Hi, Peter. I am not sure if you receive comments via email, so I decided to reply through email to your post "XML, trees and laticces" post (http://poorbuthappy.com/ease/archives/002665.html).
Actually, it is done not via XML, but using external storage for your selections. Thus you can store them in some order and quite easily manipulate. But there's no tools to do this currently, as far as I know.
It's like "selection history" (or scope history, or whatever). You store it and then you can see them all at once, or in the order of creation, or in the order of length or in any other order. Some basic example here: http://www.livejournal.com/users/urbansheep/1045767.html
The main fun thing here is that selections aren't actually "semantic mark-up", they are indeed representing content layer above the mark-up (since selections are often very subjective). And they really can overlap as ideas and concepts are so very fluid in the flow of narration, absolutely unlike such basic "semantic" elements as paragraphs or empasises.
The second fun thing is that those scopes capture actual ideas and can tag them, therefore your keywords can be bind not to the document itself, but to specific position in the document. Well known shortcoming of keywords and categories is that lengthy documents that cover broad number of topics become a horrible mess of keywords and categories when one tries to meta-process it (lard it with metadata). Therefore the need arises to split it in a couple of more narrow-topic microdocs, or handle each instance of concept or idea in the document separately, so that the document itself becomes only a container for all those different concepts.
Identified scopes help to make the search and indexing more simple and robust, and since concepts are stored in scopes, they can be manipulated in a number of ways - you can extract scopes from document as quotes, group them, show in search results or faceted list. (Though I should make a fair note that real life shows this scopes idea is only good for crazy infoarchs and data-freaks, because normal people usually just don't get the gist of the tool at all).
We (me and moedusa, if you remember him commenting some of your posts before) are trying to complete the DTD for specification document model with support for scopes and corresponding infrastructure, it progresses but very slowly.