A processing method for automatically splitting address data elements

Abstract

The application discloses a kind of processing methods for automatically splitting address data elements, belong to computer data processing technical field, it includes obtaining address string and executing character level anomaly detection, generate correction address string and parallel input to preset rule path and preset model path, generate rule splitting result and model splitting result;First confidence is calculated based on rule splitting result, and second confidence is calculated based on model splitting result;Determine whether first confidence and second confidence are higher than preset confidence threshold, and rule splitting result is consistent with model splitting result;If not, then use rule splitting result and model splitting result to carry out dynamic negotiation, generate negotiation splitting result and determine as final splitting result.The application adopts rule and model double-path parallel processing, and combines confidence judgment and dynamic negotiation mechanism, can give consideration to the accuracy, robustness and self-adapting learning ability of address splitting.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and in particular to a method for automatically splitting address data elements. Background Technology

[0002] Automated address data segmentation is a technology that analyzes unstructured address text in a computer system and automatically parses it into structured data fields (such as province, city, district, street, and house number). This technology is a crucial foundation in many application scenarios, including geographic information systems, logistics and distribution, financial risk control, and customer relationship management. Its accuracy and efficiency directly impact the operational quality of downstream businesses. By converting free-text addresses into machine-readable standardized elements, standardized management, precise location, and efficient utilization of address data can be achieved.

[0003] In existing technologies, address element segmentation mainly relies on two mainstream approaches. One is a rule-based method, which pre-constructs a detailed administrative division dictionary and address keyword database, and combines string matching algorithms such as maximum forward matching and regular expressions to identify the various components of an address. The other is a machine learning-based method, which typically models the problem as a sequence labeling task, using models such as conditional random fields or recurrent neural networks, and learns the mapping relationship from address strings to element labels by training on a large amount of labeled address data.

[0004] However, while rule-based methods achieve high accuracy when processing standardized addresses, their rule base maintenance is costly, and they are poorly adaptable to non-standard cases such as colloquial expressions, typos, missing information, or reversed order of addresses, resulting in insufficient generalization ability. Methods relying solely on machine learning models, while exhibiting strong generalization ability, are highly dependent on the quantity and quality of training data. They may produce unpredictable errors for new address patterns not present in the training set, and their decision-making process lacks interpretability, sometimes making it difficult to guarantee the stability of the results. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an automated method for splitting address data elements. It employs a dual-path parallel processing approach combining rules and models, and integrates confidence assessment and dynamic negotiation mechanisms. This approach balances the accuracy, robustness, and adaptive learning capabilities of address splitting.

[0006] The above objectives can be achieved through the following approach:

[0007] A method for automatically splitting address data elements includes: acquiring an address string and performing character-level anomaly detection on the address string to generate a corrected address string; inputting the corrected address string in parallel to a preset rule path and a preset model path, wherein the rule path uses a preset rule base for address element matching to generate a rule splitting result, and the model path uses a preset sequence labeling model for address element identification to generate a model splitting result; calculating a first confidence level based on the rule splitting result and calculating a second confidence level based on the model splitting result; determining whether both the first confidence level and the second confidence level are higher than a preset confidence level threshold, and whether the rule splitting result is consistent with the model splitting result; if so, determining the rule splitting result as the final splitting result; if not, dynamically negotiating the rule splitting result and the model splitting result to generate a negotiated splitting result, and determining the negotiated splitting result as the final splitting result.

[0008] Optionally, generating the corrected address string includes: obtaining the address string and scanning the address string based on a preset character-level attention network to generate an abnormal character identifier; obtaining the semantic environment before and after the abnormal character based on the abnormal character identifier and generating contextual semantic features; generating a candidate correction scheme set based on the abnormal character identifier and the contextual semantic features; evaluating and selecting a candidate correction scheme from the candidate correction scheme set, and applying the candidate correction scheme to the address string to generate the corrected address string.

[0009] Optionally, generating the rule splitting result includes: establishing a rule base using a preset administrative division tree and a preset street standard library; performing pattern matching between the modified address string and the rule base to extract matching address elements; and generating the rule splitting result based on the matching address elements.

[0010] Optionally, the generation of model splitting results includes: establishing a sequence labeling model for identifying address features; inputting the corrected address string into the sequence labeling model to generate a sequence labeling tag sequence; parsing the sequence labeling tag sequence to determine the address feature type and boundary; and generating model splitting results based on the address feature type and boundary.

[0011] Optionally, generating the negotiated splitting result includes: using the rule splitting result as a reinforcement learning signal to adjust the sequence labeling model; extracting key semantic features from the model splitting result and using the key semantic features to expand the rule base; and using the adjusted sequence labeling model and the expanded rule base to reprocess the corrected address string to generate the negotiated splitting result.

[0012] Optionally, expanding the rule base using the key semantic features includes: parsing a new address pattern from the key semantic features; adding the new address pattern to the fuzzy matching rule set of the rule base to generate an expanded rule base; and recording the error patterns reflected by the semantic features.

[0013] Optionally, the method further includes: filtering candidate confidence cases based on the confidence of the final splitting result; verifying the candidate confidence cases to generate labeled samples; using the labeled samples to establish a training dataset, and using the training dataset to periodically update the sequence labeling model and the rule base.

[0014] Optionally, the step of verifying the candidate confidence cases and generating labeled samples includes: verifying the candidate confidence cases based on preset verification conditions; if the verification is successful, generating a first verification split result; if the verification fails, submitting the candidate confidence cases to a manual verification process to generate a second verification split result; combining the address string of the candidate confidence cases with the first verification split result or the second verification split result, and adding version information to generate labeled samples.

