Is there any public information about the back-end? Is it doing Markov chains on some suitable corpus or something? I think the cleverest part of the project i

Is there any public information about the back-end? Is it doing Markov chains on some suitable corpus or something? I think the cleverest part of the project is choosing well-known image macros to riff on - the format is so limiting that it's easier to generate convincing text.