I'd like to train many markov text generators using lots of individual text sources. Then, the transition matrix used to generate my novel will be dynamically assigned as the novel evolves.
The program will begin to generate the novel in the style of, say, The Bible. When the name of another work in the corpus is generated (e.g, ...and whence a heart of darkness), the markov generator will use the transition matrix from that other work. Perhaps each work will be differentiated into e.g. its own chapter.
...and whence a heart of darkness
(has this been attempted before?)
Not that i am aware of. There are several generators that generate books in the style of another book ( #72 combines two books) and I think there are some that switch their generation according to various rules, but I can't recall any that are even particularly close to your plan.
In general, I think combining multiple approaches is an underutilized way to create order in novel-length generators. Probably because it can take a month to learn a single new approach, let alone figuring out how to combine multiple ones. Your idea seems to strike a balan e between a strong concept and a relatively straightforward implementation. I suspect the trick will be to have enough works to key off bits of the titles. May want to throw some Shakespeare in there: everyone is always quoting his lines as titles.
"That Hamlet is full of clichés."