| In the multimedia data of the Internet,some of the data are discriminatory.These discriminatory feelings,through the dissemination and diffusion of the network platform,may have different degrees of negative impact on individuals and society,which is not conducive to the healthy growth of individuals and social harmony and stability.From the perspective of data media,these discrimination phenomena can be divided into three categories:language discrimination,image discrimination and multimodal discrimination.Multimodal discrimination detection is of great significance in the content supervision of social media.In this paper,a deep learning-based approach is used to study the problem of discriminative emotion detection for multi-modal high-level semantics.The details are as follows:(1)For the detection and determination of multi-modal high-level semantic discrimination emotions,analyze the factors that constitute discrimination,classify the discrimination phenomenon,and use the multi-modal data collected by the artificial shooting and crawling from the network to construct a multi-modality for discrimination state dataset.(2)In view of the problem that the previous multi-modal emotion detection model is too simple for different modal feature aligrnment,this paper proposes a multi-modal feature alignment method to achieve more fine-grained alignment of image features and text features.Convolutional neural networks and long short term memory neural networks efficiently extract image and text features,and use bilinear fusion to semantically align features of images and text.(3)For the problem that the feature vector contributes differently to the discriminatory emotion,the feature vector of different modalities is used to obtain the joint representation of multimodal data.And the experimental comparison of different indicators on the constructed multimodal discrimination data set effectively improves the accuracy of multimodal discrimination detection. |
PDF Full Text Request