In many domains such as speech recognition and machine translation
  it is extremely useful to be able able to distinguish coherent from
  non-coherent sentences.  We introduce a set of word-based
  statistical features which measure semantic coherence and can be
  used to enhance any language application where coherent sentences
  need to be generated or recognized.  We train a decision tree using
  the constructed feature set to automatically classify sentences as
  coherent or not. We find that our combination of boosted decision
  trees and coherence features achieves an accuracy of 80\% when
  distinguishing trigram-generated sentences (non-coherent) from those
  in the Broadcast News dataset (coherent).