[0015] Optionally, the method further includes: the rule base is stored using a graph structure, wherein nodes represent address elements and edges represent the relationships between the address elements; the matching process of the rule path is a graph traversal process, and the graph traversal process uses the relationships represented by the edges to perform fuzzy matching and semantic association.

[0016] Based on the same inventive concept, this invention also provides an automated processing system for splitting address data elements. The system includes: a character correction module for acquiring an address string and performing character-level anomaly detection on the address string to generate a corrected address string; a character splitting module for inputting the corrected address string in parallel to a preset rule path and a preset model path, wherein the rule path uses a preset rule base for address element matching to generate a rule splitting result, and the model path uses a preset sequence labeling model for address element identification to generate a model splitting result; a confidence calculation module for calculating a first confidence level based on the rule splitting result and a second confidence level based on the model splitting result; and a splitting result determination module for determining whether both the first confidence level and the second confidence level are higher than a preset confidence threshold, and whether the rule splitting result is consistent with the model splitting result; if so, the rule splitting result is determined as the final splitting result; if not, the rule splitting result and the model splitting result are dynamically negotiated to generate a negotiated splitting result, and the negotiated splitting result is determined as the final splitting result.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. This invention effectively combines the high precision of deterministic rules with the strong generalization ability of deep learning models by constructing a framework for parallel processing of rules and models. When the results of the two are consistent and have high credibility, they are quickly adopted; when they are inconsistent, dynamic negotiation is initiated. This mechanism enables the address splitting system to efficiently process standard addresses and intelligently cope with complex and ever-changing non-standard addresses, significantly improving the overall accuracy and coverage of address resolution.

[0019] 2. This invention proposes a closed-loop mechanism of dynamic negotiation and adaptive learning. When there is a discrepancy between the rule and the model result, the system can not only generate a better negotiation result, but also use the difference information in the process to fine-tune the model with the rule result as a reinforcement learning signal, and at the same time use the new patterns discovered by the model to expand the rule base. This bidirectional learning mechanism enables the system to continuously optimize itself, continuously improve its ability to recognize new address patterns, and enhance the long-term availability and robustness of the system.

[0020] 3. This invention incorporates a character-level anomaly detection and correction step before the address parsing process. It actively identifies and corrects common writing errors through a deep learning model. This preprocessing step improves the quality of input data from the source, effectively avoiding subsequent splitting failures or errors caused by noisy data such as typos. It provides a solid foundation for the stability and accuracy of the entire address element splitting process and reduces the dependence on the standardization of the original data.

[0021] 4. This invention innovatively uses a graph structure to store and organize the rule base, explicitly representing complex relationships between address elements such as membership, adjacency, and aliases as nodes and edges of the graph; the matching process is transformed into graph traversal, enabling deeper semantic association and reasoning, thus more flexibly handling issues such as omission, abbreviation, and reversed order of elements in the address. Compared with traditional linear matching methods, it greatly enhances the intelligence level of rule paths and the ability to handle complex addresses.

[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating an automated method for splitting address data elements according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of an automated address data element splitting processing system according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1 One embodiment of the present invention proposes an automated method for splitting address data elements. It adopts a dual-path parallel processing of rules and models, and combines confidence judgment and dynamic negotiation mechanism, which can take into account the accuracy, robustness and adaptive learning ability of address splitting.

[0028] The method described in this embodiment specifically includes:

[0029] Obtain the address string, perform character-level anomaly detection on the address string, and generate a corrected address string;

[0030] Optionally, generating the corrected address string includes:

[0031] Obtain the address string and scan the address string based on a preset character-level attention network to generate an abnormal character identifier;

[0032] Based on the abnormal character identifier, the semantic environment before and after the abnormal character is obtained, and contextual semantic features are generated;

[0033] Based on the abnormal character identifier and the contextual semantic features, a set of candidate correction schemes is generated;

[0034] Evaluate and select a candidate correction scheme from the set of candidate correction schemes, apply the candidate correction scheme to the address string, and generate a corrected address string.

[0035] Specifically, this embodiment elaborates on the steps of generating the corrected address string. Its purpose is to preprocess the original address string to identify and correct potential character-level errors, thereby improving the accuracy of subsequent address splitting. The process first obtains the address string to be processed, and then uses a pre-defined character-level attention network to scan the string character by character. This character-level attention network is a deep learning model trained on a large number of address samples, capable of judging the rationality of each character's appearance in the address context based on its context. The network outputs an anomaly score for each character in the address string. When the anomaly score of a character exceeds a preset threshold, the system generates an anomaly character identifier, which records the location information of the anomaly character. After obtaining the anomaly character identifier, the system locates the specific anomaly character based on the identifier and extracts its surrounding semantic environment, i.e., the several characters around the character as context. This character sequence containing context information is converted into a numerical vector, i.e., contextual semantic features. This feature vector contains the local semantic information of the location of the anomaly character. Next, the system initiates the correction scheme generation process based on the anomaly character identifier and the generated contextual semantic features. This process can be a generative model-based module or a retrieval engine combining pinyin and character shape similarity. Based on contextual semantic features, it recommends one or more possible correct characters for abnormal characters, thus forming a set of candidate correction schemes. Finally, the system needs to select the optimal scheme from the set. To do this, the system evaluates each candidate correction scheme in the set. The evaluation process involves generating a temporary corrected address string for each candidate scheme and calculating its overall confidence score. This score can be calculated using an evaluation function, for example:

[0036] ,

[0037] in, It is the overall confidence score of the candidate correction scheme; It is a temporary address string generated after applying the candidate scheme; It is calculated using a general pre-trained language model. The language fluency probability is used to assess whether it conforms to natural language habits; It is calculated using a domain-specific address model trained on massive amounts of standard address data. The probability of address compliance is used to assess its suitability as an address; and These are preset weighting coefficients used to balance the importance of common language conventions and address domain expertise. The system selects the candidate correction scheme that results in the highest H score, applies it to the original address string, replaces the original abnormal characters, and finally outputs a high-quality corrected address string.

