Address resolution method and device, electronic equipment, storage medium and program product

Abstract

The embodiment of the invention provides an address resolution method and device, electronic equipment, a storage medium and a program product. The method comprises the following steps: acquiring an original address text, and segmenting the original address text to obtain a plurality of address fragments; performing address resolution processing on the plurality of address fragments to obtain at least one group of first candidate address elements; if the confidence score of at least one group of first candidate address elements meets a preset confidence score threshold, determining a target address element corresponding to the original address text according to the at least one group of first candidate address elements; and if the confidence scores of all the first candidate address elements do not meet a preset confidence score threshold, performing enhanced analysis processing based on the address fragment and the original address text to obtain a target address element corresponding to the original address text. The method is used for improving the accuracy of address resolution and enhancing the robustness of non-standard and noise addresses, so that the automation efficiency and reliability of business handling auditing are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an address resolution method, apparatus, electronic device, storage medium, and program product. Background Technology

[0002] In sectors such as finance, logistics, and government affairs, where standardized address information is crucial, address resolution for relevant merchants and subsequent business processing verification based on the resolution results are key steps in ensuring business security and data quality. For example, in merchant onboarding verification scenarios, it is necessary to verify whether the resolved address elements comply with national administrative division standards and whether they contain sensitive words or redundant information to prevent risks and ensure data availability.

[0003] Currently, related technologies mainly rely on rule-based automated systems to achieve address parsing. Traditional automated systems often use methods such as keyword matching, regular expressions, or static address database comparison to achieve basic parsing of address elements and noise filtering.

[0004] However, the relevant technologies struggle to effectively handle traditional characters, non-standard expressions, new place names, and maliciously filled content in address text, easily leading to parsing omissions. Furthermore, these technologies cannot accurately identify and remove semantically irrelevant noise, affecting parsing accuracy and reliability, and making them unsuitable for complex and ever-changing real-world business scenarios. Summary of the Invention

[0005] This application provides address resolution methods, apparatus, electronic devices, storage media, and program products to improve the accuracy of address resolution, enhance robustness to non-standard and noisy addresses, and thereby improve the automation efficiency and reliability of business processing and review.

[0006] In a first aspect, embodiments of this application provide an address resolution method, including:

[0007] Obtain the original address text and segment it to obtain multiple address fragments;

[0008] Multiple address segments are parsed to obtain at least one set of first candidate address elements; wherein each set of first candidate address elements includes multiple address labels and address segments corresponding to each address label;

[0009] If at least one set of first candidate address elements has a confidence score that meets the preset confidence score threshold, then the target address element corresponding to the original address text is determined based on at least one set of first candidate address elements.

[0010] If the confidence scores of all first candidate address elements do not meet the preset confidence score threshold, then enhanced parsing processing is performed based on the address fragment and the original address text to obtain the target address element corresponding to the original address text.

[0011] The target address element is used to review the original address text for business processing.

[0012] In one possible implementation, the original address text is segmented to obtain multiple address fragments, including:

[0013] Based on the preset word segmentation tool, the original address text is segmented to obtain a text sequence containing multiple words;

[0014] Based on a preset combination length threshold N, a sliding window enumeration is performed on the text sequence to generate multiple address fragments, where N is an integer greater than 1.

[0015] In one possible implementation, based on a preset combination length N, a sliding window enumeration is performed on the text sequence to generate multiple address fragments, including:

[0016] Using a combined length threshold N as the maximum size of the sliding window, the window slides sequentially starting from the beginning of the text sequence.

[0017] For any sliding operation, merge the M consecutive words covered by the current sliding window into a single address segment; where M is a variable integer satisfying 1 < M ≤ N;

[0018] Traverse the text sequence until the starting position of the sliding window covers the end of the text sequence to obtain multiple address fragments.

[0019] In one possible implementation, address resolution processing is performed on multiple address segments to obtain at least one set of first candidate address elements, including:

[0020] Retrieve a predefined set of address labels; where each address label in the set is defined according to address hierarchy rules;

[0021] For any address label in the address label set, determine at least one candidate address segment corresponding to the address label from multiple address segments;

[0022] Based on each address label in the address label set and its corresponding candidate address fragment, a directed graph is constructed. In the directed graph, each graph node represents an address label and its corresponding candidate address fragment. A directed edge indicates that the address fragments represented by two graph nodes are sequentially adjacent in the original address text, and their corresponding address labels are allowed to be adjacent in the predefined address label address structure rules.

[0023] Path search is performed based on the directed graph to obtain at least one set of first candidate address elements; wherein each set of first candidate address elements corresponds to a path from the starting address label graph node to the ending address label graph node.

[0024] In one possible implementation, determining at least one candidate address segment corresponding to an address label from a plurality of address segments includes:

[0025] Calculate the similarity between the address label and each address segment;

[0026] Based on the similarity scores, the top K address segments with the highest similarity scores are identified as candidate address segments corresponding to the address labels.

[0027] In one possible implementation, a directed graph is constructed based on each address label in the address label set and its corresponding candidate address fragment, including:

[0028] Each address label is combined with its corresponding candidate address fragment to form a graph node;

[0029] If the address fragment corresponding to the current graph node and the address fragment corresponding to the next graph node are sequentially adjacent in the original address text, and the address label corresponding to the current graph node and the address label corresponding to the next graph node are allowed to be adjacent in the predefined address label address structure rules, then a directed edge is established between the current graph node and the next graph node.

[0030] Construct a directed graph based on all created graph nodes and established directed edges.

[0031] In one possible implementation, path search is performed based on a directed graph to obtain at least one set of first candidate address elements, including:

[0032] Search for at least one path from the graph node corresponding to the predefined start address label to the graph node corresponding to the predefined end address label, and calculate the overall score for each path;

[0033] Based on the comprehensive score, at least one path is selected from all the searched paths to obtain at least one set of first candidate address elements.

[0034] In one possible implementation, a comprehensive score is calculated for each path, including:

[0035] Obtain the node weights of all graph nodes on the path, and sum the node weights to obtain the node weight sum; where the node weights represent the degree of matching between the address label and the candidate address fragment;

[0036] Obtain the transition weights of all directed edges on the path, and sum the transition weights to obtain the transition weight sum; where the transition weights are used to quantify the rationality and commonality of the address structure rules represented by the established edges connecting two different address labels in the directed graph.

[0037] The overall score is calculated based on the sum of node weights and the sum of transition weights.

[0038] In one possible implementation, the transition weights of all directed edges on the path are obtained, including:

[0039] Based on the edge identifiers of all directed edges on the path, query the predefined transition weight mapping table to obtain the corresponding transition weights;

[0040] The transfer weight mapping table is constructed based on at least one of the following methods:

[0041] Based on the address hierarchy specifications and commonality, different weights are assigned to addresses that conform to the predefined address label address structure rules.

[0042] Based on the statistical probability of conforming to predefined address label address structure rules in historical address data, calculate and assign corresponding transfer weights.

[0043] In one possible implementation, based on the comprehensive score, at least one path is selected from all the searched paths to obtain at least one set of first candidate address elements, including:

[0044] Based on the comprehensive score of each path, the top P paths with the highest comprehensive scores are selected as candidate paths;

[0045] For each candidate path, the address labels and candidate address fragments corresponding to the graph node sequence contained in the candidate path are determined as a set of first candidate address elements.

[0046] In one possible implementation, after obtaining at least one set of first candidate address elements, the method further includes:

[0047] Obtain a pre-built first prompt word template, which contains preset slots for receiving the original address text and multiple address fragments to guide the calculation of confidence scores;

[0048] Fill the original address text and multiple address fragments into the preset slots of the first prompt word template to obtain the confidence score calculation prompt words;

[0049] Based on the confidence score, the prompt words are calculated, and the large language model is called to process the confidence score to obtain the confidence score of the first candidate address element.

[0050] In one possible implementation, enhanced parsing is performed based on the address fragment and the original address text to obtain the target address elements corresponding to the original address text, including:

[0051] Obtain a pre-built second prompt word template, which contains preset slots for receiving the original address text and multiple address fragments to guide the generation of address elements;

[0052] Fill the original address text and multiple address fragments into the preset slots of the second prompt word template to obtain the address element generation prompt word;

[0053] Based on address elements, prompt words are generated, and a large language model is called to perform address parsing processing to obtain the target address elements.

[0054] In one possible implementation, a large language model is invoked for address resolution processing to obtain the target address elements, including:

[0055] Obtain the second candidate address element and its corresponding confidence score from the output of the large language model;

[0056] If the confidence score meets the preset confidence score threshold, then the second candidate address element is determined as the target address element;

[0057] If the confidence score meets the preset confidence score threshold, the target address element is determined based on the first candidate address element and the second candidate address element.

[0058] In one possible implementation, the method further includes the following steps before segmenting the original address text:

[0059] Standardize the format of the original address text;

[0060] Based on preset text verification rules, the text compliance of the address text after format standardization is verified.

[0061] If the address text passes the text compliance check, then proceed with the subsequent segmentation steps;

[0062] If the address text fails the text compliance check, it will be marked as an abnormal address and the processing flow will end.

