PAD维度下的深度情感关联模型研究

Research on deep affective association model in PAD dimension

人类的情感不是相互独立的,而是以系统的方式相互关联。为此,提出用PAD情感维度预测值作为关联认知网络的权值,构建深度情感关联模型。以TYUT2.0情感语音数据库和柏林德语情感语料库的“悲伤”、“愤怒”及“高兴”三类情感语句为基础提取情感特征;通过PAD情感维度预测模型得到各情感的PAD预测值,并用各类情感预测值的均值计算关联认知网络(ICN)情感间的权值;用遗传算法(GA)优化深度情感关联模型超参数;根据是否使用预测值作为权值以及是否使用遗传算法设置四组实验方案进行对比实验。实验结果表明,对比未使用PAD预测值及遗传算法的方案,仅使用PAD预测值的方案识别率提高4.26%,仅使用遗传算法的方案提高6.39%,同时使用PAD预测值和遗传算法的方案识别率提高8.51%;数据显示,使用PAD预测值且优化参数的深度情感关联模型具有较好的识别性能。

Human emotions are not independent of each other,but interrelated in a systematic way.Therefore,the pre-dictive value of PAD affective dimension is used as the weight of Interactive Cognitive Network(ICN)to construct the deep affective association model.Based on TYUT2.0 emotion speech database and Berlin German Emotion corpus,the emotional features were extracted.The PAD predictive value of each emotion was obtained by using the PAD affective dimension predictive model,and the weight between the ICN emotions were calculated by using the average of the predictors of various emotions.Genetic Algorithm(GA)was used to optimize the hyperparameters of the deep affective association model.Four groups of experimental schemes were compared according to whether the predicted value was used as the weight and whether the genetic algorithm was used.The experimental results showed that,compared with the scheme does not use PAD predictive value and GA,the scheme recognition rate with PAD predictive value alone increased by 4.26%,that with GA alone increased by6.39%,and that with PAD predictive value and GA increased by 8.51%.The data show that the deep affective association model used PAD predictive value and optimized parameters has better recognition performance.