Emotion Classification and Its Application on Humanoid Robot
Ning Liu · Institutional Repositories DataBase (IRDB) · 2018
劉 寧which can recognize nine emotion categories in texts as emotional trigger to generate the corresponding action labels according to the robot's response.For the robot system, running the computed basic expression and the voice at the same time, we can get an acceptable humanoid robot interaction with emotion expression.During the running interaction with Actroid REN-XIN, the fast WMD based emotional trigger system needs at least 7s to deal with the response.To make a real time interaction, the seamless user experience is a essential aspect.Thereby, for people are communicating with humanoid robot, the delayed feedback results to no long communication desire.To solve this rough gap, we propose a CNN+LSTM based DNN model.In the experiments, we utilize the same sub-data sets of the Chinese emotional corpus(Ren_CECps) used in fast WMD experiments.The experiments are proceeded in fast WMD, CNN+LSTM, CNN and LSTM respectively.The results show that CNN+LSTM gets the best result of F1 score 0.35 in 1v1 experiment, and almost the same accuracy with fast WMD of F1 scores 0.367 with 0.366 in 4v1 experiment.In the training process, our experiments show that the DNNs only need 3 epochs to finish training.This is not only the difference between minutes and weeks cost in training, but also the extended flexibility for the actroid robot.Our contributes show the CNN+LSTM model has excellent ability for emotion classification and robot control with time sensitive.