[0063] In one possible implementation, after obtaining the target address elements corresponding to the original address text, the method further includes:

[0064] Based on preset business verification rules, the target address elements are verified for business compliance.

[0065] If the target address element passes the business compliance verification, then the original address text will be reviewed for business processing based on the target address element.

[0066] If the target address element fails the business compliance verification, a verification result containing risk markers will be generated and output.

[0067] In one possible implementation, the original address text is reviewed for business processing based on the target address elements, including:

[0068] The target address elements and the original address text are output to the preset upper-level business system to trigger the upper-level business system to conduct business processing review on the original address text based on the target address elements.

[0069] Secondly, embodiments of this application provide an address resolution apparatus, comprising:

[0070] The address fragment acquisition module is used to obtain the original address text and segment the original address text to obtain multiple address fragments;

[0071] The first candidate address element acquisition module is used to perform address parsing processing on multiple address segments to obtain at least one set of first candidate address elements; wherein, each set of first candidate address elements includes multiple address tags and address segments corresponding to each address tag;

[0072] The first target address element determination module is used to determine the target address element corresponding to the original address text based on at least one set of first candidate address elements if the confidence score of any set of first candidate address elements meets the preset confidence score threshold.

[0073] The second target address element determination module is used to perform enhanced parsing processing based on the address fragment and the original address text to obtain the target address element corresponding to the original address text if the confidence scores of at least one group of first candidate address elements do not meet the preset confidence score threshold.

[0074] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0075] The memory stores the instructions that the computer executes;

[0076] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0077] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0078] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0079] The address resolution method, apparatus, electronic device, storage medium, and program product provided in this application segment the original address text to obtain address fragments, and perform preliminary parsing based on the fragments to generate first candidate address elements. If the candidate confidence score meets the standard, the target address element is directly determined; otherwise, it is parsed again in combination with the original text. Thus, while ensuring the accuracy of the parsing, the combination of segmentation and overall parsing improves the success rate and reliability of parsing complex or non-standard addresses, thereby improving the efficiency and reliability of business processing and review. Attached Figure Description

[0080] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0081] Figure 1 An application scenario diagram of the data processing method provided in this application;

[0082] Figure 2 Flowchart of the address resolution method provided in this application Figure 1 ;

[0083] Figure 3 Flowchart of the address resolution method provided in this application Figure 2 ;

[0084] Figure 4 A schematic diagram of the address resolution device provided in this application;

[0085] Figure 5 A schematic diagram of the structure of the electronic device provided in this application.

[0086] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0087] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0088] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0089] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. The collection, use and processing of the relevant data shall comply with relevant laws, regulations and standards, and corresponding operation entry points shall be provided for users to choose to authorize or refuse.

[0090] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0091] First, let me explain the terms used in this application:

[0092] Address resolution refers to the process of converting the address of an entity in one system or format into an equivalent address of the same entity in another system or format.

[0093] In the context of Chinese geographic information processing, address resolution specifically refers to the process of automatically identifying, decomposing, and converting unstructured Chinese address text into structured geographic information elements. These elements typically include hierarchical information such as province, city, district / county, street, and house number, and can be further linked to spatial data such as latitude and longitude coordinates.

[0094] As mentioned in the background technology, in fields such as finance, logistics, and government affairs that require standardized address information, address parsing of relevant merchants and reviewing business transactions based on the parsing results are crucial steps to ensure business security and data quality. For example, in merchant onboarding review scenarios, it is necessary to verify whether the parsed address elements comply with national administrative division standards and whether they contain sensitive words or redundant information to prevent risks and ensure data availability.

[0095] Currently, related technologies mainly implement address parsing based on rule-matching automated systems. Specifically, automated systems mostly use methods such as keyword matching, regular expressions, or comparison with a static address library to achieve basic parsing of address elements and noise filtering. For example, an address parsing scheme based on dictionary matching and word segmentation is adopted.

[0096] Specifically, the scheme based on dictionary matching and word segmentation relies on a predefined geographical name dictionary (such as province, city, district, county, and street name libraries) and Chinese word segmentation tools to extract address elements through rule matching. The specific process includes: segmenting the address text, matching the standard geographical names in the dictionary to extract fields such as province, city, and district, and taking the remaining part as the street or house number. Its advantages are simple implementation and no need for training data, but there are the following problems: high dictionary dependence, if the geographical name library is incomplete or does not contain new geographical names, the parsing error rate increases; poor ability to handle abbreviations and ambiguities, unable to recognize abbreviated forms such as using "Yue" instead of "Guangdong Province", or insufficient discrimination of similar fields such as "development zone" and "district"; weak noise recognition ability, unable to effectively filter meaningless characters or malicious filling content (such as "test", "haha") in the address, and manual secondary processing is required.

[0097] In summary, related technologies are difficult to effectively handle complex writing, non-standard expressions, new geographical names, and malicious filling content in address texts, which easily leads to parsing omissions. In addition, related technologies are also unable to accurately identify and eliminate irrelevant noise at the semantic level, affecting parsing accuracy and reliability, and it is difficult to adapt to complex and ever-changing actual business scenarios.

[0098] To address the above technical problems, this application provides an address parsing method aimed at solving the above technical problems. Specifically, the original address text is segmented to obtain address fragments, and based on the fragments, a preliminary parsing is performed to generate the first candidate address elements; if the candidate confidence score meets the standard, the target address elements are directly determined, otherwise, combined with the original text, parsing is performed again. Thus, on the premise of ensuring parsing accuracy, through two parsing methods of combining segmentation and overall parsing, the parsing success rate and reliability of complex or non-standard addresses are improved, and the efficiency and reliability of business processing and approval are enhanced.

[0099] The address parsing method provided by this application is applicable to various scenarios that require structured extraction and verification of unstructured address texts.

[0100] For example, in the scenario of customer address verification in financial credit or insurance business, this method can quickly and accurately parse the address information filled in by users, extract elements such as province, city, street, and house number, and assist in completing the verification of address authenticity and risk control.

[0101] In addition, the above methods can also be applied to scenarios such as address standardization in logistics and distribution, household registration or real estate registration in government systems, and address resolution for e-commerce platforms, thereby improving the availability of address data and business processing efficiency.

[0102] In summary, any field that requires converting addresses described in natural language into a standardized structure to support subsequent business judgment or review falls under the application scenarios of the address resolution technology solution in this application.

[0103] The following is combined Figure 1 The application scenarios applicable to the embodiments of this application are described below. Figure 1 This is a diagram illustrating an application scenario of the data processing method provided in this application.

[0104] See Figure 1 Taking the merchant network access application scenario as an example, this scenario mainly involves user terminals and network access review systems. The user terminal is used for users to fill in and submit application materials containing unstructured raw address text. The network access review system includes a parsing module and an review module, which are used to parse and review the raw address text in the application materials, respectively.

[0105] Based on this, the address resolution technology solution provided in this application mainly includes the following steps:

[0106] 1. The user fills in the application materials on the user terminal and submits them. The application materials include the unstructured raw address text entered by the user.

[0107] 2. The network access review system receives the application materials submitted by the user and extracts the original address text from them through the parsing module.

[0108] 3. The parsing module segments the original address text to obtain multiple address fragments.

[0109] 4. The parsing module performs preliminary parsing on multiple address fragments to generate at least one set of first candidate address elements; each set of first candidate address elements includes multiple address tags and their corresponding address fragments, such as province, city, district, street, house number, etc.

[0110] 5. The parsing module determines whether the confidence score of the first candidate address element meets the preset threshold. If it does, the candidate elements in this group are directly identified as the target address elements.

[0111] 6. If the confidence scores of all first candidate address elements do not meet the threshold, the parsing module performs a second deep parsing based on the address fragment and the complete context of the original address text to obtain the structured target address elements.

[0112] 7. The review module executes subsequent business review processes based on structured target address elements, such as address authenticity verification and network access assessment.

[0113] Through the above steps, this application embodiment can improve the accuracy and reliability of address resolution by combining the original text context when the initial resolution confidence score is insufficient, thereby providing effective support for automated business review.

[0114] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0115] Figure 2 Flowchart of the address resolution method provided in this application Figure 1 This method can be executed by an address resolution device, which can be a server or an electronic device; the following explanation uses an electronic device as an example. The method in this embodiment can be implemented through software, hardware, or a combination of both. Figure 2 As shown, the method includes:

[0116] S201. Obtain the original address text and segment the original address text to obtain multiple address fragments.

[0117] In this embodiment, users can input the relevant information required for handling business through the business processing interface on their user terminal. It should be understood that the relevant information includes, but is not limited to, unstructured raw address text.

[0118] After a user submits information, the business system receives and stores the relevant information, including the original address text, through its interface module. Subsequently, the business system's parsing module extracts the original address text from the corresponding fields of the relevant information.

[0119] During the process of obtaining the original address text, the parsing module can simultaneously perform preprocessing operations on the text. For example, filtering invalid characters, removing redundant spaces and newlines, and correcting common typos based on preset rules or dictionaries, etc., to improve the standardization and parsingability of the text, thereby improving the processing efficiency of subsequent segmentation.

