Year: 2025
Type: dissertation
Abstract: Task-oriented dialogue systems have gained significant attention in both academia and industry for their practical utility in assisting users with specific tasks within defined domains. Unlike open-domain systems, which focus on maximizing user engagement, task-oriented systems prioritize task completion. Existing research in this field can be broadly categorized into pipeline meth- ods and end-to-end approaches. While pipeline methods compartmentalize the system into modules such as Natural Language Understanding (NLU), Natural Language Generation (NLG), and User satisfaction Modeling (USM). End-to-end approaches utilize a single model for seamless processing. How- ever, integrating domain knowledge, including domain schema and domain documents, poses a significant challenge, hindering the effectiveness of these systems. With the emergence of large language models (LLMs), various strate- gies have been proposed to integrate common sense knowledge from LLMs to enhance system performance. This thesis explores different methodologies for integrating domain knowledge and common sense knowledge into task- oriented dialogue systems. Specifically, (1) we explore how to use the task schema knowledge to track dialogue state to enhance the natural language understanding ability of the task-oriented dialogue system; (2) we explore how to use task document knowledge to generate system responses to clar- ify users’ needs and user profile; (3) we explore how to effectively estimate the user satisfaction using domain knowledge to better simulate the user and evaluate the performance of the task-oriented dialogue system; (4) we explore how to use common scene knowledge in the LLMs to empower the end-to- end task-oriented dialogue system. Each methodology is discussed in detail, highlighting its contributions and experimental results.
Author: Feng, Yue
Language: English
Topic: Speech and dialogue systems
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7010941590