Multimodal Sentiment Analysis With Image-Text Interaction Network

More and more users are getting used to posting images text on social networks share their emotions or opinions. Accordingly, multimodal sentiment analysis has become a research topic of increasing interest in recent years. Typically, there exist affective regions that evoke human an image, which usually manifested by corresponding words peoples comments. Similarly, people also tend portray the image when composing descriptions. As result, relationship between associated is great significance for analysis. However, most existing approaches simply concatenate features from text, could not fully explore interaction them, leading suboptimal results. Motivated this observation, we propose new image-text network (ITIN) investigate Specifically, introduce cross-modal alignment module capture region-word correspondence, based fused through adaptive gating module. Moreover, considering complementary role context information analysis, integrate individual-modal contextual feature representations achieving reliable prediction. Extensive experimental results comparisons public datasets demonstrate proposed model superior state-of-the-art methods.