End-to-End Task-Oriented Dialog System Through Template Slot Value Generation

To overcome the limitations of conventional pipeline-based task-oriented dialog systems, an end-to-end approach has been introduced. To date, many end-to-end task-oriented dialog systems have been proposed and these have shown good performance in various domains. However, those have some limitations such as the need for dialog state annotations. And there is also room for improvement for those systems. In this paper, we examine the issues of recent end-to-end task-oriented dialog systems and present a model that can handle these issues. The proposed model classifies a system utterance template in a retrieval-based manner and then generates the slot values in the template through a decoder. Also, we propose an unsupervised learning based template generation method that allows model training even in a domain where the templates are not given and the dialog information is not tagged. Our model obtains new state-of-the-art results on a restaurant search domain.