| 研究実績の概要 |
Estimating the emotional state of a person is an important task with applications in medicine, social interaction as well as in services industry. Closely related is the task of estimating emotions induced by music which is useful in music recommendation systems. On the other hand, recognition of the person personality is even more challenging task which includes emotion estimation as one of its components. In this research, we used latest developments and technologies in signal processing and machine learning fields to build several systems fro emotion recognition from speech and music as well as text based personality recognition system. The used methods and technologies include Gaussian Processes, non-linear state-space models and deep neural networks. For example, our system for music emotion recognition based on Gaussian Processes scored first for Arousal estimation and second for Valence estimation at the MediaEval EmotionInMusic competition in 2015. Our second system for emotion recognition is based on non-linear state-space models utilizing Gaussian Processes and showed improved performance compared to the linear Kalman Filter system. The results of these two systems were reported at two international conferences and our methodology was described in the IEEE Access journal.
During the last year of this research, we focused on personality recognition task and built a system based on deep neural networks which is capable of recognizing the five personality traits with accuracy of more than 60% only from text data taken from Facebook user updates.
|
