Deep learning address word segmentation method combined with geographic information enhancement

Abstract

The invention relates to a geographic information enhanced deep learning address word segmentation method. The method comprises the following steps of: preprocessing an address text, receiving an original address text, and performing text cleaning, address standardization and semantic disambiguation processing; feature fusion extraction: extracting basic features of the text and multi-dimensional spatial features of geographic information, and performing feature fusion by using a self-defined convolutional neural network and a bidirectional long-short term memory network; model customization training: designing a special deep learning model architecture and training by using a professionally labeled address data set; and word segmentation execution and optimization: inputting the preprocessed text into the trained model to carry out word segmentation operation, and optimizing and verifying a result by utilizing a post-processing rule and a special verification system. According to the method, the precision, efficiency and adaptability of address word segmentation are effectively improved by combining geographic information and a deep learning technology, and the method is suitable for multiple fields such as map navigation, logistics distribution and public security household registration management.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to a deep learning address segmentation method combined with geographic information enhancement. Background Art

[0002] In the field of natural language processing (NLP), word segmentation is one of the basic tasks of text processing. Address segmentation technology plays an important role in many fields, such as geographic information processing, logistics distribution, and map navigation. However, traditional word segmentation methods have many limitations when processing diverse address information, such as poor adaptability to complex formats and low tolerance for errors and confusing text. In addition, manually entered address information often contains errors, confusion, and repetitions, further increasing the difficulty of word segmentation. Although existing deep learning word segmentation technology has improved, there is still room for improvement, especially in optimizing for the particularity and complexity of address text. Therefore, a more efficient, accurate, and adaptable address segmentation method is needed. Summary of the Invention

[0003] The purpose of the present invention is to provide a deep learning address segmentation method combined with geographic information enhancement, so as to solve the above-mentioned problems existing in the prior art.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] A geographic information enhanced deep learning address segmentation method, comprising the following steps:

[0006] Address text preprocessing: Receive the original address text, perform text cleaning to remove irrelevant characters, implement address standardization to ensure the uniformity of place names, and perform semantic disambiguation to clarify the meaning of the address;

[0007] Feature fusion extraction: Extracts basic text features, including word frequency, part of speech, and geographic entity features. Combined with a geographic information database, it extracts multi-dimensional spatial features of the address, including spatial coordinates, administrative division levels, and geographic entity relationships. Feature fusion is performed using a custom convolutional neural network (CNN) and a bidirectional long short-term memory (Bi-LSTM) network. The attention mechanism is used during feature fusion to highlight key address information.

[0008] Customized model training: Design a dedicated deep learning model architecture consisting of an input layer, multiple hidden layers, and an output layer. The hidden layers contain at least one multi-layer perceptron (MLP) module for feature interaction. Key features are extracted and industry feature vectors are constructed to meet the address segmentation needs of different industries. Model training is performed using professionally labeled address datasets, employing a multi-task learning strategy to simultaneously optimize segmentation accuracy and industry feature recognition accuracy. Model parameters are dynamically adjusted using an adaptive learning rate adjustment strategy and a gradient descent algorithm.

[0009] Word segmentation execution and optimization: The pre-processed address text is input into the trained deep learning model to implement word segmentation operations; the word segmentation results are optimized using customized post-processing rules, including rule-based pattern matching, statistics-based frequency adjustment, and semantic-based association correction; a dedicated word segmentation result intelligent verification system that combines the detailed topological information of national administrative divisions, geographic entity semantic association rules, and industry-specific verification rules is called to conduct a comprehensive verification and evaluation of the optimized word segmentation results, and the model parameters and optimization rules are dynamically adjusted according to the verification feedback to form a closed-loop optimization mechanism.

[0010] Preferably, in the address text preprocessing stage, text cleaning adopts a method based on rule matching combined with deep learning language model to accurately remove spaces, special characters, industry-specific filler words and ambiguous expression fragments, and automatically correct common text errors. The rule matching of text cleaning is based on regular expressions, and the deep learning language model is based on the Transformer architecture.

[0011] Preferably, the geographic information database is a multi-source heterogeneous database that integrates national administrative division standard data, high-precision geographic spatial data, and real-time updated geographic entity association data to ensure the accuracy and timeliness of geographic information. The data is updated at least once a quarter.

[0012] Preferably, the attention mechanism in the feature fusion process adopts a dynamic attention allocation strategy, which adaptively adjusts the interaction strength between text features and geographic information features according to the feature distribution of different address texts to improve the feature fusion effect. The dynamic attention allocation strategy calculates the attention weight based on the vector inner product of the input features.

[0013] Preferably, the multi-task learning strategy includes the main task of address segmentation and the auxiliary task of industry feature identification. By sharing the underlying feature extraction layer and the independent upper task-specific layer, the two tasks are jointly trained. The loss function is the weighted sum of the losses of the two tasks, and the weight coefficient is dynamically adjusted according to the task difficulty during the training process.

[0014] Preferably, the adaptive learning rate adjustment strategy adopts a dynamic adjustment algorithm based on training loss and validation loss. When the difference between training loss and validation loss exceeds a preset threshold for multiple consecutive iterations, the learning rate is reduced; when both training loss and validation loss decrease for multiple consecutive iterations, the learning rate is appropriately increased to ensure stable training and rapid convergence of the model.

[0015] Preferably, the dedicated word segmentation result intelligent verification system uses a graph neural network to construct a detailed topological map of the national administrative divisions, maps the word segmentation results to the topological map for path analysis and semantic association analysis, and combines industry-specific verification rules to accurately identify and correct word segmentation errors to ensure that the word segmentation results conform to actual geographical logic and industry specifications. The topological map nodes of the graph neural network include administrative division units and geographical entity nodes, and the edges represent the administrative affiliation and geographical adjacency between them.

[0016] Preferably, the method also includes a data update and maintenance step, regularly updating the geographic information database and the annotated address data set to ensure the timeliness of the data base. The data update frequency is configurable and supports both manual triggering and automatic timed update modes.

[0017] Preferably, the method is implemented using a distributed architecture, with the preprocessing module, feature fusion module, model training module, and word segmentation optimization module deployed on different server nodes respectively. Data transmission and task scheduling are performed through a high-speed network, supporting parallel processing and efficient calculation of large-scale address texts. The system processing capacity can be dynamically expanded according to the task load.

[0018] Preferably, the method provides a standardized API interface to support seamless integration with other geographic information systems, business management systems and applications. The API interface adopts a RESTful style, supports both JSON and XML data formats, and has detailed documentation and sample codes to facilitate rapid calling and integration by third-party developers.

[0019] The beneficial effects of the present invention are:

[0020] Improve word segmentation accuracy: By integrating geographic information features and deep learning technology, it can more accurately identify and segment address text and reduce word segmentation errors, especially when processing complex and ambiguous addresses.

[0021] Enhanced model generalization capabilities: The customized deep learning model can adapt to address texts of different formats and styles, as well as the specific needs of different industries, improving the model's generalization capabilities and scope of applicability.

[0022] Improved processing efficiency: The optimized model architecture and algorithm enable the present invention to quickly process large amounts of address data, making it suitable for real-time data processing scenarios and improving word segmentation efficiency.