[0120] After obtaining the original address text, the parsing module can further segment the text. Optionally, segmentation can be performed based on at least one of a predefined address dictionary, a statistical word segmentation model, or a sequence labeling model to divide the continuous unstructured address text into multiple address segments with independent semantics.

[0121] For example, entity boundaries such as administrative divisions, roads, and house numbers contained in the original address text can be identified through address dictionary matching, or a trained word segmentation model can be used to perform sequence annotation on the original address text to determine the start and end positions of each address segment. Then, based on the identified entity boundaries or the annotated start and end positions, the original address text is segmented to obtain multiple address segments.

[0122] In the above method, by performing segmentation processing, the original address text can be parsed into a sequence of address fragments with appropriate granularity, which can improve the efficiency and reliability of subsequent address element parsing and processing.

[0123] S202. Perform address parsing processing on multiple address segments to obtain at least one set of first candidate address elements.

[0124] In this embodiment, the parsing module can use multiple address fragments as input sequences and process them through a pre-trained address feature recognition model. This address feature recognition model can be implemented based on a deep learning architecture, such as a Bidirectional Long Short-Term Memory-Conditional Random Field (BiLSTM-CRF) model, or a Bidirectional Encoder Representations from Transformers-Conditional Random Field (BERT-CRF) model; alternatively, it can employ a statistical model trained on a large corpus of address data, such as a Hidden Markov Model or a Maximum Entropy Model.

[0125] Specifically, the input sequence is fed into the model described above, and the model then predicts the label for each address segment in the input sequence and outputs its corresponding address element category label, such as provincial, municipal, district / county, street / township, road name, house number, etc.

[0126] During the label prediction process, due to the ambiguity and blurred boundaries of the original address text, multiple reasonable label assignment schemes may arise. For example, for the address fragment "Nanjing Road", the model may label it as "road name", or it may resolve it as a composite structure of "city level" and "road name" based on the context. In this case, the parsing module will retain these different labeling possibilities and generate corresponding address element combinations based on each labeling result, thereby forming multiple sets of first candidate address elements.

[0127] It should be understood that each set of first candidate address elements in the model prediction output is a complete structured address representation, that is, it contains elements at all levels from administrative division to detailed house number information, and each element includes its category label and its corresponding address fragment.

[0128] In addition, the parsing module will calculate a confidence score for each group of first candidate address elements; this score can be obtained by comprehensively evaluating factors such as the similarity between the address labels and address fragments output by the model, and the logical rationality of the address hierarchy rules.

[0129] By retaining multiple possible first candidate address elements, the ambiguity and boundary blurring of address text can be resolved, thereby improving the coverage and robustness of the parsing results in complex address scenarios.

[0130] S203. If the confidence score of any group of first candidate address elements meets the preset confidence score threshold, then the target address element corresponding to the original address text is determined based on at least one group of first candidate address elements.

[0131] In this embodiment, after generating multiple sets of first candidate address elements and their corresponding confidence scores, the parsing module can filter each first candidate address element according to a preset confidence score threshold to obtain the target address element.

[0132] Specifically, during the filtering process, the parsing module can compare the confidence score of each group of first candidate address elements with a preset confidence score threshold. If at least one group of first candidate address elements has a confidence score that reaches or exceeds the threshold, the parsing module will select the group with the highest confidence score from these candidate groups that meet the conditions, and determine it as the target address element that ultimately corresponds to the original address text.

[0133] Through the above steps, based on the structured probability and confidence assessment of the model output, the optimal target address element can be automatically and reliably determined from multiple reasonable candidate solutions, ensuring the accuracy of the analysis.

[0134] It should be understood that the aforementioned preset confidence score threshold can be set or adjusted according to the needs of the actual application scenario. For example, a higher threshold can be set in scenarios requiring high accuracy, while the threshold can be appropriately reduced in scenarios pursuing recall.

[0135] S204. If the confidence scores of at least one group of first candidate address elements do not meet the preset confidence score threshold, then enhanced parsing processing is performed based on the address fragment and the original address text to obtain the target address element corresponding to the original address text.

[0136] In this embodiment, when the confidence scores of all first candidate address elements fail to reach the preset threshold, it indicates that the parsing results based on the initial model have high uncertainty or may contain errors. At this time, the parsing module will initiate an enhanced parsing process to obtain parsing results with higher accuracy.

[0137] Specifically, during enhanced parsing, broader contextual information or external knowledge bases (such as standard address databases and geographic information system data) can be used to perform secondary verification and adjustment of the logical relationships and hierarchical rules between address fragments; one or more backup address parsing models based on different architectures or training data can be called for parallel or serial processing, and new first candidate address elements can be generated through result fusion or voting mechanisms; preset address parsing rules (such as administrative division affiliation and road-house number order rules) can also be applied to correct existing candidate results, or to adjust the boundaries and redistribute labels for suspected erroneous fragments.

[0138] It should be understood that the processing methods in the above-described enhanced parsing process are not fixed steps, but rather combinations of one or more strategies flexibly selected based on the actual parsing scenario and the specific reasons for insufficient confidence. In other words, the above processing methods can be implemented individually, or selectively combined or executed sequentially according to the specific uncertainties and error types in the parsing process, thereby improving the accuracy and robustness of the parsing results.

[0139] Through the enhanced parsing process described above, the parsing module regenerates one or more new first candidate address elements and recalculates their confidence scores. If any of the newly generated first candidate address elements meets the confidence threshold, it is identified as the target address element. If no candidate that meets the conditions can be obtained, a parsing failure signal can be output directly, and a manual review process can be triggered or a prompt requesting clarification can be returned to the user, depending on the circumstances.

[0140] Based on the obtained target address elements, this structured address information can be associated with the business processing review process. That is, the obtained target address elements can be directly input into the review module for business processing review. This not only improves the automation of address information processing, but also enhances the accuracy and efficiency of business review through structured address data.

[0141] In the above technical solution, address fragments are obtained by segmenting the original address text, and the first candidate address element is generated by preliminary parsing based on the fragments. If the candidate confidence score meets the standard, the target address element is directly determined. Otherwise, the original text is parsed again. Thus, under the premise of ensuring the accuracy of parsing, the success rate and reliability of parsing complex or non-standard addresses are improved by combining segmentation and overall parsing, thereby improving the efficiency and reliability of business processing and review.

[0142] See Figure 3 Next, the detailed execution process of the address resolution technology solution will be further explained in detail. It should be noted that the following description is merely an exemplary implementation of the technical solution in this application and does not constitute a limitation on the technical solution of this application.

[0143] In this embodiment, after obtaining the original address text, it can be first subjected to text compliance review and preprocessing to improve the accuracy and efficiency of subsequent address parsing.

[0144] In this embodiment, an optional implementation of text compliance review and preprocessing may include: performing format standardization processing on the original address text; performing text compliance verification on the format-standardized address text based on preset text verification rules; if the address text passes the text compliance verification, then performing subsequent segmentation steps; if the address text fails the text compliance verification, then marking it as an abnormal address and ending the processing flow.

[0145] Specifically, since the original address text input by the user may have a series of problems such as inconsistent formatting, abnormal character encoding, redundant information, or non-standard expressions, in order to improve the efficiency and accuracy of subsequent segmentation, the original address text can be formatted after it is obtained. Here, formatting can be understood as the process of converting the original address text into a standardized text form that conforms to the system's processing specifications. Optionally, formatting can include, but is not limited to, the following operations: converting full-width characters in the original address text to half-width characters; converting traditional Chinese characters in the text to simplified Chinese characters; filtering and deleting various abnormal symbols contained in the text; removing extra spaces at the beginning, end, and within the text; and removing redundant characters from the text.

[0146] Through the above processing, the format of the original address text can be unified, eliminating parsing obstacles caused by factors such as character format and symbol interference, and enabling the address text to meet the unified character encoding and format standards.

[0147] After format standardization, considering that address text may still have issues such as abnormal length or incompatible character encodings, a text compliance check can be performed on the processed address text. In this embodiment, text compliance check can be understood as a process of systematically verifying the basic format and content standardization of the address text. Optionally, this check includes at least one of the following: a length check to verify whether the character length of the address text is within a preset threshold range; and a character encoding check to verify whether the character encoding of the address text conforms to a preset standard.

[0148] In some scenarios, after verification, it can be determined whether the normalized address text meets the preset compliance conditions, that is, the conditions for executing the splitting process. If so, the address text meets the parsing requirements and the subsequent address splitting steps continue to be executed. If the normalized address text still fails the compliance verification, it is marked as an abnormal address and the current processing flow is terminated. At the same time, the corresponding abnormal verification result is fed back to the system.

[0149] The above text compliance verification and preprocessing process can effectively filter address text with abnormal format or non-compliant content, reduce noise interference in subsequent parsing processes, and thus improve the overall processing efficiency and accuracy of the address resolution system.

[0150] After preprocessing the original address text, the preprocessed address text can be used as input to perform a segmentation process. Optionally, one possible implementation of the segmentation process may include: performing word segmentation on the original address text based on a preset word segmentation tool to obtain a text sequence containing multiple words; and performing sliding window enumeration on the text sequence based on a preset combination length threshold N to generate multiple address fragments, where N is an integer greater than 1.

