Research On The Method Of Constructing Knowledge Graph For Social Security Incidents

Internet is a double-edged sword,which not only satisfies people's information needs,but also raises serious network security issues.The social security event studied in this paper is spread on the Internet and belongs to information content security events.The Internet public opinion triggered by the rapid spread of social security events may cause negative social influence,destroy the image of party and government,affect social stability.Knowledge graph realizes data association analysis by integrating entities,concepts,and relationships.It can help researchers to study events with the help of relationship analysis.The knowledge graph constructing of social security events is important to explore the diversified relationship by analyzing event information.This paper studied the construct method of knowledge graph on social security events.The research contents include the following five aspects:(1)The structured data resources construction of social security eventsThe paper completed the data repository construction based on Baidubaike.We proposed the Baidubaike data processing method of acquisition,parse,and storage.Also we realized the Neo4 j storage of nodes,relationships,attributes,etc.The hierarchical design of the social security event data repository was proposed.We completed the construction of the word distributed representation,Json repository,and Neo4 j database.(2)The method of named entity recognitionThe paper based on deep learning model to study the named entity recognition and proposed two bidirectional LSTM optimization models--the center character enhancement model and the reverse sequence enhancement model.The model of center character enhancement can optimize the core character of the current sequence window,and the F1 value reached 97.16%.The model of reverse sequence enhancement can achieve the optimization of the reverse sequence,and the F1 value reached 97.27%.Experimental results showed that the two optimization models enhance the effect of named entity recognition.(3)The method of entity relation extractionThe paper proposed relation extraction optimization model--word distributed representation optimization model and parallel structure optimization model,and realized open relation extraction on dependency syntax analysis.The relation extraction optimization model solved the problem that the traditional deep learning model can not learn the POS feature.The word distributed representation optimization model added POS vector on the basis of word distributed representation.The parallel structure optimization model learned entity pairs' POS feature,which has achieved good results in the CNN,LSTM,and GRU models,the F1 value increased by 12%,6.96% and 6.07%.We realized the noun phrase recognition based on the ATT relation and the open relation extraction based on the dependency syntax analysis.Experimental results show that open relation extraction methods achieved good results.(4)The method of event summary generationThe automatic summary generation based on TextRank algorithm is studied and implemented.We proposed a summary generation model based on TextRank and sentence comprehensive similarity.Sentence comprehensive similarity integrated lexical similarity,statistical similarity and semantic similarity.The model solved the problem of abstract redundancy and improved abstract extraction effect.(5)The design and implementation of event knowledge graph systemThe paper designed the overall framework of the system and the functional architecture of the sub-modules.We designed and implemented the event knowledge graph system which based on technologies such as Webcollector crawler framework,dependency syntax analysis,TextRank algorithm,and Neo4 j graph database.