Violaine Prince
University of Montpellier and LIRMM-CNRS
161 Ada Street 34392 Montpellier cedex 5 France
Knowledge Management, Knowledge sharing, Multi-agent systems, heterogeneous Agents.
In this paper, we describle a model that relies on the following assumption; ontology negotiation and creation
is necessary to make knowledge sharing and KM successful through communication. We mostly focus on the
modifying process, i.e. dialogue, and we show a dynamic modification of agents knowledge bases could occur
through messages exchanges, messages being knowledge chunks to be mapped with agents KB . Dialogue
takes account of both success and failure in mapping. We show that the same process helps repair its own
anomalies. We describe an architecture for agents knowledge exchange through dialogue. Last we conclude
about the benefits of introducing dialogue features in knowledge management.
Knowledge has muted from a personal expertise to-
wards a collective lore. Thus, knowledge manage-
ment (KM) and modelling contains social features
and deals with society of agents. As a society, agents
interact, and interaction conventional aspects have
attracted the attention of the KM community, with
different points of view : (i) The deep relationship
between knowledge and knowledge communication
(Ravenscroft and Pilkington 2000); (ii)The power of
the communicative process as a knowledge modifier
(Parson et al. 1998), (Zhang et al. 2004 ).
Although communication is an active component in
KM, most studies have dealt with cases where agents
were sufficiently close in type and knowledge in or-
der to share common ontologies. In extensive KM
systems, knowledge sharing is hampered by the lack
of common ontologies between artificial agents. In-
tegrating different designations for the same concepts
has been tackled by (Williams 2004) in a shared en-
vironment named DOGGIE ( Distributed Ontology
Gathering Group Integration Environment) allowing
negotiation of terminology among different ontolo-
gies. The author’s approach is that of a ’peer-to-peer’
situation that allows agents to share knowledge and
learn. His aims was more to demonstrate how hetero-
geneity in representations could be overcome, than to
focus on the dynamic process that underlies it, that is,
Our approach is in the same main line of thought, but
the added value is that we focus on the properties of
dialogue as a process of incremental knowledge ad-
justment. Human dialogue occurs between interlocu-
tors distinct in state, knowledge and intents. It hap-
pens when a need for knowledge occurs. It is used
when discrepancies in representations appear.
Applying dialogue requirements to KM is an issue
we will address in section 2. Although difference is
a mandatory component for dialogue existence, too
great a distance between agents might not even let the
chance for a dialogue to occur. Section 2 states the
likely requirements for dialogue success and conse-
quently those for an adequate KM through interac-
tion. Section 3 presents an architecture implementing
the process between artificial agents, an instantiation
of which, dedicated to agent teaching, has been pre-
sented in (Yousfi and Prince, 2005). Teaching is one
of the tasks in which knowledge sharing is tracked
at best (Williams 2000). This architecture has been
evaluated through its applications: the teaching appli-
cation, dedicated to conceptual knowledge revision,
presently runs as a prototype. Another application
about risk management ontologies acquisition (Makki
et al. 2006) is developed according to both model and
architecture described in this paper.
Prince V. (2006).
In Proceedings of the First International Conference on Software and Data Technologies, pages 266-271
DOI: 10.5220/0001314002660271
Agents are considered heterogeneous when they dif-
fer in nature or in a major attribute. In this pa-
per we will restrain the definition of heterogeneity to
cognitive artificial agents with the following proper-
ties :(i)Different ontologies and knowledge bases (i.e
different world representations); (ii) Different tasks
within the system; (iii) Possibly belonging to different
applications that need to share knowledge, Web ser-
vices or activity. Heretogeneity produces variations
in building, sharing and communicating knowledge.
Theoretical Framework in Agents Knowledge
Human beings favour dialogue as a major mean for
knowledge acquisition: Each agent considers any fel-
low agent as a knowledge source ’triggered’ through
questioning. Information is acquired, from the an-
swer, as an external possible hypothesis. This is the
starting point of both an acquisition and a revision-
based process, where the external fact, the message,
is subject to confrontation with the inner knowledge
source of the requiring agent. It drives the latter
to proceed to derivation by reasoning.The feedback
observed in natural dialogue is that the knowledge
source could be addressed for understanding confir-
mation. Along with other researchers,(Finin et al.
