Chinese address verification method

Abstract

The invention relates to a Chinese address verification method applied to the technical field of data processing. The method comprises the following steps: S1, setting a label; s2, data acquisition and processing; s3, performing feature engineering; by introducing the deep learning model, the method can more accurately capture subtle differences in address texts, and compared with a traditional rule matching method, the deep learning model has higher context understanding ability, can process complex semantic structures and spelling errors, and can improve the accuracy of the method. The model is particularly excellent in performance when processing diversified and non-standardized Chinese address data, the misjudgment rate caused by fuzzy matching can be remarkably reduced, the self-adaptive algorithm designed by the invention can dynamically adjust the recognition strategy and adapt to address input forms in different scenes, and in addition, the recognition accuracy of the false address is improved. According to the invention, high-efficiency verification can be realized under the condition that manual intervention is reduced.

Description

Technical Field

[0001] The present invention relates to a verification method, in particular to a Chinese address verification method applied in the field of data processing technology. Background Art

[0002] With the rapid development of e-commerce, logistics and distribution, and geographic information systems, the quality of address data directly affects the efficiency and accuracy of these systems. At present, common address processing technologies include rule-based address parsing, address standardization, and mapping address text to specific geographic coordinates using geocoding technology. These technologies mainly rely on predefined rules and templates to segment and match addresses, which have certain effects, but often encounter challenges when processing complex and changeable Chinese addresses. Existing address matching algorithms are prone to misjudgment when dealing with addresses with spelling errors, confusing grammar, or irregular formats, and have limited ability to identify false addresses, which seriously affects the accuracy of data and the reliability of application systems.

[0003] Therefore, we propose a Chinese address verification method. Summary of the invention

[0004] In view of the above-mentioned prior art, the technical problem to be solved by the present invention is that the existing address matching algorithm is prone to misjudgment when processing addresses with spelling errors, confusing grammar or irregular formats, and has limited ability to identify false addresses.

[0005] To solve the above problems, the present invention provides a Chinese address verification method, comprising the following steps:

[0006] S1. Label setting: setting the real address label and the fake address label;

[0007] S2. Data collection and processing: collect real address data and fake address data, and perform data cleaning, data enhancement, data labeling, data balancing and sampling, and data segmentation and verification on the collected data in turn;

[0008] S3. Feature Engineering:

[0009] S31, basic feature construction;

[0010] S32, feature extraction;

[0011] S33, data enhancement and feature enhancement;

[0012] S34, feature interaction and combination;

[0013] S35, feature verification and adjustment;

[0014] S4. Model selection and optimization:

[0015] S41, model selection;

[0016] S42, model optimization;

[0017] S43, Model evaluation and validation;

[0018] S44. Model deployment and optimization.

[0019] In the above-mentioned Chinese address verification method, by introducing a deep learning model, it is possible to more accurately capture subtle differences in the address text and handle complex semantic structures and spelling errors, thereby improving the accuracy of identifying false addresses and significantly reducing the misjudgment rate caused by fuzzy matching.

[0020] As a further improvement of the present application, step S31 includes:

[0021] S31A, text features:

[0022] Character-level features: extract the character length, character frequency distribution, and character type distribution of each address, and analyze their rationality;

[0023] Word-level features: using the word segmentation tool Jieba to segment the address and extract word-level features, including but not limited to word frequency, part of speech, and word order;

[0024] S31B, Geographical features:

[0025] Calculate the longitude and latitude corresponding to the address and extract geographical features, including but not limited to the distance to the neighboring area, the distance to the center point, and the distance to the administrative division boundary;

[0026] Extract the administrative division level from the address and match it with the standard geocode;

[0027] S31C, semantic features:

[0028] Use TFIDF to represent the address text and extract the high-frequency vocabulary features of each address;

[0029] The pre-trained word vector model BERT is used to convert the address into a vector and extract the semantic similarity features of the words.

