Detection And Search Of Cross-media Emergencies In Social Networks Based On Deep Learning

With the flourish of social networks,big data in social networks show a huge growth trend in the form of rich cross-media.Since the messages released by the social networks have the characteristics of rapid dissemination and wide sharing,cross-media big data reveal abundant social phenomena and contains a lot of emergencies.In order to meet the needs of the users and organizers for effective management and trend analysis of emergencies,it is of great significance to carry out research on social networks emergencies detection and search.However,the presentation forms of messages in social networks emergencies have a certain degree of randomness and semantic sparseness,and the feature space of cross-media data is heterogeneous,which makes emergency detection and search in social networks face great challenges.This dissertation focuses on the detection and search of cross-media emergency in social networks,and studies the relevant crucial theories and technologies.It includes the acquisition and expression of the emergency semantics based on the multi-attribute features of social networks,the semantic analysis for the emergency of social networks based on the deep semantic hashing,detection and search of the social networks cross-media emergency.The main research works and contributions in this dissertation are as follows:(1)Aiming at the problem of semantic sparsity of cross-media data in social networks and insufficient mining of association relationship and extended semantics between data by existing methods,this dissertation proposed a semantic acquisition and expression method for emergencies based on multi-attribute features of social networks,including a short text semantic acquisition and expression algorithm based on social and conceptual expansion(SCSE)and an image sematics acquisition and expression algorithm based on hashtag heterogeneous graph model(HHGM),respectively.In terms of short text,SCSE algorithm obtained an explicit semantics of short text based on external knowledge base,and constructed the social and conceptual semantic graph model by integrating the multi-attribute features of social networks,such as topic hashtag and link information.The graph model mined the potential semantics relevance among short texts.The extension of short text based on the rich semantic information alleviated the sparseness of emergency text to a certain extent,and generated the semantic feature representation of short text with explicit and implicit semantics.In addition,HHGM algorithm introduced the hashtags into image data,and a heterogeneous social networks graph model was constructed to realize the analysis of semantic associations among images.Based on the characteristics of neighbor aggregation in graph convolutional networks,the image semantics were supplemented and learned through association hashtags to obtain image features with rich semantic representation.The experimental results showed that the proposed SCSE algorithm and HHGM algorithm can effectively alleviate the sparseness of multi-media data of social networks in the semantic acquisition and expression,and achieved better performance on the semantic acquisition and expression of the short text and image,respectively.(2)In order to solve the problem that the existing semantic analysis methods of social networks events are difficult to overcome the semantic limitations in deep semantic feature learning,which leads to the performance degradation of emergency detection and search,this dissertation proposed a short text semantic analysis algorithm based on double semantic extension and deep hashing model(SCSE-DH).It can achieve efficient event detection and search.Based on the proposed SCSE algorithm,a deep hash model was established,and deep semantic feature analysis and learning were performed in the short text after double semantic extension.The rich and refined semantic representation of short texts in social networks was obtained.In the combined training and learning based on the stack auto-encoder and the semantic hash network,the internal semantics of short texts were captured and compressed effectively,and the semantic information of short texts was retained by layer-by-layer reduction.The event detection and search based on deep semantic hash feature obtained remarkable results.The validity of the proposed SCSE-DH algorithm was verified by experiments in short text semantic analysis,event detection and search tasks.(3)Aiming at the problems of the existing emergency detection and search research,which focused on the study of a single feature leading to the loss of important information or the inability to effectively integrate cross-media information,a multi-view graph attention network model based on time information guidance(T-MVGAN)was proposed.A multi-view model of social networks event was established.The comprehensive presentation of social networks event through three aspects:text,image,and time factor.It realized the complementarity and correlation between the characteristics of social networks cross-media emergencies.A heterogeneous graph model of social networks cross-media emergencies was desigened.Based on the proposed HHGM algorithm,image semantic features were obtained,and in the text heterogeneous graph model,the semantic features and temporal distribution features of short texts were learned.In addition,a multi-view graph attention network model guided by temporal distribution information was established.The temporal feature was used as a bridge and the consensus information of social networks cross-media data.It fused text and image features to learn an effective representation of social networks cross-media emergencies,and achieved accurate detection and search for the emergency.Through a large number of experimental evaluations,it was verified that the T-MVGAN model can obtain rich feature representations of social networks cross-media emergencies,and achieved better performance of the detection and search for social networks emergency.(4)Combining the proposed SCSE,HHGM,SCSE-DH algorithms and T-MVGAN model,a detection and search system of social networks cross-media emergency based on deep learning is implemented.The system included the module of semantic acquisition and expression of the social networks emergency,the emergency in social networks deep semantic analysis module,cross-media emergency detection and search module.It evaluated the performance of all the algorithms proposed in this dissertation and displayed the effectiveness.The module of social networks cross-media emergency detection and search addressed the task of emergency detection and search,respectively.It provided the emergencies that occurred within a specified period of time or a given query content.The effectiveness of the proposed algorithms in this dissertation was verified and evaluated by the implememted system.