=Paper=
{{Paper
|id=Vol-225/paper-5
|storemode=property
|title=Arguing Over Ontology Alignments
|pdfUrl=https://ceur-ws.org/Vol-225/paper5.pdf
|volume=Vol-225
|dblpUrl=https://dblp.org/rec/conf/semweb/LaeraTEBP06a
}}
==Arguing Over Ontology Alignments==
https://ceur-ws.org/Vol-225/paper5.pdf
Arguing over ontology alignments
L. Laera1 , V. Tamma1 , J. Euzenat2 , T. Bench-Capon1 , and T. Payne3
1
Department of Computer Science, University of Liverpool,UK
2
INRIA Rhône-Alpes, Montbonnot, France
3
Computer Science, University of Southampton, UK
Abstract. In open and dynamic environments, agents will usually differ in the
domain ontologies they commit to and their perception of the world. The avail-
ability of Alignment Services, that are able to provide correspondences between
two ontologies, is only a partial solution to achieving interoperability between
agents, because any given candidate set of alignments is only suitable in certain
contexts. For a given context, different agents might have different and incon-
sistent perspectives that reflect their differing interests and preferences on the
acceptability of candidate mappings, each of which may be rationally acceptable.
In this paper we introduce an argumentation-based negotiation framework over
the terminology they use in order to communicate. This argumentation frame-
work relies on a formal argument manipulation schema and on an encoding of
the agents preferences between particular kinds of arguments. The former does
not vary between agents, whereas the latter depends on the interests of each agent.
Thus, this approach distinguishes clearly between the alignment rationales valid
for all agents and those specific to a particular agent.
1 Introduction
Traditionally ontologies have been used to achieve semantic interoperability between
software applications, as such applications provide the definitions of the vocabularies
they use to describe the world [12], and they have proved especially effective when sys-
tems are embedded in open, dynamic environments, such as the Web and the Semantic
Web [4]. Interoperability relies on the ability to reconcile the differences between het-
erogeneous ontologies [18]. This reconciliation usually relies on the existence of corre-
spondences (or mappings) between different ontologies (ontology alignment [11]), and
uses them in order to interpret or translate messages exchanged by applications. Such
correspondences may be generated by a variety of different matching algorithms [16] 4 ,
and their production usually requires several steps. These can include the definition of
an initial alignment, or the training over some examples, and these invariably involve
some form of interpretation of preliminary results [10]. Therefore, approaches to on-
tology alignment can only be effective when used to support semantic interoperation
at design time in closed or partially open environments, where the actors involved are
often known, where ontology changes are controlled and thus the alignments can be
established before the systems interact. However, these approaches are not sufficient
to support semantic interoperation in open environments, where systems can dynami-
cally join or leave and no prior assumption can be made on the ontologies to align. In
4
A comprehensive review can be found at http://www.ontologymatching.org
such environments, the different systems involved need to agree on the semantics of
the terms used during the interoperation, and reaching this agreement can only come
through some sort of negotiation process [1].
This paper extends the notion of reaching agreement through automated negotiation
(i.e. without human intervention) by considering the type of systems that need to inter-
operate, which can affect how the negotiation should proceed. Specifically, autonomous
agents (within an open environment) may perform different tasks depending on their
state and the service providers they interact with. Thus, such agents will differ in the
domain ontologies they commit to [12]; and their perception of the world (and hence
the choice of vocabulary used to represent concepts). Imposing a single, universally
shared ontology on agents is not only impractical because it would result in assuming
a standard communication vocabulary (and thus violate the dynamics of open environ-
ments) but it also does not take into account the conceptual requirements of services
that could appear in future. Instead, every agent assumes its own heterogeneous private
ontology, which may not be understandable by other agents. The availability of Align-
ment Services that are able to provide correspondences between two ontologies is only
the beginning of a solution to achieving interoperability between agents, as any given
candidate set of alignments is only suitable in certain contexts. For a given context,
agents might have different and inconsistent perspectives; i.e. interests and preferences,
on the acceptability of a candidate mapping, each of which may be rationally accept-
able. This may be due to the subjective nature of ontologies, to the context and the
requirement of the alignments and so on. For example, an agent may be interested in
accepting only those mappings that have linguistic similarities, since its ontology is too
structurally simple to realise any other type of mismatch. In addition, any decision on
the acceptability of these mappings has to be made dynamically (at run time), due to
the fact that the agents have no prior knowledge of either the existence or constraints of
other agents.
