Language choice models for microplanning and readability

This paper describes the construction of language choice models for the microplanning of discourse relations in a Natural Language Generation system that attempts to generate appropriate texts for users with varying levels of literacy. The models consist of constraint satisfaction problem graphs that have been derived from the results of a corpus analysis. The corpus that the models are based on was written for good readers. We adapted the models for poor readers by allowing certain constraints to be tightened, based on psycholinguistic evidence. We describe how the design of microplanner is evolving. We discuss the compromises involved in generating more readable textual output and implications of our design for NLG architectures. Finally we describe plans for future work