Facilitating Interaction Between Virtual Agents through Negotiation over Ontological Representation

It is universally acknowledged that the problem of integration of information across large communities is a difficult and pressing one, particularly when these communities are disparate, widespread and not under centralised control, such as in the Semantic Web (Berners-Lee et al, 2001). The simplest solution to this problem is the enforcement of a single ontology: a single view of the world. However, if the agents interacting are from different organisations or fields, attempting to use a single ontology is usually neither practical nor desirable. Users need to develop a representation that is best suited to their own problems and they need to maintain and update that representation locally. Even if all users do subscribe to a single ontology, integration problems still exist, as changes and updates are made and users tune their ontologies to fit their own needs.

The problem of ontology matching has been widely studied and powerful solutions are available (see (Shvaiko and Euzenat, 2013) and (Euzenat and Shvaiko, 2013) for a comprehensive survey. However, the ontologies considered are almost always taxonomies, and the problem of ontology matching is concerned with relating a single term in one ontology to one or more terms in another ontology: for example, a term car in one ontology may relate to a term automobile or a term carriage in another ontology. Much less considered is the problem of relating compound terms such as first-order terms or database entries with multiple fields: for example, a term car(make,model) in one ontology relating to a term automobile(model,year,brand) in another. In such situations we still have the problem of relating the single terms contained within these compound terms – e.g., this matching depends on knowing that car may be related to automobile and that make may be related to brand. But we must also consider the overall relation of the compound terms, which requires not only semantic but also structural matching.

Another drawback of traditional ontology matching in an online environment is that it tends to assume full knowledge of all relevant ontologies and is generally performed off-line, prior to interaction. These are the assumptions made by the main evaluation processes for Ontology Matching, such as the Ontology Alignment Evaluation Initiative (OAEI ). But in large, fast-moving agent communities, or situations where some information may be confidential, we cannot assume that we can have full knowledge of any agent or service we may interact with, nor is it possible to perform the matching off-line if we may not know prior to interaction which agents will need to interact.

In this paper, we introduce our theory of on-the-fly, structured matching and briefly describe the ORS system, which we have developed to implement this theory. Our central hypothesis is that representation – as well as vocabulary and beliefs – must be treated as a fluent and that automated, dynamic matching techniques that can map between structured terms are necessary for full integration of disparate ontologies (Bundy et al, 2006).