1997), (Zhang et al. 2004 ), we consider this process
as translatable into the software world and interesting
as an economical mean to acquire, mediate and share
The cognitive agents we are representing can be
seen as entities foregoing the following cycle (Prince
1996): (i)Capture, and symmetrically, edit data. Data
is every trace on a media; (ii)Transform the captured
data into information, as the output of an interpre-
tative process on data. This result will be either
stored as knowledge or discarded. (iii)Keep and in-
crease knowledge, which, in turn, is of various types:
(a) Stored information seen as useful; (b) Operat-
ing ’manuals’ to interpret data; (c) Deriving modes
to produce knowledge out of knowledge (reasoning
modes); (d) Strategies to organise and optimise data
capture, information storage, and knowledge deriva-
tion. Knowledge is by essence defined as explic-
itable knowledge, the only variety that could be im-
plementable. (iii) Acquire bypasses to data inter-
pretation and action on the information environment
through time-saving procedures: e.g. developping
know-how in the information field. The four parts
model is called DIKH (for Data, Information, Knowl-
edge and know-How). KM relies mostly in the two
last components : ”know-How”, and ”Know” are dis-
tinguished according to their properties involving ex-
plicitness vs implicitness, transferabilty vs non trans-
missibility. Since it is representable and conceptual-
isable, ”Know” or ”Knowledge” (K) has been mostly
investigated by KM. ”Know-how” has been claimed
as embedded in expert systems, but since it is implicit
it cannot be easily described. Therefore, know-how
has to be assumed as an important skill of cognitive
agents, but not as to be further investigated, unless in
very restricted areas. Since KM is at stake, we are
focusing on the K fourth of the model. DIKH as-
sumes a recursive modelling: Part of knowledge de-
rives from data, part of it derives from present knowl-
edge, and part of it is an economical optimisation of
its organisation. The K part puts in a nutshell sev-
eral known elements of KM. (i) Lexical Knowledge
represents tontological contents, relations, organisa-
tion. (ii)Produced knowledge is the kept part of in-
formation (interpreted data). It might rejoin either
lexical knowledge (new concepts to integrate in the
ontology), or production knowledge (rules, general
laws). (iii)Production knowledge has been mostly
investigated by research in AI: rules, reasoning, go-
ing further to meta-rules, and to strategies in organ-
ising KM is an important issue in cognitive agent
modelling. (iv) The”know-how” part of knowledge,
an originality of the DIKH model, strengthens upon
profiling, preferences, and presentation, as the agent
signature key. In single-application architectures, all
agents tend to share the same ”know-how” since the
latter belongs to the architecture designer. In hetero-
geneous systems, many architectures might be con-
fronted with each other. Web semantics and services
have integrated this aspect: Languages such as XML
have been devoted to emphasize the know-how about
knowledge presentation. The DIKH recursive mod-
elling might easily represent the KM part of a rational
natural agent (a person), as an extension from artifi-
cial agents environment, to Human-System environ-
ments. Since ”information” is a temporary status for
knowledge, the model reduces to the representation
given in figure 1. A rational cognitive agent might
be designed, from the static point of view, as: (i)A
set of lexical skills: ontological knowledge, relation-
ships between names and concepts, variables and their
domains; (ii) The core of a reasoning engine: Local
axioms, strong beliefs that help deriving other rules
and rules having a lesser status; (iii) Last, elements
of belief and knowledge that help optimising the en-
gine, those are the adaptive modes derived from ex-
A Message/Knowledge Chunk Exchange The-
A message (Jacobson and Halle 1956) can be de-
fined as a formated data set which: (i) is emited
as a sender’s intention concerning his recipients ;
(ii)follows a protocol (conventions in format and ex-
changes) ; (iii) has a content ; such that the whole
(form, intention, protocol, contents) is supposed to
Value Domains
Lexical skill
Figure 1: the K-model: A static representation of the ratio-
nal cognitive agent.
modify the internal state of the recipient agent(s).
