville, France

Keywords:

Ontology engineering, Multi-Agent System, Knowledge acquisition.

Abstract:

Ontologies have become an important means for structuring knowledge and deﬁning semantic information

retrieval systems. Ontology engineering requires a signiﬁcant effort, and recent researches show that human

language technologies are useful means to acquire or update ontologies from text. In this paper we present

DYNAMO, a tool based on a Multi-Agent System, which aims at assisting ontologists during the ontology

building and evolution processes. This work is carried out in the context of the DYNAMO project. The main

novelty of the agent system is to take advantage of text extracted terms and lexical relations together with some

quantitative features of the corpus to guide the agents when self-organizing. We exhibit the ﬁrst experiment

of ontology building that shows promising results, and helps us to identify key issues to be solved to the

DYNAMO system behavior and the resulting ontology.

1 INTRODUCTION

One way to provide efﬁcient search on a document re-

trieval system is to explicitly state the meaning of doc-

ument contents. On-going research in this area tries

to address the problem by tagging and indexing the

contents of documents thanks to an organised knowl-

edge representation called ontology.

Ontologies are often used to represent a speciﬁ-

cation of domain knowledge by providing a consen-

sual agreement on the semantics of domain concepts,

or an agreement on the concepts required for a spe-

ciﬁc knowledge intensive application. An ontology

also deﬁnes rich relationshipsbetween concepts. It al-

lows members of a community of interest to establish

a shared formal vocabulary. In short, ontologies are

deﬁned as a formal speciﬁcation of a shared concep-

tualisation (Gruber, 1993) where formal implies that

the ontology should be machine-readable and shared

that it is accepted by a human group or community.

Further, it is restricted to the concepts and relations

that are relevant for a particular task or application.

Typically, ontologies are composed of a hierarchy

of concepts the meaning of which is expressed thanks

to their relationships and to axioms or rules that may

constrain the relations or that deﬁne new concepts as

formulas. Concepts may be labelled with terms that

are their linguistic realisations or linguistic clues of

their meaning.

Originally, ontologies were deﬁned as rigid struc-

tures that are supposed to be stable over time. Nev-

ertheless, ontologies may need to evolve because do-

main knowledge changes, users’ needs may be differ-

ent or because the ontology could be used in a new

context or even reused in a new application (Haase

and Sure, 2004). Ontology maintenance may re-

sult difﬁcult especially if their structural semantics is

complex, deﬁning hundreds of concepts and relations.

Ontology engineering is a costly and complex task

(Maedche, 2002). In the last ten years, ontology engi-

neering from text has emerged as a promising way to

save time and to gain efﬁciency for building or evolv-

ing ontologies (Buitelaar et al., 2005). However, texts

do not cover all the required information to build a

relevant domain model, and human interpretation and

validation are required at several stages in this pro-

cess. So ontology engineering remains a particularly

complex task when it comes to the extraction or ob-

servation of both terminological and ontological rep-

resentations from a speciﬁc document corpus.

Sellami Z., Gleizes M., Aussenac-Gilles N. and Rougemaille S. (2009).

DYNAMIC ONTOLOGY CO-CONSTRUCTION BASED ON ADAPTIVE MULTI-AGENT TECHNOLOGY.

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development, pages 56-63

DOI: 10.5220/0002302700560063

 SciTePress

Our motivation is to propose a tool facilitating on-

tology maintenance and evolution by the ontologist.

The principle is to provide a system that automati-

cally proposes solutions to be discussed and evalu-

ated. This system learns from the user’s feedback.

It can be seen as a virtual ontologist that helps the

“real one” to carry out ontology learning and evolu-

tion from text. We call this process a co-construction

or co-evolution.

In this article, we propose DYNAMO, a Multi-

Agent System that supports ontology co-construction

and evolution. After an overview of the DYNAMO

context (section 2), we detail the principles of this

MAS (section 3). In Section 5 we exhibit preliminary

results obtained when building an ontology with DY-

NAMO. These results are discussed in the next sec-

tion. Section 7 presents the improvements we plan to

bring to our work as well as our perspectives.