[0038] The corrected address string is input in parallel to a preset rule path and a preset model path, wherein the rule path uses a preset rule base for address feature matching to generate rule splitting results, and the model path uses a preset sequence labeling model for address feature recognition to generate model splitting results.

[0039] Optionally, the generation rule splitting results include:

[0040] Establish a rule base using a pre-defined administrative division tree and a pre-defined street standard library;

[0041] The corrected address string is matched with the rule base to extract the matching address elements;

[0042] Based on the matched address elements, generate rule splitting results.

[0043] Specifically, this process is the rule path in the dual-path parallel processing of address splitting. Its core lies in using explicit and interpretable rules to parse the modified address string and generate a baseline splitting result. First, a comprehensive and structured rule base needs to be built, which is the foundation for pattern matching. This step involves the integration of two key data sources: a pre-defined administrative division tree and a pre-defined street standard library. The administrative division tree is a hierarchical data structure that organizes administrative units at all levels into a tree structure based on published administrative division codes, accurately reflecting their hierarchical relationships. The street standard library is a detailed list containing the standard spellings of street, road, alley, and hutong names within each administrative region. Combining these two forms an address knowledge graph with both macro-level hierarchical structure and micro-level specific information—this is the rule base. After the rule base is built, the system can perform pattern matching on the input modified address string. This operation uses a greedy matching strategy from front to back, namely the maximum forward matching method, combined with the hierarchical constraints of the administrative division tree. The system scans from the beginning of the string, attempting to match it with the provincial-level administrative division names in the rule base. Once a match is successful, such as "City B," the system continues scanning the string under that administrative division node to match its subordinate districts, such as "District B1." This hierarchical constraint-based matching method effectively avoids ambiguity caused by duplicate place names. After completing the matching of higher-level administrative divisions, the system continues to use a street standard library to match street and road information, and uses predefined regular expressions or keyword rules to extract more granular address elements such as house numbers and community names. To improve the robustness of the matching, the system also calculates a matching confidence score during the matching process to determine the optimal matching segment. This score can be calculated as follows:

[0044] ,

[0045] in, The overall confidence score representing a specific matching candidate; The length of the currently matched address element string reflects the level of detail in the match. The similarity score between the matched string and the standard names in the rule base can be calculated using algorithms such as edit distance to evaluate the accuracy of the match. and These are preset weighting coefficients used to adjust the importance of length and similarity in the overall evaluation. The system will select and use... The system selects the match with the highest value. Finally, it organizes all address elements successfully extracted through the pattern matching process according to a predetermined data structure. For example, it fills in information such as province, city, district, street, and house number into the corresponding fields, thereby generating a structured rule-based splitting result. This result completely preserves all address information identified through deterministic rules.

[0046] Optionally, the generative model splitting results include:

[0047] Establish a sequence labeling model for identifying address features;

[0048] The corrected address string is input into the sequence labeling model to generate a sequence of label sequences;

[0049] Parse the sequence labeling to determine the address feature type and boundaries;

[0050] Generate model splitting results based on address feature type and boundaries.

[0051] Specifically, this process, as the model path in the dual-path parallel processing of address splitting, focuses on using data-driven deep learning methods to identify address elements, thus overcoming the limitations of rule-based methods. The first step is to build a sequence labeling model for identifying address elements. Sequence labeling is a natural language processing task whose goal is to assign a predefined label to each unit in the input sequence, and in this case, each character in the address string. A composite model structure combining a bidirectional long short-term memory network and a conditional random field layer, namely BiLSTM-CRF, is adopted. This model excels at capturing long-distance dependencies and contextual information in sequence data. The model building process requires supervised training using a large-scale labeled address dataset. Each character in each address in the dataset is assigned a label, for example, using the BIO labeling system, where B represents the start of an address element, I represents the interior of an element, and O represents not belonging to any element. After the model is built and trained, the second step involves inputting the corrected address string generated in the previous stage into the trained sequence labeling model. The input string is first converted into a series of low-dimensional dense character vectors, i.e., character embeddings. Subsequently, these vector sequences are fed into a BiLSTM layer, which processes the sequences in both forward and backward directions to generate a hidden state representation for each character that incorporates its left and right context information. Finally, the CRF layer receives the output from the BiLSTM layer and uses the transition probabilities between labels learned to compute the globally optimal label sequence. This process aims to find a probabilistically most likely label sequence, whose likelihood function can be expressed as:

[0052] ,

[0053] in, It is the input sequence of corrected address strings. It is the label sequence predicted by the model. Represents all possible label sequences. It is a model for a specific label sequence The score consists of two parts: emission score and transition score. The emission score, derived from the BiLSTM layer, represents the probability that a character will be assigned a specific label in a given context; the transition score, learned from the CRF layer, represents the plausibility of transitioning from one label to the next. The model's goal is to find the probability that... Maximize the label sequence The third step is to parse the tag sequence of the sequence label. The system iterates through the tag sequence output by the model, determining the type and boundaries of each address element according to the rules of the BIO labeling system. For example, a B-Province tag marks the beginning of a province element, and the following I-Province tag belongs to that province element, until a new B or O tag appears. In this way, the system can accurately segment address fragments with independent semantics from a continuous character stream and assign them categories, such as province, city, district, street, house number, etc. Finally, based on the parsed address element types and boundaries, the system integrates this information to generate a structured model splitting result. This result is organized in the form of key-value pairs, where the key is the address element type and the value is the corresponding text content extracted from the corrected address string.

