Research On Cross-media Sentimental Analysis Based On Deep Learning

With the development of social media on the Internet,people began to use more types of devices to express their personal opinions and comments.These mixed-text comments are fragmented but contain lots of information.The emotional information contains people's emotional tendency,which is significant for the government's decision-making department to grasp the dynamics of social public opinion and merchants to obtain consumer feedback.Therefore,how to extract the implicit emotional information from these mixed data of images and texts needs to be solved urgently.In recent years,deep learning techniques have achieved remarkable results in the direction of image classification and text classification.In image classification tasks,convolutional neural networks have become the most important feature extraction method,but the performance still needs to be improved.In the same way,in the text classification task,the emergence of word embedding technology is also exciting,but how to apply it better is still the focus of attention.At the same time,for the mixed data of images and texts(even data with other more complex information),how to comprehensively utilize their features is also the pain point of various mainstream models in deep learning direction.Based on the above problems,the main work of this thesis is as follows:Firstly,for the feature extraction scheme of image data,the latest attention-based mechanisms based on channel-wise and spatial-wise are studied.Then we embedded them in VGG(Visual Geometry Group Network)and ResNet(Resual Network)separately,we found that the accuracy of the image classification model performed better.Secondly,in the feature extraction of text data,the textual convolutional neural network and the two-way long-term memory network combined with the spatial-wise attention mechanism are used to construct the classification model.Finally,in the aspect of graphic and text mixed data,we introduce the graph convolutional neural network suitable for non-Euclidean distance and use it to construct a text classification model.Compare the performance of the graph and the existing text classification model in the experiment and explain its effectiveness;Then based on the previous experiments,a hybrid data training scheme based on graph convolutional neural network is designed and proposed for the training of mixed data in the academic field.The topological connection expression ability of the convolutional neural network is different.The data characteristics of the class are mixed and trained,and the emotion classification task of the mixed data of the graphic and text is completed.The experimental results show that the graph convolutional neural network constructed in this thesis has strong feature extraction ability for mixed data of images and texts,and can also be used for various forms of data features.