[0151] Specifically, the preprocessed address text can be segmented based on a preset word segmentation tool. For example, taking the original address text "No. XX, XX Road, XX New Town, Tianhe District, Guangzhou City, Guangdong Province" as an example, the text can be segmented into a text sequence containing multiple independent semantic units by matching with address-specific dictionaries and rules: ["Guangdong", "Province", "Guangzhou", "City", "Tianhe District", "XX New Town", "XX Road", "No. XX"].

[0152] In address parsing, the word sequence obtained from processing the original address text using word segmentation tools can be directly used as address fragments for subsequent address parsing. However, in some application scenarios, especially when the word segmentation strategy is fine-grained, relying solely on the segmentation results may lead to semantically fragmented address fragments. For example, long place names or complex address elements may be broken down into multiple independent words. If these decomposed results are directly used for subsequent address element recognition and parsing, it can easily cause semantic fragmentation, affecting the accurate extraction and understanding of complete address elements.

[0153] Considering the above, this embodiment further introduces a fragment combination processing method after obtaining the basic word segmentation results. That is, by selectively merging adjacent words, address fragments with variable length and more complete semantics are generated to compensate for the information fragmentation problem that may be caused by fine-grained word segmentation, ensuring that subsequent processing steps can be based on semantically coherent text units, thereby improving the overall accuracy and robustness of address resolution.

[0154] Specifically, when combining fragments, the combination process can be performed based on a preset combination length threshold. The preset combination length threshold N can be defined as the maximum number of words that the sliding window can cover. In this embodiment, the combination length threshold N is an integer greater than 1, such as 3, 5, or 10, and its specific value can be adjusted according to the requirements for address fragment length in the actual application scenario.

[0155] Based on this, for the text sequence to be processed, an enumeration operation is performed using a sliding window technique based on the combined length threshold defined above.

[0156] Specifically, the enumeration process of the sliding window can start from the beginning of the text sequence and sequentially extract consecutive word substrings of length N as address segments. After each address segment is obtained, the window slides forward one word position and continues to extract new substrings of length N until the window slides to the end of the text sequence, so that the end position of the last window is aligned with the end of the sequence.

[0157] This combination method can generate a set of fragments covering address semantics at different granularities, which not only retains the information of the original fine-grained word segmentation, but also supplements a more complete composite address unit, thereby improving the accuracy of subsequent address element parsing.

[0158] Alternatively, in the above implementation process, an optional implementation method for generating multiple address segments by enumerating the text sequence using a sliding window may include: using a combination length threshold N as the maximum size of the sliding window, and sliding sequentially from the beginning position of the text sequence; for any sliding, merging the M consecutive words covered by the current sliding window into one address segment; where M is a variable integer satisfying 1 < M ≤ N; traversing the text sequence until the beginning position of the sliding window covers the end of the text sequence, resulting in multiple address segments.

[0159] Specifically, a combination length threshold is preset. During sliding, an initial sliding window is placed starting from the beginning position of the given original address text sequence. The starting position of the window points to the first word, and the number of words M it covers is an integer that can dynamically change during sliding, but always satisfies the condition 1 < M ≤ N.

[0160] For each swipe operation, instead of capturing a fixed N words, the system merges M consecutive words covered by the current window, starting from the beginning position, to form an address segment. It should be understood that the specific value of M can be dynamically determined each time the operation is performed, based on preset semantic integrity rules, specific delimiters, or the actual word segmentation results.

[0161] After generating the current address segment, the sliding window moves backward. The movement step can be a word or determined according to other rules. Subsequently, based on the new window position, the number of consecutive words M (M≤N) currently covered is determined again, and the next address segment is generated by merging them.

[0162] Repeat the sliding and merging process described above, traversing the entire text sequence until the starting position of the sliding window moves to the last word of the text sequence, and completes the generation of the fragment at that position, thereby obtaining a set of multiple address fragments covering different ranges and lengths.

[0163] For example, continuing with the original address text "No. XX, XX Road, XX New Town, Tianhe District, Guangzhou City, Guangdong Province", after preliminary word segmentation, this sequence is parsed into the word sequence: ["Guangdong", "Province", "Guangzhou City", "Tianhe District", "XX New Town", "XX Road", "No. XX"]. Setting N=3 indicates that any subsequently generated address fragment consists of at most 4 consecutive address words.

[0164] For the above text sequence, the following address fragments can be generated by using a sliding window enumeration:

[0165] When the window length is 2, the first segment combination is obtained: ["Guangdong Province", "Guangzhou City", "Tianhe District, Guangzhou City", "XX New Town, Tianhe District", "XX Road, XX New Town", "No. XX, XX Road"];

[0166] When the window length is 3, the second segment combination is obtained: ["Guangzhou City, Guangdong Province", "Tianhe District, Guangzhou City, Guangdong Province", "XX New Town, Tianhe District, Guangzhou City", "XX Road, XX New Town, Tianhe District", "No. XX, XX Road, XX New Town"];

[0167] Based on this, by combining the fragments generated by the different window lengths mentioned above, multiple address fragments can be obtained.

[0168] This approach generates a rich set of address fragments, including both short, basic place name units and longer, composite address descriptions, providing a multi-granular textual foundation for subsequent address resolution. Compared to a fixed-length sliding window, this method can more flexibly adapt to the natural combinations of different levels and lengths in address representations.

[0169] Based on the above implementation method, multiple address fragments are obtained, and address parsing processing can be performed on the original address text based on these multiple address fragments. Optionally, one possible implementation method for address parsing processing may include: obtaining a predefined set of address tags; wherein each address tag in the address tag set is defined according to address hierarchy rules; for any address tag in the address tag set, determining at least one candidate address fragment corresponding to the address tag from multiple address fragments; constructing a directed graph based on each address tag in the address tag set and its corresponding candidate address fragment, wherein a graph node in the directed graph represents an address tag and its corresponding candidate address fragment; an edge indicates that the address fragments represented by two graph nodes are sequentially adjacent in the original address text, and their corresponding address tags are allowed to be adjacent in the predefined address tag address structure rules; performing path search based on the directed graph to obtain at least one set of first candidate address elements; wherein each set of first candidate address elements corresponds to a path from the starting address tag graph node to the ending address tag graph node.

[0170] Specifically, it retrieves a predefined set of address tags. It should be noted that the tags in this set are not arbitrarily set, but are accurately defined according to the address specifications of countries, regions, or industries, following a hierarchical rule from top to bottom. For example, they include province (prov), city (city), district (district), development zone (devzone), street / township (town_road), location name (poi), and house number (num).

[0171] It should be noted that, in addition to the tags mentioned above, the predefined tag set in this embodiment also includes a predefined "other" tag. This tag is used to accommodate meaningless or illegal distracting fields. For example, meaningless fields such as "haha" or "test" can be categorized into the "[other]" field, thereby enabling quick detection of illegal content in addresses during review and improving review efficiency and reliability.

[0172] For each tag in the above set of address tags, matching and filtering can be performed on multiple generated address fragments to find all address fragments that may be correctly identified by that tag, which will then be used as candidate address fragments corresponding to that address tag.

[0173] Optionally, candidate address fragments can be determined by combining dictionary matching, rule-based judgment, or model prediction. For example, for the tag "city," fragments such as "Guangzhou" might be selected; for the tag "road," fragments such as "XX Road" might be selected.

[0174] It should be noted that since an address fragment may simultaneously conform to the semantics of multiple tags (which is ambiguous), it may be listed as a candidate fragment for multiple different tags.

[0175] Based on this, a directed graph is constructed to formally represent the parse space, using all address labels and their corresponding candidate address fragments.

[0176] In constructing a directed graph, graph nodes are first built based on each address label and all candidate address segments corresponding to each address label. Each graph node represents an address label and its corresponding candidate address segment. For example, a graph node can be represented as (Address label: City, Address segment: Guangzhou City).

[0177] Once the graph nodes are created, if any two graph nodes meet a preset condition, a directed edge is constructed between them. A directed edge indicates that the address segments represented by the two graph nodes are sequentially adjacent in the original address text, and their corresponding address labels are allowed to be adjacent according to predefined address label structure rules.

[0178] Based on this, and using all the graph nodes and the directed edges between at least some of them, a complete directed graph is obtained.

[0179] After constructing the directed graph, a path search operation is performed. The goal of this search is to find all complete directed paths in the graph that start from the graph node representing the top level (e.g., province) and end at the graph node representing the bottom level (e.g., house number).

[0180] In this embodiment, each path sequentially connects a series of graph nodes, and the edges between adjacent graph nodes in the path satisfy the aforementioned dual constraints of text order adjacency and structural rule compliance. Thus, each complete path realizes the mapping from the original unstructured address text to structured address elements.

[0181] Based on the obtained paths, the parsing module can retain all the searched paths, and each path corresponds to a candidate structured parsing result of the original address text; it can also evaluate and sort each path based on preset scoring rules, and then determine the first candidate address element corresponding to the original address text according to preset filtering conditions (such as selecting the highest-scoring path).

