A Realistic Success Criterion for Discourse Segmentation

In this study, compared to the existing one, a more realistic evaluation method for discourse segmentation is introduced. It is believed that discourse segmentation is a fuzzy task [Pas96]. Human subjects may agree on dierent discourse boundaries, with high agreement among them. In the existing method a threshold value is calculated and sentences that marked by that many subjects are decided as real boundaries and other marks are not been considered. Furthermore automatically discovered boundaries, in case of being misplaced, are treated as a strict failure, disregarding the proximity wrt to the human found boundaries. The proposed method overcomes these shortcomings, and credits the fuzziness of the human subjects' decisions as well as tolerates misplacements of the automated discovery