Modeling semantic applications that go beyond just triples
Semantic modeling to date has been largely an exercise in considering the whole world as triples (or at least, n-way tuples). Efforts to bridge the XML gap such as XSPARQL have focused on low-hanging syntactic fruit, and have not had much effect on deeper layers of the architecture.
In semantic modeling, common situations that are supposedly bread-and-butter for RDF, for example tracking provenance or recording facts known to a certain probability, give rise to complexity such as the much reviled RDF reification, or statements about statements. Even the W3C RDF Primer doesn't seem altogether comfortable with it:
While there are applications that successfully use reification, they do so by following some conventions, and making some assumptions, that are in addition to the actual meaning that RDF defines for the reification vocabulary, and the actual facilities that RDF provides to support it.
In a similar vein, inference is commonly viewed in the semantic world as something that produces (only) triples from (only) triples. A broader view of inference encompassing XML documents and values as inputs and outputs can make many common use cases far more straightforward. This paper discusses a (hypothetical?) world where triples and documents get along better with each other, and speculates about what future products fusing these technologies might look like.