[0023] Optimizing user experience: Accurate address segmentation provides more precise data support for applications such as map navigation and logistics distribution, significantly improving user experience.

[0024] Data update and maintenance: Regularly update the geographic information database and annotated address dataset to ensure that the system's data foundation is always kept up to date, ensuring the timeliness and accuracy of word segmentation results.

[0025] Distributed architecture support: The distributed architecture supports parallel processing and efficient computing of large-scale address texts. The system processing capacity can be dynamically expanded according to the task load to meet the application needs of different scales.

[0026] Standardized API interface: Provides a standardized API interface to facilitate seamless integration with other geographic information systems, business management systems and applications, reducing development costs and integration difficulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of the address segmentation method of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] Reference Figure 1 The geographic information enhanced deep learning address segmentation method shown includes the following steps:

[0030] Address text preprocessing: Receive the original address text, perform text cleaning to remove irrelevant characters, implement address standardization to ensure the uniformity of place names, and perform semantic disambiguation to clarify the meaning of the address.

[0031] This step is the starting point of the entire word segmentation method. Its purpose is to convert the original address text into a clean, standardized, and semantically clear form, laying a good foundation for subsequent feature extraction and analysis. First, the original address text is received. This text may come from various sources, such as manual user input or system transmission, and often contains many irrelevant characters and non-standard expressions.

[0032] Perform text cleaning to remove irrelevant characters. Irrelevant characters include spaces, special symbols (such as "@," "#," "$," etc.), industry-specific filler words (for example, non-address-related information such as "OBD" that may appear in some logistics addresses), and ambiguous expressions. By combining rule-based matching (using regular expressions and other technologies) with a deep learning language model (based on the Transformer architecture), these interfering contents can be accurately identified and removed. At the same time, common text errors such as homophone replacement and typos can be automatically corrected, making the text neater and more accurate.

[0033] Next, address standardization was implemented to ensure the uniformity of place names. Different place names were standardized into a standard form. For example, "Kunshi Road, Guandu District, Yunnan Province" and "Kunshi Road, Guandu District, Kunming City, Yunnan Province" were standardized into a standard address format that includes the complete administrative division hierarchy. This helped eliminate ambiguity caused by differences in expression and improved the consistency and accuracy of subsequent processing.

[0034] Finally, semantic disambiguation is performed to clarify the meaning of the address. Leveraging the powerful semantic understanding capabilities of knowledge graphs and deep learning language models, this process disambiguates potentially ambiguous portions of the address text. For example, in a name like "Shiba Community Residents Committee," it accurately distinguishes whether it represents a part of an administrative division or a specific institution, thereby clarifying the specific geographic location conveyed by the entire address. This ensures the semantic clarity and accuracy of the address text, providing a strong guarantee for the efficient execution of subsequent steps.

[0035] Feature fusion extraction: Extract basic features of the text, including word frequency, part of speech, and geographic entity features; combine with geographic information databases to extract multi-dimensional spatial features of addresses, including spatial coordinates, administrative division levels, and geographic entity associations; use a custom convolutional neural network (CNN) and bidirectional long short-term memory network (Bi-LSTM) for feature fusion, in which the attention mechanism is used to highlight key address information during the feature fusion process.

[0036] This step is as follows: After completing address text preprocessing, the feature fusion and extraction phase begins. The core task of this step is to extract rich and discriminative features from the processed text and effectively fuse the text features with geographic information features to enhance the comprehensive representation of address information and provide high-quality feature input for subsequent model training and word segmentation operations.

[0037] First, we extract basic text features, including word frequency, part of speech, and geographic entity features. Word frequency reflects the importance of a word in the text, while part of speech helps us understand the grammatical function of a word within a sentence. Geographic entity features identify key geographic elements in the text, such as place names and road names. These features provide essential descriptive information for the address text from different perspectives.

[0038] At the same time, the multi-dimensional spatial features of the address are extracted by combining it with a geographic information database. A geographic information database is a resource library that integrates various geographic-related data, including spatial coordinates (used to determine the precise geographic location of the address), administrative division levels (clearly defining the administrative level to which the address belongs, such as province, city, district, street, etc.), and geographic entity relationships (such as the affiliation of a residents' committee with its street, the adjacency relationship between adjacent communities, etc.). These spatial features give the address text geographic spatial semantic information.

[0039] The extracted text features and geographic information features are then fused using a custom convolutional neural network (CNN) and a bidirectional long short-term memory (Bi-LSTM) network. CNNs can capture local features and spatial structure in text through convolution operations, while Bi-LSTMs excel at processing sequential data and can model the contextual semantics of text. Combining the two networks and employing an attention mechanism during the fusion process enables the model to adaptively adjust the interaction strength between text and geographic information features based on the feature distribution of different address texts, thereby highlighting key address information and suppressing irrelevant interference. Ultimately, this generates fused features that are more discriminative for address segmentation tasks, providing strong support for accurate predictions by subsequent models.

[0040] Customized model training: Design a dedicated deep learning model architecture, which includes an input layer, multiple hidden layers, and an output layer. The hidden layers contain at least one multi-layer perceptron (MLP) module for feature interaction. Targeted at the address segmentation needs of different industries, key features are extracted and industry feature vectors are constructed. Model training is performed using professionally labeled address datasets, employing a multi-task learning strategy to simultaneously optimize segmentation accuracy and industry feature recognition accuracy. Model parameters are dynamically adjusted using an adaptive learning rate adjustment strategy and a gradient descent algorithm.

[0041] After feature fusion and extraction, the resulting fused features are used to train a dedicated deep learning model, which is a key step in achieving high-precision address segmentation. This step aims to design and train a deep learning model that can meet the specific needs of different industries and achieve excellent performance in address segmentation tasks.

[0042] A dedicated deep learning model architecture was designed, consisting of an input layer, multiple hidden layers, and an output layer. The input layer receives the fused features extracted in the previous step; the hidden layers contain at least one multi-layer perceptron (MLP) module for feature interaction. The MLP module, through its multi-layered neuron structure, performs complex nonlinear transformations and interactive operations on input features, fully exploring potential correlations between features and extracting higher-level abstract feature representations, enhancing the model's ability to understand the complex semantics of address text.

[0043] To meet the address segmentation needs of different industries, such as the logistics industry, which focuses on accurate segmentation of shipping and receiving addresses to ensure accurate delivery of goods, and public security household registration management, which focuses on address standardization and regularization, we extract key features and construct industry feature vectors. This enables the model to learn industry-specific address feature patterns, improving its adaptability and segmentation accuracy for address texts from different industries.

[0044] A professionally annotated address dataset is used for model training. This dataset has been rigorously screened and annotated to ensure that it covers address samples from different regions, formats, and styles. The annotation quality has passed multiple rounds of manual review and automated consistency checks, and the annotation accuracy is not less than 98%. During the training process, a multi-task learning strategy is adopted to combine the main task of address segmentation (accurately dividing the various components of the address) with the auxiliary task of industry feature recognition (identifying the industry characteristics of the address text). The joint training of the two tasks is achieved through a network structure that shares the underlying feature extraction layer and the independent upper-level task-specific layer. The loss function is designed to be the weighted sum of the losses of the two tasks, and the weight coefficient is dynamically adjusted according to the difficulty of the task during the training process, so that the model can effectively learn relevant knowledge of industry feature recognition while optimizing the accuracy of word segmentation, further improving the overall performance of the model.

