Research On Deep Affective Association Model Under PAD Dimension

How to make the computer correctly recognize,understand and express the emotional information in human speech is one of the main research directions in the field of artificial intelligence.In speech emotion recognition,the selection of a suitable emotion recognition model is one of the keys to achieve effective emotion classification,but most of the emotion recognition models currently used divide emotions separately and do not consider the mutual influence between emotions.Human emotions are not independent,but related to each other in a systematic way.The PAD three-dimensional description model describes different emotions from three aspects: pleasure,activation and dominance in the dimensional space.Therefore,it is proposed to use the predicted value of PAD emotion dimension to represent the degree of association between emotions,use the predicted value of PAD emotion dimension to calculate the weight of the Interactive Cognitive Network,and combine the multi-layer Restricted Boltzmann Machine to construct a Deep Emotional Association Model.The experimental results show that the recognition rate of this model in the TYUT2.0 emotional speech database reaches 80.85%,and the recognition rate in the EMO-DB emotional speech database reaches95.74%,it is an effective model for emotion classification The main research contents of this paper are as follows:(1)Use the predicted value of PAD emotion dimension to calculate the correlation between emotions.Because different databases or even different sentences in the same database have different PAD values,if only the PAD value(basic emotional PAD value)is used to express the degree of association between emotions,it will affect the recognition effect.Therefore,a method of PAD emotion dimension prediction is proposed to predict PAD for different emotion statements,and the correlation degree among emotions is calculated with the predicted value,uses the degree of emotional association as the weight of the Interactive Cognitive Network.The experimental results show that the scheme using PAD predictive value to calculate the weight of the Interactive Cognitive Network has a recognition rate of 76.60% in TYUT2.0 and91.49% in EMO-DB.For the scheme using the basic affective PAD value to calculate the weight of Interactive Cognitive Network,the recognition rates in TYUT2.0 and EMO-DB were 72.34%and 87.23%,respectively,.The experimental results proved that using the PAD emotion dimension predictive value can better represent the relationship between emotions.(2)Hyperparameter optimization of Deep Emotional Association Model.In order to avoid the blindness of the hyperparameter setting of the Deep Emotional Association Model,it is proposed to use Genetic Algorithms to optimize the hyperparameter of the model,such as the momentum factor,the learning rate,the number of the hidden layer nodes,the training times of the multi-layer RBM and the ICN.The validity of the method is proved by comparing the scheme.The experimental results show that the recognition rates in TYUT2.0 and EMO-DB are 78.72% and 93.62% for the scheme using Genetic Algorithm to optimize the model's hyperparameter,and 72.34% and 87.23% for the scheme using experience tuning.The experimental results proved that the use of genetic algorithm to optimize the hyperparameters of the Deep Emotional Association Model can effectively improve the model recognition effect.(3)Improvement of Deep Emotional Association Model.To further optimize the structure of the Deep Emotional Association Model,starting from the temporal information of the speech signal,the Gate Recurrent Unit that can capture the temporal information of the speech signal is introduced to construct the GRUs-ICN model.The experimental results show that the recognition rate of GRUS-ICN model is 76.71%,and the recognition rate of Deep Emotional Association Model is 76.60% under the condition of using PAD prediction value to calculate the emotion correlation degree and manually adjusting the parameters.The experimental results proved that the temporal information of speech emotion is beneficial to speech emotion recognition.