[0054] A first confidence level is calculated based on the rule splitting results, and a second confidence level is calculated based on the model splitting results;

[0055] Determine whether both the first confidence level and the second confidence level are higher than a preset confidence level threshold, and whether the rule splitting result is consistent with the model splitting result;

[0056] If so, the result of the rule splitting is determined as the final splitting result;

[0057] Specifically, a dual-path verification mechanism is used to assess the confidence and verify the consistency of the initial splitting results generated by the rule path and model path, thereby determining the final splitting result. After parallel processing, the system first calculates the first confidence score based on the rule splitting results. This confidence score is a quantitative indicator designed to evaluate the comprehensiveness and accuracy of rule matching. Its calculation method can be expressed by the following formula:

[0058] ,

[0059] in, Represents the first confidence level; Coverage is calculated by dividing the total character length of all address elements covered in the rule split by the total length of the corrected address string. It is used to measure the degree to which the rule interprets the input address. It is the matching quality score, which is the weighted average of the preset quality scores of all successfully matched address elements. Elements that are precisely matched directly from standard libraries such as administrative division trees are given high scores, while elements extracted through fuzzy matching or regular expressions receive relatively low scores. and These are preset weighting coefficients used to balance the importance of coverage and matching quality. All symbols in this formula are dimensionless values. It is obtained by calculating the length ratio. The second confidence score is calculated using a preset quality score, ensuring the logical compliance of the operation. Subsequently, the system calculates the second confidence score based on the model splitting results. For sequence labeling models, the second confidence score... Typically derived from the normalized probability score of the Conditional Random Field (CRF) layer output, this score reflects the likelihood of the entire predicted label sequence given the input modified address string. It directly reflects the model's overall confidence in its prediction results; a higher value indicates that the model believes the current splitting scheme is the globally optimal solution. After obtaining the first and second confidence scores, the system enters the judgment process. This process executes a double-condition AND gate judgment, checking whether both the first and second confidence scores are higher than a preset confidence threshold, and simultaneously checking whether the rule-based splitting result is completely consistent with the model's splitting result. The confidence threshold here is an empirical value derived from statistical analysis of a large amount of historical data, designed to filter out low-quality splitting results. The consistency judgment is a rigorous structured comparison process. The system compares the address element types in the two splitting results one by one, as well as the text content corresponding to each type and its start and end positions in the original string. Only when all address elements are completely identical in type, content, and position are the two considered consistent. If both of the above conditions are met, it indicates that the two distinct methods—rule-based and model-based approaches—have a highly consistent understanding of the address string and both possess high confidence. This is a very strong signal, proving the correctness and reliability of the splitting result. In this case, the system determines the rule-based splitting result as the final splitting result because the rule-based method has better interpretability and stability.

[0060] If not, then the rule-based splitting result and the model splitting result are used to perform dynamic negotiation to generate a negotiated splitting result, and the negotiated splitting result is determined as the final splitting result.

[0061] Optionally, the generation of the negotiated split result includes:

[0062] The result of the rule splitting is used as a reinforcement learning signal to adjust the sequence labeling model;

[0063] Key semantic features are extracted from the model splitting results, and the rule base is expanded using the key semantic features;

[0064] The corrected address string is reprocessed using the adjusted sequence labeling model and the expanded rule base to generate a negotiated splitting result.

[0065] Specifically, this process is initiated when the rule-based splitting result differs from the model-based splitting result or when the confidence level is insufficient. Its core is a dynamic, bidirectional adaptive negotiation mechanism designed to integrate the advantages of both paths to generate a better splitting scheme. The process includes two parallel adjustment steps and a final reprocessing step. First, the system uses the rule-based splitting result as a reinforcement learning signal to fine-tune the sequence labeling model online. In this context, the reinforcement learning signal does not involve the traditional state-action space but is simplified to a reward mechanism. The system treats the rule-based splitting result as a highly reliable judgment. When the model-based splitting result differs from the rule-based splitting result, the system calculates the difference between the two and transforms it into a penalty or reward. For example, a reward function can be defined as:

[0066] ,

[0067] in, It is the reward value obtained by the sequence labeling model; It is the set of entities extracted from the model splitting results; It is the set of entities extracted from the result of rule splitting; This is a function that calculates the similarity between two entity sets, such as a metric based on the F1 score. The value ranges from 0 to 1, with 1 indicating complete similarity and 0 indicating complete dissimilarity. This is the reward value. Used to guide model parameter updates, the model fine-tunes its internal parameters through optimization algorithms such as gradient ascent, making its prediction of the current corrected address string closer to the rule-based splitting result. This maintains generalization ability while incorporating deterministic knowledge from the rules. Simultaneously, the system extracts key semantic features missing from the rule base from the model's splitting results and uses these features to dynamically expand the rule base. After the sequence labeling model is adjusted and the rule base is expanded, the system uses this dual-optimization configuration to perform a new, collaborative processing on the original corrected address string. That is, the adjusted sequence labeling model and the expanded rule base parse the string in parallel again. Since both have absorbed some information from the other, the likelihood of the new model splitting result and the rule-based splitting result being consistent is greatly increased. The system identifies the highly consistent or finalized splitting scheme generated after this collaborative processing as the negotiated splitting result and designates it as the final splitting result.

