One of the main tasks of mediators is to fuse information from heterogeneous
information sources. This may involve, for example, removing redundancies,
and resolving inconsistencies in favor of the most reliable source.
The problem becomes harder when the sources are unstructured/semistructured
and we do not have complete knowledge of their contents and structure.
In this paper we show how many common fusion operations can be specified
non-procedurally and succinctly. The key to our approach is to assign
semantically meaningful object ids to objects as they are ``imported''
into the mediator. These semantic ids can then be used to specify how various
objects are combined or merged into objects ``exported'' by the mediator.
In this paper we also discuss the implementation of a mediation system
based on these principles. In particular, we present key optimization
techniques that significantly reduce the processing costs associated
with information fusion.