I would look at the most recent work by Hoifung Poon:
- Poon and Domingos (2010). "Unsupervised Ontological Induction from Text." [paper]
- Poon and Domingos (2009). "Unsupervised Semantic Parsing." [paper | slides | code]
They have a completely unsupervised technique for semantic parsing, and they test it on question-answering, getting twice the recall of competing approaches. The Java code for the 2009 approach (USP) is available at the link above. I am asking Hoifung about the 2010 (LOFT) code.
When I read the 2009 work, it completely blew my mind. You read this paper and you're like: "Whoa, you really just did that. You attempted to actually solve NLP. I forgot that we're allowed (supposed) to do that." And the technique is so elegant, small wonder it got best paper award.