Implicit semantic transmission in social learning Analysis and modeling

Project Leaders: Katharina Rohlfing, Frank Joublin, Martin Heckmann
PhD Student: Juana Salas Poblete, Nikolas Hemion

The social context of learning has increasingly gained attention in developmental psychology, cognitive science and robotics. It has been proposed that an agent – in order to learn – needs to be grounded in a meaningful embodied activity. The robotic research has just started to benefit from the use of developmental approaches: Orienting towards ‘learning by communicating’ offers new learning paradigms, within which it can be analyzed how semantic information is transmitted, and which effect the way of transmission has onto learning. So far this paradigm involves face-to-face scenarios, where a tutor is focusing on a student. However, this learning situation is not offered in every culture. Instead, developmental research has shown that children are likely to benefit also from other scenarios. Motivated by animal studies by e.g. Irene Pepperberg on grey parrots which were trained in a social learning paradigm (model-rival-paradigm), it is our goal to investigate multi-party learning scenarios, in which the tutor does not address the student directly but the student is learning while observing a tutoring behaviour towards another person. Thus, our assumption is that learning can take place from both, direct and indirect teaching.

With this project, we will investigate the behaviour of tutors and students and study the achieved learning effects in different situations of social learning. Based on the data gathered in psychophysical experiments on both, direct and indirect teaching scenarios, we aim to identify different verbal and non-verbal patterns, e.g. denominating objects, showing an object. Following the identification and classification of these patterns, we aim to develop a generative model for their production. The purpose of this model is twofold. Firstly, it will allow setting up a virtual tutor. A virtual tutor can be used to create simulated dialogues with the virtual tutor replacing the real tutor or tutors and an additional avatar, which replaces the child. Secondly, building a generative model for the behaviour of the tutor will allow us to understand the underlying principles of learning in a social context better and the insights from the modelling will provide valuable feedback on the design of the psychophysical experiments.

The results of this research should enable the setup of a social interaction simulation environment, where reproducible experiments between tutor avatars and a robotic artefact could be performed. These experiments will allow testing new hypotheses on how social learning takes place.