[0030] As a further improvement of the present application, step S32 includes:

[0031] S32A, address structure characteristics:

[0032] Address hierarchical analysis: decompose addresses into multiple levels, analyze the integrity and hierarchical relationship of each level, and mark addresses that do not conform to the hierarchical relationship as abnormal;

[0033] Nested and repeated features, detect repeated and nested structures in addresses;

[0034] S32B, Logic and Context Features:

[0035] Address logical consistency: Use geographic information systems and map APIs to verify whether each part of the address is logically consistent;

[0036] Contextual semantic consistency: BERT is used to perform contextual semantic analysis on the address text to determine whether the various parts of the address are semantically coherent;

[0037] S32C, Dynamic Feature Generation:

[0038] Dynamic dictionary expansion, based on the address dictionary that is constantly updated in actual applications, for feature extraction and model training;

[0039] Feature crossover, by cross-calculating basic features to generate high-order features to capture more complex patterns.

[0040] As a further improvement of the present application, step S33 includes:

[0041] S33A, data enhancement features:

[0042] Perturbation enhancement features: based on the perturbed address generated by the data enhancement part, extract the difference features between it and the original address to distinguish between true and false addresses;

[0043] For addresses with mixed languages, extract multilingual features and detect their rationality;

[0044] Analyze the style of the address and extract style features that are different from standard modern addresses to assist in determining the authenticity of the address;

[0045] S33B, Dynamic Feature Optimization:

[0046] Feature screening and weight adjustment: use the feature importance of feature selection algorithm XGBoost to screen out the most important features, and dynamically adjust feature weights to improve the robustness of the model;

[0047] Automated feature engineering: With the help of the automatic feature engineering tool FeatureTools, features are automatically generated and screened, reducing manual intervention and improving the efficiency of feature extraction.

[0048] As a further improvement of the present application, step S34 includes:

[0049] Multi-feature interaction, exploring the interactive relationship between different features and generating new composite features;

[0050] Nonlinear feature combination, through polynomial feature combination or deep learning feature interaction layer, extracts nonlinear feature combination and improves the expressiveness of the model.

[0051] As a further improvement of the present application, step S35 includes:

[0052] S35A, Feature Importance Analysis:

[0053] Use the SHAP model interpretation tool to analyze the model's dependence on each feature, identify the features that have the greatest impact on the model, and optimize feature engineering;

[0054] S35B, Feature extension and update:

[0055] Dynamically expand and update the feature library based on the latest data in the application scenario to improve the adaptability of the model;

[0056] S35C, Feature Selection and Optimization:

[0057] Greedy algorithm and genetic algorithm: Apply greedy algorithm or genetic algorithm to automatically select the optimal feature combination to further improve the generalization ability of the model.

[0058] As another improvement of the present application, the step S41 includes:

[0059] S41A, pre-trained language model selection: select BART as the pre-trained language model;

[0060] S41B, Model architecture modification:

[0061] Multi-task learning: design a multi-task learning architecture to handle both address classification and geocoding tasks in the same model, and improve the overall performance of the model by sharing feature extraction layers;

[0062] Hierarchical attention mechanism,The introduction of hierarchical attention mechanism enables the model to focus on different levels in the address text and improve the understanding of hierarchical relationships.

[0063] As another improved supplement of the present application, the step S42 includes:

[0064] S42A, domain-adaptive fine-tuning:

[0065] Domain adaptive training: Based on the existing pre-trained model, a large amount of Chinese address data is used for further fine-tuning, so that the model can better understand the specialized knowledge in the address field;

[0066] Data enhancement optimization: using the perturbed data generated by the data enhancement part for adversarial training to improve the model’s robustness to false addresses and reduce misclassification;

[0067] S42B, Model Compression and Acceleration:

[0068] Knowledge distillation, which distills the knowledge of the large BERT model into a smaller and more efficient model DistilBERT, reducing computing resource consumption while ensuring performance;

[0069] Pruning and quantization: Through model pruning and quantization technology, the number of model parameters and computational complexity are reduced, the reasoning speed is improved, and it is suitable for deployment environments with limited resources;

[0070] S42C, Innovative training strategies:

[0071] Dynamic learning rate adjustment: introduces an adaptive learning rate adjustment strategy, combined with Warmup and CosineAnnealing, so that the model can converge efficiently at different stages of training;

[0072] Mixed precision training: Use mixed precision training to improve training efficiency;

[0073] S42D, Self-supervised learning and contrastive learning:

[0074] Self-supervised learning: In the absence of labels, the self-supervised learning method Masked Language Modeling is used to further improve the model's understanding of address data;

[0075] Through comparative learning, the model can learn the semantic differences between different addresses and enhance the classification ability of the model.