Related to artificial agents communication, the mes-
sage properties are the following: (i) presentation
: Formatting properties of the message as a meta-
format. (ii) content : it is in itself a complex system
that can be decomposed into: 1. how content is for-
malised; i.e.: (i) The selected elements in the chosen
language to designate different items ; (ii) composi-
tion rules used for the message.
2. The semantic content of the message: lexical data
meaning and formal compositions.
3. The informational content: What the recipient
agent has been able to understand from the received
4. The Intentional content: what the sender agent
has wanted to transmit.
Definition : The formal structure of a message can
be described as a ternary structure composed of:(i)
Data: The lexical and syntaxic items composing the
message strata, equivalent to the rheme in the Speech
Act Theory (SAT)(Searle 1969).
(ii) Knowledge: which is itself decomposed into: 1.
the necessary knowledge to encode/decode data (1).
2. the message semantic content is knowledge (2),
equivalent to the topic in SAT. 3. the knowledge to
embed / derive the semantic content (intentional / in-
formational content) (3)
(iii) Formulation (the adapted terminology for ”pre-
sentation” in the formal structure) : Style and formal-
ism are qualitative indicators.
The preceding definition has a striking resemblance
with agents K-models. Hence, the formal structure
of a message and the K-model of the agent could be
seen as related by a strong morphism. (i) Data in
message definition is provided by the lexical skills of
both sender and recipient agents. (ii) The semantic
content is the knowledge chunk exchanged between
agents: It is used to enhance the recipient lexical skills
in ontological knowledge building or updating (iii)the
intentional versus informational contents of the mes-
sage, tackles the issue of confronting the knowledge
production modes or engine of both locutors. (iv)
Last, message formulation is the result of applying the
sender’s preferences about message exchanges, and
triggers the recipient’s adaptive modes to accept the
message or reject it if it is not properly designed. This
explains why the exchange of messages is a natural
mode of knowledge enhancement. If the message is
structurally compatible with the K-model of an agent,
Let A
be the K-model of agent µ. Let m be the for-
mal structure of an incoming message. The question
is: is A
m a new possible state of µs K-model ?
For this, we need an interpreted form of the message.
Definition : An interpreted form of a message is
obtained through the following process: (i) Applying
decoding knowledge: unification algorithms and ab-
ductive rules are used to initiate this phase; (ii) Se-
mantic interpretation of the decoded form (informa-
tional content): Deductive and inductive reasoning is
used. (iii)If formulation is not a liability for interpre-
tation, the informational content should be equal or
close to semantic content of the message.
Given the preceding results, a message, seen from
its formal structure point of view, could be designed
as a knowledge bridge between agents. Its purpose
is to: (i)Allow agents to update their knowledge
through other agents knowledge; (ii)Fix knowledge
discrepancies between agents. Let m be the formal
structure of the message to be exchanged between
two agents µ and ν. The three components of m ,
data, knowledge and formulation, could be
designed as following; (i) The data of m belongs to
the lexicon of µ as a sender, and should also belong
to the lexicon of ν. (ii) The knowledge in m needs
the corresponding items of µs and νs knowledge en-
gines. However, if the message is supposed to in-
crease the recipient’s knowledge, this part also com-
prises knowledge that is either new or which exten-
sions are new to the recipient. (iii) Last, the formula-
tion is the formalism used for the bridge (language, or
protocol). It requires adaptation from the sender to the
recipient, and vice-versa.Figure 2 shows a representa-
tion of the formal structure of a message as bridge
between cognitive agents.
Dialogue as Knowledge Adjustment Seeking :
A theory of messages suggests dealing with the fol-
lowing cases: (i) Wether the message has been mis-
interpreted, or not decoded at all, which is a failure
in communication (an issue we will not tackle here);
(ii) Or if the message, being correctly interpreted has
roused contradictions and thus, has failed to reach its
goal; (iii) A combination of both cases, a common
situation in natural dialogues. We will focus here on
contradiction in knowledge or belief revision.
