Modelling control in generation

Abstract

In this paper we present a view of natural language generation in which the control structure of the generator is clearly separated from the content decisions made during generation, allowing us to explore and compare different control strategies in a systematic way. Our approach factors control into two components, a ‘generation tree’ which maps out the relationships between different decisions, and an algorithm for traversing such a tree which determines which choices are actually made. We illustrate the approach with examples of stylistic control and automatic text revision using both generative and empirical techniques. We argue that this approach provides a useful basis for the theoretical study of control in generation, and a framework for implementing generators with a range of control strategies. We also suggest that this approach can be developed into tool for analysing and adapting control aspects of other advanced wide-coverage generation systems.