PUB - Publikationen an der Universität Bielefeld

In recent years, research has moved towards the learning by interaction paradigm 1 suggesting that interaction with an artificial agent is facilitated when characteristics of a social interaction are considered. It is envisoned that agents will learn from humans by simply interacting with each other. In such a scenario, learning by interaction {\textquoteright}goes beyond common supervised or unsupervised strategies by taking into account wider feedback and assessments for the learning processes{\textquoteright} (1 p.140). So far, little is known about interactional processes and feedback strategies involved. Yet, in order to learn, a learner will typically need to be provided with information given by a teacher who not only gives certain structure to the interaction but also instructs for and demonstrates the learning contents. The given information can only be effective, if the learner is receptive. To assure this, the tutor makes use of interactive regularities checking the learner{\textquoteright}s behavior. The term contingency has been suggested to encompass such regu- larities in interaction. More specifically, it refers to a temporal sequence of behavior and reaction (2, 3). It has been shown that contingency is an important factor in interactions with infants and contributes to the cognitive development of infants 4. In the interaction with an artificial agent, contingency has been operationalized by eye-gaze bouts 5. So far, it has been shown that while in a situation with a child, eye-gaze bouts in total, average and frequency is much higher as in interaction with an adult as learner. In a situation with an artificial agent, a decrease of eye-gaze bouts could be observed 5. It could thus be reasoned that tutor{\textquoteright}s monitoring behavior is impaired when interacting with a robot. However, so far only an interaction with a virtual robot has been investigated. In contrast, an embodied robot could evoke a more natural tutor behavior. In this study, we therefore investigated a tutoring situation with an embodied robot and focused on tutor{\textquoteright}s monitoring behavior. We followed the minimal definition of embodiment by K. Dautenhahn et al. to quantify the difference in embodiment between these two systems 6. Accordingly, the Degrees of Embodiment (DOM) are calculated as:DOMS,E = f (x, y, t), where system S in respect to an environment E is calculated by a function f of the vectors x and y and the time t. x describes the number of sensors, the detected modalities of the sensors and the channels of information provided by the sensors, y describes the degree of freedom (DoF) of the robot. In our experiments E remained unchanged and thus the DOM is not affected by this factor. We argue that in our case, the function f is only dependent on the degrees of freedom of the robot. Possibly, we can give a value which represents the difference in the degree of embodiment (DDOM). Since our goal was to study some characteristics of contingency (7, 5), our dependent variable was the tutor{\textquoteright}s monitoring behavior operationalized by the eye-gaze.