2 DYNAMIC ONTOLOGY

CO-CONSTRUCTION

2.1 Context

The study of ontology evolution is part of the DY-

NAMO

(DYNAMic Ontology for information re-

trieval) ANR

(Agence Nationale de la Recherche)

funded research project. DYNAMO addresses the im-

provementof semantic information retrieval drivenby

user satisfaction in a dynamic context. One of the

project originality is to take into account the potential

dynamics of the searched document collection, of the

domain knowledge as well as the evolution of users’

needs.

The DYNAMO project aims at proposing a

methodological approach and a set of tools that allow

the deﬁnition and the maintenance of ontological re-

sources from a set of unstructured documents. These

resources are used to facilitate information retrieval

within the corpus by means of a semantic indexing.

Several project partners propose domain speciﬁc

document collections. ACTIA

provides documents

covering the area of automobile diagnosis (automo-

bile components, symptom, engine failure, etc.) and

written in French, while ARTAL

corpus consists in

software bug reports written in English, and the part-

ners in charge of the ARKEOTEK

project are con-

http://www.irit.fr/DYNAMO/

http://www.agence-nationale-recherche.fr/

http://www.ACTIA.com

http://www.ARTAL.fr

http://www.ARKEOTEK.org

cerned by archaeological scientiﬁc research papers

structured as a set of rules. Thus, one point of im-

portance in DYNAMO tools, is the handling of a vari-

ety of heterogeneous document collections among the

projects (as for language support both in French and

English) for the co-evolution of ontological resources.

2.2 The Ontology Model

In DYNAMO, the ontology and its lexical component

form what we call a Terminological and Ontological

Resource (TOR). Such a resource is represented using

the OWL

-based TOR model proposed in (Reymonet

et al., 2007). This model recently evolved to become a

meta-model, where concepts and terms are two meta-

classes adapted from owl:class. In this TOR, model

ontological elements (concepts) are related to their

linguistic manifestations in documents (terms): a term

“denotate” at least one concept. This models form

the core of the DYNAMO project as longs as term in-

stances (which represent term occurrences), concept

instances and relations between instances are used to

represent document annotations.

The problem addressed in this paper is how a

multi-agent system can be used to build and update

a TOR represented with this meta-model and using

documents as information sources.

2.3 The Adaptive Multi-Agent System

(AMAS) Theory

Because of their local computation and openness,

MAS are known to be particularly well ﬁtted to dy-

namic and complex problems. The design of an on-

tology from the analysis of a corpus is an obviously

complex task (as we discussed in the introduction).

The AMAS Theory (Capera et al., 2003) empha-

sises on the cooperation between agents to achieve

global adequacy by the way of self-organisation.

Each agent in the system tries to maintain a coop-

erative state, more precisely, it tries to avoid and re-

pair harmful situations (Non Cooperative Situations).

According to the AMAS principles, the agent co-

operative behaviour ensures that, during times, the

function realised by the system is always adapted to

the problem (functional adequacy). The main idea

of the work presented here is to take advantage of

the AMAS properties to propose an ontology co-

construction system that uses as information sources

the documents as well as the ontologist interactions.

As a result, the DYNAMO MAS proposes some

modiﬁcations to the ontologist because it evaluates

Web Ontology Language

http://www.w3.org/2004/OWL/

DYNAMIC ONTOLOGY CO-CONSTRUCTION BASED ON ADAPTIVE MULTI-AGENT TECHNOLOGY

that these modiﬁcations improve the ontology. The

system also beneﬁts from the ontologist answers to

its proposals (basically acceptance or rejection), it al-

lows to strengthen or weaken the conﬁdence in the

position of the involved agents.

