Abstract
The increasing integration of conversational agents in real-world applications demands robust systems capable of adapting to continuous domain changes. This thesis focuses on addressing these challenges in Task-Oriented Dialogue systems by proposing Dialogue Domain Adaptation as a method to generate new dialogues that are adherent to domain changes. The work centers on mitigating performance degradation caused by changes in knowledge bases due to domain evolution. A rule-guided slot-substitution method is introduced as a solution to domain changes that affect slot-value availability and distribution. The approach adapts system behavior without requiring extensive retraining, maintaining dialogue coherence and domain adherence. The thesis further explores generative slot-substitution using fine-tuned language models like BERT and LLaMA, demonstrating how they can automatically generate new slot-values that align with the evolving domain. Additionally, the thesis assesses the effectiveness of using Large Language Models as end-to-end dialogue systems in dynamic domain settings. Through experiments on the MultiWOZ dataset, it is shown that modern LLMs outperform traditional methods in maintaining conversational consistency and handling unseen slot-values. In addition, a domain change simulator is introduced as a practical tool to model domain changes, assess their impact on system components, and estimate the retraining effort required for system adaptation. The key findings of this work provide practical methodologies and theoretical insights for building more resilient and adaptable TOD systems, capable of handling complex, real-world domain shifts while minimizing resource costs.