Research On Address Standardization And Geospatial-Semantic Model Construction Based On Deep Neural Network

With the improvement of GIS cognitive ability and application ability,address information has gradually become the core resource of the smart city era.Its semantic and spatial connotation is the basis for constructing the geographic ontology and spatio-temporal semantic framework in smart cities.Therefore,the standardized construction and social application of addresses has become a hot topic in the current academic world.From the point of view of understanding address text,it is the key premise for the above task to make the computer extract the comprehensive features of addresses and form quantitative expression in numerical form,so that it has the processing ability of human cognitive level.It has important theoretical value and practical significance for the integration and understanding of urban semantics and geospatial content.However,due to the lack of in-depth mining of text features,the current researchs centered on unstructured text management or address coding are facing such prominent problems as information islands,independent technical routes,dependence on additional data and poor versatility.The use of address data in smart cities is limited.Aiming at the existing problems in address research,this paper uses the deep neural network structure of modern artificial intelligence to transform the tasks of text feature extraction,address standardization,semantic geospatial fusion into quantifiable deep neural network model construction and training.In this paper,the characters in the address are used as the input unit,and the language model is designed to be vectorized.On this basis,the key technologies for the standardized construction of addresses are realized through the target tasks of neural networks.At the same time,considering the expression characteristics of geospace,the feature fusion scheme of address semantics-geospace is proposed,the weighted clustering method and feature fusion model are designed.This study establishes a set of address theory framework and production application system of "Semantic Expression-standardized construction-semantic-geospatial Fusion-downstream task".In this study,the address data of Hangzhou Shangcheng District and Hangzhou Xiacheng District are taken as the research objects,and the core methods are tested,applied and demonstrated.The main contents of this thesis are as follows:(1)An address language model based on text self-attention mechanism is proposed to quantify the linguistic meaning of each character.This paper designs a deep neural network with self-learning ability,establishes a training framework suitable for address characteristics,and forms a set of address language model construction theory.Finally,the experiment proves the validity of the theory for the extraction of semantic features of address texts.(2)The deep neural network structure and training framework of unsupervised word segmentation are established.The optimal training scheme of "micro-supervised"is proposed.The composite geographic entity annotation method with self-learning ability is designed.Finally,the standard address is output by unified specification.Taking the address texts as examples,this paper demonstrates the accuracy and efficiency of this method in terms of word segmentation,annotation and production,and proves that it has more generalization than previous methods based on rules,databases or supervised learning.(3)The definition and design of the theory of semantic-spatial fusion of place names and addresses are given.On the basis of address language model and data standardization,this paper proposes a clustering method combining semantic and geospatial weighting,defines a semantic-geospatial address model and constructs a training framework for classification tasks,then designs the downstream tasks of geocoding regression for verification.Taking the address texts and spatial coordinates as an example,the feasibility and validity of this theory are fully proved,and the semantic-geospatial address model can unify the task framework,effectively combine the address text semantic and geospatial information,and significantly reduce the calculation errors of downstream tasks such as spatial location prediction.This research is expected to achieve the theoretical innovation and model breakthrough of address information standardization,numeral ization,spatialization and intellectualization modeling methods,expected to improve the operational efficiency and generalization ability of address construction and application,and promote the research and development of smart city spatial information construction system methods.