Research On Visual And Textual Emotional Analysis Based On Convolutional Neural Network

With the development of technology and the popularity of hardware and software,reading and commenting on the Internet has become people's daily activities,therefore the Internet contains a variety of valuable information.Emotional analysis is a basic and important research direction in the field of Natural Language Processing,it use emotional analysis techniques to analysis people's emotional attitude toward a specific thing,and the attitudes mainly include two emotions of positive and negative.Traditional and extensive research work mainly focus on text data,with Microblogging,Twitter,Facebook and other platform development,people can easily publish pictures,video and other forms of data,so the researchers are not limited to the text emotion classification,they also take the pictures,video and other multimedia information into account to do cross-modality sentiment classification.This article uses the word vector representation tool word2 vec and the Glove model to represent the text,and use the deep learning model to do sentiment classification from three aspects of the text,image and image-text fusion.Details as follows.1.Sentiment classification with multiscale convolutional recurrent neural network.In this section,this model utilize convolutional neural network(CNN)to extract word feature which contain a rich context information,and then utilizes the long short-term memory(LSTM)model's advantages which can deal with arbitrary sentence length sequences and have long-time dependencies,finally using the extracted features do the English sentiment classification task.The experimental results show: this model we proposed has some promotion with the results obtained using other models.2.Image global and local combination of sentiment classification.CNN has a significant advantage in extracting image features,this article use CNN model to do image sentiment classification.This method can be divided into four parts of the neural network,the first part of the network extract the low-level features of the image,the second and third part of the network utilize the extracted low-level features,and then use convolution,full connection operation to extract the global features and local features of the image,the fourth part fuses the local features and global features,and then execute the final sentiment classification.Experiments show that the model comparing with the existing deep learning emotion classification method obtains the better emotion classification effect.3.Image and text cross-modal sentiment classification based on multimodal compact bilinear pooling(MCB).In this part,using the pre-training 152-layer residualneural network to extract image features,using LSTM model to extract text feature.And then generating a soft attention map by combining two features and get the final visual representation.In the end,using MCB algorithm to fuse visual representation and text features to execute cross-modal sentiment classification.In the experimental comparative analysis,the model obtained the better classification results compared with other methods.