3 PRINCIPLES

DYNAMO is a tool, based on an Adaptive Multi-

Agent System (AMAS) presented in Section 2.3, en-

abling the co-construction and the maintenance of an

ontology starting from a textual corpus and resulting

an OWL ﬁle. DYNAMO is a semi-automatic tool be-

cause the ontologist has to validate, reﬁne or mod-

ify the organisation of concepts, terms and relations

between concepts until it reaches a satisfying state.

Figure 1 gives an overview of the DYNAMO system

components : DYNAMO Corpus Analyzer and DY-

NAMO MAS.

The DYNAMO Corpus Analyzer prepares the input of

the DYNAMO MAS. It contains the Corpus, the Pat-

tern Base, the Candidate Term Base and a set of NLP

tools. Those tools process the lexical relations extrac-

tion mechanism whose result is used to determine po-

tential semantic relations between candidate terms.

The DYNAMO MAS is composed of two agent

types: TermAgent and ConceptAgent which are de-

tailed further in Section 3.3. Thanks to the extracted

relations these agents try to self-organise in order to

ﬁnd their proper location in the TOR hierarchy.

3.1 Syntactic Patterns

Many approaches for ontology learning from text are

based on Natural Language Processing (NLP) tech-

niques. We can quote two main groups: on the one

hand, statistical approaches (Harris, 1968), like clus-

tering, are interested in ﬁnding a semantic interpre-

tation to several kinds of term co-occurrences in cor-

pora; on the other hand, approaches based on linguis-

tics rely on a more or less ﬁne-grained linguistic de-

scription of the language used in text to derive an in-

terpretation at the semantic level. Recent ontology

learning processes combine both approaches (Cimi-

ano, 2006).

For instance, lexico-syntactic patterns can be used ei-

ther for concept or semantic relation extraction, but

what they actually identify in text are terms or lexi-

cal relations (Hearst, 1992). The extraction process

then includes pattern adaptation to the corpus to be

parsed, lexical relations extraction on each document,

phrase interpretation and ﬁnally term and relation ex-

traction (Barri`ere and Akakpo, 2006). An extra step

would be to deﬁne concepts and semantic relations

from those items. Systems like Text-to-Onto (Cimi-

ano and V¨olker, 2005) or OntoLearn (Velardi et al.,

2005) propose a fully automatic run from pattern-

matching to ontology learning, while systems like

Prom´eth´ee (Morin, 1999) and Cam´el´eon (Chagnoux

et al., 2008) support a supervised process when the

ontologist may validate or modify the concepts and

relations proposed by the analysis tool.

In keeping with results established by V. Malais´e

(Malaise, 2005) we have experimentallyobservedthat

a statistical processing is not very effective on small

corpora with little redundancy, which is the case for

the three DYNAMO speciﬁc applications. Not only

the corpora have a relatively modest size (ACTIA cor-

pus: 46000 words, ARTAL’s corpus: 13000 words,

ARKEOTEK corpus: 106000 words), but each doc-

ument is of very short length and deals with a spe-

ciﬁc subject. For all these reasons, we adopted a

pattern-based approach to obtain relevant information

on terms and their relationships, and then to deﬁne

concepts and semantic relations from these evidences.

3.2 Semantic Relations

In DYNAMO, we are interested in four types of lexi-

cal relations:

1. Hyperonymy expresses a generic-speciﬁc relation

between terms. This may lead to deﬁne a class-

subclass (is a) relation between the concepts de-

noted by these terms.

2. Meronymy means a parthood relation between

terms, which may lead to deﬁne a part of seman-

tic relation between concepts, or an ingredient of

relation or domain-speciﬁc adaptations of part-

hood like has members in biology for instance.

3. Synonymy relates semantically close terms that

should denote the same concept.

4. Functional relations: which are any other kind

of lexical relations that will lead to a speciﬁc

set of semantic relations, either general ones like

causes, leads to, ... or task speciﬁc relations like

has fault, is an evidence for or domain speciﬁc re-

lations like has skills in archaeology.

In our system, linguistics manifestation of seman-

tic relations are used by agents as clues for self-

organisation. We call them triggers.