In order to address this problem, we present a framework to support agents to nego-
tiate agreement on the terminology they use in order to communicate, by allowing them
to express their preferred choices over candidate correspondences. This is achieved by
adapting argument-based negotiation to deal specifically with arguments that support
or oppose the proposed correspondences between ontologies. The set of potential argu-
ments are clearly identified and grounded on the underlying ontology languages, and
the kinds of mapping that can be supported by any such argument are clearly specified.
Specifically, we use a value-based argumentation framework [2], allowing each agent to
express its preferences between the categories of arguments that are clearly identified in
the context of ontology alignment. Our approach is able to give a formal motivation for
the selection of any correspondence, and enables consideration of an agents’ interests
and preferences that may influence the selection of a given correspondence. Therefore,
this work provides a concrete instantiation of the ”meaning negotiation” process that
we would like agents to achieve. Moreover, in contrast to current ontology matching
procedures, the choice of alignment is based on two clearly identified elements: (i) the
argumentation framework, which is common to all agents, and (ii) the preference rela-
tions which are private to each agent.
The remainder of this paper is structured as follows. Section 2 presents the argu-
mentation framework and how it can be used. Section 3 defines the various categories
of arguments that can support or attack mappings. Section 4 describes our agent model
and discusses how agents should reach agreement. An example illustrating the argu-
mentation process is given in Section 5, followed concluding remarks in Section 65 .
2 Argumentation Framework
This paper focuses on autonomous agents situated within an open system. Each agent
has a knowledge base, expressed using one of several possible ontologies. The mental
attitudes of an agent towards correspondences are represented in terms of interests and
preferences, which represent the motivations of the agent, and thus determine whether
a mapping is accepted or rejected. The preferences are represented as a (partial or total)
pre-ordering of preferences over different types of ontology mismatches (Pref )6 .
For agents to communicate, they first need to establish a mutually acceptable set
of alignments between their ontologies. Potential alignments are generated at design
time (by a variety of different ontology-matching approaches [16]), and provided at
run-time by a dedicated agent, called an Ontology Alignment Service (OAS) (Figure 1).
An alignment consists of a set of correspondences between the two ontologies. A corre-
spondence (or a mapping) can be described as a tuple: m = he, e0 , n, Ri, where e and e0
are the entities (concepts, relations or individuals) between which a relation is asserted
by the correspondence; n is a degree of confidence in that correspondence; and R is
the relation (e.g., equivalence, more general, etc.) holding between e and e0 asserted
by the correspondence [16]. A candidate mapping is a correspondence (provided by an
OAS) that could be used by the agents to align their ontologies. Each correspondence
m is accompanied by a set of justifications G, which provide an explanation as to why
the correspondence was generated7 . This information is used by the agents when gen-
erating and exchanging arguments, for and against a candidate mapping. In addition,
every agent has a private threshold value ε which will be compared to the degree of
confidence, n, of a mapping, to decide whether it should be considered.
In order for the agents to consider potential mappings and the reasons for and against
accepting them, we use an argumentation framework based on Value-based Argument
Frameworks (VAFs) [2], that extends Dong’s classical argument system [7]8 .
Definition 1. An Argumentation Framework (AF ) is a pair AF = hAR, Ai, where
AR is a set of arguments and A ⊂ AR × AR is the attack relationship for AF . A
comprises a set of ordered pairs of distinct arguments in AR. A pair hx, yi is referred
to as ”x attacks y”. We also say that a set of arguments S attacks an argument y if y is
attacked by an argument in S.
5
A survey of related work is given in an extended version of this paper [13].
6
Although the agents’ ontologies may differ, we eliminate the problem of integrating different
ontology languages by assuming that ontologies are encoded in the same language, i.e. OWL.
