Real Mood Detection Using Denoising Autoencoder and LSTM

論文翻譯標題: 應用降噪自動編碼器及長短期記憶模型於真實心情之偵測
  • 傅 翔祺

學生論文: Master's Thesis


In a rapidly changing social environment emotions are more and more difficult to handle for human beings Sometimes people do not even know that they have negative emotions As a result the accumulation of negative emotions become a mental illness Thus it is essential to develop an emotion tracking system to help users manage their emotions In current study an extended subjective self-report method is generally used for measuring emotions Even though it is commonly accepted that the emotion perceived by the listener is close to the intended emotion conveyed by the speaker several research indicated that there still remains a mismatch between them In addition the individuals with different personalities generally have different expressed emotions Based on this investigation this thesis proposes an emotion conversion model which characterizes the relationship between the perceived emotion and the expressed emotion of the user for a specific personality Emotion conversion from perceived to expressed emotions is applied based on the personality traits of the user This thesis considers mood swing as a long-term accumulation of emotions A database containing user’s long-term speech data and mood annotation is collected This database is used for constructing the temporal relationships between emotion and mood In order to reflect the real mood from people an SVM-based emotion model is developed to generate multiple probabilistic class labels Moreover a Gaussian distribution is built to generate noisy data since there is a difference between expressed and perceived emotions The input is the expressed emotion value contaminated by the generated noise and the target is the expressed emotion for denoising autoencoder (DAE) training Finally for modeling the temporal fluctuation of emotions a long short-term memory (LSTM)-based mood model is constructed for mood detection In mood detection experiments the mood database was provided by 10 participants There were 104 positive moods and 96 negative moods Leave-one-speaker-out cross validation was employed for evaluation Experimental results show that the proposed method achieved a detection accuracy of 64 5% which improves 5% comparing to the HMM-based method In the future the tracking of the dialog content and blog of the users can be applied to obtain a better performance
獎項日期2016 8月 25
監督員Chung-Hsien Wu (Supervisor)