Research on Emotion Analysis of Multimodal Learning Supported by Deep Learning

Faced with the increasing demand of multimedia data analysis, traditional sentiment analysis methods are limited because they only rely on text information. Therefore, the paper introduces a novel multimodal fusion technique aimed at enhancing the precision of sentiment analysis. In terms of text feature modeling, this paper uses bidirectional gated loop unit (BiGRU) to extract deep context features of single-mode data, so as to better capture semantic information and emotional tendencies in text. For the extraction of video and audio modal features, considering the uniqueness of time series data, we apply 1D Temporal Convolutional Layer on the input sequence to enhance the perception of modal characteristics. This structure efficiently captures the relationships across various time points, subsequently enhancing the model's comprehension of emotional expressions. Experimental outcomes from several publicly accessible datasets demonstrate that the introduced multi-modal fusion model surpasses single-modal models in the task of emotion classification.