[0045] To improve model training efficiency and optimization results, an adaptive learning rate adjustment strategy and gradient descent algorithm are used to dynamically adjust model parameters. This adaptive learning rate adjustment strategy automatically adjusts the learning rate in real time based on changes in loss during training. When the difference between the training loss and validation loss exceeds a preset threshold for multiple consecutive iterations, the learning rate is reduced to prevent overfitting and other issues during optimization. When both the training loss and validation loss decrease over multiple consecutive iterations, the learning rate is appropriately increased to accelerate model convergence, ensuring that the model can stably and quickly learn effective feature representations and word segmentation rules, ultimately achieving optimal word segmentation results.

[0046] Word segmentation execution and optimization: The pre-processed address text is input into the trained deep learning model to implement word segmentation operations; the word segmentation results are optimized using customized post-processing rules, including rule-based pattern matching, statistics-based frequency adjustment, and semantic-based association correction; a dedicated word segmentation result intelligent verification system that combines the detailed topological information of national administrative divisions, geographic entity semantic association rules, and industry-specific verification rules is called to conduct a comprehensive verification and evaluation of the optimized word segmentation results, and the model parameters and optimization rules are dynamically adjusted according to the verification feedback to form a closed-loop optimization mechanism.

[0047] After completing the customized model training, the trained deep learning model will be used for actual address segmentation operations, and a series of optimization methods will be used to further improve the accuracy and reliability of the segmentation results, ensuring that the final output segmentation results can meet the high requirements of address information accuracy in different application scenarios.

[0048] First, the preprocessed address text is fed into the trained deep learning model. Based on the feature representations and word segmentation rules learned during the training phase, the model performs word segmentation on the input address text, generating preliminary word segmentation results. This process automatically parses and segments the address text, quickly generating basic word segmentation output.

[0049] Then, the preliminary word segmentation results are optimized using customized post-processing rules. These post-processing rules include rule-based pattern matching, statistical-based frequency adjustment, and semantic-based association correction. Rule-based pattern matching can match and correct some common address formats and expressions, such as correctly combining words such as "road", "street", and "avenue" with the following numbers or place names; statistical-based frequency adjustment is to optimize and adjust some potentially ambiguous or erroneous word segmentation results in a probabilistic statistical sense based on word frequency statistics, and is more inclined to choose word segmentation methods with high frequency and more in line with conventional expression habits; semantic-based association correction uses semantic analysis technology to ensure that the various parts after word segmentation are semantically related to each other and logically smooth, such as avoiding the mistaken division of a complete place name into multiple meaningless parts, thereby further improving the accuracy and rationality of the word segmentation results.

[0050] Then, a dedicated intelligent verification system for word segmentation results is called upon, combining detailed topological information of national administrative divisions, semantic association rules for geographic entities, and industry-specific verification rules. This system uses a graph neural network to construct a detailed topological map of national administrative divisions, mapping the word segmentation results onto the topological map for path analysis and semantic association analysis. By comprehensively comparing and verifying the word segmentation results with the administrative division topological structure, semantic relationships between geographic entities, and industry-specific rules, it can accurately identify and correct various errors that may occur during the word segmentation process, such as errors in administrative division hierarchical relationships, errors in geographic entity associations, and industry-specific formatting errors, ensuring that the word segmentation results strictly conform to actual geographic logic, semantic rules, and industry specifications.

[0051] Finally, based on feedback from the verification system, a closed-loop optimization mechanism is formed. Problems discovered during the verification process and the corresponding corrections are fed back to the model and optimization rules, dynamically adjusting model parameters and optimization rules. In this way, the model and optimization rules can continuously learn and adapt to new address patterns and features, continuously improving segmentation performance, and enhancing the accuracy and adaptability of segmentation results. This ensures that the entire address segmentation method remains efficient and accurate in practical applications, providing solid and reliable address data support for subsequent geographic information processing, map navigation, logistics and distribution, and other businesses.

[0052] Preferably, in the address text preprocessing stage, text cleaning adopts a method based on rule matching combined with deep learning language model to accurately remove spaces, special characters, industry-specific filler words and ambiguous expression fragments, and automatically correct common text errors. The rule matching of text cleaning is based on regular expressions, and the deep learning language model is based on the Transformer architecture.

[0053] Text cleaning is a key step in address text preprocessing. It aims to remove all kinds of irrelevant characters and erroneous content in the original address text, correct common text errors, and thus improve the accuracy and efficiency of subsequent processing.

[0054] Text cleaning based on rule matching:

[0055] Rule matching principle: Utilizing technologies such as regular expressions, a series of rule patterns are defined to identify and match irrelevant characters and erroneous content in specific formats within address text. For example, regular expressions can be used to match consecutive spaces, common special symbols (such as "@," "#," "$," "%," "^," "&," "*," etc.), industry-specific filler words (such as specific identifiers before and after logistics order numbers), and ambiguous expressions (such as ambiguous abbreviations or colloquial expressions).

[0056] Removing Irrelevant Characters: Once these irrelevant characters and content are matched, they will be removed from the text. For example, clean "Kunming Stone Road, Guandu District, Yunnan Province @ Phase II of Fontainebleau ## Building 12, Unit 1, Room 1802" to "Kunming Stone Road, Guandu District, Yunnan Province, Phase II of Fontainebleau, Building 12, Unit 1, Room 1802".

[0057] Correcting Common Text Errors: Based on rule matching, some common text errors can also be corrected, such as homophone substitution errors (e.g., confusion between "road" and "land") and fixed format errors (e.g., incorrect telephone number or postal code format). By defining corresponding rules, identify these error patterns and replace them with the correct forms.

[0058] Text Cleaning Based on Deep Learning Language Models:

[0059] Deep Learning Language Model Architecture: Adopt a deep learning language model based on the Transformer architecture. The Transformer architecture performs excellently in natural language processing tasks due to its powerful parallel computing ability and effective capture of long-term dependencies in sequential data.

[0060] Pre-training and Fine-tuning: First, pre-train the language model on a large-scale text data to enable it to learn general text features and semantic information. Then, fine-tune the pre-trained model on the address text dataset to make it better adapt to the characteristics and cleaning requirements of address text.

[0061] Text Error Correction: Input the original address text into the trained language model. The model can automatically identify and correct some complex text errors based on the semantic knowledge and text patterns it has learned. For example, for some misspelled, semantically incoherent or non-standard address expressions, the language model can generate more reasonable and standardized text forms.

[0062] Removing Vague Expression Fragments: The language model can also identify vague or unreasonable expression fragments in the text and replace them with clearer descriptions or directly remove them, thereby improving the readability and accuracy of address text.

