Tika looks interesting too (“detects and extracts metadata and structured text content from various documents using existing parser libraries”). Again, does anyone/anywhere offer it as an open, hosted service? How might it work in a Scraperwiki context, for example?

This reminds me of something else I came across recently that helps regenerate structured markup from template created free text: Boomerang, “a bidirectional programming language for ad-hoc, textual data.” [ http://www.seas.upenn.edu/~harmony/ ].
As with all these things, the problem is keeping up with what’s possible whilst maintaining enough skills with any particular toolset to be able to make appropriate use of it…