It almost certainly builds Hidden Markov Models (Wikipedia has a decent, though perhaps overly technical article if you're interested) for the transition probability of each set of answers to a particular target. In such a case, all it needs is a database of target nouns and a large database of associated descriptors (question answers) to build the statistical models. It could index certain areas of Wikipedia for nouns and simply model the nouns as a function of descriptors in the articles (it would be especially easy to extract article groupings and standard sidebar information for certain sections) or even in Google searches for the term. At each step, it takes the path with the highest probability of yielding a definitive transition probability, which it eventually returns as the answer.
Alternatively, because I'm not sure how it was when it started, they got a huge database of targets, created a stock set of questions, and then presented them approximately at random to the first users and modeled transition probabilities from that with each subsequent user refining the model (in other words, the system self-optimizes). It's also very possible that they're using both models and adding their transition probabilities in some way.
The thing that often makes such systems so neat is that you can completely automate model building and, given a large enough data pool either from spidering or crowdsourcing, you can build extremely robust models that often prove surprising accurate to non-machine minds.
