Research On Algorithms For Multimodal Sentiment Analysis Based On Interaction Fusion

Sentiment analysis is an active research field in natural language processing,which mainly focused on mining the sentiment polarity from textual content of the past.With the development of the mobile Internet and the popularization of mobile devices,the trend of message content has changed from text to the combination of text and picture.And the research target has also changed from text to picture and text.The process of recognizing the sentiment expressed by multimodal data is multimodal sentiment analysis.And how to fusion the information on multimodal data then improves the effect of emotion recognition is the mainly research content of multimodal sentiment analysis.Previous works do not effectively utilize the relationship and influence between texts and images.We proposes a multimodal sentiment analysis algorithm based on interactive fusion.It uses a fine-grained attention mechanism to learn the relationship and catch the influence between images and text.And it extracts the key information which uses to recognition the emotion from multimodal fusion features by self-attention mechanism.When the attention mechanism models the relationship of the multimodal data and the self-attention mechanism extracts features of the multimodal data,there are a problem of information loss.We optimize the multimodal sentiment analysis algorithm based on interactive fusion,then we obtain a new method which calls multimodal sentiment analysis algorithm based on gated interactive fusion.This method uses a gating mechanism to overcome the noise of multimodal fusion features,and uses a convolution neural network to enhance the information extraction effect of the self-attention feature extraction layer.We conduct evaluations on two public multimodal datasets,namely MVSA-Single and MVSA-Multiple.The result of experiment shows that our algorithms outperform existing algorithms,all the modules of our algorithms play a role in improving the performance of sentiment recognition.