Dual-channel Recurrent Neural Network Model for Target-oriented Multimodal Sentiment Analysis
The task of the target-oriented multimodal sentiment analysis is to classify sentiment for a given target word in a multi-modal post or comment.Aiming at the problems that current models incorporating recurrent neural networks in this field only focus on general text and image representations,but never take intra-modal and inter-modal information interactions into account,and ig-nore noise in image information,in this paper,we propose a dual-channel recurrent neural network model(DRNN).The model de-signs a recurrent neural network module based on the attention mechanism,which first uses gate recurrent unit(GRU)to filter the noise of the image,then fuses the text and image through the attention mechanism,and finally adds the fused information to the tar-get information step by step to obtain the dynamic representation between the modalities.In addition,we propose an recurrent neural network module for target-text interaction that learns the contextual representation within a modality by computing target informa-tion with the weight of each word in the context.Finally,we stitch together the information obtained from the two modules and send it to the fully connected and softmax layers to predict the sentiment polarity.Extensive experiments are conducted on two bench-mark datasets,Twitter-15 and Twitter-17,which showed that the model is effective in enhancing target-oriented multimodal senti-ment classification compared to current state-of-the-art models.