Abstract
This paper presents a novel dataset for training end-to-end task oriented conversational agents. The dataset contains conversations between an operator - a task expert, and a client who seeks information about the task.
Along with the conversation transcriptions, we record database API calls performed by the operator, which capture a distilled meaning of the user query. We expect that the easy-to-get supervision of database calls will allow us to train end-to-end dialogue agents with significantly less training data.
The dataset is collected using crowdsourcing and the conversations cover the well-known restaurant domain. Quality of the data is enforced by mutual control among contributors.
The dataset is available for download under the Creative Commons 4.0 BY-SA license.