Research On Construction Of Geographic Knowledge Graph Driven By Natural Language

"How to express,construct and storage geographic knowledge scientifically?"is always one of the fundamental key issues to geographers.Knowledge is the symbol of wisdom that is aggregated by all the results of both the material and spiritual cognitions.Geographic knowledge is a broad boundary concept,which includes types of concept statements,evolutionary mechanisms,constraint relationships and et.al,.In order to use geographic knowledge in current high-tech situation with Internet,Thing of Internet,Big Data and Artificial intelligent,the relatively isolated and complicated geographic knowledge needs to be extracted and organized.Knowledge graph is a structure of knowledge representation,which represent concepts by nodes and relations by the links between the nodes.Using the graph-based networks,geographic knowledge could be structured with well representation,organization and correlation.Additionally,natural language is the fundamental representation forms to the knowledge.Therefore,this dissertation systematically studied the construction of geographic knowledge graph driven by natural language which follows the principle line of "knowledge representation modeling,geographic information extraction,geographic knowledge generation,knowledge graph construction" from the perspective of geography,linguistics and computer science.The main research topics and conclusions in this dissertation are discussed as follows.(1)The Construction and the formalization of the conceptual model of geographical knowledge representationA conceptualized model of geographic knowledge representation was proposed in the view of the basic issues of geography,which is aggregated with geographic entities.And then the issues of knowledge representation on the logical foundation were analyzed in order to make geographic knowledge computable.According to the issue of semantic loss of descriptive logic of geographic knowledge graph,the construction operators of ALC descriptive logic were supplemented including four aspects of inclusion,reverse,transfer and value limitation.These improvements implement geographic knowledge formalization,especially the representations of states,evolutions and processes of geographic knowledge.(2)Geographic entity recognition in natural languageAs geographic entity is the key part to geographic information extraction,geographic entity recognition is mainly focused on.In this dissertation,a context related word representation was proposed based on the characteristics of geographical entity descriptions in natural language.And then,a deep belief networks based geographic entity recognition model using this word representation was proposed.The results indicate that the performances of two models have almost the same statistics results.However,the recognition results are significantly complementary,The combination of the two models can be applied to the recognition of geographic named entity to improve the accuracy of information acquisition to a great extent.Moreover,spatial information,temporal information,attribute information and relations were also extracted by pervious methods,which made the foundation for geographic knowledge generation.(3)The generation and storage of geographic knowledgeAfter the representation of three representation levels of geographic knowledge(feature layer,composition layer,and mechanism layer),this dissertation proposed the construction method of geographic knowledge including temporal knowledge,spatial knowledge,attribute knowledge,expression state knowledge,changing knowledge,relationship knowledge and evolutionary knowledge by using bottom-up correlation method.Moreover,the differences between graph database and the conceptual model of geographic knowledge were analyzed.And the elements of conceptualized model were mapping to nodes and relations,which are the elements of graph database.By using this way,the geographic knowledge can store in a graph database.(4)Geographic knowledge graph construction and experimental verification and analysesBased on the data of "Encyclopedia of China,Chinese Geography",the geographic knowledge graph is constructed,and the statistical performance index and the ability of geographic knowledge expression are evaluated and analyzed.The research shows that the geographic knowledge graph constructed in this dissertation is more prominent than the other domain knowledge graph in the statistical performance index.In the aspect of knowledge expression ability,geographic knowledge graph is perfect in temporal knowledge,spatial knowledge,attribute knowledge and evolutionary knowledge,and it lacks in expression state knowledge and relationship knowledge.In conclusion,the research indicate four main issues.In the stage of the conception and formalization,adapted description logic constructor set(SHIQ)can support geographic knowledge formalization that avoid the weaknesses of the representation of geographic knowledge of the change and the relation on the current description logic of ALC.In the stage of acquisition,Deep Belief Network Model represents the validity on the recognition of geographic entity name and deep learning could have a great foreground on geographic information retrieval.In the stage of storage,geographic knowledge cannot comprehensively adapt to general graph based database,geographic knowledge based database warrant further studied.In the stage of validation,there does not exist the whole methodology of geographic knowledge evaluation,the geographic knowledge evaluation system should be a key issue in the future.