An Information-Providing Closed-Domain Human-Agent Interaction Corpus

HAI Alice - An Information-Providing Closed-Domain Dialog Corpus Jelte van Waterschoot1�, Guillaume Dubuisson Duplessis2, Lorenzo Gatti3, Merijn Bruijnes1 and Dirk Heylen1 1Human Media Interaction, University of Twente, 7522 NB, Enschede, The Netherlands 2CNRS, ISIR, Sorbonne Universit´e, 75005, Paris, France 3FBK, University of Trento, 38123, Trento, Italy l.gatti@fbk.eu, gdubuisson@telecom-paristech.fr {j.b.vanwaterschoot, m.bruijnes, d.k.j.heylen}@utwente.nl Abstract The contribution of this paper is twofold: 1) we provide a public corpus for Human-Agent Interaction (where the agent is controlled by a Wizard of Oz) and 2) we show a study on verbal alignment in Human-Agent Interaction, to exemplify the corpus’ use. In our recordings for the Human-Agent Interaction Alice-corpus (HAI Alice-corpus), participants talked to a wizarded agent, who provided them with information about the book Alice in Wonderland and its author. The wizard had immediate and almost full control over the agent’s verbal and nonverbal behavior, as the wizard provided the agent’s speech through his own voice and his facial expressions were directly copied onto the agent. The agent’s hand gestures were controlled through a button interface. Data was collected to create a corpus with unexpected situations, such as misunderstandings, (accidental) false information, and interruptions. The HAI Alice-corpus consists of transcribed audio-video recordings of 15 conversations (more than 900 utterances) between users and the wizarded agent. As a use-case example, we measured the verbal alignment between the user and the agent. The paper contains information about the setup of the data collection, the unexpected situations and a description of our verbal alignment study. Keywords: Corpus, Human-Agent Interaction, Wizard of Oz, Closed Domain, Information-Providing, Unexpected Situations, Verbal Alignment 1 Introduction This paper presents the Human-Agent Interaction Alice- corpus (HAI Alice-corpus), a corpus of conversations be- tween a user and an embodied conversational agent (ECA) operated by a wizard via the Wizard of Oz (WOz) method (Dahlb¨ack et al., 1993). The corpus has been collected as part of the ARIA-VALUSPA1 project, in which a multi- modal virtual agent for information retrieval that can deal with unexpected situations is being developed. The goal of the project is to provide interested parties, such as fellow re- searchers and industry, with a toolkit for building their own virtual agent (Valstar et al., 2016; Bruijnes et al., 2013). We collected the corpus to investigate how users react to un- expected situations in a conversation with an autonomous state-of-the-art virtual agent. However, in a classic WOz approach where the wizard uses a button interface, it is nearly impossible to improvise in unexpected situations. This is why we gave our wizard the freedom to choose his own words and facial expressions, to create and respond to unexpected situations. To the best of our knowledge most state-of-the-art agents cannot yet cope with interrup- tions effectively, nor perform human-level verbal alignment (Dubuisson Duplessis et al., 2017b). However, we believe we are approaching such capabilities in agents, thanks to recent efforts in agent development. Therefore, we wanted to simulate an agent as closely as possible with this type of WOz setup. We analyzed the corpus for verbal …

Content truncated to 500 words