7
Although few approaches for ontology alignment provide justifications [17, 5], tools such as
[9] combine different similarity metrics which can be used to provide necessary justifications.
8
More details can be found in an extended version of this paper [13].
Ontology Alignment Service
OWL Ontology OWL Ontology
Argumentation
Agent Agent
Agreed and agreeable
alignments
Fig. 1. Reaching agreement over ontology alignments
An argumentation framework can be simply represented as a directed graph whose
vertices are the arguments and whose edges correspond to the elements of A. In this
paper, we are concerned only with arguments about mappings. We can therefore define
arguments as follows:
Definition 2. An argument x ∈ AF is a triple x = hG, m, σi where m is a correspon-
dence he, e0 , n, Ri; G is the grounds justifying a prima facie belief that the correspon-
dence does, or does not hold; σ is one of {+, −} depending on whether the argument
is that m does or does not hold.
An argument x is attacked by the assertion of its negation ¬x, namely the counter-
argument, defined as follows:
Definition 3. An argument y ∈ AF rebuts an argument x ∈ AF if x and y are ar-
guments for the same mapping but with different signs, e.g. if x and y are in the form
x = hG1 , m, +i and y = hG2 , m, −i, x counter-argues y and vice-versa.
Moreover, if an argument x supports an argument y, they form the argument (x →
y) that attacks an argument ¬y and is attacked by argument ¬x.
When the set of such arguments and counter arguments have been produced, it is
necessary for the agents to consider which of them they should accept.
Definition 4. Let hAR, Ai be an argumentation framework. Let R, S, subsets of AR.
An argument s ∈ S is attacked by R if there is some r ∈ R such that hr, si ∈ A. An
argument x ∈ AR is acceptable with respect to S if for every y ∈ AR that attacks x
there is some z ∈ S that attacks y. S is conflict free if no argument in S is attacked by
any other argument in S. A conflict free set S is admissible if every argument in S is
acceptable with respect to S. S is a preferred extension if it is a maximal (with respect
to set inclusion) admissible subset of AR.
In addition, an argument x is credulously accepted if there is some preferred exten-
sion containing it; whereas x is sceptically accepted if it is a member of every preferred
extension. The key notion here is the preferred extension which represents a consistent
position within AF , which is defensible against all attacks and which cannot be further
extended without becoming inconsistent or open to attack.
In order to take into account that, for a given situation, agents might have different
point of view, we are concerned by a set of audiences, which adhere to different argu-
ment with a different strengths. Therefore we use a Value-based Argumentation Frame-
work , which prescribes different strengths to arguments on the basis of the values they
promote and the ranking given to these values by the audience for the argument. This
allows us to systematically relate strengths of arguments to their motivations, and to
accommodate different audiences with different interests and preferences.
Definition 5. A Value-Based Argumentation Framework (V AF ) is defined as hAR, A, V, ηi,
where (AR, A) is an argumentation framework, V is a set of k values which represent
the types of arguments and η: AR → V is a mapping that associates a value η(x) ∈ V
with each argument x ∈ AR
In section 3, the set of values V will be defined as the different types of ontology mis-
match, which we use to define the categories of arguments and to assign to each argu-
ment one category.
Definition 6. An audience for a V AF is a binary relation R ⊂ V × V whose (irreflex-
ive) transitive closure, R∗ , is asymmetric, i.e. at most one of (v, v 0 ), (v 0 , v) are members
of R∗ for any distinct v, v 0 ∈ V. We say that vi is preferred to vj in the audience R,
denoted vi R vj , if (vi , vj ) ∈ R∗ .
