Recently IBM Watson demonstrated the effectiveness of a natural language question answering algorithm. Of course, beyond the game of Jeopardy, this problem becomes more difficult. The accumulated body of mathematical knowledge seems to be a good target for domain specific natural language question answering algorithms, something like an automated Math Overflow. Is this being pursued?
I do not know whether there is anyone pursuing this, but I believe that there should be. All LaTeX files on the arxiv are publicly visible, and of course they contain richly structured markup with cross-references between different parts of the same document and links to other documents. MathSciNet and Zentralblatt also have a lot of richly structured data. I think that a Watson-like system could digest enough of this to give much better answers to certain types of questions than current search engines can. I don't think you could expect to provide actual answers to mathematical questions, but I think you could do a good job of pointing people to papers where answers could be found.