> > The answer is to save the context node in a variable before 
> going into 
> > the for-each.
> 
>   I don't see how this could help here.  Actually, I don't see any
> *simple* way to solve this.  Could you please explain a 
> little bit further?

OK, the problem's a bit deeper than I realized.

Fact is, if you extract the string value of a node with complex/mixed
content then you're throwing away any information about which parts of the
string came from which node - and that's before you start tokenizing. If you
want to retain this information, you're going to have to process the nodes
one at a time.

Michael Kay
http://www.saxonica.com/