SFNN: Semantic Features Fusion Neural Network for Multimodal Sentiment Analysis

Detecting sentiment in online reviews is a key task, and effective analysis of sentiment in online reviews is the foundation of applications such as user preference modeling, consumer behavior monitoring, and public opinion analysis. In previous studies, the sentiment analysis task mainly relied on text content and ignored the effective modeling of visual information in comments. This paper proposes a neural network SFNN based on semantic feature fusion. The model first uses convolutional neural networks and attention mechanism to obtain the effective emotional feature expressions of the image, and then maps the emotional feature expressions to the semantic feature level. Then, the semantic features of the visual modal is combined with the semantic features of the text modal, and finally the emotional polarity of the comment is effectively analyzed by combining the emotional features of the physical level of the image. Feature fusion based on semantic level can reduce the difference of heterogeneous data. Experimental results show that our model could achieve better performance than the existing methods in the benchmark dataset.