(英) |
This paper analyzes how empathy and antipathy, aroused between people while interacting in face-to-face conversation, are perceived by external observers. The present study aims to investigate the process of their perception and to develop a computational model that realizes the automatic estimation of perceived empathy/antipathy based on the process. The present study first finds that an observer's perception of an interacting pair is affected both by the time lags found their actions and reactions in facial expression and by whether their expressions are congruent or not; e.g. a rapid congruent reaction is likely to produce perceived empathy, while delayed reaction is likely to produce perceived antipathy. Based on the findings, we propose a Bayesian network that relates the perceived empathy/antipathy of external observers to the actions and reactions of conversation participants in facial expression, gaze and head gesture. An experiment is conducted on nine external observers for conversations over 70 min. The results demonstrate that dynamic cues are helpful in improving the estimation performance of empathy perception, especially for antipathy, compared to the use of only static cues, i.e. how behaviors co-occur between a pair at that moment. |