Let R be an audience, α is a specific audience (compatible with R) if α is a total
ordering of V and ∀ v, v 0 ∈ V, (v, v 0 ) ∈ α ⇒ (v 0 , v) 6∈ R∗
In this way, we take into account that different agents (represented by different au-
diences) can have different perspectives on the same candidate mapping. Acceptability
of an argument is defined in the following way: 9
Definition 7. Let hAR, A, V, ηi be a V AF and R an audience.
a. For arguments x, y in AR, x is a successful attack on y (or x defeats y) with respect
to the audience R if: (x, y) ∈ A and it is not the case that η(y) R η(x).
b. An argument x is acceptable to the subset S with respect to an audience R if: for
every y ∈ AR that successfully attacks x with respect to R, there is some z ∈ S
that successfully attacks y with respect to R.
c. A subset S of AR is conflict-free with respect to the audience R if: for each (x, y) ∈
S × S, either (x, y) 6∈ A or η(y) R η(x).
d. A subset S of AR is admissible with respect to the audience R if: S is conflict free
with respect to R and every x ∈ S is acceptable to S with respect to R.
e. A subset S is a preferred extension for the audience R if it is a maximal admissible
set with respect to R.
f. A subset S is a stable extension for the audience R if S is admissible with respect to
R and for all y 6∈ S there is some x ∈ S which successfully attacks y with respect
to R.
In order to determine whether the dispute is resolvable, and if it is, to determine the
preferred extension with respect to a value ordering promoted by distinct audiences, [2]
introduces the notion of objective and subjective acceptance as follows:
9
Note that all these notions are now relative to some audience.
Definition 8. Given a V AF , hAR, A, V, ηi, an argument x ∈ AR is subjectively ac-
ceptable if and only if, x appears in the preferred extension for some specific audiences
but not all. An argument x ∈ AR is objectively acceptable if and only if, x appears
in the preferred extension for every specific audience. An argument which is neither
objectively nor subjectively acceptable is said to be indefensible.
Next, we define the various types of arguments that can be distinguished for sup-
porting or attacking correspondences.
3 Arguments for Correspondences
Potential arguments are clearly identified and grounded on the underlying ontology lan-
guage OWL. Therefore, the grounds justifying correspondences can be extracted from
the knowledge in ontologies10 . Our classification of the grounds justifying correspon-
dences is the following:
semantic (M ): the sets of models of two entities do or do not compare;
internal structural (IS): two entities share more or less internal structure (e.g., the
value range or cardinality of their attributes);
external structural (ES): the set of relations, each of two entities have, with other
entities do or do not compare;
terminological (T ): the names of two entities share more or less lexical features;
extensional (E): the known extension of two entities do or do not compare.
These categories correspond to the type of categorizations underlying ontology match-
ing algorithms [18]. In our framework, we will use the types of arguments described
above as types for the V AF ; hence V = {M, IS, ES, T, E}. For example, an audi-
ence may specify that terminological arguments are preferred to semantic arguments,
or vice versa. Note that this may vary according to the nature of the ontologies being
aligned. Semantic arguments will be given more weight in a fully axiomatised ontology,
compared to that in a lightweight ontology where there is very little reliable semantic
information on which to base such arguments.
Table 1 presents a sample set of argument schemes, instantiations of which will
comprise AR. Attacks between these arguments will arise when we have arguments
for the same mapping but with conflicting values of σ, thus yielding attacks that can
be considered symmetric. Moreover, the relations in the mappings can also give rise to
attacks: if relations are not deemed exclusive, an argument against inclusion is a fortiori
an argument against equivalence (which is more general).
Example 1. Consider a candidate mapping m = hc, c0 , , ≡i between two OWL ontolo-
gies O1 and O2 , with concepts c and c0 respectively. An argument for accepting the
mapping m may be that the labels of c and c0 are synonymous. An argument against
may be that some of their super-concepts are not mapped.
In V AF s, arguments against or in favour of a candidate mapping are seen as grounded
on their type. In this way, we are able to motivate the choice between preferred ex-
tensions by reference to the type ordering of the audience concerned. Moreover, the
10
This knowledge includes both the extensional and intensional OWL ontology definitions.
pre-ordering of preferences Pref for each agent will be over V, that corresponds to the
determination of an audience.