3.3 Agents Behaviour

The Multi-Agent System is constituted of two differ-

ent types of agent: one representing the terminolog-

ical part of the TOR (TermAgent) and the other, the

KEOD 2009 - International Conference on Knowledge Engineering and Ontology Development

Figure 1: Relations between the DYNAMO Corpus Analyzer, the DYNAMO MAS and the ontologist.

conceptual part (ConceptAgent). Considering an on-

tology as the MAS organisation, the aim of our system

is to reach a stable organisation of terms and concepts

according to the semantic relations extracted from the

considered corpus. This is mainly achieved through

a cooperative self-organisation process between the

agents encapsulating all these elements.

3.3.1 TermAgent Behaviour

TermAgents represent terms that have been extracted

from the corpus. Each extracted term is related to one

or many other terms by either syntactic relations or

lexical relations, which are linguistic manifestations

of semantic relations. These relations have been de-

tected thanks to the lexical relations extraction mech-

anism (essentially using speciﬁc triggers).

The system is initialised by creating one TermA-

gent from each extracted term. The agent behaviour

consists then in processing all the extracted relations

that connect it with other terms. Each relation has a

conﬁdence degree which is computed from the fre-

quency of occurrences of the corresponding pattern.

Using this conﬁdence degree TermAgents are related

with each other (see 3.2).

Initially the MAS is only composed of TermA-

gents which are linked by these valued relations. Each

TermAgent takes into account the most important re-

lation (which has the greatest conﬁdence value) ac-

cording to its type:

1. Synonymy relation between two TermAgents

leads to the creation of a ConceptAgent linked to

the corresponding TermAgents.

2. Hyperonymy relation between two TermAgents

causes the creation of two ConceptAgents linked

to the corresponding TermAgents. The two new

ConceptAgents are related by an is a relation.

3. Meronymy relation between two TermAgents im-

plies the creation of two ConceptAgents linked to

the corresponding TermAgents. The new Concep-

tAgents are associated with a part of relation.

4. Functional relation between two TermAgents

leads to the creation of two ConceptAgent linked

to their corresponding TermAgents. The two new

ConceptAgents are then related by this relation.

TermAgents send a request to create ConceptAgents

and link with it. Additionally, when a ConceptAgent

previously exists, it is not created a second time.

3.3.2 ConceptAgent Behaviour

This agent type represents concepts that have been

created by TermAgents. ConceptAgents behaviour

leads to the optimisation of the ontology by treating

a set of Non Cooperative Situations (NCS) derived

from speciﬁc data about the three relations type (hy-

peronymy, meronymy and functional) as well as their

link to terminological data (TermAgent denotation).

An example of these situations can be found in the

election of a concept label. Typically, each Concep-

tAgent has to choose among its related terms the one

that is the more representative to become its label. To

do so, a ConceptAgent selects the denotation relation

on which it is the most conﬁdent, and proposes to the

designated TermAgent to become its label. However,

conﬂicts may appear in this process. As a single Ter-

mAgent could denote several ConceptAgents it may

DYNAMIC ONTOLOGY CO-CONSTRUCTION BASED ON ADAPTIVE MULTI-AGENT TECHNOLOGY

receive several label requests. This situation is quoted

as an NCS (a conﬂict one), it is detected by a TermA-

gent and should be treated by a ConceptAgent. Sev-

eral solutions could be adopted at this stage:

• If there is only one common TermAgent (the la-

bel) linked to the concerned ConceptAgents, they

should be merged.

• If there are several other TermAgents linked to the

ConceptAgents, ConceptAgents have to chose an-

other label in their TermAgents pool. The TermA-

gent that detects the situation kept as a label by

one of the ConceptAgents depending on its simi-

larity to other designated TermAgents.

Between these two cases a mid-term solution have to

be found, by considering the number of TermAgents

linked to each ConceptAgents, the similarity between

these TermAgents, the relation held by the denoted

ConceptAgent, etc. This is achieved through coop-

eration and thanks to the ontologist actions. The co-

operation at the system level is the purpose of the fol-

lowing section.

