How Servi
e Choreography Statisti
s Redu
e the
Ontology Mapping Problem
Paolo Besana and Dave Robertson
S
hool of Informati
s, University of Edinburgh
Abstra
t In open and distributed environments ontology mapping pro-
vides interoperability between intera
ting a
tors. However,
onventional
mapping systems fo
us on a
quiring stati
information, and on map-
ping whole ontologies, whi
h is infeasible in open systems. This paper
shows that the intera
tions themselves between the a
tors
an be used to
predi
t mappings, simplifying dynami
ontology mapping. The intuitive
idea is that similar intera
tions follow similar
onventions and patterns,
whi
h
an be analysed. The
omputed model
an be used to suggest the
possible mappings for the ex
hanged messages in new intera
tions. The
suggestions
an be evaluate by any standard ontology mat
her: if they
are a

urate, the mat
hers avoid evaluating mappings unrelated to the
intera
tion.
The minimal requirement in order to use this system is that it is possible
to des
ribe and identify the intera
tion sequen
es: the OpenKnowledge
proje
t has produ
ed an implementation that demonstrates this is pos-
sible in a fully peer-to-peer environment.
1 Introdu
tion
Most ontology mapping systems [9,15℄ available for the semanti
web and for
semanti
web servi
es fo
us on a
quiring stati
, a priori information about map-
pings. Depending on the approa
h, mat
hers
ompare labels, ontology stru
tures
[10℄, use external di
tionaries like WordNet to prove similarity between nodes
in hierar
hies [7℄, learn how instan
es are
lassied to nd similarities between
on
epts [5℄ or
ombine information from dierent sour
es [4,6℄. In an open and
distributed environment ontology mapping systems aim at providing interoper-
ability between intera
ting a
tors, ea
h with possibly a dierent ontology. Map-
ping in advan
e, before the intera
tions, is unfeasible, as the agents may be still
unknown. Mapping during the intera
tions may be
omputationally di
ult, as
many intera
tions with dierent a
tors
an go on simultaneously.
This paper shows that the intera
tions between the a
tors
an be used to
predi
t the mappings, making the problems related to dynami
ontology map-
ping more tra
table. The intuitive idea is that intera
tions follow
onventions and
patterns, and these patterns are repeated when similar situations arise. The pat-
terns are extra
ted by analysing the intera
tions in order to model the relations
between the terms that appear in them. If the
omputed model is representative
of a
lass of intera
tions, then it
an provide the basis for predi
ting the
ontent
of ex
hanged messages in future intera
tions. A predi
tion is a list of suggestions

for the mapping, that any standard ontology mat
her
an evaluate. If the predi
-
tions are a

urate, the mat
hers
an avoid evaluating mappings unrelated to the
intera
tion, improving e
ien
y and de
reasing ambiguity. In fa
t, the
ontext
of the intera
tion provides additional information, that
an help in those
ases
in whi
h the mat
hers do not have enough stati
information to distinguish the
orre
t mapping among many possible ones.
This paper shows that, after a reasonably small number of intera
tions the
predi
tor
onsistently provides reliable suggestions. The minimal requirement
in order to use this system is that it is possible to des
ribe and identify the
intera
tion sequen
es. In prin
iple, any system based on workow language
an
provide this. Workow systems normally are
entralised, but we have re
ently
shown, in the EU funded OpenKnowledge proje
t
1
that it is possible to a
hieve
peer-to-peer based workow systems, as a means of web servi
e
horeography.
In this paper, we rst des
ribe, in Se
tion 2, the intuitive notions of dia-
logue and intera
tion behind our work; then, in Se
tion 3, we briey dis
uss the
alternative approa
hes for agents
ommuni
ation, introdu
ing the OpenKnowl-
edge peer-to-peer framework for dening and exe
uting intera
tions. Se
tion 4
denes the
on
epts and terms used in modelling the
ontext of intera
tions,
while Se
tion 5 des
ribes what needs to be modelled, and how to model it with
an example. Se
tion 6 denes what needs to be evaluated, then reports how the
testing was stru
tured and next presents and interprets the results.
2 Servi
es' intera
tions
Many a
tivities require intera
tion between dierent a
tors: for example, in order
to book a room for a
onferen
e an inquirer needs to
onta
t a travel agen
y (or
more than one) or dire
tly a number of hotels.
In the simplest version,
ommuni
ation between two agents is a message
transmitted from a sender to a re
eiver. A

ording to the spee
h a
t theory,
a message is a performative a
t that
hanges the state of the world [14℄. For
example, a message sent from agents i to agent j to inform about φ will likely
hange the beliefs of j, adding the belief about φ. In our example, the following
message, sent from Mr Smith's agent to the agent representing the hotel Y:
inform(booking, 11 Nov 2007, 15 Nov 2007, Mr Smith, single)
should make the hotel agent believe that a single room must be reserved for Mr
Smith from the 11th to the 15th of November. Or at least this is what Mr Smith
thinks. But, for example, the hotel agent may not know the meaning of booking
or single, or it may use a dierent format for dates. To over
ome this problem,
either all agents that
onta
t the hotel servi
e must share the same ontology,
whi
h is not feasible in an open environment where agents from dierent ba
k-
grounds may intera
t, or the agents must have a

ess to the mappings between
dierent ontologies.
Unfortunately, it is infeasible in an open system to pre
ompute all mappings,
as it is impossible to fore
ast whi
h agents will
onta
t the hotel servi
e, so some
1
www.openk.org