[0076] As another improved supplement of the present application, the step S43 includes:

[0077] S43A, Diversified Evaluation Indicators:

[0078] Classification accuracy and contrast loss: In addition to conventional accuracy, recall, and F1score, contrast loss is added as an evaluation indicator to verify the model's ability to distinguish between positive and negative samples.

[0079] Conduct robustness tests on the model by introducing disturbances to test its performance in dealing with uncertainty and incomplete address data;

[0080] S43B, Model interpretation and explainability:

[0081] Attention weight visualization: using the model’s attention mechanism, we analyze the key parts that the model focuses on when processing addresses, and verify whether the model correctly understands the address structure.

[0082] Adversarial sample analysis: Generate adversarial samples to test the robustness of the model, analyze the model’s weaknesses during misclassification, and further optimize the model.

[0083] As another improved supplement of the present application, the step S44 includes:

[0084] S44A, lightweight deployment:

[0085] Distributed reasoning: Using a distributed reasoning architecture, the model's reasoning tasks are decomposed into multiple computing nodes to improve the efficiency of processing large-scale address data;

[0086] S44B, Continuous Learning and Model Updating:

[0087] Incremental learning: After the model is deployed, the model is updated in real time through incremental learning methods to handle emerging address patterns and language changes, ensuring the long-term adaptability of the model;

[0088] User feedback mechanism: Establish a user feedback mechanism to collect data from users in actual use, continuously optimize the model, and update the feature library to adapt to the latest needs.

[0089] In summary, the present invention, by introducing a deep learning model, can more accurately capture subtle differences in address texts, and can handle complex semantic structures and spelling errors, thereby improving the accuracy of identifying false addresses, and further significantly reducing the misjudgment rate caused by fuzzy matching. The adaptive algorithm designed by the present invention can dynamically adjust the recognition strategy to adapt to address input forms in different scenarios. The present invention learns various address expressions through training models, thereby reducing dependence on rule library updates and enabling the system to maintain efficient and accurate matching capabilities when facing new forms of address input. In addition, with the help of deep learning models and automated data processing procedures, the present invention can achieve efficient verification with reduced manual intervention, and the present invention can continuously improve the verification effect and reduce the frequency and workload of manual intervention through automated model training and optimization, which not only improves the operating efficiency of the system, but also reduces maintenance costs. This advantage is even more obvious in large-scale address data processing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 This is a method flow chart of the first implementation mode of this application; DETAILED DESCRIPTION

[0091] An implementation of the present application is described in detail below with reference to the accompanying drawings.

[0092] The first implementation method:

[0093] Figure 1 A Chinese address verification method is shown, comprising the following steps:

[0094] S1. Label setting: setting the real address label and the fake address label;

[0095] True address labels include the following:

[0096] The administrative divisions of the address at all levels (such as province, city, district, township, street, village, community, etc.) are correct;

[0097] The words are spelled correctly and have the possibility of real existence;

[0098] Fake address labels include the following:

[0099] Administrative division attribution errors: cross-administrative region errors, hierarchical mismatches, and non-existent administrative divisions;

[0100] Spelling and grammatical errors: common typos, homophone substitutions, missing or extra characters, confusing grammatical structures, and phonetic errors;

[0101] Non-address data or illegal characters: all numbers or letters, mixed meaningless characters, phone numbers or ID numbers, illegal characters, special symbols or emoticons;

[0102] Exaggerated or fictitious addresses: obviously fictitious place names, addresses that are too long or too short, repeated or circular addresses, and addresses generated by fake address generators;

[0103] Formatting and hierarchy errors: incorrect separators, confusing address hierarchies, incorrect description of administrative levels;

[0104] Incomplete address: missing important information, only contains the upper administrative division;

[0105] Language mixing: mixing Chinese and English, and mixing pinyin and Chinese characters;

[0106] Historical or abandoned addresses: old or abandoned place names, obsolete place names;

[0107] Logical errors: geographical contradictions, impossible combinations, semantic contradictions;

