| dc.description.abstract | The conversational capabilities of a dialog system have a direct impact on the tasks it can accomplish. Solving all conversational issues in dialog systems have the potential to make them serve in complex domains. While this is not achievable, addressing fundamental aspects in conversation is desired to make a task-oriented dialog system (TODS) serve in new domains where they are needed, besides increasing their usefulness. One such aspect is the ability to advance a conversation logically. The primary aim of this study was to develop a novel architecture that will guarantee advancing conversations in TODS. To realize this aim, theories and literature were interrogated that informed the formulation of an agent-based architecture for dialog management. Then implementation of the architecture previously realized in a dialog system prototype. Followed by training the dialog system on initial domain-specific data. And evaluating its performances in a specific domain. The study used exploratory methodology to provide the theories that justified the construction of the multi-agent system (MAS_DM) architecture, while the experimental design was explored to synthesize and train the prototype.The design involved the fusion of agent-based architecture with reinforcement learning technique to enable tracking of context, structure and policy without depending on handcrafted rules. MAS_DM architecture explores learning agents in an unknown environment, where each agent is endowed with the ability to learn and select a policy. Learning and policy selection is sustained through reinforcement learning, eliminating the need for handcrafted rules. The architectural model was evaluated and validated in a prototype Chabot system. The Chatbot system was trained and tested in the maternal healthcare domain and was evaluated by human users. In this context, each user filled out an online questionnaire after successful interaction with the Chatbot. The evaluation parameters were coherence, task success, general performance, user satisfaction and goal achievement. This evaluation adheres to the specifications of Goal Question Metrics and PARAdigm for DIalog System Evaluation frameworks. The key findings were that Chabot’s ability to advance the conversation scored 0.8903, and achieved an overall performance score of 0.553. It achieved a task success rate of 0.936. with a user satisfaction score of 0.775. Based on global acceptable measures, interpreted this task success as substantial, coherence score as substantial, user satisfaction as excellent and the overall performance as good. Where machine learning is involved kappa statistic values above 0.40 are considered exceptional. The results suggest that it is reasonable to conclude that the MAS_DM architecture can be trusted to guarantee conversations that advance logically. The study contributes to the body of knowledge of conversational artificial intelligence by; - developing a novel agent-based architectural model for TODS, demonstrating the practicability of combining multi-agent systems and machine learning toward solving conversational issues and enhancing the capability of TODS. | en_US |
