Abstract: Aimed at some problems in multimodal sentiment analysis methods, such as equal treatment of different modal features, insufficient mining and fusion of information carried by each modal, and low accuracy of sentiment classification, a multimodal sentiment analysis model based on information enhanced multi-head attention is proposed. We used multi-task learning to mine the shared semantics of other two non-text features (speech features and visual features) relative to text features, and enhanced the representation of text emotion features. Moreover, since the interaction information of any two models made different contributions to the final sentiment prediction, we designed a multi-head attention fusion network, which could effectively fuse the effective information carried by different model through the reasonable allocation of speech-text, visual-text and speech-visual features. The experimental results show that the performance of the model in multimodal emotion classification tasks is better than the existing methods.