IBM's Unstructured Information Management Architecture
A while back I had the pleasure of chatting with some really bright people at IBM's Watson labs who were doing some really cool things in NLP/Unstructured Information. They'd come up with a really nice model and architecture for hooking in all kinds of analyzers and tools, attaching annotations, and piping things around.
Anyway, they made it available on AlphaWorks just recently! Part of what made it really interesting was that most of the researchers ae using any old language -- perl is popular, C is popular, and some Fortran is, apparently, still in use for doing hardcore number crunching. Building a plugin architecture for that kind of things has got to be fun!
Reading the blurb they posted with the SDK, it doesn't look like they included their existing analysis engines, which would have been extremely cool, but is still fun to see stuff appearing which was last seen in researcher's labs =)