Simulated Task Oriented Dialogues for Developing Versatile Conversational Agents

Task-Oriented Dialogue (TOD) Systems are increasingly important for managing a variety of daily tasks, yet often underperform in unfamiliar scenarios due to limitations in existing training datasets. This study addresses the challenge of generating robust and versatile TOD systems by transforming instructional task descriptions into natural user-system dialogues to serve as enhanced pre-training data. We explore three strategies for synthetic dialogue generation: crowdsourcing, encoder-decoder models, and in-context learning with large language models. The evaluation of these approaches, based on a comprehensive user study employing 10 different metrics, reveals the top quality of the dialogues generated by learning an encoder-decoder model as per human evaluation. Notably, employing this synthetic dialogue further improves the performance of advanced TOD models, especially in unfamiliar domains, with improvements spanning 5.5% to as much as 20.9% in combined evaluation scores. Our findings advocate for the use of specialised, task-oriented knowledge bases and step-wise dialogue generation techniques to advance the capabilities and generalizability of TOD systems.