[0063] Combination and Advantages of the Two Methods:

[0064] Complementary: The text cleaning method based on rule matching has the characteristics of clear rules and high execution efficiency, and can quickly remove a large number of irrelevant characters and common errors. The method based on deep learning language models has stronger semantic understanding and text generation capabilities, can handle complex text errors and vague expressions, and makes up for the deficiencies of the rule matching method in dealing with complex semantic scenarios.

[0065] Accurately remove irrelevant characters and correct errors: By combining these two methods, spaces, special characters, industry-specific filler words, and ambiguous expressions in address text can be more accurately removed, and common text errors can be automatically corrected, thereby improving the quality and standardization of address text, laying a solid foundation for subsequent address standardization, feature extraction, and other steps.

[0066] Preferably, the geographic information database is a multi-source heterogeneous database that integrates national administrative division standard data, high-precision geographic spatial data, and real-time updated geographic entity association data to ensure the accuracy and timeliness of geographic information. The data is updated at least once a quarter.

[0067] This geographic information database adopts a multi-source heterogeneous architecture, aiming to integrate and consolidate geographic-related data of various types and sources to build a comprehensive, rich and accurate geographic information resource center, providing solid geographic data support for address segmentation methods.

[0068] Data source and type:

[0069] National administrative division standard data: This is a core component of the database, derived from administrative division information released by national authorities. It covers detailed division data at multiple administrative levels, including national, provincial, municipal, district / county, township / town / street, and village / community levels, including administrative division codes, names, and boundary information. This data ensures accurate identification and standardized division of administrative divisions during address segmentation, enabling segmentation results to be output strictly according to the standard administrative division hierarchy. For example, it can accurately distinguish between administrative division information at different levels, such as "Yunnan Province," "Kunming City," and "Guandu District."

[0070] High-precision geospatial data: This primarily comes from professional geographic surveying and mapping organizations, satellite remote sensing data processing vendors, and other sources. It includes detailed geographic coordinate information (such as longitude and latitude), road network data, topographic data, and building distribution data. In address segmentation, geographic coordinate information can be used to precisely determine the geographic location of an address. Road network data helps identify and parse information such as road names and house numbers within an address. Topographic and building data can aid in understanding and refining address descriptions with unique geographic features, such as a village at the foot of a mountain or a specific unit number in a high-rise building.

[0071] Real-time updated geographic entity-related data: This data is updated and collected in real time or near real time through various channels, including geographic information service platforms, government business systems, and field research. It covers the relationships between geographic entities, such as the affiliation of a residents' committee with its affiliated street, the adjacency of adjacent communities, and the connectivity between a shopping mall and surrounding roads. This linked data enables the database to promptly reflect the latest changes in the relationships between geographic entities, providing dynamic geographic semantic support for address segmentation. This helps resolve issues such as ambiguous address representations or delayed updates caused by changes in geographic entity relationships, such as the relationship between newly established communities or newly constructed roads and surrounding geographic entities.

[0072] Data integration and fusion:

[0073] To effectively integrate this multi-source, heterogeneous data, the database employs data integration and fusion technologies. By establishing unified data models and data standards, data from different sources and types are converted, cleaned, matched, and integrated. For example, administrative division boundary information from national administrative division standard data is spatially matched and fused with geographic coordinates and road networks from high-precision geospatial data to achieve a precise correspondence between administrative division information and geospatial locations. Simultaneously, real-time geographic entity association data is linked and updated with underlying administrative division and geospatial data to ensure the logical consistency and relevance of geographic information across the database. This data integration and fusion approach enables the database to provide comprehensive, accurate, and consistent geographic information support for address segmentation methods, improving the geographic accuracy and semantic relevance of segmentation results.

[0074] Data update mechanism:

[0075] To ensure the accuracy and timeliness of geographic information, the database has established a strict data update mechanism. The frequency of data update is no less than once a quarter, that is, all types of data in the database are fully or partially updated and maintained at least every three months.

[0076] Updates to national administrative division standards: Closely monitor administrative division adjustment information released by national authorities. Once new administrative division standards are released, timely update the corresponding parts of the database to ensure that administrative division codes, names, boundaries, and other information are consistent with the latest standards. For example, when a region undergoes township mergers or street rezoning, timely update the administrative division hierarchy and related boundary information in the database to ensure that address segmentation can accurately identify and classify the newly adjusted administrative divisions.

[0077] High-precision geospatial data updates: We establish long-term, stable partnerships with geographic surveying and mapping organizations and data providers to regularly obtain the latest geospatial data updates. As cities develop and the geographical environment changes, road construction, building construction, and demolition are constantly occurring. Timely updates of geographic coordinates, road networks, and topographic data ensure that the geospatial information in the database always reflects the latest real-world conditions. For example, a newly constructed urban arterial road's route, name, and associated geographic coordinates are promptly updated in the database to accurately identify the road name and corresponding location during address segmentation.

[0078] Real-time updating of geographic entity-related data: By establishing data sharing and updating mechanisms with government business systems (such as the community management system of the civil affairs department and the land use system of the planning department), geographic information service platforms (such as online map service providers), and field research and collection teams, real-time or near-real-time updating of geographic entity-related data is achieved. For example, when a new resident group is established in a community or a major adjustment is made to the commercial layout of a shopping mall, this updated information can be promptly fed back to the database, updating the associations between geographic entities. This ensures that the geographic entity-related data in the database is always up to date, providing dynamic and accurate geographic semantic support for address segmentation.

[0079] Data quality assurance:

[0080] During the data integration and update process, a strict data quality assurance system has been established. Using a combination of data quality assessment tools and manual review, we conduct quality checks and verification on all types of data entered into the database. Data that does not meet data standards and quality requirements is cleaned, corrected, or re-collected to ensure the accuracy, completeness, and consistency of the geographic information in the database. For example, administrative division data is checked for boundary integrity and name accuracy, geospatial data is verified for geometric accuracy and attribute information integrity, and geographic entity-related data is checked for logical consistency and field verification. This ensures that the geographic information database can provide reliable, high-quality geographic data resources for address segmentation methods, improving the performance and stability of the entire address segmentation system.

[0081] Preferably, the attention mechanism in the feature fusion process adopts a dynamic attention allocation strategy, which adaptively adjusts the interaction strength between text features and geographic information features according to the feature distribution of different address texts to improve the feature fusion effect. The dynamic attention allocation strategy calculates the attention weight based on the vector inner product of the input features.

[0082] The dynamic attention allocation strategy is the core mechanism in the feature fusion process. It aims to adaptively adjust the interaction strength between text features and geographic information features based on the feature distribution of different address texts. By calculating attention weights based on the inner product of the input feature vectors, the model automatically focuses on the features most critical to the current address text segmentation task, thereby improving the effectiveness of feature fusion and segmentation accuracy.

[0083] Text features and geographic information features come from different aspects of the address text. Text features mainly include word frequency, part of speech, word position information, and semantic features extracted by deep learning language models. These features can reflect the grammatical structure and semantic meaning of the address text. For example, the part of speech and semantic information of words such as "road", "community", and "building" help determine their roles in the address. Geographic information features include multi-dimensional data such as spatial coordinates, administrative division levels, and geographic entity associations. Spatial coordinates accurately identify the geographical location of the address, administrative division levels clearly define the administrative level to which the address belongs, and geographic entity associations reveal logical connections such as affiliation and adjacency between different geographic entities.

