Changes of users' discursive behavior in HRI

Whether interacting with a colleague from another department, a child with distinct cognitive and linguistic skills or a foreigner with different cultural background, humans try to adapt to their communication partners. The adaptation allows for a flow in communication and thereby for successful turn-taking. In contrast, when users communicate with artificial systems often reduced flow and turn-overtaking 4 are observable. The reasons for these insufficiencies may lie in the fact that little is known about the turn-taking strategies that dialogue systems should pursue 4. In our approach, we postulate that it is the feedback of the system and the interpretation of the feedback by the user that are in the center of an adaptation process As a result, one way to improve HRI, and especially its communicational flow, is to focus on different types of feedback and how they affect the user{\textquoteright} discursive behavior.

In contrast to former work concentrating on telephone based speech systems or multimodal systems consisting of a touch screen and speech 1, 2, 5, we study an embodied interaction with the service robot BIRON (BIelefeld RObot CompanioN). The general goal of the studies is to develop a situated system that can learn the spatial environment as well as the names and visual appearance of objects. Therefore, BIRON can not only understand spoken speech but also co-verbal deictic references to objects in the scene and carry out mixed-initiative dialogues.

In our study, the subjects were asked to introduce objects to the robot by showing and pointing. This very restricted goal allowed us to compare verbal and gestural behavior across subjects. We analyzed the interactions of 15 native German speaking users communicating with BIRON. The interaction was carried out with the fully autonomous mode of BIRON, except for the speech recognition, which was simulated by keyboard input in order to avoid speech recognition errors. By using the autonomous interaction system, we were able to produce realistic communication sequences including problems caused by the complex interaction of the diverse perceptual system components.

In analyzing the users{\textquoteright} discursive behavior, we noticed different task-related interaction strategies. Focusing on subject{\textquoteright}s consistency and changes of strategies in the course of the ongoing turn-taking, we analyzed which feedback of the system caused a change of a strategy within a user. We found that subjects decided for one strategy addressing one perceptual channel (vision or speech) of the system and used it as long as they did not receive any feedback about the failure of this specific channel. We observed a change in strategy when it became obvious to the subject via feedback that a particular perceptual channel was not working appropriately. The change in strategy was likely to maintain the interaction flow and thereby the user satisfaction.

While these results support previous findings indicating when and how users change strategies 1, 5, they give new insights into the discursive behavior, i.e. into the repertoire of strategies in embodied and situated interaction. The results imply that within turn-taking, users interpret the system{\textquoteright}s feedback and thus verify their model of the capabilities of the interlocutor. Thus, the change in discursive strategy is an indicator of users{\textquoteright} expectation of how the robot functions 3.

In further studies we plan to deliberately vary the robots misunderstandings and integrate the personality traits of the subjects into the research on strategies.