Creating Dynamic Robot Utterances in Human-Robot Social Interaction : Comparison of a Selection-Based Approach and a Neural Network Approach on Giving Robot Responses in Conversations

Abstract: This study examines two different approaches to dialogue management system in order to achieve dynamic utterances in human-robot social interactions. This was done in order to determine whether Robot Assisted Language Learning is viable as a solution to the current teacher shortage situation in Sweden. One approach is a selection-based approach with the use of a dialogue tree with sentence embedding while the other approach is a neural network approach with two different models; transformer and seq2seq. Results from the user testing show that both implementations were not particulary successful, although the selection-based method performed better than neural network approaches and shows promise for future research. Due to the results not reaching a satisfying level of performance and the existence of cheaper virtual education tools RALL might not be the most appropriate solution for current day Sweden but is promising as a long-term solution to the continuing trend of teacher declination and increasing labor costs.