[0182] In the above scheme, the address resolution task is transformed into a path search problem on a directed graph. By integrating the sequence information in the original text with the prior address hierarchy rules, the accuracy and overall robustness of the resolution process are improved.

[0183] In the implementation of the above scheme, an optional implementation method for determining the candidate address segments corresponding to the address tags may include: calculating the similarity between the address tags and each address segment; and based on each similarity, determining the top K address segments with the highest similarity as the candidate address segments corresponding to the address tags.

[0184] In this embodiment, a parallel similarity calculation method can be used to determine the corresponding candidate address fragment for each address tag. For example, for each address tag to be extracted, its similarity can be calculated simultaneously with each address fragment line extracted from the original address text.

[0185] It can also be explained that the similarity calculation in this embodiment can adopt a multi-channel fusion mechanism, which integrates multiple information such as semantic similarity, character suffix rule matching, place name dictionary matching, contextual consistency, and length regularization for weighted calculation.

[0186] First, obtain the semantic vector representations of the address tags and each address fragment. Then, calculate the similarity result by calculating the cosine similarity between the two vectors, as shown in the following formula:

[0187]

[0188] in, A vector representation of the address label; A vector representation of an address segment; This represents the cosine similarity between address labels and address fragments.

[0189] After calculating the similarity between all address tags and address segments, for each address tag, it can be sorted according to the similarity score, and the top K address segments with the highest similarity can be selected as candidate address segments corresponding to that tag.

[0190] The above implementation method, through parallel computing and multi-dimensional similarity fusion, can quickly filter out candidate segments that are semantically related to each address tag and have a reasonable structure while ensuring recall, thereby improving the processing quality and efficiency of subsequent directed graph construction and path search.

[0191] In the process of constructing the directed graph described above, an optional implementation may include: creating a graph node by combining each address label with each corresponding candidate address segment; if the address segment corresponding to the current graph node and the address segment corresponding to the next graph node are sequentially adjacent in the original address text, and the address label corresponding to the current graph node and the address label corresponding to the next graph node are allowed to be adjacent in the predefined address label address structure rules, then a directed edge is established between the current graph node and the next graph node; and a directed graph is constructed based on all the created graph nodes and the established directed edges.

[0192] In the process of creating a directed graph, the creation of graph nodes is performed first. Specifically, for predefined address labels such as province, city, district / county, street, house number, and "other," and the candidate address segments corresponding to each address label, each address label is combined with each corresponding candidate address segment, resulting in an independent graph node for each combination. It should be understood that each graph node carries dual attribute information: the type of address label corresponding to the graph node and the content of the candidate address segment corresponding to the graph node. These dual attributes clearly define the location of each graph node in the address hierarchy and the corresponding address segment.

[0193] After creating all graph nodes, traverse all created graph nodes. For any two graph nodes to which a directed edge needs to be established (denoted as the current graph node and the next graph node), based on preset judgment conditions, determine whether a directed edge is established between them. A directed edge is established only if both graph nodes simultaneously meet the following two conditions. Optionally, the preset judgment conditions specifically include: First, the candidate address fragments corresponding to the current graph node and the candidate address fragments corresponding to the next graph node are adjacent in character order in the original address text. Second, the address label corresponding to the current graph node and the address label corresponding to the next graph node meet the predefined address label structure rules regarding label adjacency. The predefined address label structure rules are set based on the hierarchical rules of Chinese addresses. For example, it allows nodes corresponding to province labels to be adjacent to nodes corresponding to city labels, and nodes corresponding to city labels to be adjacent to nodes corresponding to district / county labels, while prohibiting nodes corresponding to house number labels from being adjacent to nodes corresponding to province labels. When two graph nodes simultaneously satisfy both of the above conditions, a directed edge is established between the current graph node and the next graph node. The direction of this directed edge is consistent with the order of the address fragments in the original address text.

[0194] Based on this, a directed graph is formed by integrating all created graph nodes and all directed edges that meet the conditions. This directed graph can fully represent the text order relationship of candidate address fragments and the hierarchical relationship of address labels, providing data support for subsequent selection of the optimal combination of address elements through graph traversal algorithms.

[0195] Based on the constructed directed graph, a path search operation can be performed on the graph, and at least one set of first candidate address elements can be determined according to the search results.

[0196] In this embodiment, an optional implementation of obtaining at least one set of first candidate address elements based on path search may include: searching for at least one path from the graph node corresponding to the predefined start address label to the graph node corresponding to the predefined end address label, and calculating the comprehensive score of each path; and selecting at least one path from all the searched paths based on the comprehensive score to obtain at least one set of first candidate address elements.

[0197] Specifically, the starting and ending address labels for path searching can be determined first based on the hierarchical rules corresponding to each address label in the address label set. The starting address label is the top-level label in the address structure (e.g., the province label), and the ending address label is the bottom-level label in the hierarchical rules (e.g., the house number label).

[0198] Using the constructed directed graph as the search space, graph path search algorithms such as depth-first search, breadth-first search, or shortest path search are invoked to search from all graph nodes corresponding to the starting address label to all graph nodes corresponding to the ending address label, obtaining at least one path. Here, each path consists of several sequentially connected graph nodes and directed edges. The nodes on the path correspond sequentially to the address label-candidate address fragment combinations in the original address text that have hierarchical and grammatical relationships.

[0199] For each path obtained from the search, its comprehensive score is calculated to quantitatively evaluate the rationality and reliability of the path as an address resolution result.

[0200] Optionally, the comprehensive score can be obtained by weighted fusion of multiple preset scoring dimensions, such as: the degree of matching between the address fragments corresponding to each node in the path and the original text, the completeness of the address label hierarchy covered by the path, and the strength of the structural association between labels reflected by the directed edges in the path. Further, appropriate weights are assigned to each of the above dimensions, and the scores of each dimension are integrated into a comprehensive score for a single path through weighted summation. Further, based on the comprehensive score, at least one path is selected from all the searched paths to obtain at least one set of first candidate address elements.

[0201] The above implementation method determines the first candidate address element by combining directed graph path search and comprehensive scoring screening. It can accurately match the optimal address combination that conforms to the address hierarchy rules, thereby improving the accuracy and reliability of address resolution.

[0202] It should be noted that in this embodiment, the calculation of node weights and transition weights is performed simultaneously with the graph construction process.

[0203] Specifically, when creating each graph node, its node weight can be calculated and stored in the node weight mapping table based on the degree of matching between the address label corresponding to the node and the candidate address fragment; similarly, when establishing each directed edge, its transfer weight can be calculated and stored in the transfer weight mapping table based on the rationality and commonness of the structural relationship between the address labels represented by the edge.

[0204] Therefore, when calculating the comprehensive score of the path in the subsequent process, the pre-calculated weights of each node and edge can be directly read from the above mapping table, without having to repeat similarity or compliance calculations, thereby improving the processing efficiency of the path evaluation stage and the overall performance of the system.

[0205] Based on this, in the process of calculating the comprehensive score of the above-mentioned path, an optional implementation method may include: obtaining the node weights of all graph nodes on the path, and summing the node weights to obtain the node weight sum; wherein, the node weights characterize the degree of matching between the address label and the candidate address fragment; obtaining the transition weights of all edges on the path, and summing the transition weights to obtain the transition weight sum; wherein, the transition weights are used to quantify the rationality and commonness of the address structure rules represented by the established edges connecting two different address labels in the directed graph; and calculating the comprehensive score based on the node weight sum and the transition weight sum.

[0206] Specifically, taking any path as an example, based on all the graph nodes contained in the path, the weight value corresponding to each graph node is retrieved from the pre-generated node weight mapping table. It should be understood that the node weight mapping table is established synchronously when constructing the directed graph, and it records the node weight corresponding to each graph node. The retrieved node weights are then summed to obtain the total node weights of the path.

[0207] Simultaneously, based on all directed edges traversed by the path, the transfer weights corresponding to each edge are queried from a pre-generated transfer weight mapping table. It should be understood that the transfer weight mapping table is also established synchronously during the construction of the directed graph, recording the reasonableness score of the transfer relationship between address labels represented by each directed edge. The queried transfer weights are then summed to obtain the total transfer weights for the path.

[0208] Based on the calculated sum of node weights and the sum of transition weights, the comprehensive score of the path can be calculated through weighted summation, weighted average, or other preset fusion functions.

[0209] For example, when calculating the overall score based on the above method, the following expression can be used: The expression is:

[0210]

[0211] in, Characteristic composite score; Node weight; Characterize the transfer weights; The influence parameter characterizing the control transfer weight.

[0212] In the above scheme, since the node weights and transition weights are pre-calculated and stored in the mapping table during the graph construction stage, only efficient weight query and accumulation operations are needed during the path scoring stage. There is no need to repeatedly perform complex similarity calculations or rule matching, thereby improving the efficiency of path score evaluation and reducing the response time of the overall address resolution process.

[0213] In the above process, an optional implementation of calculating the transfer weight of directed edges may include: querying a predefined transfer weight mapping table based on the edge identifiers corresponding to all directed edges on the path to obtain the corresponding transfer weights; wherein, the address label transfer weight mapping table is constructed based on at least one of the following methods: assigning different weights to address labels that conform to predefined address structure rules based on address hierarchy specifications and commonness; calculating and assigning corresponding transfer weights based on the statistical probability of conforming to predefined address label address structure rules in historical address data.