[0108] Cultural or custom incompatibility: cultural background mismatch, dialect or local terms, use of unconventional address descriptions;

[0109] Artificial tampering or mischief: maliciously tampered addresses, mischievous or mocking place names;

[0110] Location-related errors: geographic deviation, location conversion errors;

[0111] S2. Data collection and processing:

[0112] S21. Data collection:

[0113] S21A, collect real address data:

[0114] Including but not limited to collecting real address data from authoritative address databases, geographic information systems, and map service APIs;

[0115] S21B, collect fake address data:

[0116] Collect public fake address data sets (such as incorrect user input addresses, addresses generated by pranks, etc.), generate various types of fake addresses through strategies (strategies include but are not limited to rule combinations and random perturbations), and use user input data in actual applications to obtain incorrect address data in real scenarios;

[0117] S21C, multi-regional samples and historical data collection:

[0118] Cover addresses in different provinces, cities and districts to ensure that the model can handle addresses in different regions, dialects or cultural backgrounds. Collect address data from different time periods, including outdated place names and historical changes, to improve the model's ability to handle old addresses.

[0119] S22. Data cleaning:

[0120] S22A, multi-level cleaning:

[0121] Unified format: format all address data, unify separators, spaces, upper and lower case, etc. to ensure data consistency;

[0122] Remove redundant information and clean up invalid information in the address, such as comments, advertising words, special characters, etc.

[0123] Spelling and grammar checking: Use the spelling checker to check the spelling and grammar of the address text and correct obvious errors;

[0124] S22B, Geographic Location Verification:

[0125] Longitude and latitude matching: using GIS and map API to verify whether the address data matches the actual longitude and latitude;

[0126] Administrative division verification to ensure that the administrative level information such as province, city, district in the address is correct and consistent with official data;

[0127] S22C, Anomaly Detection:

[0128] Rule detection: use preset rules (address length, separator usage) to detect and filter abnormal data;

[0129] Abnormal character filtering: through character set analysis, filter out abnormal data containing garbled characters, illegal characters, etc.

[0130] S23, data enhancement:

[0131] S23A, rule-based generation:

[0132] Rule combination generation, generating fake addresses through a combination of spelling errors and administrative division confusion;

[0133] Randomly perturb the real address, including replacing characters, disrupting the order, inserting special symbols, and generating a variety of false addresses;

[0134] S23B, Templated Generation and Adversarial Generation:

[0135] Templated generation: based on templates of common errors, batches of fake addresses with specific error characteristics are generated;

[0136] S23C, data driven generation:

[0137] Analyze common error patterns in real data and generate fake data in a targeted manner based on these patterns;

[0138] Cluster generation: cluster analysis is performed on the error addresses to generate representative samples;

[0139] S23D, dynamic enhancement and iterative optimization:

[0140] Iterative generation: initially generate some fake addresses. After training the model, generate new fake addresses based on the performance of the model to optimize the model.

[0141] Real-time generation: after discovering new error patterns in actual applications, new false address data is dynamically generated to supplement;

[0142] S24. Data Annotation

[0143] S24A, multi-label annotation:

[0144] Layered labeling: multi-level labeling for each address data, not only distinguishing true from false but also marking the specific error type;

[0145] S24B, crowdsourcing annotation:

[0146] Manual review, through crowdsourcing platforms or communities, especially for complex or borderline addresses;

[0147] Dynamic annotation, combining model predictions and user feedback to dynamically update annotations to ensure the accuracy of the data set;

[0148] S25. Data balancing and sampling:

[0149] S25A, sample balance:

[0150] False address oversampling: Balance the dataset by oversampling or generating false addresses;

[0151] Undersampling of real addresses: Undersampling of real addresses with a large number of real addresses to reduce redundancy and prevent overfitting of the model;

[0152] S25B, Intelligent Sampling:

[0153] Cluster sampling, using cluster analysis to group data and conduct sampling within each group to ensure sample diversity;

[0154] Semantic distribution sampling: distributes data based on semantic analysis and selects representative samples;

[0155] S26. Data segmentation and verification:

[0156] S26A, cross-domain data segmentation:

[0157] Regional segmentation: data segmentation based on different geographical regions to ensure the cross-regional generalization ability of the model;

