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

This paper brings a contribution to the field of discourse annotation of corpora. Using ANNODIS, a french corpus annotated with discourse relations by naive and expert annotators, we focus on two of them, Elaboration and Entity-Elaboration. These two very frequent relations are (a) often confused by naive annotators (b) difficult to detect automatically as their signalling is poorly studied. We propose to use lexical cohesion to differentiate between them, and show that Elaboration is more cohesive than Entity-Elaboration. We then integrate lexical cohesion cues in a classification experiment, obtaining highly satisfying results.