Research On Semantic Address Matching And Semantic-Geospatial Fusion Model Based On Pretrained Deep Learning Architecture

With the continuous development of China's digital city and smart city construction,address information,as the strategic basis of geographic information and spatial data resource,plays the pushing role in daily life of human,and also plays a crucial role in the country's economic construction,cultural development and social management.Driven by the demand of big data application based on location service in all walks of life,China's relevant departments have collected and accumulated massive amounts of address data.However,due to the various address standards in China and the confusion of manual collection and management order,the analysis and understanding of address information has become a big problem,which greatly limits its application in all walks of life.Therefore,it is necessary to form the Address Semantic Model from the perspective of recognizing address text information and understanding address semantics knowledge,so that it can dig into the semantic features of address information and be applicable to the high-speed computing of computers.In addition,the unique addressing property of address makes it of great theoretical and practical significance to understand and merge its semantic and spatial information,and it is also a hot academic topic.Aiming at the dilemma of incomplete expression of semantic information,insufficient application of intelligent information,and weak generalization of related task scenarios in the existing research on address models,this paper uses the attention mechanism and pre-training fine-tuning mode in deep neural networks,to address semantic understanding,semantic matching and semantic space fusion and so on correlation large tasks into can calculate the depth of the construction of the neural network model and optimization problem.In view of the characteristics of address corpus,a deep learning framework using self-supervised learning is constructed to form the Address Semantic Model to support all related address tasks.On this basis,the model is fine-tuned by generating supervised matching dataset,so that the model can identify the semantic similarity between addresses and realize the high-precision address matching task.At the same time,considering the unique spatial attributes of addresses,a dataset that follows the rules of spatial similarity was designed and the spatial distance and address semantics were deeply correlated and fused by fine-tuning the Address Semantic Model.The study systematically constructed a theoretical system and method framework of Chinese address supported by semantic cognitive comprehension-address exact matching-spatial semantic deep fusion-downstream application task verification,and tested,applied and verified the method system with the million-magnitude address corpus of Shang Cheng district,Hang Zhou city,Zhe Jiang province as the experimental data.The research content of this paper is summarized as follows:(1)The Address Semantics Model based on the generalized autoregressive pretraining method is constructed under the structure of deep neural network to realize the automatic acquisition of address semantic information,and the new paradigm of "pretraining-fine-tuning" is introduced into address research.Considering the interaction between each character in the address and the bidirectional contextual word order relation,the permutation address language model is proposed.This model uses the dual-track self-attention mechanism to solve the problem of missing target location information in the permutation address modeling.Finally,a novel deep multi-level neural network with self-learning ability and the ability to provide the transfer learning ability of all related applications is designed.The experimental results show that it has realized the semantic information understanding and representation of the massive multi-source heterogeneous address data set,which lays a solid foundation for the excellent performance of the follow-up application task research based on this model.(2)The design and implementation of the Semantic Address Matching method based on the Address Semantic Model is presented.Based on geospatial information,the dataset of supervised address pairs with labels is constructed,and a deep neural network architecture and training framework for effective address matching task is established.The experiment is carried out with the constructed semantic address matching data set as the object,and the result proves that this work can effectively solve the redundancy,incomplete or abnormal expression in address matching,and has the performance of "high precision and light flow".It is proved that the two-stage paradigm of "unsupervised pre-training first-supervised fine-tuning later" can greatly improve the accuracy and effectiveness of the task.(3)In this paper,a novel theoretical design of address semantic-spatial deep fusion is proposed,and the study of address semantic-spatial characteristics has realized the transformation from the past "physical combination" to the "deep fusion" in this study.A training dataset is constructed which integrates geospatial location information with address semantic information,and on the basis of Address Semantic Model,a SemanticGeospatial Fusion Model is built by fine-tuning training of regression task.The address representation containing address specific addressing properties that can be understood by the computer is implemented and the downstream task of address location prediction is set up for verification.The experiment takes the address strings and their corresponding geospatial coordinates as the object,the experimental results fully prove the effectiveness of the design,and demonstrate that the model can integrate the address semantic information and geospatial information more effectively than the previous models,which greatly improves the accuracy of the address location prediction task.This study is looking forward to achieving the theoretical and model's innovations in intelligentization,structured,and numerical modeling methods for address information,and solve the problem of low quality and excessive quantity of address data driven by big data.It improves the address information analysis and mining operation efficiency and migration ability to learn,to promote the research on urban address model,promote the wisdom of urban space information construction method in the research and development and the popularization service system.