3.3.3 Collective Behaviour

The DYNAMO MAS is a real time system that uses a

corpus that implies the creation of several hundreds of

TermAgents and ConceptAgents. We need to be sure

that the system converges and outputs at least a solu-

tion. This convergence is guaranteed by the AMAS

theory. In short, because agents are implemented in

such a way that they can be considered as coopera-

tive, their cooperation ensures that the whole set will

stabilize after a large set of iterations for information

exchange. In fact, the collective process stops when

each agent reaches a local equilibrium. This equilib-

rium occurs when its remaining NCS levels are lower

than the NCS levels of its neighbourhood (agents that

are related to it). For example, let us consider a given

TermAgent looking at its relations with other TermA-

gents which are not currently in the TOR:

• If a neighbour agrees to take their relationship into

account, the MAS changes the ontology by adding

the new relationship. This modiﬁcation is pro-

posed later to the ontologist for agreement. If the

ontologist disagrees the MAS stores this informa-

tion to avoid the same request one more time.

• If related neighbours disagrees, about this modiﬁ-

cation because of contradictions with other more

critical situations in its own neighbourhood, no

change occurs.

Furthermore, we considered a minimum conﬁdence

threshold in the algorithm, in order to dismiss a

large number of non-signiﬁcantsemantic relations ex-

tracted from the corpus. The conﬁdence degree of

any relation could evolve when new documents are

analysed. By this means, relations that were set apart

could be later taken into account providing that their

conﬁdence degree reaches the threshold. This sim-

ple rule also avoids processing relations considered

as analysis noise.

The collective solving process of agents is also

very efﬁcient for algorithmic reasons:

• the AMAS algorithm assumes a monotonic de-

creasing of NCS level: typically three or four

agent activations are sufﬁcient to obtain a local

equilibrium;

• when new information arrives in the MAS (com-

ing from new corpus analysis or the ontologist)

only the considered agents work. Thus the pertur-

bation process is very limited inside the MAS.

4 EXPERIMENTS

We experimented our system using the ARTAL cor-

pus deﬁned in the DYNAMO project. The objective

of the experiment is to evaluate the ontology built up

with Agents, and, from this analyse, to improve the

TermAgent and ConceptAgent behaviour.

The pattern base was fed with some of the trig-

gers deﬁned in the TerminoWeb project (Barri`ere and

Akakpo, 2006). These triggers are used to extract re-

lations between terms from a corpus (see 3.1). The

following list presents some of the chosen triggers:

• for hyperonymy relations: such as, and other, in-

cluding, especially;

• for meronymy relations: is a part of, elements of,

components of;

• for synonymy relations: another term for, also

called, also known as, synonym.

The precise numbers of triggers used for each relation

type are presented in Table 1. We also used speciﬁc

triggers for functional relations such as: when, if, at,

on, before, after.

Table 1: Number of triggers for relation extraction.

Semantic Relation Number of Triggers

SYNONYMY 12

HYPERONYMY 21

MERONYMY 23

FUNCTIONAL 16

Due to the small number of triggers, we estimate

potential semantic relations between terms by the fol-

KEOD 2009 - International Conference on Knowledge Engineering and Ontology Development

lowing rule: if it is possible to ﬁnd two terms sharing

the same contexts in the corpus (i.e. often situated

among same words), then these terms could be con-

sidered as potential semantically related. The conﬁ-

dence of this relation is calculated using the Jacquard

Formula based on the shared context cardinality. For

example, the DYNAMO Corpus Analyser considered

“Web”, “Applet” and “Service” as potential hyper-

onymy of the concept “Application” because they

share same context.

Finally, we ﬁlled the Candidate terms base with

the terminological part of the ARTAL TOR which

contains 692 terms.

