Abstract
Rapport is a harmonious relationship with others. We aim to automatically estimate a speaker's subjective rapport using their nonverbal behavior during a conversation. Rapport estimation is generally formulated as a regression. However, it is difficult to learn a mapping between nonverbal behaviors and rapport ratings during a conversation because of individual differences in rapport ratings. To alleviate this problem, we formulate rapport estimation as learning to rank. Learning to rank avoids the problem of individual differences in rapport ratings using preference learning, which learns the ordinal relationship between two conversations based on rapport reported by the same user. To evaluate the proposed method, we used a dataset consisting of first-meeting conversations and friend conversations that includes subjective rapport ratings. We compared the proposed model with the regression model using metrics for ranking. The result indicates that the proposed model is more suitable than the regression model for rapport estimations.
