If you spend a weekend or two watching machine learning Google tech talks you can start with a roar.
There's also 'Programming Collective Intelligence', Norvig's own book 'Artificial Intelligence: A Modern Approach', and the growing Netflix literature.
One of the interesting things people discovered while attacking the netflix problem is that the rating graph encodes features like 'heroism', 'horror', 'absurdity', 'romance', 'political protest', and 'not monty python' -- you only need a little ingenuity to assign words to said features by looking at extreme feature exemplars.
