نبذة مختصرة : In practical applications, dialog systems often comprise multiple conversational agents. There are multiple reasons for such a combination: The developers of a dialog system might want to combine the capabilities of different agents because each of them is specialized in one task. Another reason is that the individual dialog agents already exist, and a combination is easier than a reimplementation of the two into a single system. A third reason might be the requirement to separate different bot functionalities, e.g., for software architecture or data privacy reasons. Although this is a common practical requirement, there is little research on such multibot systems. This thesis addresses this gap and examines three research questions: First, what are the characteristics of multibot systems? Which types of multibot systems perform better, such as combining many small or a few large agents? What is the difference in performance between a multibot system and the same dialog system implemented as a single one? The second research question investigates machine learning models that can decide which dialog agent should respond to the question for each user utterance. Multibot systems use considerable computational resources because many systems are implemented multiple times. Therefore, research question three is how to implement memory-efficient multibot systems. The thesis introduces the term Modular Dialog System. Modular Dialog Systems are one type of multibot system in which every agent is an independent chatbot. Then, it elaborates on the Research Questions in three experiments. The first experiment implements multiple combinations of task-oriented dialog systems and sheds light on Research Questions one and two. Then, the second experiment investigates the combinations of task-oriented dialog and question answering. Afterward, a third experiment focuses on using adapters for the resource-efficient deployment of natural language understanding models. Regarding research question one, this thesis shows that multibot ...
No Comments.