I wonder if there's any article with more details on the internals. From reading some of the other stuff it seems they have competing models and Watson picks the answer from the model with the highest confidence. But I haven't found anything more detailed than that unfortunately.
Wild guess: Neural net processing old Jeopardy questions w/Baysian updates?