Probabilistic Dialogue Models for Dynamic
Ontology Mapping
Paolo Besana and Dave Robertson
Centre for Intelligent Systems and Applications
University of Edinburgh
Abstract. Agents need to communicate in order to accomplish tasks
that they are unable to perform alone. Communication requires agents
to share a common ontology, a strong assumption in open environments
where agents from different backgrounds meet briefly, making it impos-
sible to map all the ontologies in advance. An agent, when it receives a
message, needs to compare the foreign terms in the message with all the
terms in its own local ontology, searching for the most similar one. Ho-
wever, the content of a message may be described using an interaction
model: the entities to which the terms refer are correlated with other
entities in the interaction, and they may also have prior probabilities
determined by earlier, similar interactions. Within the context of an in-
teraction it is possible to predict the set of possible entities a received
message may contain, and it is possible to sacrifice recall for efficiency
by comparing the foreign terms only with the most probable local ones.
This allows a novel form of dynamic ontology matching.
1 Introduction
Agents collaborate and communicate to perform tasks that they cannot accom-
plish alone. To communicate means to exchange messages, that convey meanings
encoded into signs for transmission. To understand a message, a receiver should
be able to map the signs in the message to meanings aligned with those intended
by the transmitter.
Therefore agents should agree on the terminology used to describe the domain
of the interaction: for example, if an agent wants to buy a particular product from
a seller, it must be able to specify the properties of the products unambiguously.
Ontologies specify the terminology used to describe a domain [4].
However, a shared ontology can be a strong assumption in an open envi-
ronment, such as a Peer-to-Peer system: agents may come from different back-
grounds, and have different ontologies, designed for their specific needs [13].
In this sort of environment, communication implies translation. The standard
approach is to find mappings between the ontologies, creating a sort of bilingual
dictionary. Many different techniques have been developed for ontology mapping,
but in an open environment it is impossible to know which agents will take part in
the interactions; therefore it is impossible to anticipate which ontologies should
be mapped.
P.C.G. da Costa et al. (Eds.): URSW 2005-2007, LNAI 5327, pp. 41–51, 2008.
c
© Springer-Verlag Berlin Heidelberg 2008

42 P. Besana and D. Robertson
product
consumer electronicscomputers ...
photographymobiles ...pc laptop
digital analog
digital_SLRdigital_compact
has_brand
has_sensor_resolution
has_optical_zoom
has_cpu
has_ram has_hard_disk
has_weight
Fig. 1. Fragment of buyer ab ontology
Thing
Cell_Phones Cameras Computers...
Analog_CamerasDigital_Cameras ... Desktop_PCsNotebooks ...
BrandLensResolution Hard_Drive_CapacityMemory_Ram Processor_Speed
Weight
Fig. 2. Fragment of seller as ontology
Agents have to map ontologies dynamically when needed. Mapping full on-
tologies is a time-consuming task: in the standard process, each term in one
ontology is compared with all the terms in the other ontology, and the most
similar term is the mapping.
However, agents may meet infrequently, for a single interaction on a specific
topic. A full ontology mapping would be a waste of resources: mapping only
“foreign” terms that have appeared in the conversation can be more convenient.
Comparing a foreign term in a message with all the terms in the ontology can
still be costly. Yet, the entities referred by the signs in the message are not ran-
domly chosen: the dialogue has a meaning because entities are related. For exam-
ple, if the conversation is about the purchase of a laptop, entities related to cars
are unlikely to appear. It is reasonable to compare the signs in the message with
entities about laptops, rather than compare with all the entities indiscriminately.
This paper shows how to extract, represent, and use knowledge about the
relations and properties of the entities in an interaction to support dynamic
ontology mapping.
2 Example Scenario
The example scenario is a purchase interaction between the buyer and seller
agents ab and as. In the dialogue, agent ab asks as about a laptop he needs. The
seller as inquires about properties of the product in order to make an offer.
The two agents do not share the same ontology: the buyer uses the one in
Figure 1 and the seller the one in Figure 2. In the figures the ovals are classes
and the grey boxes are properties. The classes are structured in taxonomies, and
the domains of the properties are shown by grey arrows.