I would try an n-gram model with dynamic programming.

e.g.

logp(_, "") = 0

logp(word0, text) = max(logp_bigram(word0, word1) + logp(word1, rest) for word1, rest in prefix_words(text))