[0214] In this embodiment, the transfer weight mapping table can be stored using a key-value pair structure during construction, where the edge identifier of the directed edge is used as the key and the corresponding transfer weight is used as the value. Optionally, the edge identifier can be represented based on a string or based on the node identifiers corresponding to the nodes at both ends of the directed edge; there is no limitation on this.

[0215] In subsequent queries, the edge identifier of the directed edge to be queried can be used as the index key to quickly retrieve the data in the mapping table and directly obtain the transition weight corresponding to that edge identifier as the index value.

[0216] It should be noted that the transfer weights used as values ​​in the above mapping table can be determined in various ways to reflect the rationality and frequency of transfer relationships between different address tags.

[0217] In one alternative implementation, directed edges between address labels conforming to predefined hierarchical rules defined in national or industry standards can be assigned corresponding weights based on domain knowledge. For example, directed edges between common address labels such as "province → city" and "city → district," which conform to the usual hierarchical order, are assigned higher transfer weights; transitions such as "district → road," which conform to the rules but occur less frequently, are assigned medium transfer weights; and transitions such as "house number → province," which clearly violate hierarchical logic, are assigned zero or very low weights. This method relies on prior rules to ensure the structural regularity of address resolution.

[0218] In another alternative implementation, a large-scale, labeled historical address dataset can be used to statistically analyze the frequency or conditional probability of adjacent occurrences of each address tag pair in the actual corpus. For example, the probability of the next address tag appearing given the previous address tag can be calculated, and this probability value can be directly used as or converted into the corresponding transition weight. The higher the frequency and statistical probability of the tag pair, the higher its transition weight.

[0219] In practical applications, the two methods mentioned above can also be used in combination. For example, based on the rule weights, statistical probabilities can be dynamically adjusted to construct a transfer weight mapping table that conforms to the standards and is close to actual usage habits.

[0220] Based on the comprehensive score obtained from the above implementation method, an optional implementation method for obtaining the first candidate address element by filtering the path according to the comprehensive score may include: selecting the top P paths with the highest comprehensive scores as candidate paths based on the comprehensive scores of each path; for each candidate path, determining the address labels and candidate address fragments corresponding to the graph node sequences contained in the candidate path as a corresponding set of first candidate address elements.

[0221] Specifically, based on the comprehensive score of each path, they can be sorted in descending order of score, and the top P paths with the highest comprehensive scores can be selected as candidate paths, where P is a preset positive integer. Alternatively, the system can set a score threshold and select all paths with comprehensive scores exceeding the threshold as candidate paths.

[0222] Furthermore, for each selected candidate path, its sequence of graph nodes can be parsed to extract the address label and corresponding candidate address fragment for each node. Then, according to the order of the nodes in the path, these address labels and candidate address fragments are combined to form a complete set of first candidate address elements. Each candidate path generates a corresponding set of first candidate address elements.

[0223] In the above method, by prioritizing the path with the highest comprehensive score, it can be ensured that the output first candidate address element performs better in terms of structural rationality and text matching degree, thereby improving the accuracy and reliability of address parsing results.

[0224] After obtaining multiple sets of first candidate address elements based on the above implementation method, the confidence score corresponding to each set of candidate address elements can be further determined. When the confidence score meets the preset confidence threshold condition, the final target address element is determined from the candidate groups that meet the condition. In this way, by introducing confidence assessment and threshold screening, the reliability and accuracy of the target address elements can be improved, thereby enhancing the pass rate and overall efficiency of the review process in subsequent business processing.

[0225] Based on this, an optional implementation of confidence assessment may include: obtaining a pre-constructed first prompt word template, the first prompt word template containing preset slots for receiving the original address text and multiple address fragments to guide confidence score calculation; filling the original address text and multiple address fragments into the preset slots of the first prompt word template to obtain confidence score calculation prompt words; and based on the confidence score calculation prompt words, calling a large language model to perform confidence score calculation processing to obtain the confidence score of the first candidate address element.

[0226] Specifically, after obtaining multiple sets of first candidate address elements, they can be evaluated using a pre-trained or fine-tuned large language model to obtain their respective confidence scores.

[0227] Taking the process of calculating an arbitrary confidence score as an example, before the calculation, a first prompt word template can be obtained to guide the large language model to calculate the confidence score. In this embodiment, the first prompt word template contains preset slots for receiving the original address text and multiple address fragments.

[0228] The received original address text, along with the parsed and evaluated first candidate address elements, are input into the preset slots in the first prompt word template to generate structured confidence score calculation prompt words.

[0229] The large language model is invoked, and the aforementioned confidence score calculation prompts are input into it for confidence score calculation. During the calculation process, the large language model comprehensively evaluates the parsing results based on its understanding of address semantics, structural regularity, and contextual consistency, and outputs a quantified confidence score.

[0230] Based on this, the confidence score is used as the confidence assessment result of the first candidate address element in this group, and is used for subsequent threshold judgment and target element screening.

[0231] In the above implementation, the semantic understanding and reasoning capabilities of the large language model are utilized to evaluate the confidence level of the address resolution results, thereby further improving the reliability of the address resolution process.

[0232] Having obtained the confidence scores of each first candidate address element, each confidence score is compared with a preset confidence score threshold.

[0233] If at least one set of first candidate address elements has a confidence score that meets the confidence score threshold, the set of candidate address elements with the highest confidence score among the first candidate address elements that meet the confidence score threshold shall be selected as the target address element corresponding to the original address text.

[0234] Optionally, if the confidence scores of all first candidate address elements fail to meet the preset confidence score threshold, it indicates that the parsing results based on the initial model have high uncertainty or may contain errors. In this case, a large language model can be used to enhance the parsing of the original address text to obtain more accurate parsing results.

[0235] Optionally, one possible implementation of enhanced parsing of the original address text may include: obtaining a pre-constructed second prompt word template, the second prompt word template containing preset slots for receiving the original address text and multiple address fragments to guide the generation of address elements; filling the original address text and multiple address fragments into the preset slots of the second prompt word template to obtain address element generation prompt words; and based on the address element generation prompt words, calling a large language model for address parsing processing to obtain the target address elements.

[0236] Specifically, the available area can be used to guide the large language model to generate address elements. In this embodiment, the second prompt word template includes preset slots for receiving the original address text and multiple address fragments.

[0237] The received original address text and the multiple address fragments obtained from the aforementioned implementation are filled into the preset slots in the second prompt word template to generate structured address element prompt words.

[0238] The large language model is invoked, and the above address element generation prompts are input into the large language model for address element generation processing.

[0239] During the generation process, the large language model comprehensively analyzes the semantics of the original address text, the logical relationships between each address fragment, and the standard address structure specifications. That is, through semantic understanding and reasoning, it reassembles the discrete address fragments and maps them to the standardized address element hierarchy to obtain a set of processed address elements, which are then used as the target address elements.

[0240] In the above implementation, structured prompts guide the large language model to perform semantic understanding and reasoning, thereby accurately recombining discrete address fragments and mapping them to a standardized address element hierarchy, thus improving the accuracy of address resolution.

[0241] Considering that large language models may misjudge address text containing malicious padding or garbled formatting due to noise interference, such as misidentifying meaningless symbols or garbled characters as valid address fields, this embodiment further verifies the address elements generated by the large language model after obtaining them to improve the accuracy and reliability of the output results.

[0242] Based on this, an optional implementation method for verifying the output results of a large model to obtain address elements with higher reliability may include: obtaining the second candidate address element output by the large language model and its corresponding confidence score; if the confidence score meets a preset confidence score threshold, then the second candidate address element is determined as the target address element; if the confidence score meets the preset confidence score threshold, then the target address element is determined based on the first candidate address element and the second candidate address element.

[0243] In this embodiment, the pre-constructed second prompt word template, in addition to including logic for generating multiple sets of address elements, may also include logic for generating confidence scores corresponding to each address element. Thus, the large model can output multiple sets of address elements during processing, namely, second candidate address elements and their corresponding confidence scores. These confidence scores can be output synchronously by the large language model when generating address elements, or they can be calculated by a separate confidence evaluation module based on preset rules or a model.

[0244] Based on this, it is determined whether each confidence score meets the preset confidence score threshold. If it does, it indicates that the output of the large language model has high credibility, and the second candidate address element can be directly determined as the target address element.

[0245] If the confidence score does not meet the preset confidence score threshold, it indicates that the output of the large language model may have uncertainty or noise interference. In this case, the system does not directly adopt the result. Instead, it can comprehensively determine the final target address element based on the first and second candidate address elements through rule matching, voting mechanisms, or weighted fusion. For example, the second candidate address element, which is structurally consistent with or highly overlaps with the first candidate address element, can be prioritized as the target. Alternatively, the two types of candidate results can be compared and corrected at the element level to generate a more reliable combination of address elements.

[0246] The above implementation method, by introducing confidence threshold judgment and multi-source result fusion, can identify and filter noise output that may be generated by large language models, thereby enhancing parsing capabilities while ensuring the reliability of address element results.

