Evaluating the Robstness and Scalability of Revision-Based Natural Language Generation

This paper presents the rst quantitative, corpus-based evaluation of the same-domain robustness and scalability of a new revision-based language generation model, in comparison to the traditional single pass pipeline model. Robustness is deened as the proportion of sentences, in a given corpus test sample that can be generated using only knowledge structures abstracted from another sample. Scalability is deened as the proportion of knowledge structures, among those needed to generate all sentences from a given corpus test sample, of those already acquired from another sample. Results show that the incremental revision model is far more robust and signiicantly more scalable than the single pass model.