A Flexible and Reusable Framework for Dialogue and Action Management in Multi-Party Discourse
Fachrichtung Informatik, Universität des Saarlandes, Saarbrücken, Dissertation, (2008)DOI: 10.22028/D291-25921
Abstract
This thesis describes a model for goal-directed dialogue and activity control in real-time for multiple conversation participants that can be human users or virtual characters in multimodal dialogue systems and a framework implementing the model. It is concerned with two genres: task-oriented systems and interactive narratives. The model is based on a representation of participant behavior on three hierarchical levels: dialogue acts, dialogue games, and activities. Dialogue games allow to take advantage of social conventions and obligations to model the basic structure of dialogues. The interactions can be specified and implemented using reoccurring elementary building blocks. Expectations about future behavior of other participants are derived from the state of active dialogue games; this can be useful for, e. g., input disambiguation. The knowledge base of the system is defined in an ontological format and allows individual knowledge and personal traits for the characters. The Conversational Behavior Generation Framework implements the model. It coordinates a set of conversational dialogue engines (CDEs), where each participant is represented by one CDE. The virtual characters can act autonomously, or semi-autonomously follow goals assigned by an external story module (Narrative Mode). The framework allows combining alternative specification methods for the virtual characters' activities (implementation in a general-purpose programming language, by plan operators, or in the specification language Lisa that was developed for the model). The practical viability of the framework was tested and demonstrated via the realization of three systems with different purposes and scope.