Belief revision appears as a compulsory process when:
(i) The recipient finds its knowledge contradicted by
Agent µ Agent ν
dge engine
dge engine
Figure 2: Modelling the message as a bridge between two
what the others know (this happens when launching
the primitive DetectConflict described in next
section) ; (ii)The informational content the recipient
has derived from the message seems irrelevant with
the dialogue situation. The first case is a pure KM
problem. An agent needs to receive a message, i.e., a
future knowledge chunk of its own K-model, and this
leeds to a major revision of its beliefs and its lexical
skills. Therefore a ”why” -type of dialogue is initi-
ated on the recipient’s behalf, and revision could be
undertaken as soon as the recipient is convinced of
the quality of the received knowledge. This has been
mostly investigated in (Parson et al. 1998) and (Zhang
et al. 2004 ). They have shown that not only the con-
tradicted agent is likely to change, but also its contra-
dictor, since the latter needs to restructure its own KB
in order to convince the other. Our theory reaches the
same results since the ”why-questions”, seen as mes-
sages from the former recipient, force the sender to
use its skills in formulation and in knowledge deriva-
tion. In (Yousfi and Prince, 2005), we have shown
how an artificial ”teacher” agent modifies its KB dur-
ing the process, at least at the student model level, but
furthermore, while teaching, it might spot its own de-
faults in knowledge.
The other case is the knowledgeable problem of mis-
understanding. Wether misunderstanding bursts out
from decoding errors or mistaken reasoning, the fact
is that the following equation : informational content
(INFC) = semantic content(SC) = intentional con-
tent (INTC) is sometimes not achieved between for-
mal agents. Until now, the latter tended to abort in-
teraction whenever it failed. However, since artifi-
cial agents need to be more robust, they have to be
provided with fonctionalities helping them to pursue
dialogue further. Our theory provides some heuris-
tics for approximating this double equality: (i) Re-
formulation dialogues tend to achieve the first equal-
ity : INFC= SC; by using other knowledge items
or other laws maybe closer its present K-model, the
recipient agent might reach a state where it finally un-
derstands the message ; (ii) Every careful choice of
formulation, on the sender’s behalf might help pro-
viding the second equality SC = INTC. Argumenta-
tion/explanation dialogues play an important role in
trying to reach a good approximation.
In conclusion, whenever agents need to interact, it is
not a problem if interaction is not a successful one
shot process. Dialogue is an incremental process act-
ing as a mutual adjustment mechanism, that repairs
both its own failures and agents mutual discrepan-
cies. However time consuming, dialogue is less costly
than a wrong action based on false beliefs. It is thus
very important in decision-making tasks where cru-
cial stakes are involved.
The general sofware architecture modelling a
knowledge sharing and revision activity between
two rational artificial agents is presented in figure
3. Components that have to be implemented are
the agents K-models comprizing the other agent
model, communication primitive allowing messages
exchanges, and dialogue strategies underlying the
communication protocol. World and activity models
are provided either from each agent environment or
activity, or from applications and services they need
to address or from which they are issued.
Implementing Agents:
Implementing the K-model is an easy task since it
involves : (i) The local agent ontology and lexical
skills; (ii) A set of Reasoning Primitives, and an
access to Dialogue Strategies (sharable between
rational agents) and a set of local rules for knowledge
and message construction; (iii) One or many repre-
sentation formalisms with which the agent processes
( e.g.XML for structure, KQML for communication
acts ((Finin et al. 1997) etc... ).
Sharable Reasoning Primitives are the following:
1.Add(K) ; adds knowledge to the K-Model
2. Revise (K,O); triggers an algorithm trying to
attach parts of K to the Ontology. Described in (Yousfi
and Prince, 2005). Provides a flag REVISEF indicating
success (or not) in attachment, and where it happens.
3. DetectConflict (K) ; if
REVISEF = false detects conflictual attachments. Its
result is issuing a CreateMessage(M,OtherAgent)
(see next subsection);
Two other modal primitives are necessary for a situa-
tion where knowledge has to be shared: (i) Wanted
; applied on knowledge (from O or A parts), is the
situation that launches dialogue; (ii) Believed :
applied on knowledge from O, A or P parts, is what
the agent assumes as true without have checked it.