[0068] Optionally, expanding the rule base using the key semantic features includes:

[0069] A new address pattern is obtained by parsing the key semantic features;

[0070] The new address pattern is added to the fuzzy matching rule set of the rule base to generate an expanded rule base;

[0071] Record the error patterns reflected by the semantic features.

[0072] Specifically, when the sequence labeling model identifies address elements that the rule path fails to match, these elements are considered key semantic features. The first step is to parse new address patterns from these key semantic features. This parsing is not a simple text copying process, but rather an abstraction and generalization process. The system analyzes the textual content of the key semantic feature, the label type assigned to it by the model, and its context within the original corrected address string. For example, if the model identifies "Future City Phase III" as a community name, but this record is not in the rule base, the system will parse it as a new address entity and may extract a new naming pattern based on its structure, such as name plus phase number. Next, the system adds the parsed new address pattern to the fuzzy matching rule set in the rule base. The rule base is not a static knowledge base that only supports exact matching; it contains a fuzzy matching rule set specifically designed to handle non-standard, colloquial expressions. This rule set allows for a certain degree of textual variation, for example, through thesaurus, near-synonym dictionaries, or character edit distance-based algorithms. Newly parsed address patterns, such as "Future City Phase III," are added to this set as new terms or rules, thus completing the real-time update of the rule base. For example, the system can use "Future City" as the core term, "Phase III" as a suffix, and classify the whole thing under the category of community address elements. In this way, when encountering addresses containing "Future City Phase III" or its variations, such as "Future City Phase 3," in the future, the expanded rule base can successfully match them, thereby generating the expanded rule base. Finally, this method also records error patterns reflected by key semantic features. This is not merely a one-time supplement to unmatched content, but a systematic diagnostic and learning process. Whenever the model discovers a knowledge blind spot in the rule base, the system generates an error pattern log. This log records in detail the original address string, the failure point of the rule path, the key semantic features identified by the model, and the context of those features. These logs are accumulated to form an error knowledge base. By performing offline analysis on this knowledge base, systemic defects in the rule base can be discovered, such as insensitivity to naming rules for new buildings in a certain area, or insufficient ability to recognize certain colloquial abbreviations. These analytical results will guide deeper structural optimization of the rule base, rather than simply adding entries.

[0073] Optionally, the method further includes:

[0074] Based on the confidence level of the final splitting result, candidate confidence level cases are selected;

[0075] The candidate confidence cases are validated to generate labeled samples;

[0076] A training dataset is established using the labeled samples, and the sequence labeling model and the rule base are periodically updated using the training dataset.

[0077] Specifically, this embodiment details the self-learning and iterative optimization closed loop included in the automated address data element splitting method. This mechanism ensures continuous improvement in system performance and adaptability to new data. First, after each address split, the system obtains a final splitting result and its corresponding confidence level. Instead of processing all results, the system intelligently selects the most valuable learning cases, i.e., candidate confidence level cases, based on the confidence level. The selection strategy typically focuses on a preset intermediate confidence level range. Cases with excessively high confidence levels are usually patterns that the system has mastered, while cases with excessively low confidence levels may contain too much noise or be completely unresolvable. In contrast, cases within the intermediate confidence level range represent a fuzzy area where the system has uncertainty but still has the potential to correctly resolve them. After selecting candidate confidence level cases, the system initiates a verification process. The purpose of this process is to provide an accurate and correct answer for these uncertain cases, thereby generating labeled samples that can be used for model training. Then, the continuously generated labeled samples are accumulated to construct or expand a dedicated training dataset. This dataset is a core resource for system evolution, containing a large number of validated, and particularly challenging, address splitting instances. Finally, the system periodically, such as daily or weekly, uses this ever-growing and enriched training dataset to update its core components: the sequence labeling model and the rule base. For the sequence labeling model, the system retrains or incrementally fine-tunes it using the entire training dataset, enabling it to learn new address representations and correct previously inaccurate predictions. For the rule base, the system mines new, frequently occurring address entities, such as new streets, new neighborhood names, and structural patterns from the training dataset and automatically adds them to the rule base, thereby expanding the coverage of the rules.

[0078] Optionally, the step of validating the candidate confidence cases and generating labeled samples includes:

[0079] The candidate confidence cases are verified based on preset verification conditions.

[0080] If the verification is successful, the first verification split result will be generated;

[0081] If the verification fails, the candidate confidence cases will be submitted to the manual verification process to generate a second verification split result.

[0082] Combine the address string of the candidate confidence case with the first verification split result or the second verification split result, and add version information to generate labeled samples.

