Multimodal emotion recognition based on fusion with residual connection

Abstract

To improve the accuracy of emotional identification, a multi-mode fusion emotional recognition method of two-way long and short-term memory network (Bi-LSTM), Multi-Head Attention and Residual Connection blended emotional identification methods. This method performs long-term memory through LSTM, then uses the Attention mechanism to screen out important information, and finally improves the ability of network information transmission through the residual connection. Through this method of concentrated verification, the accuracy rate of emotional classification reaches 61.7%. The experimental results show that compared with models such as CNN, CMN, BC-LSTM, the accuracy and F1 value of the proposed model are effectively improved.