Conversational Strategies for Robustly Managing Dialog in Public Spaces

Proceedings of the of the EACL 2014 Workshop on Dialogue in Motion (DM), pages 63–67, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Conversational Strategies for Robustly Managing Dialog in Public Spaces Aasish Pappu Ming Sun Language Technologies Institute Carnegie Mellon University Pittsburgh PA, USA {aasish,mings,seshadrs,air}@cs.cmu.edu Seshadri Sridharan Alexander I. Rudnicky Abstract Open environments present an attention management challenge for conversational systems. We describe a kiosk system (based on Ravenclaw–Olympus) that uses simple auditory and visual information to interpret human presence and manage the system’s attention. The system robustly differentiates intended interactions from unintended ones at an accuracy of 93% and provides similar task completion rates in both a quiet room and a public space. 1 Introduction Dialog systems designers try to minimize disrup- tive influences by introducing physical and be- havioral constraints to create predictable environ- ments. This includes using a closed-talking mi- crophone or limiting interaction to one user at a time. But such constraints are difficult to apply in public environments such as kiosks (Bohus and Horvitz, 2010; Foster et al., 2012; Nakashima et al., 2014), in-car assistants (Kun et al., 2007; Hof- mann et al., 2013; Misu et al., 2013) or on mo- bile robots (Haasch et al., 2004; Sabanovic et al., 2006; Kollar et al., 2012). To implement dialog systems that operate in public spaces, we have to relax some of these constraints and deal with addi- tional challenges. For example, the system needs to select the correct interlocutor, who may be only one of several possible ones in the vicinity, then determine whether they are initiating the process of engaging with the system. In this paper we focus on the problems of identifying a potential interlocutor in the environ- ment, engaging them in conversation and provid- ing suitable channel-maintenance cues (Bruce et al., 2002; Fukuda et al., 2002; Al Moubayed and Skantze, 2011). We address these problems in the context of a simple application, a kiosk agent that Figure 1: Ravenclaw–Olympus augmented with multimodal input and output functions. accepts tasks such as taking a message to a named recipient. To evaluate the effectiveness of our ap- proach we compared the system’s ability to man- age conversations in a quiet room and in a public area. The remainder of this paper is organized as fol- lows: we first describe the system architecture, then present the evaluation setup and the results, then review related work and finally conclude with an analysis of the study. 2 System Architecture Figure 1 shows the architecture; it incorporates Ravenclaw/Olympus (Bohus et al., 2007) stan- dard components (in white), new components (in black) and modified ones (shaded). In the system pipeline, the Audio Server receives audio from a microphone, endpoints it and sends it to the ASR 63 Figure 2: Face states; some are animations. engine (PocketSphinx); the decoding is passed to NLU (Phoenix parser). ICE (Input Confidence Es- timation) (Helios) assigns confidence scores for the input concepts. Based on user’s input and the context, the Dialog Manager (DM) …

Content truncated to 500 words