[0083] Specifically, the process of validating candidate confidence cases to generate labeled samples begins with validating the selected candidate confidence cases. The first step is to attempt automated validation. The system verifies the splitting results of candidate cases based on a series of preset validation conditions. These validation conditions may include, for example, cross-referencing the split administrative divisions with authoritative geographic information systems to verify their existence and correct hierarchical relationships; querying the split detailed addresses using an external postal code database to confirm their validity; or checking whether the combination of the split address elements conforms to a known, high-confidence address paradigm. If the splitting result of a candidate case passes all preset validation conditions, i.e., the validation is successful, the system considers the splitting result accurate and marks it as the first validated splitting result. If a candidate confidence case fails any preset validation condition, i.e., the validation fails, it indicates that the complexity or ambiguity of the case exceeds the capabilities of automated validation. In this case, the system automatically submits the candidate confidence case, along with its original address string and the system-generated preliminary splitting result, to an integrated manual validation process. In this process, data annotation experts manually review and correct the case, providing an accurate address element splitting scheme based on their expertise and judgment. This manually confirmed and corrected result is defined as the second verification splitting result. After obtaining either the first or second verification splitting result, the system executes the final step: constructing a standard annotation sample. This process pairs the original address string in the candidate confidence case with the verified splitting result obtained in the previous step, whether it's the first or second verification splitting result. Furthermore, to ensure data traceability and the effectiveness of version management, the system adds version information to this newly generated pairing data. This version information may include the model version number used to generate the case, the rule base version number, and the timestamp of annotation completion. Ultimately, this structured data unit, containing the original address string, the authoritative splitting result, and complete version information, constitutes a high-quality annotation sample that can be directly used for model training and rule base updates.

[0084] Optionally, the method further includes:

[0085] The rule base is stored using a graph structure, where nodes represent address elements and edges represent the relationships between the address elements.

[0086] The matching process of the rule path is a graph traversal process, which uses the relationships represented by edges to perform fuzzy matching and semantic association.

[0087] Specifically, when constructing the rule base, the system no longer uses simple lists or strict tree structures, but rather a more flexible graph structure. In this graph, each node represents an independent address element, such as "Beijing," "Haidian District," "Zhongguancun Street," and "Yinke Building." These elements can be administrative divisions, streets, residential areas, buildings, or points of interest. The edges in the graph represent the diverse relationships between these address elements. These relationships are not limited to hierarchical "belonging" but can also be geographical "adjacency," functional "aliases," or semantic "location." For example, the "Haidian District" node has a "belonging" edge pointing to the "Beijing" node, while the "Yinke Building" node has a "location" edge pointing to the "Zhongguancun Street" node. When the rule path receives the corrected address string and begins matching, the process is no longer a linear string scan but a graph traversal process within the address knowledge graph. The system first attempts to identify some high-confidence entities from the address string as "anchor points" or starting nodes for the graph traversal. For example, for the address "Beijing Haidian Zhongguancun Yinke", the system might first identify potential nodes such as "Beijing", "Haidian", "Zhongguancun", and "Yinke" through keyword matching. Then, based on these anchor points, the system explores the optimal path connecting them in the graph. This traversal process fully utilizes the relationships represented by edges for fuzzy matching and semantic association. For instance, when the system locates the node "Zhongguancun", it doesn't process the subsequent "Yinke" in isolation, but explores along the edges radiating outwards from the "Zhongguancun" node. If there exists a "located" edge in the graph pointing from the "Zhongguancun Street" node to the "Yinke Building" node, and there is a high text similarity or alias relationship between "Zhongguancun" and "Zhongguancun Street", and between "Yinke" and "Yinke Building", the system will infer a reasonable semantic path. This relationship-based traversal is essentially a form of semantic association; it can intelligently infer and complete other relevant elements based on the context of identified elements, effectively solving problems such as abbreviations, aliases, or element omissions in addresses. Meanwhile, fuzzy matching is also reflected in node matching and path selection. When traversing, the system will calculate a comprehensive score for each path, taking into account factors such as text matching degree and relationship strength, and finally select the path with the highest score as the best matching result. All nodes on this path constitute the final rule splitting result.

[0088] Based on the same inventive concept, the present invention also provides an automated processing system for splitting address data elements, the system comprising:

[0089] A character correction module is used to obtain an address string, perform character-level anomaly detection on the address string, and generate a corrected address string;

[0090] A character splitting module, which is used to parallelly input the corrected address string into a preset rule path and a preset model path. The rule path uses a preset rule library for address element matching to generate a rule splitting result, and the model path uses a preset sequence labeling model for address element recognition to generate a model splitting result;

[0091] A confidence calculation module, which is used to calculate a first confidence based on the rule splitting result and calculate a second confidence based on the model splitting result;

[0092] A splitting result determination module, which is used to determine whether both the first confidence and the second confidence are higher than a preset confidence threshold, and the rule splitting result is consistent with the model splitting result; if so, the rule splitting result is determined as the final splitting result; if not, the rule splitting result and the model splitting result are used for dynamic negotiation to generate a negotiated splitting result, and the negotiated splitting result is determined as the final splitting result.

[0093] To verify the feasibility of the present invention in implementation, the present invention is applied to the address data processing center of a certain logistics company. This logistics company needs to process millions of address data from online orders, handwritten waybills and third-party systems every day. Among them, there are a large number of addresses with inconsistent formats, misspelled words, missing information and colloquial descriptions. The traditional address cleaning and splitting methods have a high error rate, resulting in sorting errors and delivery delays, seriously affecting the operation efficiency. The company hopes to adopt the method of the present invention to automate the upgrade of its address data processing process to improve the accuracy and robustness of address element splitting.

