From Conversational Adaptability to Response Reliability: Advancing Flexible and Trustworthy Conversational AI

This work advances conversational AI by improving both adaptability and reliability. The first part focuses on enhancing dialogue flexibility by unifying tasks that are traditionally treated separately, including task-oriented dialogue, question answering, and open-domain conversation. One line of work integrates structured task completion with retrieval-augmented question answering to support more informed and context aware responses. Another system enables smooth transitions between task-based interactions and open-ended dialogue, allowing for more natural and dynamic user experiences.

The second part addresses the reliability of large language model outputs by focusing on hallucination, which refers to content that is factually incorrect or inconsistent with the context. The first line of work introduces a modular fact checking framework that verifies generated responses using plug-and-play components, without requiring model fine-tuning. To support broader research, a dataset is developed for evaluating fact checking in realistic generation scenarios. The subsequent work presents a comprehensive benchmark for evaluating faithful hallucinations in summarization. This benchmark captures the diversity of hallucinations across different language model families and reveals limitations in existing detection methods.

Committee: Dr. Qi Li (major professor), Dr. Hongyang Gao, Dr. Mengdi Huai, Dr. Forrest Bao, and Dr. Zhu Zhang