MaLM: Machine Learning Middleware to Tackle Ontology Heterogeneity

Abstract

We envisage pervasive computing applications to be predominantly engaged in knowledge-based interactions, where services and information will be found and exchanged based on some formal knowledge representation. To enable knowledge sharing and reuse, current middleware make the assumption that a single, universally accepted, ontology exists with which queries and assertions are exchanged. We argue that such an assumption is unrealistic. Rather, different communities will speak different `dialects'; in order to enable cross-community interactions, thus increasing the range of services and information available to users, on-the-fly translations are required. In this paper we introduce MaLM, a middleware for pervasive computing devices that exploits an unsupervised machine learning technique called self-organising map to tackle the problem of ontology heterogeneity. At any given time, the MaLM instance running on a device operates in one of two possible modes: `training', that is, MaLM is autonomically learning how to group together semantically closed concepts; and `expert', that is, given in input a query or assertion expressed in a foreign dialect, MaLM identifies the concept, expressed in the device mother-tongue, that most closely represents it