5 RESULTS AND ANALYSIS

For the experiment, we have used a strict matching

of terms on the corpus, we haven’t looked for any of

their orthographic variations. This option explains the

relatively small number of matched relations and fa-

cilitates the readability of MAS results. Using DY-

NAMO Corpus Analyser, we have extracted 476 rela-

tions between 503 terms. Table 2 gives the number of

relations detected using triggers.

Table 2: Number of matched relations in the ARTAL cor-

pus.

Semantic Relation Number of matched relations

SYNONYMY 2

HYPERONYMY 68

MERONYMY 36

FUNCTIONAL 370

The small amount of synonymy, hyperonymy

and meronymy relations is due to the use of generic

triggers that are obviously not ﬁt to the corpus. On the

opposite, we use speciﬁc triggers to extract functional

relations and this is why we get the best result. This

point highlights the main feature of pattern based

approaches which is the strong dependencies of the

results on the deﬁnition of pattern.

Thanks to the Jaccard formula we have extracted

650 potential semantic relations between terms. The

MAS input is formed by 657 candidate terms and

1124 instance of relations.

Firstly, all terms are agentiﬁed and every agent

tries to be related with a ConceptAgent. In Figure

2, links without label represent denotation relations

between TermAgents and ConceptAgents. In the on-

tology model, each term can denote several concepts.

Each labelled link represents the functional relation

Figure 2: Functional relations established between Concep-

tAgents.

discovered between terms. For example, Action and

Activation concepts are related by a link labelled On.

In the considered domain (software bug reports)

the presented subgraph means that an Action trig-

gers two kinds of activation. The Activation pos-

sessed several modes which can be selected and

modiﬁes the status of the associated component

(Activation status ﬁeld). This representation is close

to the one proposed by the ontologist, but expresses

also some new functional relations.

Figure 3 presents a result obtained with Ter-

mAgent reasoning only with Jaccard Formula.

Five TermAgents are related to the “Application”

ConceptAgent who is related to the top concept.

This ﬁrst graph represents a potential hyperonymy

relation between Agents. In the second graph ﬁve

TermAgents (Impossible to send, Not activated,

Have the same name, can not be change,

Https port number) are related to the

“Https port number” ConceptAgent. The rela-

tion established between this ConceptAgent and the

ﬁrst four TermAgent represents a potential functionnal

relation (affects). In the considered domain (software

bug reports), the relation means that http port can

have several bugs like a non-activation problem.

To improve our results, we need to deﬁne more

precise patterns by using regular expressions rather

than triggers which only detect possible relations be-

tween words situated on their both sides. We are also

investigating on the use of external information re-

sources (dictionary, generic ontology such as Word-

Net

, other domain related corpus, etc.) to deal with

the limitation of corpus based approaches.

To improve the agents behaviour, we can also

combine the Pattern-based approach with a statistical

approach by using the Jaccard Formula. For example,

a TermAgent can improve its own trust in a relation

http://wordnet.princeton.edu/

DYNAMIC ONTOLOGY CO-CONSTRUCTION BASED ON ADAPTIVE MULTI-AGENT TECHNOLOGY

Figure 3: Example of semantic relation between TermAgents and ConceptAgents identiﬁed with jaccard Formula.

when the same relation is also detected using Jaccard

Formula.

As it has been expressed in Section 3.3, we do

not have completely speciﬁed the ConceptAgent be-

haviour, this prototyping phase is part of the ADELFE

methodology (Bernon et al., 2005). The aim of this

speciﬁc task is to obtain the ﬁrst draft of the sys-

tem that allows to highlight more efﬁciently the NCS.

Thanks to these ﬁrst experiments, we have been able

to quote some NCS as, for instance, the label conﬂict

described in Section 3.3.2.

