Research On Multi-Modality Semantic Space Learning And National Security Emergency Detecting In Social Network

Social network is a user relationship-based information creation and dissemination platfom,which has sprung up in recent years.With the rapid development of the Mobile Internet,social networks have gradually become an indispensable part of people's lives.While users publish the text on social network platforms such as Weibo,they often add some images as a supplement to the description of the event.Therefore,extracting deep feature from social network multi-modality data and learning multi-modality semantic space to achieve information fusion and complementation is of great significance for improving the effect of topic discovery and emergency detection.Detailed contributions of this thesis are summarized as follows:(1)For text in the social network,we propose a user interaction based short text semantic extension(UISSE)algorithm,and then use the deep denoising self-encoding network for deep feature extraction.For images in the social network,we propose a method which used spatial pyramid pooling to refactor traditional convolutional neural networks.Experiments show that compared with the baseline methods,the precision,recall,and F-measure of the proposed deep feature extraction and expression method are improved.(2)A multi-modality deep quantization(MDQS)algorithm is proposed for the social network.The MDQS algorithm first maps deep features of text and images into continuous potential semantic space,then maps them into discrete hash semantic space to improve retrieval efficiency.Experiments show that the MDQS algorithm can effectively learn multi-modality semantic space from social network dataset.Comparing with baseline methods,MDQS achieves better results on same-modality retrieval and cross-modality retrieval.(3)A multi-modality semantic space-based topic detection(MSSTD)algorithm is proposed.The MSSTD algorithm constructs a text graph and a visual graph for the social network in each time slides,and performs multi-modality graph fusion according to the semantic similarity and the time decay coefficient.Finally,the topic discovery and the emergency detection result are obtained through the topic recovery algorithm.Experiments show that the MSSTD algorithm can effectively detect the emergencies from social network multi-modality big data.Comparing with baseline methods,the quality and bursty of the topics detected by MSSTD are better.(4)A multi-modality semantic space learning and national security emergency detecting system is designed and deployed for social networks.The system includes four modules:data collection,deep feature extraction,emergency detection,and system exhibition.The system is very useful and effectively validates the algorithms proposed in this thesis.