[0084] Preferably, the multi-task learning strategy includes the main task of address segmentation and the auxiliary task of industry feature identification. By sharing the underlying feature extraction layer and the independent upper task-specific layer, the two tasks are jointly trained. The loss function is the weighted sum of the losses of the two tasks, and the weight coefficient is dynamically adjusted according to the task difficulty during the training process.

[0085] The multi-task learning strategy adopted in this invention cleverly combines the main task of address segmentation with the auxiliary task of industry feature identification, innovatively shares the underlying feature extraction layer, and establishes an independent upper-level task-specific layer for each task, realizing joint training of the two tasks and effectively improving the overall performance of the model.

[0086] The specific content and relationship between the main task and auxiliary tasks

[0087] The main task focuses on address segmentation, that is, accurately dividing the complete address text, such as "Unit 1, Building 12, 1802, Fontainebleau Phase II, Kunshi Road, Guandu District, Yunnan Province", into multiple constituent elements: "Yunnan Province", "Guandu District", "Kunshi Road", "Fontainebleau Phase II", "Building 12", "Unit 1", "Room 1802".

[0088] The auxiliary task is industry feature recognition, aiming to identify the industry sector to which the address text belongs, such as logistics, public security household registration, or insurance. Address texts in different industries vary in format, keywords, and other aspects. For example, addresses in the logistics industry often include information such as express delivery service points, while public security household registration addresses focus more on the completeness and standardization of hierarchical information such as administrative divisions and communities.

[0089] The two tasks are closely related. Industry feature recognition results assist the address segmentation task, optimizing the segmentation strategy based on industry characteristics. The feedback from address segmentation results also helps to more accurately identify industry features, forming a two-way promotion mechanism.

[0090] The role and advantages of sharing the underlying feature extraction layer

[0091] The shared underlying feature extraction layer is a core component of the multi-task learning strategy. It is responsible for extracting common basic features from the raw address text, including word frequency, part of speech, geographic entity features, and comprehensive features derived from the fusion of text and geographic information. By sharing parameters, both tasks benefit from the same underlying features, reducing the total number of model parameters, mitigating the risk of overfitting, and improving training efficiency. For example, part-of-speech features: nouns typically represent place names or road names, while numbers may represent house numbers or unit numbers. These features are critical for both tasks.

[0092] Establishing independent, upper-level task-specific layers precisely addresses the personalized needs of each task. The primary task-specific layer focuses on optimizing the accuracy of address segmentation, applying techniques like sequence labeling to assign the correct label to each address segment. The auxiliary task-specific layer, on the other hand, focuses on identifying industry characteristics, using classification algorithms to determine the industry to which the address text belongs. This architecture maintains the efficiency of common features while also meeting personalized needs.

[0093] Loss function design and dynamic adjustment mechanism

[0094] The design of the loss function is the key to achieving multi-task learning. The loss function is set as the weighted sum of the losses of the two tasks, and the formula is as follows:

[0095] Ltotal=αLmain+(1-α)Laux

[0096] Among them, Lmain is the loss of the main task of address segmentation, which uses the loss function of the sequence labeling task, such as the negative log-likelihood loss of the conditional random field (CRF) layer, to measure the difference between the predicted label sequence and the true label sequence; Laux is the loss of the auxiliary task of industry feature recognition, usually the cross-entropy loss, which measures the difference between the predicted industry category and the true category; the weight coefficient α is used to balance the losses of the two tasks. In the early stages of training, the model is unfamiliar with both tasks, so α is set to 0.7, focusing more on the main task to ensure that the model first masters the core address segmentation capabilities.

[0097] As training progresses, α is dynamically adjusted based on the changes in the losses of the two tasks. For example, if the loss of the primary task decreases rapidly while the loss of the auxiliary task is high, this indicates that the model is making rapid progress in address segmentation but is weak in identifying industry characteristics. In this case, α is reduced to around 0.5, increasing the weight of the auxiliary task to encourage the model to better learn industry characteristics.

[0098] Through this dynamic adjustment mechanism, the model can automatically balance the attention of the two tasks during joint training, give full play to the advantages of multi-task learning, and achieve dual optimization of address segmentation and industry feature recognition.

[0099] Preferably, the adaptive learning rate adjustment strategy adopts a dynamic adjustment algorithm based on training loss and validation loss. When the difference between training loss and validation loss exceeds a preset threshold for multiple consecutive iterations, the learning rate is reduced; when both training loss and validation loss decrease for multiple consecutive iterations, the learning rate is appropriately increased to ensure stable training and rapid convergence of the model.

[0100] In this step, after each iteration of model training, training loss and validation loss are calculated. Training loss reflects how well the model fits the training set, while validation loss measures the model's generalization ability on the validation set (a dataset not used in training). By comparing these two loss values, we can understand the training status of the model. Loss calculations are typically performed using methods such as mean squared error or cross-entropy loss, but the specific details are complex.

[0101] Dynamically adjust learning rate

[0102] Conditions and operations for reducing the learning rate

[0103] If the difference between the training loss and validation loss exceeds a preset threshold (such as 0.1) for multiple consecutive iterations (for example, three iterations), this indicates that there is a significant difference in the model's performance on the training and validation sets, possibly indicating overfitting. In this case, reducing the learning rate helps the model learn the patterns of the data more carefully, enhancing its generalization ability and avoiding overfitting to the training data that can lead to poor performance on the validation set. Reducing the learning rate multiplies the current learning rate by a coefficient less than 1 (such as 0.5), thereby slowing down parameter updates.

[0104] Conditions and operations for increasing learning rate

[0105] When both the training loss and validation loss show a downward trend for multiple iterations (for example, two), this indicates that the model is improving and the current learning rate may not be sufficient for the model to converge quickly to the optimal state. In this case, appropriately increasing the learning rate can help the model speed up parameter updates, thereby accelerating the convergence process. Increasing the learning rate is to multiply the current learning rate by a factor greater than 1 (such as 1.5).

[0106] Ensure stable model training and fast convergence

[0107] Selection of preset parameters

[0108] The preset threshold is used to determine whether the difference between the training loss and validation loss is excessive. Its value is typically determined through experience or experimentation. A smaller threshold is suitable for scenarios requiring high generalization, while a larger threshold is suitable for scenarios with high data noise. The learning rate adjustment coefficient determines the magnitude of the learning rate adjustment. The balance between the adjustment magnitude and training stability should be determined based on the model's training status. A smaller decay coefficient and a larger growth coefficient are suitable for initial model training, while a larger decay coefficient and a smaller growth coefficient are suitable for fine-tuning the model.

[0109] Prevent frequent changes in learning rate

[0110] To avoid frequent changes in the learning rate between successive rounds, a cool-down period (e.g., 2 rounds) can be set. After adjusting the learning rate, you need to wait for at least the number of rounds in the cool-down period before making the next adjustment. For example, if you lower the learning rate in round 10, you are not allowed to adjust it again in rounds 11 and 12, and the adjustment conditions will not be re-evaluated until round 13.