By default, all knowledge in the K-model that is nei-
ther wanted nor believed is considered ’known’
until contradicted. Accessible Dialogue Strategies
are common and shared rules about dialogue protocol
(cf Dialogue Strategies component). Whenever an
agent is in a situation to communicate, it launches the
creation of the ’Other Agent Model’. The latter is a
subcomponent of the K-model created and updated
with a procedure inspired from the following (we
have chosen here an application where agents accept
several programming languages):
Create OtherAgentModel(Ag) ;
Proc main
New O; New A; New P;
Add (wanted (K), O);
Believed (A) : <-Add; Revise;
Believed (P) : <- KQML; XML; Java
Comment: The agent creates a new K-model of its
interlocutor agent with its three parts : Ontology
(O), Axioms (A) and P (presentation). It boot-
straps the ontology with the knowledge it wants
(it assumes that the other has it). By default, the
OtherAgentModel A part gets the sharable Rea-
soning Primitives, and the P part gets the languages
the agent itself understands. Message sending and
especially message reply will make it know if its as-
sumptions are right or wrong. An agent creates such
a model to trigger communication. The next logical
step is to create a message asking the interlocutor
about the wanted knowledge.
Communication Primitives
Three basic communication primitives are
necessary: One is CreateMessage the
second is AcceptMessage, and the last
isRejectMessage. All deal with three vari-
ables: D standing for Data, K standing for knowledge
and F for formulation. The first receives receives the
lexical choice after the K part has been expressed in
a logical form the INTC (intentional content) and
translated in the language instantiating the F part. As
an example, we present CreateMessage below.
The other primitives follow a similar description.
RejectMessage returns the value of the part re-
sponsible for rejection (language, or unknown data).
Whereas AcceptMessage triggers a matching
and revision procedure within the recipient agent
K-model, that might in turn end up with another
CreateMessage primitive. These primitives
use also basic functions Sendto and Ack (for
acknowledge) that deal directly in interfacing with
the other agent.
CreateMessage(M,Ag, L) ; arguments are the
Message, to the Agent, in a Language
Proc main New D; New K; New F;
Set F to L; sets the formulation in a given language.
Helps adjusting to the other agent language if L value is not
INTC <- Wanted (Kn) ; The intentional content of
the message is the wanted knowledge
F <- TranslateIn (L, INTC)
D <- D(F) ; Message Data is the data part of the trans-
lated message
K <- INTC ; Message knowledge is the wanted knowl-
Sendto (M,Ag, R); sends the message to the agent
typing it with a role’ (to be explained in next subsection)
Dialogue Strategies : Message-level and Scripts:
Messages Roles: The exchange of communication
primitives follow dialogue strategies available to ev-
ery artificial agent. A strategy is related to the agent
goal and satisfaction of its needs. It can be typed in
order to extract information from the other agent or
to make it perform an action or a task.Thus, every
message plays a role’ in a dialogue instantiation,
according to a strategy. Roles have been labelled
after the Speech Act Theory illocutionary functions
(Searle 1969) or according to the functional roles
theory (as in (Yousfi and Prince, 2005)) and depend
on the task or activity type.At the message level,
roles are the materialisation of Dialogue Strategies.
For instance, the most used speech acts in agents
modelling and communication are performative
(i.e the message runs an applet) or directive ( the
message is a command to the other agent). The
functional roles we have used atmost in agents
learning are askfor-knowkedge, askfor-explanation,
give-knowledge, give-explanation, assert-satisfaction
or assert-unsatisfaction. Those were modals applied
to the CreateMessage primitive and transmited with
the message. In this paper, we present a generali-
sation of the architecture and components, and one
can notice that the role R is sent as an argument of
the Sendto command. Dialogue Strategies from
expectations: When an agent issues a message with
a given role, then it expects in turn a reply with a
compatible role. The adjustment script available to
agents follows these guidelines:
Expect(AcceptMessage(M’,OtherAgent, L),
R); the agents expects an understandable reply with a
given role
If no (AcceptMessage(M’,OtherAgent, L))
or RejectMessage(M’,Reason) then when no
answer is or an answer with decoding problems is provided
Call RepairCommunication else the strategy
Agent A
Model of B
Agent B
Model of A
Web Services
Model of the world
And activity (A)
Model of the world
And activity (B)
Dialogue strategies:
Adjustment and repair
Web Services
Figure 3: Architecture of Communicating Agents Sharing
Dialogue Strategies, but Addressing Different Applications
and Activities.
shifts to the other script
TransformIn (M, O, A) , the message is trans-
formed into its parts and matched with ontology and
Revise (O, A) reasoning is applied on the added
if Wanted(Kn) then R<- ’ok’
CreateMessage (M, OtherAgent, L) the
agent has found the wanted knowledge and asserts its
else R <- need else the agent sets the role to its
Call Adjusment and calls recursively the strategy
Let us note that their is a timeout associated to re-
cursive calls ie if no replies are given or if the dia-
logue enters an endless loop, then the dialogue strat-
egy component stops communication.
