DiaBuLI – Building Dialogues for Human-Robot Interaction by Learning from Object Information

Abstract

We report on preliminary results of our efforts of building a human-robot dialogue by using decision tree learning. The system is particularly designed for dialogues that go along with decision processes. In our application, the self-developed robot MoBi should assist young children in the classroom with waste management. To do so, MoBi asks a couple of yes/no-questions about a waste item to dispose in order to decide on the correct bin. We take a collection of instances of the classification task where we know the right bin decision for and we characterize these examples by a set of attributes that describe features like material and usage properties. Then, we perform decision tree learning to generate a tree which builds the basis for the dialogue with MoBi. While many existing work aim for a more open-ended form of dialogue we focus on a specific setting where the robot assists in a decision process. Existing approaches have investigated the use of learning to optimize the length or number of turns in an interaction dialogue. We are also interested in optimizing our dialogue but we look into find-ing other interesting qualities as well. We compare trees resulting from different configurations of our decision tree learning both with one another as well as with hand-crafted dialogues used for the robot MoBi so far.