[0247] It should be noted that in some optional embodiments, after obtaining multiple address fragments, end-to-end address parsing can be directly performed based on the large language model, and the target address elements can be output. However, considering that the large language model usually adopts a token-by-token autoregressive generation method during inference, its processing latency is relatively high. For some business scenarios with strict requirements for real-time performance and throughput, such as high-concurrency scenarios like merchant onboarding and real-time risk control audits, it may not meet business needs.

[0248] Based on this, the technical solution described in this embodiment can generate multiple sets of candidate address elements simultaneously by constructing a directed graph based on address fragments and their corresponding address tags, and performing path search based on this directed graph. This method transforms the address resolution task into a multi-path search problem on a graph structure, generating multiple candidate paths at once through parallelization or efficient search algorithms. This alleviates the computational latency and inefficiency of the token-by-token generation method, better meeting the requirements for real-time processing and system throughput in high-concurrency scenarios. Simultaneously, the directed graph-based search process effectively utilizes address hierarchy constraints and transition rules, ensuring that the generated candidate address elements are logically consistent in structure, balancing resolution efficiency and result reliability.

[0249] Once the target address elements are obtained based on the above implementation method, they can be further audited for business compliance based on business needs to ensure that the parsing results meet the specifications and risk control requirements of the specific business scenario, thereby improving the accuracy and security of subsequent business processing.

[0250] Based on this, one possible implementation method for conducting business compliance audits may include: performing business compliance audits on target address elements based on preset business verification rules; if the target address elements pass the business compliance audit, then performing business processing audits on the original address text based on the target address elements; if the target address elements fail the business compliance audit, then generating and outputting a verification result containing risk markers.

[0251] In this embodiment, pre-built business verification rules are obtained. During the rule building phase, different business verification rules can be pre-set according to different business scenarios and compliance requirements. For example, in the merchant onboarding review scenario, if the address element is identified to contain special area types such as development zone or new area, the rule can require that the address element must also contain specific house number information to ensure that the address is deliverable. In addition, business verification rules that are common to all businesses can also be built during the construction process. For example, for field content marked as "other" or other unclassified fields during the parsing process, the rule can trigger a secondary semantic analysis or manual review process to avoid misjudging meaningless noise or unconventional expressions as valid address information.

[0252] By combining general rules with scenario-based rules, the system can ensure comprehensive verification coverage while flexibly adapting to the specific address compliance requirements of different businesses.

[0253] Optionally, when verifying the target address element based on the above business verification rules, if the target address element passes the above business compliance verification, then the original address text is processed according to the verified target address element, and subsequent business processing review procedures are performed, such as merchant network access review and address information entry into the database.

[0254] If the target address element fails the business compliance verification, a verification result containing risk markers is generated and output. In this embodiment, the result can specifically indicate the rule type that failed the verification, the content of the abnormal field, and the relevant risk level, and can trigger subsequent processing actions such as warnings, manual review, or automatic rejection based on configuration.

[0255] In the above approach, the address resolution results are deeply integrated with specific business rules through business compliance audits. This not only enhances the reliability and usability of the resolution results in practical applications, but also enables the automated identification and control of abnormal or high-risk addresses, thereby effectively supporting business risk control and compliance management.

[0256] In the above verification process, one possible implementation of the subsequent business processing review process may include: outputting the target address element and the original address text to a preset upper-level business system to trigger the upper-level business system to conduct business processing review on the original address text based on the target address element.

[0257] Specifically, when the current system is only capable of performing address resolution tasks, it can output the target address elements and the original address text in structured data form to the preset upper-layer business system through a predefined API interface.

[0258] The upper-level business system receives this structured data and triggers an automated business processing review process based on the target address elements and the original address text. For example, in the merchant onboarding scenario, the review platform can automatically complete steps such as regional qualification verification and business premises authenticity verification based on standardized address information, thereby improving review efficiency and accuracy.

[0259] Figure 4 A schematic diagram of the address resolution device provided in this application is shown below. Figure 4 As shown, the address resolution device 40 provided in this embodiment includes:

[0260] Address fragment acquisition module 401 is used to acquire the original address text and segment the original address text to obtain multiple address fragments;

[0261] The first candidate address element acquisition module 402 is used to perform address parsing processing on multiple address segments to obtain at least one set of first candidate address elements; wherein, each set of first candidate address elements includes multiple address tags and address segments corresponding to each address tag;

[0262] The first target address element determination module 403 is used to determine the target address element corresponding to the original address text based on at least one set of first candidate address elements if the confidence score of any set of first candidate address elements meets the preset confidence score threshold.

[0263] The second target address element determination module 404 is used to perform enhanced parsing processing based on the address fragment and the original address text to obtain the target address element corresponding to the original address text if the confidence scores of at least one group of first candidate address elements do not meet the preset confidence score threshold.

[0264] In one possible implementation, the address fragment acquisition module 401, when segmenting the original address text to obtain multiple address fragments, is specifically used for:

[0265] Based on the preset word segmentation tool, the original address text is segmented to obtain a text sequence containing multiple words;

[0266] Based on a preset combination length threshold N, a sliding window enumeration is performed on the text sequence to generate multiple address fragments, where N is an integer greater than 1.

[0267] In one possible implementation, the address fragment acquisition module 401, when generating multiple address fragments by performing sliding window enumeration on the text sequence based on a preset combination length N, is specifically used for:

[0268] Using a combined length threshold N as the maximum size of the sliding window, the window slides sequentially starting from the beginning of the text sequence.

[0269] For any sliding operation, merge the M consecutive words covered by the current sliding window into a single address segment; where M is a variable integer satisfying 1 < M ≤ N;

[0270] Traverse the text sequence until the starting position of the sliding window covers the end of the text sequence to obtain multiple address fragments.

[0271] In one possible implementation, the first candidate address element obtaining module 402, when performing address parsing processing on multiple address segments to obtain at least one set of first candidate address elements, is specifically used for:

[0272] Retrieve a predefined set of address labels; where each address label in the set is defined according to address hierarchy rules;

[0273] For any address label in the address label set, determine at least one candidate address segment corresponding to the address label from multiple address segments;

[0274] Based on each address label in the address label set and its corresponding candidate address fragment, a directed graph is constructed. In the directed graph, each graph node represents an address label and its corresponding candidate address fragment. A directed edge indicates that the address fragments represented by two graph nodes are sequentially adjacent in the original address text, and their corresponding address labels are allowed to be adjacent in the predefined address label address structure rules.

[0275] Path search is performed based on the directed graph to obtain at least one set of first candidate address elements; wherein each set of first candidate address elements corresponds to a path from the starting address label graph node to the ending address label graph node.

[0276] In one possible implementation, the first candidate address element obtaining module 402, when determining at least one candidate address segment corresponding to an address tag from multiple address segments, is specifically used for:

[0277] Calculate the similarity between the address label and each address segment;

[0278] Based on the similarity scores, the top K address segments with the highest similarity scores are identified as candidate address segments corresponding to the address labels.

[0279] In one possible implementation, the first candidate address element acquisition module 402, when constructing a directed graph based on each address label in the address label set and its corresponding candidate address fragment, is specifically used for:

[0280] Each address label is combined with its corresponding candidate address fragment to form a graph node;

[0281] If the address fragment corresponding to the current graph node and the address fragment corresponding to the next graph node are sequentially adjacent in the original address text, and the address label corresponding to the current graph node and the address label corresponding to the next graph node are allowed to be adjacent in the predefined address label address structure rules, then a directed edge is established between the current graph node and the next graph node.

[0282] Construct a directed graph based on all created graph nodes and established directed edges.

[0283] In one possible implementation, the first candidate address element acquisition module 402, when performing path search based on the directed graph and obtaining at least one set of first candidate address elements, is specifically used for:

[0284] Search for at least one path from the graph node corresponding to the predefined start address label to the graph node corresponding to the predefined end address label, and calculate the overall score for each path;

[0285] Based on the comprehensive score, at least one path is selected from all the searched paths to obtain at least one set of first candidate address elements.

[0286] In one possible implementation, the first candidate address element acquisition module 402, when calculating the comprehensive score for each path, is specifically used for:

[0287] Obtain the node weights of all graph nodes on the path, and sum the node weights to obtain the node weight sum; where the node weights represent the degree of matching between the address label and the candidate address fragment;

[0288] Obtain the transition weights of all directed edges on the path, and sum the transition weights to obtain the transition weight sum; where the transition weights are used to quantify the rationality and commonality of the address structure rules represented by the established edges connecting two different address labels in the directed graph.

[0289] The overall score is calculated based on the sum of node weights and the sum of transition weights.

[0290] In one possible implementation, the first candidate address element acquisition module 402, when acquiring the transition weights of all directed edges on the path, is specifically used for:

[0291] Based on the edge identifiers of all directed edges on the path, query the predefined transition weight mapping table to obtain the corresponding transition weights;

[0292] The transfer weight mapping table is constructed based on at least one of the following methods:

[0293] Based on the address hierarchy specifications and commonality, different weights are assigned to addresses that conform to the predefined address label address structure rules.