[0094] In this embodiment, the company processes and verifies 10 million pieces of original address data collected in the fourth quarter of 2024. The system first performs character-level anomaly detection on the obtained address string. For example, for an input address "Haidin District, Zhongguancun Street No. 27, Beijing", the system scans through a preset character-level attention network and identifies that the character "ding" is an abnormal character in its context. The system then obtains its context semantic features and generates a candidate correction solution set including "dian". By evaluating the After calculating the comprehensive confidence of each solution, the system selects the optimal solution and corrects the address to "Haidian District, Zhongguancun Street No. 27, Beijing".

[0095] The corrected address string is parallelly sent to the rule path and the model path. For a standard address "No. 1, Keji South Road, Nanshan District, Shenzhen, Guangdong Province", the rule path uses a preset administrative division tree and street standard library, through the maximum forward matching method, and combines the matching confidence score The system successfully matched elements such as "Guangdong Province," "Shenzhen City," "Nanshan District," and "Keji South Road," generating a highly reliable rule splitting result. Simultaneously, the model path was processed using a pre-trained BiLSTM-CRF sequence labeling model, which also generated a completely consistent rule splitting result. Since the results from the two paths were identical and both the first and second confidence scores were higher than the preset threshold of 0.95, the system directly determined the rule splitting result as the final splitting result.

[0096] For a non-standard address, "Beijing Chaoyang Wangjing SOHO Tower 2A", the rule path fails to match or is incomplete because the rule base lacks the entity "Wangjing SOHO". However, the model path, thanks to its generalization ability, successfully identifies "Wangjing SOHO" as a landmark / building name and "Tower 2" and "A" as building numbers. At this point, the two path results are inconsistent, and the system initiates a dynamic negotiation mechanism. First, the rule splitting result is used as a reward signal. The sequence labeling model was fine-tuned to enhance its recognition of administrative divisions such as "Beijing" and "Chaoyang". Secondly, the key semantic feature "Wangjing SOHO" was extracted from the model's splitting results, parsed into a new address pattern, and added to the fuzzy matching rule set in the rule base. In this process, the graph structure storage method of the rule base played a crucial role. The system treated "Wangjing SOHO" as a new node and established its "location" relationship with the "Wangjing" street node, achieving semantic association through graph traversal. After adjustment and expansion, the system reprocessed the address, ultimately generating an accurate negotiated splitting result.

[0097] This invention also includes a self-learning and iterative optimization closed loop. For cases with confidence levels in the middle range, such as "convenience store opposite the east gate of Renmin University," the system will filter them as candidate confidence cases. After the automated verification process fails to cross-reference through a geographic information system, the case is submitted to the manual verification process, generating a labeled sample marked "No. 59 Zhongguancun Street, Haidian District, Beijing, opposite the east gate of Renmin University of China," with version information attached. These samples are accumulated in the training dataset to periodically update the sequence labeling model and rule base, enabling the system to continuously learn new address representations and achieve self-iterative performance improvement.

[0098] Data comparison shows that the automated processing method of this invention has significant advantages in both the accuracy and robustness of address splitting. The overall accuracy of the traditional system is about 85%, while the accuracy of the system of this invention has improved to 98.5% after one quarter of operation and self-learning. Especially when dealing with non-standard and erroneous addresses, the success rate of this system is nearly 40 percentage points higher than that of the traditional method. The dynamic negotiation mechanism successfully resolves about 5,000 difficult addresses on average per day, while the self-learning closed-loop mechanism can generate about 2,000 high-quality labeled samples per week, driving continuous optimization of the model and rule base.

[0099] Table 1 Comparison of Address Splitting Processing Accuracy

[0100] Processing month Total number of addresses (ten thousand) Accuracy rate of traditional system (%) Accuracy rate of the system of the present application (%) October 2024 320 85.2 97.1 November 2024 350 84.9 98.3 December 2024 330 85.5 98.8

[0101] Table 2 Comparison of processing success rates for different address types (Q4 2024)

[0102] Address type Sample size (ten thousand) Success rate of traditional system (%) Success rate of the system of the present application (%) Standard structure address 680 99.1 99.8 Address containing character errors 150 62.5 97.2 Non-standard / colloquial address 170 51.3 94.6

[0103] Table 3 Evaluation of the effectiveness of dynamic negotiation and self-learning mechanisms

[0104] Evaluation period Number of difficult address cases Success rate of negotiation mechanism (%) Model accuracy improvement (after period) Rule base coverage improvement (after period) 1st week 35,102 92.1 +0.2% +0.5% 2nd week 34,550 93.5 +0.18% +0.45% 3rd week 33,980 94.2 +0.15% +0.4% 4th week 32,760 95.0 +0.12% +0.38%

[0105] Tables 1-3 above record the actual application data of the present invention in the address processing center of a logistics company, and demonstrate in detail the system’s outstanding performance in terms of processing accuracy, adaptability to complex addresses, and self-optimization capabilities.

[0106] As shown in Table 1, compared to traditional systems, the accuracy of the system of this invention remains stable at over 98%, demonstrating a significant performance advantage. This is attributed to the combined effect of character correction, dual-path collaborative processing, and dynamic negotiation mechanisms, which effectively ensure the accuracy of the processing results.

[0107] The data in Table 2 further demonstrates the robustness of this invention in handling diverse address types. In particular, for addresses containing character errors and non-standard colloquialisms, the success rates of this invention reached 97.2% and 94.6%, respectively, far exceeding traditional systems. This fully illustrates the crucial value of the character-level anomaly detection module and the data-driven model path in processing complex real-world data.