[0111] This adaptive learning rate adjustment strategy allows the model to dynamically adjust its learning rate during training based on changes in training and validation losses, ensuring stable training and rapid convergence. Initially, the learning rate is relatively high, allowing the model to quickly fit the data and rapidly reduce the loss. When signs of overfitting appear in the middle, the learning rate is reduced, enhancing the model's generalization ability and stabilizing the validation loss. Later in the fine-tuning phase, the learning rate is moderately increased, allowing the model to converge quickly around the optimal parameters. Ultimately, while maintaining high training efficiency, the model achieves a significant improvement in word segmentation accuracy on the validation set.

[0112] Preferably, the dedicated word segmentation result intelligent verification system uses a graph neural network to construct a detailed topological map of the national administrative divisions, maps the word segmentation results to the topological map for path analysis and semantic association analysis, and combines industry-specific verification rules to accurately identify and correct word segmentation errors to ensure that the word segmentation results conform to actual geographical logic and industry specifications. The topological map nodes of the graph neural network include administrative division units and geographical entity nodes, and the edges represent the administrative affiliation and geographical adjacency between them.

[0113] Graph neural network topology construction

[0114] The intelligent verification system for specialized word segmentation results innovatively uses graph neural networks (GNNs) to construct a detailed topological map of national administrative divisions. GNNs are a deep learning architecture specifically designed to process graph-structured data, effectively capturing complex relationships between nodes and enabling information transfer.

[0115] Topological graph structure and node edge definition

[0116] Node Type: Topological map nodes include administrative unit nodes and geographic entity nodes. Administrative unit nodes cover administrative divisions at the national, provincial, municipal, district / county, township / town / street, and village / community levels. Each node contains attributes such as a unique identification code, name, and geographic location. Geographic entity nodes represent geographical features such as mountains, rivers, roads, and buildings. Attributes include entity type, coordinates, and name.

[0117] Edge Definition: Edges represent administrative affiliation and geographic adjacency between nodes. Administrative affiliation edges point from higher-level administrative units to lower-level units, for example, from "Yunnan Province" to "Kunming City." Geographic adjacency edges connect geographically adjacent administrative units or entities, such as the adjacency edge between "Guandu District" and "Chenggong District." The weight reflects the degree of adjacency or distance.

[0118] Word segmentation result mapping and path analysis

[0119] Mapping process: Map each address element in the word segmentation result to the corresponding node in the topology map. For example, if the word segmentation result is "Yunnan Province / Kunming City / Guandu District / Kunshi Road", the Yunnan Province node, Kunming City node, Guandu District node, and Kunshi Road node are found in sequence.

[0120] Path analysis: Checks whether these nodes on the topology map form a valid path of administrative affiliation and geographic adjacency. A valid path requires the correct administrative hierarchy and geographic adjacency between adjacent nodes. For example, the analysis checks whether the administrative affiliation path from provincial nodes to municipal nodes and then to district nodes is coherent, and whether the adjacency between geographic entity nodes and administrative division nodes is reasonable.

[0121] Semantic association analysis

[0122] Semantic association analysis technology: Combine natural language processing technology to conduct semantic association analysis on the word segmentation results. On the one hand, check whether the semantics between address elements conform to conventional expression habits. For example, "Kunshi Road" as a road name is usually paired with administrative divisions such as "Guandu District". If the word segmentation results show obvious semantic mismatches, such as unreasonable combinations such as "Kunshi Road" and "Xishan District", there may be problems. On the other hand, analyze the semantic association between address elements and industry characteristics. For example, in the logistics industry, geographical entity nodes such as "logistics park" and "freight station" appear in the address, and the word segmentation results must conform to the semantic logic of the logistics address.

[0123] Industry-specific validation rule integration

[0124] The system integrates industry-specific verification rules into a rule library. For example, public security household registration management requires that address segmentation strictly adhere to the standard administrative hierarchy, without skipping or staggering levels. The logistics industry prioritizes address accessibility and relevance to logistics outlets. During the verification process, the corresponding verification rules are applied based on the industry to which the segmentation results belong.

[0125] Error identification and correction mechanism

[0126] Error pattern recognition: Based on path analysis, semantic association analysis, and industry verification rules, it identifies common error patterns in word segmentation results. These include administrative affiliation errors (e.g., cross-level affiliation), geographic adjacency errors (e.g., address elements are not spatially adjacent), and semantic mismatches (e.g., roads are associated with areas they do not belong to).

[0127] Correction strategy: Take appropriate corrective measures based on the error type and topology information. For administrative affiliation errors, adjust the word segmentation results to the correct administrative division level based on the administrative hierarchy of the topology. For geographic adjacency errors, refer to the geographic adjacency edges to correct the order of address elements or replace the incorrect geographic entity. For semantic mismatches, re-match appropriate address elements based on semantic association rules and industry characteristics.

[0128] Verification process and feedback mechanism

[0129] Verification process: After the word segmentation results are generated, they are first mapped to a topology map for path analysis and semantic association analysis. Then, industry-specific verification rules are called for review. Finally, all analysis results are combined to determine whether the word segmentation is correct.

[0130] Feedback mechanism: Verification results are fed back to the word segmentation model and optimization rules. If errors are found, the error type and location are recorded and used to adjust model parameters and optimization rules, forming a closed-loop optimization and continuously improving word segmentation quality.

[0131] The topological map constructed by the graph neural network can accurately identify and correct word segmentation errors, ensuring that the word segmentation results conform to actual geographical logic and industry standards.

[0132] Preferably, the method also includes a data update and maintenance step, regularly updating the geographic information database and the annotated address data set to ensure the timeliness of the data base. The data update frequency is configurable and supports both manual triggering and automatic timed update modes.

[0133] Update content and source

[0134] Geographic information database update:

[0135] Administrative division adjustment information: Regularly obtain the latest administrative division adjustment data from authoritative agencies such as the national civil affairs department, including information on newly created, abolished, and merged administrative division units, as well as boundary changes. For example, when a township is abolished and merged into an adjacent township, the administrative division hierarchy and corresponding boundary data in the database are updated in a timely manner.

[0136] Geographic entity change data: This data is collected from geographic surveying and mapping agencies, urban planning departments, and other organizations regarding the addition, disappearance, and attribute changes of geographic entities. This includes information such as newly built roads and bridges, abandoned buildings, and changes in geographic entity names. This data is used to promptly update the attributes and connections of geographic entity nodes.

[0137] Update of the label address dataset:

[0138] Collect new address samples: Collect new address text data from actual business scenarios, such as the latest delivery addresses on logistics platforms and address marks added by users in map applications, to ensure that the dataset covers the latest address formats and expressions.

[0139] Manual labeling and review: Organize a professional labeling team to label newly collected address samples, clarify the boundaries and categories of each address component, and ensure labeling quality through multiple rounds of review.

[0140] Update frequency configuration mechanism

[0141] Set a default update frequency: Based on the general pattern of address information changes, set a default update frequency. Typically, geographic information databases are updated quarterly, and annotated address datasets are updated monthly. This is because large-scale changes to administrative divisions and geographic entities are relatively rare, while the format and presentation of address text can evolve rapidly with business development and user habits.