6 RELATED WORKS

6.1 The Earlier DYNAMO Prototype

The objective of DYNAMO is to facilitate ontology

engineering from text thanks to a combination of Nat-

ural Language Processing and a cooperative Multi-

Agent System. Our research is inspired from DY-

NAMO ﬁrst prototype (Ottens et al., 2008) that used

a statistical approach to build up a taxonomy from

large text corpora. In this prototype, agents imple-

ment a distributed clustering algorithm that identify

term clusters. These clusters lead to the deﬁnition of

concepts as well as their organisation into a hierar-

chy. Each agent represents a candidate term extracted

from the corpus and estimates its similarity with oth-

ers thanks to statistical features. Several evaluation

tests conducted with this DYNAMO ﬁrst prototype

proved its ability to build the kernel of a domain on-

tology from a textual corpus.

6.2 Ontology Engineering from Text in

Dynamic Environments

Two on-going major IST European projects, SEKT

and NEON

aim at similar goals with a more am-

http://www.sekt-project.com/

http://www.neon-project.org/web-content/

bitious scope. Both of them are building up tool-

kits that should give access to a panel of technolo-

gies, including several Human Language Technolo-

gies among which NLP plays a major role. SEKT

and NEON want to advance the state of the art in us-

ing ontologies for large-scale semantic applications in

distributed organisations and dynamic environments.

Particularly, they aim at improving the capability to

handle multiple networked ontologies that exist in a

particular context, they are created collaboratively,

and might be highly dynamic and constantly evolv-

ing. Human Language and Ontology Technologies

are combined to produce semi-automatic tools for the

creation of ontologies, the population of those ontolo-

gies with metadata, and the maintenance and evolu-

tion of the ontologies and associated metadata. Al-

though agents technology is not used at all in these

projects, their scope is very similar to the one of DY-

NAMO.

The ambition of SEKT is to offer this variety of

technologies to develop not only ontologies and an-

notations, but full knowledge management or knowl-

edge intensive applications. Argumentation among

ontology authors who locally update an ontology is

considered as a key stage of the evolution process

of shared ontologies. The NEON project highlights

the role of NLP when updating ontologies in dy-

namic environments together with their related se-

mantic metadata. NEON toolkit offers a tool suite

that extends OntoStudio baseline and connects it with

GATE. GATE used to propose Prot´eg´e as a plug-in

for ontology development from text analysis. The

new GATE version includes a module that manages

its own ontology representation, a plug-in for ontol-

ogy population and text annotation.

7 CONCLUSIONS

DYNAMO is a Multi-Agent System allowing the co-

construction and evolution of ontologies from text.

DYNAMO MAS uses result from lexical relations

extraction mechanism to construct ontologies. We

KEOD 2009 - International Conference on Knowledge Engineering and Ontology Development

shown in Section 5 that the system is able to create an

ontology draft containing both ontological and termi-

nological elements and enriches several dimensions

of the previous prototype (see Section 6):

1. It is able to deal with richer linguistic information

as long as agents take into account lexical rela-

tions found by matching patterns on texts.

2. The result is much richer: DYNAMO builds up a

TOR which includes a hierarchy of concepts with

their related terms, and labelled semantic relations

between concepts. A set of terms denotes each

particular concept, which is useful for the docu-

ment annotation activity.

3. The current DYNAMO system is able to deal

either with French or English language text,

whereas the ﬁrst prototype was previously limited

to French language.

According to the project schedule we need to improve

the software during the next year. To do so we plan

to:

• introduce the cooperative behaviour of ConceptA-

gents (speciﬁcation of NCS and their treatment);

• provide an adaptive patterns learning process

based on the AMAS theory;

• provide speciﬁc interfaces to enable ontologists

collaboration;

• apply the DYNAMO MAS to all the project do-

mains (archeology, car diagnosis, software bug re-

ports).

REFERENCES

Barri`ere, C. and Akakpo, A. (2006). Terminoweb: A soft-

ware environment for term study in rich contexts. In

Proceedings of International Conference on Terminol-

ogy, Standardization and Technology Transfer, pages

103–113, Beijing. Encyclopedia of China Publishing

House.

Bernon, C., Camps, V., Gleizes, M.-P., and Picard, G.

