Towards a Conversational Corpus for Human-Robot Conversations

Abstract

Conversational corpora based on human-human dialogues have often been used for training of data-driven dialogue systems. However, human-human conversations might not be the optimal inputs for machine learning training aims used in HRI. This paper suggests the creation of a conversational corpus based on Human-Robot conversations as input for the training of a dialogue system used in future conversational robots. We propose that the significant differences between Human-Human Conversation (HHC) and Human-Robot Conversation (HRC) in terms of used language and other aspects (e.g., humanlikeness, embodiment, etc.) might affect the quality of the responses from a conversational robot. Hence, the use of HRCs as an input could improve the responses of the robots when the conversational machine learning system is trained using a more realistic model of HRI conversations rather than a HHI model. Future applications of conversational robots in education and health care could be enhanced by using an appropriate HRC corpus.