[0158] Time segmentation: split the data by time axis to test the model's ability to process historical data;

[0159] S26B, Online Verification:

[0160] Through the streaming data processing system, the authenticity of new addresses is verified in real time and the model is continuously updated;

[0161] Establish a user feedback mechanism to collect and use the addresses entered by users and their verification results to optimize the model;

[0162] S3. Feature Engineering:

[0163] S31. Basic feature construction:

[0164] S31A, text features:

[0165] Character-level features: extract the character length, character frequency distribution, and character type distribution (Chinese characters, numbers, letters, special symbols, etc.) of each address, and analyze their rationality;

[0166] Word-level features: Use the word segmentation tool Jieba to segment the address and extract word-level features such as word frequency, part of speech, and word order.

[0167] S31B, Geographical features:

[0168] Calculate the longitude and latitude corresponding to the address, and extract address features such as the distance to the neighboring area, the distance to the center point, and the distance to the administrative boundary;

[0169] Extract the administrative division level in the address (such as the completeness and correctness of "province, city, district" and match it with the standard geocoding;

[0170] S31C, semantic features:

[0171] Use TFIDF to represent the address text and extract the high-frequency vocabulary features of each address;

[0172] Use the pre-trained word vector model BERT to convert the address into a vector and extract the semantic similarity features of the words;

[0173] S32, feature extraction:

[0174] S32A, address structure characteristics:

[0175] Address hierarchical analysis: decompose addresses into multiple levels (such as country, province, city, district, street, house number, etc.), analyze the integrity and hierarchical relationship of each level, and mark addresses that do not conform to the hierarchical relationship as abnormal;

[0176] Nesting and duplication features detect duplication and nested structures in addresses, which are often signs of fake addresses;

[0177] S32B, Logic and Context Features:

[0178] Address logical consistency: Use geographic information system (GIS) and map API to verify whether the various parts of the address are logically consistent, such as whether the "street" is located in the "district", whether the "district" belongs to the "city", etc.

[0179] Contextual semantic consistency: BERT is used to perform contextual semantic analysis on the address text to determine whether the various parts of the address are semantically coherent;

[0180] S32C, Dynamic Feature Generation:

[0181] Dynamic dictionary expansion: based on the constantly updated address dictionary in actual applications, new dictionaries such as place names and street names are dynamically generated for feature extraction and model training;

[0182] Feature crossover, by cross-calculating basic features to generate high-order features (such as cross-features of character length and word frequency) to capture more complex patterns;

[0183] S33, data enhancement and feature enhancement:

[0184] S33A, data enhancement features:

[0185] Perturbation enhancement features: based on the perturbed address generated by the data enhancement part, extract the difference features between it and the original address, such as edit distance, similarity score, etc., to distinguish between true and false addresses;

[0186] For multilingual addresses, such as those that are a mixture of Chinese and English or Pinyin and Chinese characters, extract multilingual features and test their rationality;

[0187] Analyze the style of the address, such as ancient place names, foreign place names, etc., extract style features that are different from standard modern addresses, and assist in determining the authenticity of the address;

[0188] S33B, Dynamic Feature Optimization:

[0189] Feature screening and weight adjustment: use the feature importance of feature selection algorithm XGBoost to screen out the most important features, and dynamically adjust feature weights to improve the robustness of the model;

[0190] Automated feature engineering: With the help of the automatic feature engineering tool FeatureTools, features are automatically generated and screened, reducing manual intervention and improving the efficiency of feature extraction;

[0191] S34. Feature interaction and combination:

[0192] Multi-feature interaction: exploring the interactive relationship between different features, such as the combination of text features and geographic features, semantic features and logical features, to generate new composite features;

[0193] Nonlinear feature combination: extract nonlinear feature combinations through polynomial feature combination or deep learning feature interaction layer to improve the expression ability of the model;

[0194] S35, Feature verification and adjustment:

[0195] S35A, Feature Importance Analysis:

[0196] Use the SHAP model interpretation tool to analyze the model's dependence on each feature, identify the features that have the greatest impact on the model, and optimize feature engineering;

[0197] S35B, Feature extension and update:

[0198] Dynamically expand and update the feature library based on the latest data in the application scenario, such as adding new street names, place names and other features to improve the adaptability of the model;

[0199] S35C, Feature Selection and Optimization:

[0200] Greedy algorithm and genetic algorithm: Apply greedy algorithm or genetic algorithm to automatically select the optimal feature combination to further improve the generalization ability of the model;

[0201] S4. Model selection and optimization:

[0202] S41. Model selection:

[0203] S41A, pre-trained language model selection:

[0204] BART is selected as the pre-trained language model. This type of model can not only handle classification tasks, but also generate text. It is suitable for scenarios that require text repair or enhancement, such as generating reasonable fake addresses.

