Exploiting naive vs expert discourse annotations: an experiment using lexical cohesion to predict Elaboration / Entity-Elaboration confusions.