Cognitive architecture of multimodal multidimensional dialogue management

Abstract

Numerous studies show that participants of real-life dialogues happen to get involved in rather dynamic non-sequential interactions. This challenges the dialogue system designs based on a reactive interlocutor paradigm and calls for dialog systems that can be characterised as a proactive learner, accomplished multitasking planner and adaptive decision maker. Addressing this call, the thesis brings innovative integration of cognitive models into the human-computer dialogue systems. This work utilises recent advances in Instance-Based Learning of Theory of Mind skills and the established Cognitive Task Analysis and ACT-R models. Cognitive Task Agents, producing detailed simulation of human learning, prediction, adaption and decision making, are integrated in the multi-agent Dialogue Man-ager. The manager operates on the multidimensional information state enriched with representations based on domain- and modality-specific semantics and performs context-driven dialogue acts interpretation and generation. The flexible technical framework for modular distributed dialogue system integration is designed and tested. The implemented multitasking Interactive Cognitive Tutor is evaluated as showing human-like proactive and adaptive behaviour in setting goals, choosing appropriate strategies and monitoring processes across contexts, and encouraging the user exhibit similar metacognitive competences.