[0205] S41B, Model architecture modification:

[0206] Multi-task learning: design a multi-task learning architecture to handle both address classification and geocoding tasks in the same model, and improve the overall performance of the model by sharing feature extraction layers;

[0207] Hierarchical attention mechanism: The hierarchical attention mechanism is introduced to enable the model to focus on different levels in the address text and improve the understanding of hierarchical relationships;

[0208] S42. Model optimization:

[0209] S42A, domain-adaptive fine-tuning:

[0210] Domain adaptive training: Based on the existing pre-trained model, a large amount of Chinese address data is used for further fine-tuning, so that the model can better understand the specialized knowledge in the address field;

[0211] Data enhancement optimization: using the perturbed data generated by the data enhancement part for adversarial training to improve the model’s robustness to false addresses and reduce misclassification;

[0212] S42B, Model Compression and Acceleration:

[0213] Knowledge distillation, which distills the knowledge of the large BERT model into a smaller and more efficient model DistilBERT, reducing computing resource consumption while ensuring performance;

[0214] Pruning and quantization: Through model pruning and quantization technology, the number of model parameters and computational complexity are reduced, the reasoning speed is improved, and it is suitable for deployment environments with limited resources;

[0215] S42C, Innovative training strategies:

[0216] Dynamic learning rate adjustment: introduces an adaptive learning rate adjustment strategy, combined with Warmup and CosineAnnealing, so that the model can converge efficiently at different stages of training;

[0217] Mixed precision training: Use mixed precision training (FP16 / FP32) to improve training efficiency, especially on large-scale data sets, which can accelerate model convergence;

[0218] S42D, Self-supervised learning and contrastive learning:

[0219] Self-supervised learning: In the absence of labels, the self-supervised learning method Masked Language Modeling is used to further improve the model's understanding of address data;

[0220] Through contrastive learning, the model can learn the semantic differences between different addresses and strengthen the classification ability of the model;

[0221] S43. Model evaluation and validation:

[0222] S43A, Diversified Evaluation Indicators:

[0223] Classification accuracy and contrast loss: In addition to conventional accuracy, recall, and F1score, contrast loss is added as an evaluation indicator to verify the model's ability to distinguish between positive and negative samples.

[0224] Conduct robustness tests on the model by introducing disturbances such as noise and spelling errors to test how the model handles uncertainty and incomplete address data;

[0225] S43B, Model interpretation and explainability:

[0226] Attention weight visualization: using the model’s attention mechanism, we analyze the key parts that the model focuses on when processing addresses, and verify whether the model correctly understands the address structure.

[0227] Adversarial sample analysis: Generate adversarial samples to test the robustness of the model, analyze the model's weaknesses when it is misclassified, and further optimize the model;

[0228] S44. Model deployment and optimization:

[0229] S44A, lightweight deployment:

[0230] Distributed reasoning: Using a distributed reasoning architecture, the model's reasoning tasks are decomposed into multiple computing nodes to improve the efficiency of processing large-scale address data;

[0231] S44B, Continuous Learning and Model Updating:

[0232] Incremental learning: After the model is deployed, the model is updated in real time through incremental learning methods to handle emerging address patterns and language changes, ensuring the long-term adaptability of the model;

[0233] User feedback mechanism: Establish a user feedback mechanism to collect data from users in actual use, continuously optimize the model, and update the feature library to adapt to the latest needs.

