Towards Intelligent Conversational Assistants: Enhancing Task-Oriented Dialogue Systems with Knowledge Integration
Author(s)
Su, Ruolin
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This thesis presents a comprehensive study on enhancing Task-Oriented Dialogue (TOD) systems by integrating domain-specific, dialogue-level, and cross-lingual knowledge to improve intelligent assistants’ ability to understand user intent and execute structured tasks. While recent advancements in Natural Language Processing (NLP) and Large Language Models (LLMs) have significantly improved dialogue systems, TOD systems still face critical challenges in knowledge integration, scalability, and generalization. To address these, this work introduces methods such as choice-fusion dialogue state tracking, schema graph-guide prompt learning, dialogue-level modeling using dialogue acts and a soft mixture-of-experts framework for TOD systems. These innovations enhance dialogue management accuracy, contextual understanding, and system modularity. Additionally, a cross-lingual knowledge transfer mechanism is proposed to improve commonsense reasoning in low-resource languages, expanding the multilingual adaptability of TOD systems. By integrating structured knowledge across different levels, it enhances system performance in user intent recognition, dialogue state tracking, and response generation. The findings provide scalable and adaptable solutions for intelligent assistants in diverse domains and multilingual settings. Together, these contributions form a unified framework for knowledge-enhanced TOD systems, enabling more robust, scalable, and effective conversational agents across diverse application domains and languages.
Sponsor
Date
2025-04-16
Extent
Resource Type
Text
Resource Subtype
Dissertation