(2005). Engineering Adaptive Multi-Agent Systems:

The ADELFE Methodology . In Henderson-Sellers,

B. and Giorgini, P., editors, Agent-Oriented Method-

ologies, volume ISBN1-59140-581-5, pages 172–202.

Idea Group Pub, NY, USA.

Buitelaar, P., Cimiano, P., and Magnini, B. (2005). On-

tology Learning From Text: Methods, Evaluation and

Applications. IOS Press.

Capera, D., Georg´e, J.-P., Gleizes, M.-P., and Glize, P.

(2003). The AMAS Theory for Complex Problem

Solving Based on Self-organizing Cooperative Agents

. In TAPOCS 2003 at WETICE 2003, Linz, Austria,

09/06/03-11/06/03. IEEE CS.

Chagnoux, M., Hernandez, N., and Aussenac-Gilles, N.

(2008). An interactive pattern based approach for ex-

tracting non-taxonomic relations from texts. In Buite-

laar, P., Cimiano, P., Paliouras, G., and Spilliopoulou,

M., editors, Workshop on Ontology Learning and

Population (associated to ECAI 2008) (OLP), Patras

(Greece), 22/07/08, pages 1–6.

Cimiano, P. (2006). Ontology Learning and Population

from Text: Algorithms, Evaluation and Applications.

Springer.

Cimiano, P. and V¨olker, J. (2005). Text2onto - a framework

for ontology learning and data-driven change discov-

ery. In Montoyo, A., Munoz, R., and Metais, E., ed-

itors, Proceedings of the 10th International Confer-

ence on Applications of Natural Language to Informa-

tion Systems (NLDB), volume 3513 of Lecture Notes

in Computer Science, pages 227–238, Alicante, Spain.

Springer.

Gruber, T. R. (1993). A Translation Approach to Portable

Ontology Speciﬁcations. Knowledge Acquisition,

5(2):199–221.

Haase, P. and Sure, Y. (2004). D3.1.1.b state-of-the-art on

ontology evolution. Technical report, Institute AIFB,

University of Karlsruhe.

Harris, Z. S. (1968). Mathematical Structures of Language.

Wiley, New York.

Hearst, M. A. (1992). Automatic acquisition of hyponyms

from large text corpora. In COLING, pages 539–545.

Maedche, A. (2002). Ontology learning for the Semantic

Web, volume 665. Kluwer Academic Publisher.

Malaise, V. (2005). M´ethodologie linguistique et

terminologique pour la structuration d’ontologies

diff´erentielles `a partir de corpus textuels. PhD thesis,

Paris 7 Denis Diderot University.

Morin, E. (1999). Using lexico-syntactic patterns to extract

semantic relations between terms from technical cor-

pus. In 5th International Congress on Terminology

and Knowledge Engineering (TKE), pages 268–278,

Innsbruck, Austria. TermNet.

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

Gilles, N. (2008). A Multi-Agent System for Dynamic

Ontologies. Journal of Logic and Computation, Spe-

cial Issue on Ontology Dynamics, 19:1–28.

Reymonet, A., Thomas, J., and Aussenac-Gilles, N. (2007).

Modelling ontological and terminological resources in

OWL DL. In Buitelaar, P., Choi, K.-S., Gangemi,

A., and Huang, C.-R., editors, OntoLex07 - From Text

to Knowledge: The Lexicon/Ontology Interface Work-

shop at ISWC07 6th International Semantic Web Con-

ference, Busan (South Korea), 11/11/07.

Velardi, P., Navigli, R., Cucchiarelli, A., and Neri, F.

(2005). Evaluation of OntoLearn, a Methodology for

Automatic Learning of Domain Ontologies. In Buite-

laar, P., O. Cimiano & B. Magnini (eds.). IOS Press,

Amsterdam.

DYNAMIC ONTOLOGY CO-CONSTRUCTION BASED ON ADAPTIVE MULTI-AGENT TECHNOLOGY