Abstract: In this paper, firstly,the vector representations of text words in comment data are obtained by using Glove word embedding and bidirectional long short-term memory networks. Then, the syntactic information, semantic information, and contextual information of the words are used to construct a heterogeneous fusion graph and combined with graph attention mechanism to learn the text features in the comments. Furthermore, the expression symbol features obtained on the basis of the auto-encoder method are fused through a multi-layer perceptron, and an emotional polarity vector about the disinformation comment content is output. The research results indicate that compared with all baseline models, the multimodal sentiment analysis model proposed in this paper can fully explore the potential interaction between emoticon and text when social network users express their opinions and emotions. This can more effectively evaluate the emotional tendency of disinformation comments, characterize the emotional attitudes or attributes of network users, and achieve effective classification of user emotions. Furthermore, the proposed model provides theoretical support for proposing strategies to suppress false propagation in social networks based on sentiment analysis results, in terms of user sentiment classification.

Key words: graph neural network, auto-encoder, disinformation, online comments, sentiment inclination