[0142] Flexible adjustment of update frequency: Users can flexibly adjust the update frequency based on actual needs for different regions and business areas. For example, in areas with rapid urban development and frequent changes in geographic information, the update frequency of the geographic information database can be increased to once a month. For industries with rapid changes in address formats during peak business periods, the update frequency of the annotated address dataset can be temporarily increased.

[0143] Manually triggered update mode

[0144] Operation interface and permission management: Provide data updaters with an intuitive and easy-to-use update operation interface, while also setting up a strict permission management system. Only authorized administrators can perform manual update operations on the interface, ensuring the security and controllability of data updates.

[0145] Update Process and Monitoring: When an update is manually triggered, the system performs data collection, cleaning, conversion, and loading according to the pre-set update process. Throughout the update process, the system monitors the update progress and data quality in real time. If anomalies (such as incorrect data format or update failure) are detected, prompt alerts are issued and detailed error information is provided to facilitate investigation and resolution by the update personnel.

[0146] Automatic scheduled update mode

[0147] Task Scheduling and Execution: Use task scheduling tools (such as the Linux cron daemon or Windows Task Scheduler) to automatically initiate data update tasks at set intervals. The system pre-configures the update task script and parameters, and the task scheduling tool triggers the script execution at the specified time, automatically completing data collection, processing, and updating operations.

[0148] Update status logging and alerts: Automatically records the execution status and results of each scheduled update task, including update time, updated data volume, success or failure indicators, and other information. If a scheduled update task fails to execute as planned or an error occurs during execution, the system automatically sends an alert notification to the administrator, ensuring timely identification and appropriate action.

[0149] Update effect evaluation and feedback

[0150] Evaluate updated data quality: After each update, perform a comprehensive check on the updated geographic information database and annotated address dataset using a data quality assessment tool. Evaluation criteria include data accuracy, completeness, and consistency. For example, check whether administrative boundaries are closed, whether geographic entity relationships are reasonable, and whether the various labels for annotated addresses are accurate.

[0151] Feedback to model training and optimization: The updated data quality assessment results are fed back to the model training phase to provide a basis for model retraining and optimization. If any quality issues are found in the updated data, the data preprocessing and cleaning strategies are adjusted promptly to regenerate high-quality training data to ensure the model can achieve optimal performance in subsequent word segmentation tasks.

[0152] Through the above detailed data update and maintenance steps, it is possible to ensure that the geographic information database and the annotated address dataset always maintain timeliness and accuracy, providing a solid data foundation for the address segmentation method, thereby improving the reliability and adaptability of the segmentation results.

[0153] Preferably, the method is implemented using a distributed architecture, with the preprocessing module, feature fusion module, model training module, and word segmentation optimization module deployed on different server nodes respectively. Data transmission and task scheduling are performed through a high-speed network, supporting parallel processing and efficient calculation of large-scale address texts. The system processing capacity can be dynamically expanded according to the task load.

[0154] Module division and deployment strategy

[0155] The preprocessing module is deployed on server nodes equipped with high-performance CPUs and large memory capacities. Because it handles large-scale address text cleaning, standardization, and semantic disambiguation, it places high demands on computing resources and memory read / write speeds. For example, when processing data containing millions of logistics addresses, this module can quickly remove irrelevant characters and standardize place names.

[0156] The feature fusion module, which requires simultaneous extraction and fusion of text and geographic features, involves extensive matrix operations and deep learning model inference. Therefore, it is deployed on GPU-accelerated server nodes. The parallel computing capabilities of GPUs significantly accelerate the feature extraction process of convolutional neural networks (CNNs) and bidirectional long short-term memory networks (Bi-LSTMs), improving feature fusion efficiency.

[0157] Model training module: This module is also deployed on GPU server nodes to meet the high computational demands of deep learning model training. When training models containing tens of millions of address samples, GPU acceleration can reduce training time from weeks to days. GPU acceleration also leverages high-speed networks to communicate with other modules, enabling timely access to pre-processed data and updates to training parameters.

[0158] Word Segmentation Optimization Module: This module performs multiple rounds of optimization and verification on word segmentation results, involving frequent read and write operations and logical judgment. It is deployed on server nodes with high-speed storage I / O performance. It interacts with a dedicated word segmentation result intelligent verification system via a high-speed network to quickly obtain verification feedback and adjust the word segmentation results.

[0159] High-speed network communication mechanism

[0160] Each server node is connected via a high-speed Ethernet network (such as 10Gbps or 25Gbps Ethernet) to ensure low latency and high throughput for data transmission. A distributed message queue (such as Apache Kafka or RabbitMQ) is used for task scheduling and data transmission. After the preprocessing module completes processing of a batch of address texts, it sends the data to the feature fusion module via the message queue. The message queue is responsible for buffering and orderly delivering data to ensure the reliability of data transmission, even in the event of temporary network failures or inconsistent module processing speeds, it can avoid data loss and backlogs. At the same time, remote direct memory access (RDMA) technology is used to allow different server nodes to directly access data in each other's memory, further reducing data transmission latency and improving overall system efficiency.

[0161] Parallel processing and load balancing

[0162] Data parallelism: During the preprocessing phase, large-scale address text data is sliced ​​into fixed-size or fixed-byte slices and assigned to multiple preprocessing module instances for parallel processing. For example, data containing one million address texts can be split into 10 batches of 100,000 addresses each, and assigned to 10 preprocessing module instances for simultaneous text cleaning, standardization, and semantic disambiguation, significantly reducing preprocessing time.

[0163] Model parallelism strategy: During the feature fusion and model training phases, different layers or computing tasks of the deep learning model are distributed to multiple GPU server nodes for parallel computing. For example, the convolutional and pooling layers of a CNN can be placed on one GPU node, while the forward and backward layers of a Bi-LSTM can be placed on another GPU node. This allows for synchronized gradient and parameter updates over a high-speed network, accelerating model training.

[0164] Load balancing: Deploy a distributed load balancer (such as Nginx or HAProxy) to monitor resource metrics such as CPU usage, memory usage, and network bandwidth across each server node. Dynamically allocate tasks based on task queue length and node resource status. When a node becomes overloaded, some of its tasks are transferred to other idle nodes, ensuring balanced operation across the system and avoiding performance bottlenecks.

[0165] Dynamic expansion mechanism

[0166] Horizontal scalability: When the task load increases, the system can be scaled horizontally by adding new server nodes. For example, during peak business periods such as e-commerce shopping festivals, the address segmentation task volume increases significantly. New virtual or physical machine nodes can be quickly launched on the cloud platform and added to clusters such as the preprocessing module and feature fusion module. These new nodes automatically join the task scheduling system by obtaining updated configuration files and model parameters, helping to share the load.

[0167] Vertical expansion: For modules with extremely high computing resource requirements (such as model training modules), when a single node's memory and CPU resources are insufficient, vertical expansion can be achieved by upgrading the server hardware configuration. For example, upgrading a GPU server from one GPU card to two, or increasing memory capacity, can improve the processing power of a single node.

