As a side note if you want an "industrial-strength" applied version, at the cost of losing the small/elegant implementation, there's been considerable work in the past years on building robust Markov-language-modeling packages. I mention it here mostly because I didn't realize until recently myself that there was much to gain on top of the vanilla versions that are fairly straightforward to DIY.

One active area of research is smoothing the models, so they capture regularities without just memorizing the input text. The vanilla approach chooses a fixed order of Markov chain (e.g. 2-grams or 3-grams), where you choose n either ad-hoc, or (ideally) to balance underfitting (miss context you could use) against overfitting (end up memorizing verbatim text strings due to insufficient data). But the right "n" to choose can vary based on context as well, since different words show up more often in the training data, and are more or less variable in what kinds of words follow them. So there are a number of approaches to choose n adaptively and/or to smooth the counts ("Katz's backoff model", "Kneser-Ney smoothing", etc.).

From a more engineering point of view, there's also considerable work on pruning the models and representing them efficiently, if you want to work with things of the size of, say, the Google Books ngram dataset.

A few free-software packages that implement a good portion of the recent work:

http://kheafield.com/code/kenlm/

http://code.google.com/p/berkeleylm/