Unfortunately we have no room here to present the
FailureCommunication script but let us say that it
deals with reformulation (shifting languages) and ex-
planation roles and strategies in messages exchanges.
The model presented here and some elements of its ar-
chitecture have instantiated in a learning environment
for cognitive artificial agents. It is sufficiently general
to be implemented within different applications and
activities, as long as they need an advanced commu-
nication framework for knowledge sharing, revision
and for communication. The originality of the model
relies in modelling the dynamic process in KM, ie,
dialogue, as the crucial component in knowledge re-
vision, and not only considering the static dimension
of KM. What has not been explicitely detailed here is
that the same theory applies to Human-Computer in-
teraction and to Collective vs Individual agents KM.
The issue that is dealt with goes much further than ar-
tificial agents programming. But what we have shown
here is that even restricted to a formal and decidable
framework, the theroy takes into account knowledge
conflict and provides it with solutions inspired from
natural agents’ behaviour.
J.L. Austin. How to Do Things with Words ed. J. O.
Urmson and Marina Sbis
a, Mass: Harvard University
Press.Cambridge.1975 (3rd edition).
J. Castelfranchi. Through the Minds of the Agents Jour-
nal of Artificial Societies and Social Simulation , 1(1).
T. Finin, Y. Labrou, and J. Mayfield. KQML as an agent
communication language in Jeff Bradshaw (Ed.), Soft-
ware Agents. MIT Press, Cambridge. 1997.
S. Franklin , A.Graesser. Is it an Agent, or just a Program?:
A Taxonomy for Autonomous Agents Proceedings of
the Third International Workshop on Agent Theories,
Architectures, and Languages Springer-Verlag, 1996.
R. Jacobson and M. Halle Fundamentals of Language Mou-
ton and co. the Hague 1956.
J. Makki, A.M. Alquier and V. Prince Learning risk ontolo-
gies from both corpora and cognitive agents. LIRMM-
CNRS research report. Montpellier France. 2006
S. Parsons , C. Sierra and N. Jennings. Agents that rea-
son and negotiate by arguing. Journal of Logic and
Computation, 8(3), pp 261–292. 1998
V. Prince. Pour une informatique cognitive dans les organ-
isations: le r
ole central du langage (Towards Cogni-
tive Informatics in Organisations: the Central Role of
Language) Editions Masson. Paris, France. 1996.
A. Ravenscroft and R.M. Pilkington. Investigation by De-
sign: Developing Dialogue Models to Support Rea-
soning and Conceptual Change. International Journal
of Artificial Intelligence in Education ,11(1), pp.273–
298, 2000.
J. Searle. Speech Acts: An Essay in the Philosophy of Lan-
guage Cambridge University Press. Cambridge.1969
Andrew B. Williams. The Role of Multiagent Learning
in Ontology-based Knowledge Management. AAAI
2000 Spring Symposium on Bringing Knowledge to
Business Processes. 2000 pp175-177,
Andrew B. Williams. Learning to Share Meaning in a
Multi-Agent System. Autonomous Agents and Multi-
Agent Systems, volume 8, number 2,2004 pp-165-193
M. Wooldridge and S. Parsons. Languages for negotiation.
Proceedings of ECAI2000 pp 393–400,2000.
M. Yousfi-Monod and V. Prince. Knowledge acquisition
modelling through dialogue between cognitive agents.
Proceedings of ICEIS05, International Conference on
Enterprise Information Systems. pp. 201-206.
D.Zhang , N.Foo ,T. Meyer and R. Kwok. Negotiation as
Mutual Belief Revision. Proceedings of AAAI 2004,