Research On Several Key Technologies Of Knowledge Graph Construction And Their Application In Public Security Field

Public security is closely related to everyone,but the massive amount of security information that exists in the Web cannot be effectively used.Knowledge graph,as a semantic web describing the relationships between objects in the objective world,is an efficient way of expressing knowledge.In public security,complex semantic associa-tions have brought huge challenges to named entity recognition and knowledge repre-sentation,resulting in greater information loss and limited model accuracy.Therefore,this dissertation studies the key technologies of knowledge graph such as named entity recognition and knowledge representation,and applies them to the actual public security knowledge graph construction.The main work of this dissertation is as follows:1.We propose a Chinese named entity recognition method based on multi-embeddings and self-attention.In public security,the data volume is large but the quality is poor,and the grammatical rules are confusing.In addition,the organiza-tional structure and usage rules are chaotic.Existing entity recognition models often require a large amount of feature design and domain-specific knowledge,with a large workload and poor versatility.Therefore,they cannot fully extract sequence seman-tic features and describe global dependencies,lose more semantics,and are difficult to deal with complex situations.This dissertation integrates multiple embeddings and self-attention mechanism into the Bidirectional GRU-CRF model(BiGRU-CRF)for the first time,combined with neural networks such as Convolutional Neural Network(CNN),to fully learn the semantic information and long-distance dependencies con-tained in sentences,and does not depend on hand-crafted features.Experimental results show that this method yields the best or competitive performance.2.We propose a knowledge representation method based on translation principle and attention mechanism.In the field of public security,the knowledge is deeply pro-fessional,with a wide variety of entities and relations,detailed descriptions,and more complicated language scenarios.The traditional translation models ignore other seman-tic information of the object,and only focus on the structural characteristics of the triples in the knowledge graph,and suffer from data sparsity.However,the injection of too much corpus information leads to difficult data acquisition and high model complex-ity.According to the translation idea,this dissertation uses global attention and Gated Recurrent Unit-Convolutional Neural Network(GRU-CNN)to integrate attribute fea-tures and entity description information for entity embedding,and utilizes multi-head attention and CNN to incorporate tag word information into relation embedding,which alleviates the problem of data sparseness and enriches the semantics of entities and re-lations.Experimental results prove that this method markedly enhances the quality of knowledge representation.3.We propose a method for constructing the public security knowledge graph that supports search.The current public security situation is not optimistic.However,most of the existing knowledge graphs are oriented to general domains,and there is a lack of a professional and reliable public security knowledge system to support security services.For public security,this dissertation applies these technologies such as entity recognition and knowledge representation,and combines bottom-up and top-down approaches to implement these modules such as data acquisition,fact extraction,knowledge fusion and graph storage.Finally,it builds a knowledge graph to efficiently manage these multi-source and heterogeneous security information.