Natural Language Dialogue Architectures for Tactical Questioning Characters

In this paper we contrast three architectures for natural language questioning characters. We contrast the relative costs and benefits of each approach in building characters for tactical questioning. The first architecture works purely at the textual level, using cross-language information retrieval techniques to learn the best output for any input from a training set of linked questions and answers. The second architecture adds a global emotional model and computes a compliance model, which can result in different outputs for different levels, given the same inputs. The third architecture works at a semantic level and allows authoring of different policies for response for different kinds of information. We describe these architectures and their strengths and weaknesses with respect to expressive capacity, performance, and authoring demands.