How Good is Automatic Segmentation as a Multimodal Discourse Annotation Aid?