[0234] First, by introducing the deep learning model, the present invention can more accurately capture the subtle differences in the address text. Compared with the traditional rule matching method, the deep learning model has a stronger context understanding ability and can handle complex semantic structures and spelling errors, thereby improving the recognition accuracy of false addresses. This model performs particularly well when processing diverse and non-standardized Chinese address data, and can significantly reduce the misjudgment rate caused by fuzzy matching; secondly, the adaptive algorithm designed by the present invention can dynamically adjust the recognition strategy to adapt to the address input form in different scenarios. The traditional rule matching method often needs to frequently update the rule base when dealing with the ever-changing address expression, which is difficult to cope with massive and dynamically changing The present invention learns various address expressions through training models, thereby reducing the dependence on rule base updates and enabling the system to maintain efficient and accurate matching capabilities when facing new forms of address input; in addition, with the help of deep learning models and automated data processing processes, the present invention can achieve efficient verification with reduced manual intervention. Traditional methods often require a large amount of manual work to formulate and maintain address rules, while the present invention can continuously improve the verification effect and reduce the frequency and workload of manual intervention through automated model training and optimization, which not only improves the operating efficiency of the system, but also reduces maintenance costs. This advantage is even more obvious in large-scale address data processing scenarios.

[0235] In view of current practical needs, the above-mentioned implementation mode adopted in this application is not limited to the scope of protection. Various changes made within the knowledge scope of technical personnel in this field without departing from the concept of this application still fall within the scope of protection of the present invention.

Claims

1. A Chinese address verification method, characterized in that: The following steps are involved: S1. Label setting: setting the real address label and the fake address label; S2. Data collection and processing: collect real address data and fake address data, and perform data cleaning, data enhancement, data labeling, data balancing and sampling, and data segmentation and verification on the collected data in turn; S3. Feature Engineering: S31, basic feature construction; S32, feature extraction; S33, data enhancement and feature enhancement; S34, feature interaction and combination; S35, feature verification and adjustment; S4. Model selection and optimization: S41, model selection; S42, model optimization; S43, Model evaluation and validation; S44. Model deployment and optimization.

2. A Chinese address verification method according to claim 1, characterized in that: The step S31 comprises: S31A, text features: Character-level features: extract the character length, character frequency distribution, and character type distribution of each address, and analyze their rationality; Word-level features: using the word segmentation tool Jieba to segment the address and extract word-level features, including but not limited to word frequency, part of speech, and word order; S31B, Geographical features: Calculate the longitude and latitude corresponding to the address and extract geographical features, including but not limited to the distance to the neighboring area, the distance to the center point, and the distance to the administrative division boundary; Extract the administrative division level from the address and match it with the standard geocode; S31C, semantic features: Use TFIDF to represent the address text and extract the high-frequency vocabulary features of each address; The pre-trained word vector model BERT is used to convert the address into a vector and extract the semantic similarity features of the words.

3. A Chinese address verification method according to claim 1, characterized in that: The step S32 comprises: S32A, address structure characteristics: Address hierarchical analysis: decompose addresses into multiple levels, analyze the integrity and hierarchical relationship of each level, and mark addresses that do not conform to the hierarchical relationship as abnormal; Nested and repeated features, detect repeated and nested structures in addresses; S32B, Logic and Context Features: Address logical consistency: Use geographic information systems and map APIs to verify whether each part of the address is logically consistent; Contextual semantic consistency: BERT is used to perform contextual semantic analysis on the address text to determine whether the various parts of the address are semantically coherent; S32C, Dynamic Feature Generation: Dynamic dictionary expansion, based on the address dictionary that is constantly updated in actual applications, for feature extraction and model training; Feature crossover, by cross-calculating basic features to generate high-order features to capture more complex patterns.

4. A Chinese address verification method according to claim 1, characterized in that: The step S33 comprises: S33A, data enhancement features: Perturbation enhancement features: based on the perturbed address generated by the data enhancement part, extract the difference features between it and the original address to distinguish between true and false addresses; For addresses with mixed languages, extract multilingual features and detect their rationality; Analyze the style of the address and extract style features that are different from standard modern addresses to assist in determining the authenticity of the address; S33B, Dynamic Feature Optimization: Feature screening and weight adjustment: use the feature importance of feature selection algorithm XGBoost to screen out the most important features, and dynamically adjust feature weights to improve the robustness of the model; Automated feature engineering: With the help of the automatic feature engineering tool FeatureTools, features are automatically generated and screened, reducing manual intervention and improving the efficiency of feature extraction.

