Generating under global constraints: the case of scripted dialogue

Recently, the view of Natural Language Generation (NLG) as a Constraint Satisfaction Problem (CSP) has seen something of a revival. The aim of this paper is to examine the issues that arise when nlg is viewed as a CSP, and to introduce a novel application of constraint-based NLG, namely the ScriptedDialogue. ScriptedDialogue shares a number of crucial features with discourse, which make it possible to control the global properties of a computer-generated dialogue in the same way as those of a generated discourse. We pay particular attention to the use of soft constraints for enforcing global properties of text and dialogue. Because there has been little research into the formal properties of soft constraints in relation to generation, we start out with a theoretical exploration. We argue that, when multiple constraints are involved, it is important to define properly what is being optimised before proposing specific algorithms, and we argue that such definitions are often lacking in csp-based nlg. We show that it can be difficult (and sometimes even impossible) to guarantee satisfaction of global constraints by following local strategies. Based on these difficulties, we propose a novel approach to the generation of discourse and dialogue which combines csp solving with revision. Scripted Dialogue is used to illustrate this approach, which is compared with alternatives such as monitoring and estimation