Adaptive Dialogue Strategy for a Teachable Social Robot

Social robots have been increasingly used in education, where they have the ability to deliver personalized, multimodal, one-on-one interactions with students. When social robots are deployed in the classroom, they most commonly act in the role of a tutor, though allowing the robot to act as a novice has been found to support learning outcomes, as well as drive higher engagement in the learning process due to the Protégé effect, where the effort to learn and organise the material is higher when done for the benefit of someone else. Teachable agents rarely engage in conversations with their student teachers, relying instead on limited natural language understanding, or only in interaction with buttons of a web interface to communicate.

This research builds upon the Curiosity Notebook, a web platform that supports teaching either a virtual or embodied learner about a classification task, through natural language dialogue.

The aim of this project is to develop and experimentally validate an adaptive dialogue strategy for a teachable robot. A reinforcement learning algorithm will select dialogue acts that attempt to maximise a reward based on student learning outcomes engagement. The robot will be capable of adapting its behaviours to suit individual students. This will require research to identify the behaviours within a teaching interaction that affect learning outcomes and engagement, the design and implementation of a Dialogue Management System and Natural Language Model to facilitate these behaviours in conversation, and an adaptive reinforcement learning algorithm to select optimal robot dialogue acts. This interdisciplinary approach combines research from the fields of language modelling, dialogue modelling, social robotics, and reinforcement learning to develop a novel dialogue strategy for a teachable robot.