The Penn Discourse TreeBank as a Resource for Natural Language Generation
Rashmi Prasad, Aravind K. Joshi, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Bonnie Webber · 2005
While many advances have been made in Natural Language Generation (NLG), the scope of the field has been somewhat restricted because of the lack of annotated corpora from which properties of texts can be automatically acquired and applied towards the development of generation systems. In this paper, we describe how the Penn Discourse Tree-Bank (PDTB) can serve as a valuable large scale annotated corpus resource for furthering research in NLG and for inducing models for the development of NLG systems. The PDTB is annotated for discourse relations, and encodes explicitly the elements of these relations: explicit and implicit discourse connectives, denoting the predicates of the relations, and text spans, denoting the arguments of the relations. Connectives and arguments are also annotated with features and spans related to attribution, and each connective will be annotated with labels standing for the projected discourse relation, including sense distinctions for polysemous connectives. We exemplify the use of the corpus for two tasks in NLG: the realization of discourse relations during sentence planning, and the representation and realization of attribution. 1