We propose a method of detecting suppressed/concealed negative emotion during compliment utterance. When people suppress/conceal emotions, very brief facial expressions called “micro expression” often appear. In order to detect such short-duration facial expression, we propose 90 features calculated from the contours of likelihood ratios for each of the five emotions (happiness, sadness, surprise, anger, and neutral). Likelihood ratios are calculated from still images in a video every 100 milliseconds. Furthermore, 384 acoustic features are calculated for multimodal analysis. Three machine learning classifiers by Support Vector Machines were constructed by using feature sets consist of facial-expression-transition, voice, and both of them, and the classifiers were evaluated how they can detect insincere compliments in Japanese. The experimental results indicate that the feature set including both of facial-expression-transition and voice was the most superior. Its precision and recall of insincerity detection and the total accuracy rate were 0.50, 0.44, and 0.64, respectively. The results were better than the annotation results by non-expert participants.

The Second Workshop on Processing Emotions, Decisions and Opinions (EDO 2017) The 8th Language and Technology Conference (LTC) 2017/11/17, Poznań, Poland

最優秀論文賞（Best Paper Award）受賞論文