5. A Chinese address verification method according to claim 1, characterized in that: The step S34 comprises: Multi-feature interaction, exploring the interactive relationship between different features and generating new composite features; Nonlinear feature combination, through polynomial feature combination or deep learning feature interaction layer, extracts nonlinear feature combination and improves the expressiveness of the model.

6. A Chinese address verification method according to claim 1, characterized in that: The step S35 comprises: S35A, Feature Importance Analysis: Use the SHAP model interpretation tool to analyze the model's dependence on each feature, identify the features that have the greatest impact on the model, and optimize feature engineering; S35B, Feature extension and update: Dynamically expand and update the feature library based on the latest data in the application scenario to improve the adaptability of the model; S35C, Feature Selection and Optimization: Greedy algorithm and genetic algorithm: Apply greedy algorithm or genetic algorithm to automatically select the optimal feature combination to further improve the generalization ability of the model.

7. A Chinese address verification method according to claim 1, characterized in that: The step S41 comprises: S41A, pre-trained language model selection: select BART as the pre-trained language model; S41B, Model architecture modification: Multi-task learning: design a multi-task learning architecture to handle both address classification and geocoding tasks in the same model, and improve the overall performance of the model by sharing feature extraction layers; Hierarchical attention mechanism,The introduction of hierarchical attention mechanism enables the model to focus on different levels in the address text and improve the understanding of hierarchical relationships.

8. A Chinese address verification method according to claim 1, characterized in that: The step S42 comprises: S42A, domain-adaptive fine-tuning: Domain adaptive training: Based on the existing pre-trained model, a large amount of Chinese address data is used for further fine-tuning, so that the model can better understand the specialized knowledge in the address field; Data enhancement optimization: using the perturbed data generated by the data enhancement part for adversarial training to improve the model’s robustness to false addresses and reduce misclassification; S42B, Model Compression and Acceleration: Knowledge distillation, which distills the knowledge of the large BERT model into a smaller and more efficient model DistilBERT, reducing computing resource consumption while ensuring performance; Pruning and quantization: Through model pruning and quantization technology, the number of model parameters and computational complexity are reduced, the reasoning speed is improved, and it is suitable for deployment environments with limited resources; S42C, Innovative training strategies: Dynamic learning rate adjustment: introduces an adaptive learning rate adjustment strategy, combined with Warmup and Cosine Annealing, so that the model can converge efficiently at different stages of training; Mixed precision training: Use mixed precision training to improve training efficiency; S42D, Self-supervised learning and contrastive learning: Self-supervised learning: In the absence of labels, the self-supervised learning method Masked Language Modeling is used to further improve the model's understanding of address data; Through comparative learning, the model can learn the semantic differences between different addresses and enhance the classification ability of the model.

9. A Chinese address verification method according to claim 1, characterized in that: The step S43 comprises: S43A, Diversified Evaluation Indicators: Classification accuracy and contrast loss: In addition to conventional accuracy, recall, and F1score, contrast loss is added as an evaluation indicator to verify the model's ability to distinguish between positive and negative samples. Conduct robustness tests on the model by introducing disturbances to test its performance in dealing with uncertainty and incomplete address data; S43B, Model interpretation and explainability: Attention weight visualization: using the model’s attention mechanism, we analyze the key parts that the model focuses on when processing addresses, and verify whether the model correctly understands the address structure. Adversarial sample analysis: Generate adversarial samples to test the robustness of the model, analyze the model’s weaknesses during misclassification, and further optimize the model.

10. A Chinese address verification method according to claim 1, characterized in that: The step S44 comprises: S44A, lightweight deployment: Distributed reasoning: Using a distributed reasoning architecture, the model's reasoning tasks are decomposed into multiple computing nodes to improve the efficiency of processing large-scale address data; S44B, Continuous Learning and Model Updating: Incremental learning: After the model is deployed, the model is updated in real time through incremental learning methods to handle emerging address patterns and language changes, ensuring the long-term adaptability of the model; User feedback mechanism: Establish a user feedback mechanism to collect data from users in actual use, continuously optimize the model, and update the feature library to adapt to the latest needs.

Images