Table 1. Argument scheme for OWL ontological alignments
Mapping σ Grounds Comment
he, e0 , n, ≡i + ∃mi = hES(e), ES(e0 ), n0 , ≡i e and e0 have mapped neighbours (e.g., super-entities,
sibling-entities, etc.) of e are mapped in those of e0
he, e0 , n, vi + ∃mi = hES(e), ES(e0 ), n0 , ≡i (some or all) Neighbours (e.g., super-entities, sibling-entities,
etc.) of e are mapped in those of e0
hc, c0 , n, vi + ∃mi = hIS(c), IS(c0 ), n0 , ≡i (some or all) Properties of concept c are mapped to those
of concept c0
hc, c0 , n, vi - 6 ∃mi = hIS(c), IS(c0 ), n0 , ≡i No properties of c are mapped to those of c0
he, e0 , n, ≡i + ∃mi = hE(e), E(e0 ), n0 , ≡i (some or all) Instances of e and e0 are mapped
he, e0 , n, vi + ∃mi = hE(e), E(e0 ), n0 , ≡i (some or all) Instances of e are mapped to those of e0
he, e0 , n, ≡i + label(e) ≈T label(e0 ) Entities’s labels share lexical features (e.g., synonyms
and lexical variants)
0
he, e , n, vi
he, e0 , n, ≡i - label(e) 6≈T label(e0 ) Entities’ labels do not share lexical features (e.g., homonyms)
he, e0 , n, vi
Although in V AF s there is always a unique non-empty preferred extension with
respect to a specific audience, provided the AF does not contain any cycles in a sin-
gle argument type, an agent may have multiple preferred extensions either because no
preference between two values in a cycle has been expressed, or because a cycle in a
single value exists. The first may be eliminated by committing to a specific audience,
but the second cannot be eliminated in this way. In our domain, where many attacks are
symmetric, two cycles will be frequent and in general an audience may have multiple
preferred extensions.
Thus, given a set of arguments justifying mappings organised into an argumenta-
tion framework, an agent will be able to determine which mappings are acceptable by
computing the preferred extensions with respect to its preferences. If there are multiple
preferred extensions, the agent must commit to the arguments present in all preferred
extensions, but it has some freedom of choice with respect to those in some but not all
of them.
Based on the above considerations, we thus define an agreed correspondence and an
agreeable correspondence as follows. An agreed correspondence is the correspondence
supported11 by those arguments which are in every preferred extension of every agent.
An agreeable correspondence is the correspondence supported by arguments which are
in some preferred extension of every agent. Thus, the agents will reach a common con-
sensus over a specific mapping m only if the mapping m is an agreed correspondence.
However, if a mapping m is an agreeable correspondence for a given agent Ag, this
mean that such mapping can only be considered valid and consensual for that agent.
In the next section, we present a model of agents which put forward arguments and
take into account other arguments coming from their interlocutors.
4 Model of Persuasive Agents
In this paper, we are assuming a multi-agent setting containing persuasive agents that
do not use the same ontology. Each agent considers the repertoire of argument schemes
11
Note that a correspondence m is supported by an argument x if x is hG, m, +i
available to it, and is able to generate a set of arguments and counter-arguments by
instantiating these schemes with respect to its interests. Moreover, the agents can record
their interlocutors arguments in a commitment store CS [14] and individually evaluate
them. Therefore, our persuasive agent can be defined as follows:
Definition 9. An agent Agi is defined by a 5-tuple hOi , V AFi , CSji , Pref, εi where Oi
is the private ontology; V AFi = hARi , Ai , V, ηi is the Valued-based Argumentation
Framework of the agent Agi ; CS ij is a commitment store, i.e. a set of arguments where
CSji (t) contains propositional commitments taken before or at time t between the Agi
and other interlocutors; P ref is the private pre-ordering of preferences over V and ε
is the private threshold value.
The set of arguments are not necessarily disjoint.
T The set of arguments shared by all
agents are called common arguments: ARc ⊆ x∈ARi ARi ∈ V AF i . Instead, the
values V = {M, IS, ES, T, E} are common and shared by all audiences.
In order to take into account the arguments notified in the commitment stores, we
extend the definition of valued-based argumentation framework with the following:
Definition 10. An extended Value-Based Argumentation Framework V AF + is defined
+ + i
as hAR , A , V, η i, where AR = AR ∪ { j6=i CS j }. The definition of A+ and η +
+ +
S
are now related to AR+
Now, we can define the notion of conviction as follows:
Definition 11. Let Agi be an agent associated with the extended valued-based argu-
mentation framework, V AF + and x be an argument provided by another agent Agj .
