Research And Application Of Cross-modal Sensitive Information Retrieval

With the popularity of the network,more and more users in Xinjiang begin to obtain information from the Internet.Due to the regional particularity,information dissemination and access in the network make users more convenient to obtain information.However,many negative effects also follow.The sensitive false information and the dissemination of information related to Xinjiang and violence in Xinjiang rapidly form social public opinion and produce huge public opinion Pressure.Therefore,it is urgent to take effective measures to control and deal with these sensitive information in time.Most of the first mock exam vectors are distributed in the network by means of text,image and video.Traditional detection methods are based on single mode and can not handle the heterogeneous data in the network.The introduction of cross modal retrieval is helpful to multi-modal data processing and cross modal matching in the network,so as to label sensitive information more comprehensively and accurately,which is of great significance and value for research.In general,the main work of this thesis is as follows:(1)A cross modal correlation analysis model integrating attention mechanism is proposed.The model extracts text features through bidirectional GRU network,and extracts image features by VGGNet-LSTM network,which is conducive to the deep finegrained expression of text and image features.The attention mechanism is used to establish asymmetric links between text and image,which provides a powerful text representation at the same time as image,and reduces the semantic gap between image and text.(2)The cross modal semantic matching model based on improved double coding is proposed.By combining multi-level coding,the network can learn to represent global,local and temporal patterns in video and text.At the same time,the whole model is trained in an end-to-end way,which alleviates the semantic heterogeneity between text and video.(3)The sensitive information retrieval subsystem based on cross modal is designed and implemented.The system verifies the validity and practicability of the model proposed in this thesis,including three functions: sensitive information annotation,sensitive information retrieval and sensitive information database establishment.It effectively helps staff to control and process sensitive information in real time,and meets the requirements of the project.