I see "add semantics" to documents, I hear "add content"...
Actually I have an inkling that some "going down the wrong path" is afoot here.
I've been tinkering, still quite in the theoretical / philosophical sense, with N-triples for thingamy lately, so let me have a go at it. How I skew the situation so to speak. (BTW, here's another triples tutorial)
I see three levels of "information" bearers, or rather three ways to go about supplying information:
1) Documents and forms
This is where we are today, the classic event documentation, ideas summaries or collection of information.
As such they need much on the side to make real use of them. To find them and make use of them they need much knowledge added for which we use tree-structures, tags and links.
If we were to use the semantic web ideas for these invented-for-paper information bearers it would be as advanced tags, added-content-snippets. Neat for sure, but only neat.
Alchemy, no atoms yet.
Atomise a document or form and you can find objects.
This is were we stand with thingamy, cutting up the documents and forms into separate objects each representing a finite real world object. "Address" being a single object representing your house but linked to you and then to your bank account and life insurance.
A few objects can create many, many different documents. Patient object + all the different condition objects and x-ray objects and so forth can create a Patient Journal or some of them can be part of a specific research paper. And so forth.
Having thus one single object per real-world object there is much less need for additional knowledge about the object. Your address is a house. Your broken arm is a condition object.
But still not there, like early nuclear science.
3) (Semantic) Triples
Now, triples consisting of subject + predicate + object can sub-atomise the atomic object.
The beauty of these sub-atoms is that they in themselves are semantic (duh) thus requiring even less knowledge to be added for usefulness. In fact the semantics are much closer to human ways of knowledge transfer than anything else - stories, relationships.
In other words, to use the conceptual ideas stemming from the semantic web to add knowledge, content or semantics to documents or forms would be a waste of something intrinsically more valuable methinks.
Objects can be mashups of N-triples, documents and forms are mashups of objects.
Thus if all is built from bottom up - using say N-triples - then all the objects and documents could be built from fewer particles adding much knowledge enabling simpler finding and better use all in one go.
There is a reason why they invest so much at CERN these days, the days of alchemy should officially be over ;)