The agent Agi is convinced by the argument x iff x is acceptable with respect to all
audience R, with Agi ∈ R .
Given this model, in order to determine the acceptability of a potential correspon-
dence, it needs to proceed by means of a dialectical exchange, in which a mapping is
proposed, challenged and defended. Many argument protocols have been proposed, e.g.
[15]. Particular dialogue games have been proposed based on Dung’s Argumentation
Frameworks, e.g. [8], and on VAFs [3].
In this paper, we are not considering any specific protocol or persuasive dialogue.
However, the idea of a dialogue is that agents reply to each other in order to reach the
interaction goal, i.e. an agreement. Thus, given a set of social and autonomous agents,
and a set of potential correspondences {m1 , . . . , mi , . . .}, an agent initiates a persuasion
dialogue when it wants present its viewpoint to the other agents. Specifically, for each
mapping mi , if the agent wants to accept that mapping, it will put forward arguments
for mi . In the negative case, it will put forward arguments against. If the other agents
have no arguments against/for the mapping, it closes the dialogue. If the players have
the same convictions, the the arguments is acceped and the dialogue closes. Otherwise,
the goal of the dialogue is the resolution of the conflict by verbal means, and thus with
an exchange of arguments and counter-arguments.
The dialogue between agents can thus consist simply of the exchange of individual
arguments, from which they can compute acceptable mappings over the CS, by com-
puting the preferred extensions. If necessary and desirable, these can then be reconciled
into a mutually acceptable position through a process of negotiation, as suggested in
[6], which defines a dialogue process for evaluating the status of arguments in a V AF ,
and shows how this process can be used to identify mutually acceptable arguments.
In [13] a detailed approach to argue over alignments and complete argumentation
framework, with a common set of arguments, is proposed.
5 A Walk through Example
Let us assume that some agents or services need to interact with each other using two
independent but overlapping ontologies. The first agent, Ag1 uses the bibliographic on-
tology12 from the University of Toronto, based on bibTeX; whereas the second agent,
Ag2 , uses the General University Ontology13 from Mondeca14 . For space reasons, we
only consider a subset of these ontologies, shown in Table 2, where the first and second
ontologies are represented by O1 and O2 respectively.
We will assume that the set of candidate mappings, provided by the Ontology Align-
ment Service (OAS), is the following::
m1 =hO1 : P ress, O2 : P eriodical, n, =i; 15
m2 =hO1 : publication, O2 : P ublication, n, =i;
m3 =hO1 : hasP ublisher, O2 : publishedBy, n, =i;
m4 =hO1 : M agazine, O2 : M agazine, n, =i;
m5 =hO1 : N ewspaper, O2 : N ewspaper, n, =i;
m6 =hO1 : Organization, O2 : Organization, n, =i.
The generation of the arguments and counter-arguments of the Ag1 and Ag2 are achieved
by instantiating the argumentation schemes, discussed previously, with respect to the
agent’s preferences and threshold. However, here we assume a degree of confidence n
that is above the threshold of both agent, and so will not influence their acceptability.
Assume now that there are two possible audiences, R1 , which prefers terminology to
external structure, (T R1 ES), and R2 , which prefers external structure to terminol-
ogy (ES R2 T ). The pre-ordering of preference Pref will correspond to the agent’s
audience. The agents Ag1 and Ag2 take on the part, respectively, of the audience R1
and R2 . For space reasons, we will only evaluate the mapping m1 16 .
The argumentation starts, with the agent Ag1 that wants to reject the mapping m1
and will thus argue against it, forwarding an argument A. A states that none of the super-
concepts of the concept O1 : P ress are mapped to any super-concept of O2 : P eriodical.
The agent Ag2 , instead, does not agree and counter-argues with an argument B. B ar-
gues for m1 , because two sub-concepts of O1 : P ress, O1 : M agazine and O1 : N ewspaper,
are mapped to two sub-concepts of O2 : P eriodical, O2 : M agazine and O2 : N ewspaper,
as established by m4 and m5 . The agent Ag1 attacks B with the argument C, be-
cause O1 : P ress and O2 : P eriodical do not have any lexical similarity. The agent Ag2
12
http://www.cs.toronto.edu/semanticweb/maponto/ontologies/BibTex.owl
13
http://www.mondeca.com/owl/moses/univ.owl
14
Note that ontology O2 has been slightly modified for the purposes of this example.