[0168] Automatic scaling strategies: Combine business traffic forecasts with real-time load monitoring to set automatic scaling strategies. For example, if the task queue length exceeds a preset threshold for 10 consecutive minutes and the average CPU utilization of a server node exceeds 80%, the horizontal scaling process is automatically triggered to add new nodes. After peak business hours, idle nodes are automatically reclaimed based on load conditions, reducing resource costs.

[0169] Implemented through a distributed architecture, the address segmentation method can efficiently process large-scale address texts, has strong scalability and flexibility, and can meet the address segmentation needs in different business scenarios.

[0170] Preferably, the method provides a standardized API interface to support seamless integration with other geographic information systems, business management systems and applications. The API interface adopts a RESTful style, supports both JSON and XML data formats, and has detailed documentation and sample codes to facilitate rapid calling and integration by third-party developers.

[0171] In summary, the standardized API interface provided by this address segmentation method achieves seamless integration with other geographic information systems, business management systems and applications by adopting a RESTful style design, supporting multiple data formats, detailed documentation and sample codes, as well as complete security and authentication mechanisms, error handling and monitoring mechanisms, providing third-party developers with convenient, efficient and secure interface services, and promoting the widespread application of address segmentation technology in different fields and the interconnection between systems.

[0172] The beneficial effects of the present invention are:

[0173] Improve word segmentation accuracy: By integrating geographic information features and deep learning technology, it can more accurately identify and segment address text and reduce word segmentation errors, especially when processing complex and ambiguous addresses.

[0174] Enhanced model generalization capabilities: The customized deep learning model can adapt to address texts of different formats and styles, as well as the specific needs of different industries, improving the model's generalization capabilities and scope of applicability.

[0175] Improved processing efficiency: The optimized model architecture and algorithm enable the present invention to quickly process large amounts of address data, making it suitable for real-time data processing scenarios and improving word segmentation efficiency.

[0176] Optimizing user experience: Accurate address segmentation provides more precise data support for applications such as map navigation and logistics distribution, significantly improving user experience.

[0177] Data update and maintenance: Regularly update the geographic information database and annotated address dataset to ensure that the system's data foundation is always kept up to date, ensuring the timeliness and accuracy of word segmentation results.

[0178] Distributed architecture support: The distributed architecture supports parallel processing and efficient computing of large-scale address texts. The system processing capacity can be dynamically expanded according to the task load to meet the application needs of different scales.

[0179] Standardized API interface: Provides a standardized API interface to facilitate seamless integration with other geographic information systems, business management systems and applications, reducing development costs and integration difficulty.

Claims

1. A geographic information enhanced deep learning address segmentation method, characterized in that: The following steps are involved: Address text preprocessing: Receive the original address text, perform text cleaning to remove irrelevant characters, implement address standardization to ensure the uniformity of place names, and perform semantic disambiguation to clarify the meaning of the address; Feature fusion extraction: Extracts basic text features, including word frequency, part of speech, and geographic entity features. Combined with a geographic information database, it extracts multi-dimensional spatial features of the address, including spatial coordinates, administrative division levels, and geographic entity relationships. Feature fusion is performed using a custom convolutional neural network (CNN) and a bidirectional long short-term memory (Bi-LSTM) network. The attention mechanism is used during feature fusion to highlight key address information. Customized model training: Design a dedicated deep learning model architecture consisting of an input layer, multiple hidden layers, and an output layer. The hidden layers include at least one multi-layer perceptron (MLP) module for feature interaction. Key features are extracted and industry feature vectors are constructed to meet the address segmentation needs of different industries. Model training is performed using professionally labeled address datasets, employing a multi-task learning strategy to simultaneously optimize segmentation accuracy and industry feature recognition accuracy. Model parameters are dynamically adjusted using an adaptive learning rate adjustment strategy and gradient descent algorithm. Word segmentation execution and optimization: The pre-processed address text is input into the trained deep learning model to implement word segmentation operations; the word segmentation results are optimized using customized post-processing rules, including rule-based pattern matching, statistics-based frequency adjustment, and semantic-based association correction; a dedicated word segmentation result intelligent verification system that combines the detailed topological information of national administrative divisions, geographic entity semantic association rules, and industry-specific verification rules is called to conduct a comprehensive verification and evaluation of the optimized word segmentation results, and the model parameters and optimization rules are dynamically adjusted according to the verification feedback to form a closed-loop optimization mechanism.

2. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: During the address text preprocessing stage, the text cleaning uses a combination of rule-based matching and deep learning language models to accurately remove spaces, special characters, industry-specific filler words, and ambiguous expression fragments, and automatically correct common text errors. The rule matching of the text cleaning is based on regular expressions, and the deep learning language model is based on the Transformer architecture.

3. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The geographic information database is a multi-source heterogeneous database that integrates national administrative division standard data, high-precision geographic spatial data and real-time updated geographic entity association data to ensure the accuracy and timeliness of geographic information. The data is updated no less frequently than once a quarter.

4. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The attention mechanism in the feature fusion process adopts a dynamic attention allocation strategy, which adaptively adjusts the interaction strength between text features and geographic information features according to the feature distribution of different address texts to improve the feature fusion effect. The dynamic attention allocation strategy calculates the attention weight based on the vector inner product of the input features.

5. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The multi-task learning strategy includes the main task of address segmentation and the auxiliary task of industry feature identification. By sharing the underlying feature extraction layer and the independent upper-level task-specific layer, the two tasks are jointly trained. The loss function is the weighted sum of the losses of the two tasks, and the weight coefficient is dynamically adjusted according to the task difficulty during the training process.

6. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The adaptive learning rate adjustment strategy adopts a dynamic adjustment algorithm based on training loss and validation loss. When the difference between the training loss and validation loss exceeds a preset threshold for multiple consecutive iterations, the learning rate is reduced; when both the training loss and validation loss decrease for multiple consecutive iterations, the learning rate is appropriately increased to ensure stable training and rapid convergence of the model.

7. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The dedicated word segmentation result intelligent verification system uses a graph neural network to construct a detailed topological map of the national administrative divisions, maps the word segmentation results to the topological map for path analysis and semantic association analysis, and combines industry-specific verification rules to accurately identify and correct word segmentation errors, ensuring that the word segmentation results conform to actual geographical logic and industry specifications. The topological map nodes of the graph neural network include administrative division units and geographical entity nodes, and the edges represent the administrative affiliation and geographical adjacency between them.

8. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The method also includes data updating and maintenance steps, regularly updating the geographic information database and the annotated address data set to ensure the timeliness of the data foundation. The data update frequency is configurable and supports both manual triggering and automatic timed update modes.

9. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The method is implemented using a distributed architecture. The preprocessing module, feature fusion module, model training module, and word segmentation optimization module are deployed on different server nodes respectively. Data transmission and task scheduling are carried out through a high-speed network. This supports parallel processing and efficient calculation of large-scale address texts. The system processing capacity can be dynamically expanded according to the task load.

10. The geographic information enhanced deep learning address segmentation method according to claim 1, characterized in that: The method provides a standardized API interface to support seamless integration with other geographic information systems, business management systems and applications. The API interface adopts a RESTful style, supports both JSON and XML data formats, and has detailed documentation and sample code to facilitate rapid calling and integration by third-party developers.

Images