On Wed, 4 Jun 2008, John Melesky wrote:
So you use those occurrence statistics to pick a feasible next word (let's choose "system", since it's the highest probability here -- in practice you'd probably choose one randomly based on a weighted likelihood). Then you look for all the word pairs which start with "system", and choose the next word in the same fashion. Repeat for as long as you want.
"Markov chain" means, that you have a sequence of random experiments, where the outcome of each experiment depends exclusively on a fixed number (the level) of experiments immediately before the current one.
Those word-pair statistics, when you have them for all the words in your vocabulary, comprise the first-level Markov data for your corpus.
When you extend it to word triplets, it's second-level Markov data (and it will generate more reasonable fake text). You can build higher and higher Markov levels if you'd like.
If the level is too high, you will just reproduce the training text.