Style evolution in NLP

Referring to a previous discussion, style is often associated with the most common n-grams an author uses.

Is there any work on dynamic style?, I was thinking on doing some kind of basic implementation to see whether one could increase the performance by doing this:

1. Using a 2 time slice Bayesian Network, we can evaluate the most common n-grams by year
2. With this, we are actually doing a filter over time, so we are considering the author's style to be evolving (thus using different words over time)

Perhaps this is kinda fart fetched, since I'm not sure how to involve dynamic variables so the Markovian assumptions make sense. Although I'm willing to bet that previous states of the style would not necessarily effect future states.

Any observations, ideas or suggestions?

Thanks