Towards Asking Clarification Questions in Task-Oriented Dialogue

Anonymous

Abstract: Task-oriented dialogues aim at providing users with task-specific services. To provide satisfactory services, two major challenges exist: 1) users are not able to fully describe their complex needs due to lack of task knowledge, and; 2) systems need to personalize the service to their users since different users have different profiles and preferences. In order to solve these challenges, systems need to be able to ask questions so as to clarify the user's profile and needs. However, existing task-oriented dialogue systems ignore this aspect. In this paper, we formulate the problem of asking clarification questions in task-oriented dialogue systems. To this end, we propose a dialogue-based user simulator to collect a dataset, called TaskClariQ\footnote{To foster research in this area, the dataset and code will be made public upon paper’s acceptance.}. We further propose a new System Ask paradigm and a Multi-Attention Seq2Seq Networks (MAS2S) that implements it. Experimental results on TaskClariQ show that MAS2S outperforms competitive baselines.