15
m1 states an equivalence correspondence with confidence n between the concept P ress in
the ontology O1 and the concept P eriodical in the ontology O2
16
An extended version of this example is provided in [13] .
Table 2. Excerpts of O1 and O2 ontologies
O1 Ontology O2 Ontology
Artif act v > Document v >
P rint M edia v Artif act P ublication v Document
P ress v P rint M edia P eriodical v P ublication
M agazine v P ress M agazine v P eriodical
N ewspaper v P ress N ewspaper v P eriodical
publication v ∀hasP ublisher.P ublisher N ewsletter v P eriodical
publication v P rint M edia Journal v P eriodical
P ublisher v Organization P ublication v Document
P ublication v ∀publishedBy.Organization
does not have any other argument to reply to C but it supports the correspondences
m4 , m5 and m6 by six arguments. K, L and M justify the mapping m4 , since, re-
spectively, the labels of O1 : M agazine and O2 : M agazine are lexically similar; their
siblings are mapped, as established by m5 , and their super-concepts; O1 : P ress and
O2 : P eriodical are mapped by m1 . There is a similar situation for the arguments M ,
N and O. Clearly, argument A attacks the arguments D and I.
This position is illustrated in Figure 2, where nodes represent arguments (labelled
with their Id) with the respective type value V. The arcs represent the attacks A, whereas
the direction of the arcs represents the direction of the attack.
Table 3 shows these arguments, labelled with an identifier Id, its type V, and the
attacks A that can be made on it by opposing arguments.
Table 3. Arguments for and against the correspondences m1 , m4 and m5
Id Argument A V
A h6 ∃m = hsuperconcept(P ress), superconcept(P eriodical), n, ≡, i, m1 , −i B,D,I ES
B h∃m = hsubconcept(P ress), subconcept(P eriodical), n, ≡, i, m1 , +i A,C ES
C hLabel(P ress) 6≈T Label(P eriodical), m1 , −i B T
D hLabel(M agazine) ≈T Label(M agazine), m4 , +i T
E h∃m = hsiblingConcept(M agazine), siblingConcept(M agazine), n, ≡, i, m4 , +i ES
F h∃m = hsuperconcept(M agazine), superconcept(M agazine), n, ≡, i, m4 , +i ES
G hLabel(N ewspaper) ≈T Label(N ewspaper), m5 , +i T
H h∃m = hsiblingConcept(N ewspaper), siblingConcept(N ewspaper), m5 , +i ES
I h∃m = hsuperconcept(N ewspaper), superconcept(N ewspaper), n, ≡, i, m5 , +i ES
Fig. 2. Value-Based Argumentation Frameworks
Finally, we can compute the acceptability of the arguments, computing the pre-
ferred extensions (see Table 5). Therefore, the arguments accepted by both audiences
Table 4. Preferred Extensions
Preferred Extensions for the framework (a) Audience
{A, C, D, E, G, H} R1
{A, C, D, E, F, G}, {B, I, D, E, F, G} R2
{A, C, D, E, F, G}, {B, I, D, E, F, G}
are {D, E, G, H}. Arguments A, C are, however, both potentially acceptable, since
both audiences can choose to accept them, as they appear in some preferred extension
for each audience. This means that the mapping m1 will be rejected for the agent Ag1
(since B is unacceptable to R1 ), while the mappings m4 and m5 will both be accepted
(they are both accepted by R1 and both acceptable to R2 ). The agreed correspondence
are then m4 and m5 .
6 Summary and Outlook
In this paper we have outlined a framework that provides a novel way for agents, who
use different ontologies, to argue and reach agreement over ontology alignment. This is
achieved using an argumentation process in which candidate correspondences are ac-
cepted or rejected, based on the ontological knowledge and the agent’s preferences. Ar-
gumentation is based on the exchange of arguments, against or in favour of a correspon-
dence, that interact with each other using an attack relation. Each argument instantiates
an argumentation schema, and utilises domain knowledge, extracted from extensional
and intensional ontology definitions.
