Generating a Variety of Backchannels Based on Linguistic and Prosodic Features for Attentive Listening Agents

Abstract

There is a growing interest in conversation agents which conduct attentive listening. However, the current conversation agents always generate the same or limited form of backchannels every time, giving a monotonous impression. We have investigated generation of a variety of backchannels according to the dialogue context using the corpus of counseling dialogue. At first, we annotate all acceptable backchannel form categories considering the arbitrary nature of backchannels. Then, we conduct machine learning to predict a backchannel form from the linguistic and prosodic features of the preceding context. This model outperformed the method which always outputs the same form of backchannels and also the method which randomly generates backchannels. Finally, subjective evaluations by human listeners show that the proposed method generates backchannels more naturally giving a feeling of understanding and empathy.