[0108] Table 3 visually reflects the effectiveness of the dynamic negotiation and self-learning closed loop. Over time, the negotiation mechanism gradually improves its resolution rate for difficult addresses. Simultaneously, through continuous learning, the model accuracy and rule base coverage also steadily increase. This demonstrates that the present invention constructs an intelligent system capable of self-evolution and continuous optimization, able to constantly adapt to new address patterns and maintain high performance over the long term.

[0109] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0110] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for automatically splitting address data elements, characterized in that, The method includes: Obtain the address string, perform character-level anomaly detection on the address string, and generate a corrected address string; The corrected address string is input in parallel to a preset rule path and a preset model path, wherein the rule path uses a preset rule base for address feature matching to generate rule splitting results, and the model path uses a preset sequence labeling model for address feature recognition to generate model splitting results. A first confidence level is calculated based on the rule splitting results, and a second confidence level is calculated based on the model splitting results; Determine whether both the first confidence level and the second confidence level are higher than a preset confidence level threshold, and whether the rule splitting result is consistent with the model splitting result; If so, the result of the rule splitting is determined as the final splitting result; If not, then the rule-based splitting result and the model-based splitting result are used to perform dynamic negotiation to generate a negotiated splitting result, and the negotiated splitting result is determined as the final splitting result; The process of generating negotiated splitting results includes: using the rule splitting results as reinforcement learning signals to adjust the sequence labeling model; extracting key semantic features from the model splitting results and expanding the rule base using the key semantic features; and reprocessing the corrected address string using the adjusted sequence labeling model and the expanded rule base to generate negotiated splitting results.

2. The method for automatically splitting address data elements according to claim 1, characterized in that, The generated corrected address string includes: Obtain the address string and scan the address string based on a preset character-level attention network to generate an abnormal character identifier; Based on the abnormal character identifier, the semantic environment before and after the abnormal character is obtained, and contextual semantic features are generated. Based on the abnormal character identifier and the contextual semantic features, a set of candidate correction schemes is generated; Evaluate and select a candidate correction scheme from the set of candidate correction schemes, apply the candidate correction scheme to the address string, and generate a corrected address string.

3. The method for automatically splitting address data elements according to claim 2, characterized in that, The generation rule splitting results include: Establish a rule base using a pre-defined administrative division tree and a pre-defined street standard library; The corrected address string is matched with the rule base to extract the matching address elements; Based on the matched address elements, generate rule splitting results.

4. The method for automatically splitting address data elements according to claim 3, characterized in that, The generative model splitting results include: Establish a sequence labeling model for identifying address features; The corrected address string is input into the sequence labeling model to generate a sequence of label sequences; Parse the sequence labeling to determine the address feature type and boundaries; Generate model splitting results based on address feature type and boundaries.

5. The method for automatically splitting address data elements according to claim 1, characterized in that, The step of expanding the rule base using the key semantic features includes: A new address pattern is obtained by parsing the key semantic features; The new address pattern is added to the fuzzy matching rule set of the rule base to generate an expanded rule base; Record the error patterns reflected by the semantic features.

6. The method for automatically splitting address data elements according to claim 1, characterized in that, The method further includes: Based on the confidence level of the final splitting result, candidate confidence level cases are selected; The candidate confidence cases are validated to generate labeled samples; A training dataset is established using the labeled samples, and the sequence labeling model and the rule base are periodically updated using the training dataset.

7. The method for automatically splitting address data elements according to claim 6, characterized in that, The process of validating the candidate confidence cases and generating labeled samples includes: The candidate confidence cases are verified based on preset verification conditions. If the verification is successful, the first verification split result will be generated; If the verification fails, the candidate confidence cases will be submitted to the manual verification process to generate a second verification split result. Combine the address string of the candidate confidence case with the first verification split result or the second verification split result, and add version information to generate labeled samples.

8. The method for automatically splitting address data elements according to claim 1, characterized in that, The method further includes: The rule base is stored using a graph structure, where nodes represent address elements and edges represent the relationships between the address elements. The matching process of the rule path is a graph traversal process, which uses the relationships represented by edges to perform fuzzy matching and semantic association.

9. A processing system for automatically splitting address data elements, applied to the processing method for automatically splitting address data elements as described in any one of claims 1-8, characterized in that, The system includes: A character correction module is used to obtain an address string, perform character-level anomaly detection on the address string, and generate a corrected address string; The character splitting module is used to input the corrected address string in parallel to a preset rule path and a preset model path, wherein the rule path uses a preset rule library for address element matching to generate rule splitting results, and the model path uses a preset sequence labeling model for address element recognition to generate model splitting results. The confidence calculation module is used to calculate a first confidence level based on the rule splitting result and a second confidence level based on the model splitting result. The splitting result determination module is used to determine whether both the first confidence level and the second confidence level are higher than a preset confidence level threshold, and whether the rule splitting result is consistent with the model splitting result; if so, the rule splitting result is determined as the final splitting result; if not, the rule splitting result and the model splitting result are used to perform dynamic negotiation to generate a negotiated splitting result, and the negotiated splitting result is determined as the final splitting result. The process of generating negotiated splitting results includes: using the rule splitting results as reinforcement learning signals to adjust the sequence labeling model; extracting key semantic features from the model splitting results and expanding the rule base using the key semantic features; and reprocessing the corrected address string using the adjusted sequence labeling model and the expanded rule base to generate negotiated splitting results.

Images