Our approach is able to give a formal motivation for the selection of a correspon-
dence, and enables consideration of an agent’s interests and preferences that may in-
fluence the selection of a correspondence. We believe that this approach will aim at
reaching mutual understanding and communicative work in agents system more sound
and effective. Future work will include experimental testing in order to demonstrate the
practicality of our approach. An interesting topic for future work would be to investi-
gate how to argue about the whole alignments, and not only the individual candidate
mapping.
7 Acknowledgements
The research has been partially supported by Knowledge Web (FP6-IST 2004-507482)
and PIPS (FP6-IST 2004-507019). Special thanks to Floriana Grasso and Ian Blacoe
for their comments.
.
References
1. K. Aberer and et al. Emergent semantics principles and issues. In Proceedings of Database
Systems for Advances Applications, 9th International Conference, DASFAA 2004, 2004.
2. T. Bench-Capon. Persuasion in practical argument using value-based argumentation frame-
works. In Journal of Logic and Computation, volume 13, pages 429–448, 2003.
3. T. J. M. Bench-Capon. Agreeing to differ: Modelling persuasive dialogue between parties
without a consensus about values. In Informal Logic, volume 22, pages 231–245, 2002.
4. T. Berners-Lee, J. Hendler, and O. Lassila. The semantic web. Scientific American,
284(5):34–43, 2001.
5. R. Dhamankar, Y. Lee, A. Doan, A. Halevy, and P. Domingos. imap. In Proceedings of the
International Conference on Management of Data (SIGMOD), pages 383–394.
6. S. Doutre, T. Bench-Capon, and P. E. Dunne. Determining preferences through argumenta-
tion. In Proceedings of AI*IA’05, pages 98–109, 2005.
7. P. Dung. On the acceptability of arguments and its fundamental role in nonmonotonic rea-
soning, logic programming and n-person games. In Artificial Intelligence, volume 77, pages
321–358, 1995.
8. P. Dunne and T. J. M. Bench-Capon. Two party immediate response disputes: Properties and
efficiency. In Artificial Intelligence, volume 149, pages 221–250, 2003.
9. M. Ehrig and S. Staab. Qom - quick ontology mapping. In Proceedings of the International
Semantic Web Conference, 2004.
10. J. Euzenat. Alignment infrastructure for ontology mediation and other applications. In
Hepp, editor, Proceedings of the First International workshop on Mediation in semantic web
services, 2005.
11. J. Euzenat and P. Valtchev. Similarity-based ontology alignment in owl-lite. In Proceedings
of European Conference on Artificial Intelligence (ECAI 04), 2004.
12. T. R. Gruber. A translation approach to portable ontology specifications. Knowledge Acqui-
sition, 5(2):199–220, 1993.
13. L. Laera, V. Tamma, J. Euzenat, T. Bench-Capon, and T. Payne. Reaching agreements over
ontology alignments. In Proceedings of the Fifth International Semantic Web Conference
(ISWC’06), 2006.
14. N. Maudet and B. Chaib-draa. Commitment-based and dialogue-game based protocols-news
trends in agent communication language. Knowledge Engineering Review, 17(2):157–179,
2003.
15. P. McBurney and S. Parsons. Locutions for argumentation in agent interaction protocols. In
Proceedings of International Workshop on Agent Communication, New-York (NY US), pages
209–225, 2004.
16. P. Shvaiko and J. Euzenat. A survey of schema-based matching approaches. Journal on data
semantics, 4:146–171, 2005.
17. P. Shvaiko, F. Giunchiglia, P. Pinheiro da Silva, and D. McGuinness. Web explanations for
semantic heterogeneity discovery. In Proceedings of ESWC, pages 303–317, 2005.
18. P. Visser, D. Jones, T. Bench-Capon, and M. Shave. Assessing heterogeneity by classifying
ontology mismatches. In N. Guarino, editor, Proceedings of the FOIS’98, 1998.