[0294] Based on the statistical probability of conforming to predefined address label address structure rules in historical address data, calculate and assign corresponding transfer weights.

[0295] In one possible implementation, the first candidate address element obtaining module 402, when selecting at least one path from all searched paths based on the comprehensive score to obtain at least one set of first candidate address elements, is specifically used for:

[0296] Based on the comprehensive score of each path, the top P paths with the highest comprehensive scores are selected as candidate paths;

[0297] For each candidate path, the address labels and candidate address fragments corresponding to the graph node sequence contained in the candidate path are determined as a set of first candidate address elements.

[0298] In one possible implementation, after obtaining at least one set of first candidate address elements, the device is further used to:

[0299] Obtain a pre-built first prompt word template, which contains preset slots for receiving the original address text and multiple address fragments to guide the calculation of confidence scores;

[0300] Fill the original address text and multiple address fragments into the preset slots of the first prompt word template to obtain the confidence score calculation prompt words;

[0301] Based on the confidence score, the prompt words are calculated, and the large language model is called to process the confidence score to obtain the confidence score of the first candidate address element.

[0302] In one possible implementation, the second target address element determination module 404, when performing enhanced parsing processing based on address fragments and original address text to obtain the target address element corresponding to the original address text, is specifically used to: obtain a pre-constructed second prompt word template, the second prompt word template containing preset slots for receiving the original address text and multiple address fragments to guide the generation of address elements;

[0303] Fill the original address text and multiple address fragments into the preset slots of the second prompt word template to obtain the address element generation prompt word;

[0304] Based on address elements, prompt words are generated, and a large language model is called to perform address parsing processing to obtain the target address elements.

[0305] In one possible implementation, the second target address element determination module 404, when calling the large language model for address resolution processing to obtain the target address elements, is specifically used for:

[0306] Obtain the second candidate address element and its corresponding confidence score from the output of the large language model;

[0307] If the confidence score meets the preset confidence score threshold, then the second candidate address element is determined as the target address element;

[0308] If the confidence score meets the preset confidence score threshold, the target address element is determined based on the first candidate address element and the second candidate address element.

[0309] In one possible implementation, before segmenting the original address text, the device is also used to:

[0310] Standardize the format of the original address text;

[0311] Based on preset text verification rules, the text compliance of the address text after format standardization is verified.

[0312] If the address text passes the text compliance check, then proceed with the subsequent segmentation steps;

[0313] If the address text fails the text compliance check, it will be marked as an abnormal address and the processing flow will end.

[0314] In one possible implementation, after obtaining the target address elements corresponding to the original address text, the device is further used to:

[0315] Based on preset business verification rules, the target address elements are verified for business compliance.

[0316] If the target address element passes the business compliance verification, then the original address text will be reviewed for business processing based on the target address element.

[0317] If the target address element fails the business compliance verification, a verification result containing risk markers will be generated and output.

[0318] In one possible implementation, when reviewing the original address text for business processing based on the target address elements, the device is also specifically used for:

[0319] The target address elements and the original address text are output to the preset upper-level business system to trigger the upper-level business system to conduct business processing review on the original address text based on the target address elements.

[0320] The address resolution device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0321] Figure 5 This is a block diagram of an electronic device provided in this application. The device can be a client terminal device or a cloud server, etc. See also... Figure 5 The device 500 may include one or more of the following components: processing component 502, memory 504, power supply component 506, multimedia component 508, audio component 510, input / output interface 512, sensor component 514, and communication component 516.

[0322] Processing component 502 typically controls the overall operation of device 500, such as operations associated with display, telephone calls, data communication, camera operation, and recording. Processing component 502 may include one or more processors 520 to execute instructions to perform all or part of the steps of the methods described above. Furthermore, processing component 502 may include one or more modules to facilitate interaction between processing component 502 and other components. For example, processing component 502 may include a multimedia module to facilitate interaction between multimedia component 508 and processing component 502.

[0323] Memory 504 is configured to store various types of data to support the operation of device 500. Examples of this data include instructions for any application or method operating on device 500, contact data, phonebook data, messages, pictures, videos, etc. Memory 504 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0324] Power supply component 506 provides power to various components of device 500. Power supply component 506 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to device 500.

[0325] Multimedia component 508 includes a screen that provides an output interface between device 500 and the user. In some embodiments, the screen may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen may be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors may sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 508 includes a front-facing camera and / or a rear-facing camera. When device 500 is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0326] Audio component 510 is configured to output and / or input audio signals. For example, audio component 510 includes a microphone (MIC) configured to receive external audio signals when device 500 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 504 or transmitted via communication component 516. In some embodiments, audio component 510 also includes a speaker for outputting audio signals.

[0327] Input / output interface 512 provides an interface between processing component 502 and peripheral interface modules, which may be keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, start buttons, and lock buttons.

[0328] Sensor assembly 514 includes one or more sensors for providing state assessments of various aspects of device 500. For example, sensor assembly 514 may detect the on / off state of device 500, the relative positioning of components such as the display and keypad of device 500, changes in the position of device 500 or a component of device 500, the presence or absence of user contact with device 500, the orientation or acceleration / deceleration of device 500, and temperature changes of device 500. Sensor assembly 514 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 514 may also include optical sensors, such as complementary metal-oxide-semiconductor (CMOS) sensors or charge-coupled device (CCD) sensors, for use in imaging applications. In some embodiments, sensor assembly 514 may also include an accelerometer, a gyroscope, a magnetometer, a pressure sensor, or a temperature sensor.

[0329] Communication component 516 is configured to facilitate wired or wireless communication between device 500 and other devices. Device 500 can access wireless networks based on communication standards, such as WiFi, 4G, or 5G, or combinations thereof. In one exemplary embodiment, communication component 516 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 516 also includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra Wide Band (UWB), Bluetooth (BT), and other technologies.

[0330] In an exemplary embodiment, device 500 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processors (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0331] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 504 including instructions, which can be executed by a processor 520 of device 500 to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0332] A non-transitory computer-readable storage medium, wherein instructions in the storage medium, when executed by a server's processor, enable the server to perform the aforementioned large model data processing method.

[0333] This application also provides a chip for executing instructions, which is used to execute the technical solution of the large model data processing method in the above embodiments.

[0334] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed on a computer, the computer performs the technical solution of the large model data processing method described in the above embodiments.

[0335] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the large model data processing method described in the above embodiments.

[0336] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0337] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

[0338] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0339] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An address resolution method, characterized in that, include: Obtain the original address text and segment the original address text to obtain multiple address fragments; Address parsing processing is performed on the plurality of address segments to obtain at least one set of first candidate address elements; wherein, each set of first candidate address elements includes a plurality of address tags and address segments corresponding to each address tag; If at least one set of first candidate address elements has a confidence score that meets a preset confidence score threshold, then the target address element corresponding to the original address text is determined based on the at least one set of first candidate address elements. If the confidence scores of all first candidate address elements do not meet the preset confidence score threshold, then enhanced parsing processing is performed based on the address fragment and the original address text to obtain the target address element corresponding to the original address text. The target address element is used to conduct business processing review of the original address text.

2. The method according to claim 1, characterized in that, The original address text is segmented to obtain multiple address fragments, including: Based on a preset word segmentation tool, the original address text is segmented to obtain a text sequence containing multiple words; Based on a preset combination length threshold N, a sliding window enumeration is performed on the text sequence to generate multiple address fragments, where N is an integer greater than 1.

3. The method according to claim 2, characterized in that, Based on a preset combination length N, a sliding window enumeration is performed on the text sequence to generate multiple address fragments, including: Using the combined length threshold N as the maximum size of the sliding window, the sliding window is sequentially slid from the starting position of the text sequence; For any sliding operation, merge the M consecutive words covered by the current sliding window into a single address segment; where M is a variable integer satisfying 1 < M ≤ N; Traverse the text sequence until the starting position of the sliding window covers the end of the text sequence to obtain the multiple address fragments.

4. The method according to claim 1, characterized in that, Address parsing is performed on the multiple address segments to obtain at least one set of first candidate address elements, including: Obtain a predefined set of address tags; wherein each address tag in the set of address tags is defined according to address hierarchy rules; For any address tag in the address tag set, at least one candidate address segment corresponding to the address tag is determined from the plurality of address segments; Based on each address tag in the address tag set and its corresponding candidate address fragment, a directed graph is constructed, wherein each graph node of the directed graph represents an address tag and its corresponding candidate address fragment; a directed edge indicates that the address fragments represented by two graph nodes are sequentially adjacent in the original address text, and their corresponding address tags are allowed to be adjacent in the predefined address tag address structure rules. Based on the directed graph, a path search is performed to obtain at least one set of first candidate address elements; wherein each set of first candidate address elements corresponds to a path from the starting address label graph node to the ending address label graph node.

5. The method according to claim 4, characterized in that, Determining at least one candidate address segment corresponding to the address label from the plurality of address segments includes: Calculate the similarity between the address tag and each address segment; Based on the aforementioned similarity scores, the top K address segments with the highest similarity scores are determined as candidate address segments corresponding to the address labels.

6. The method according to claim 4, characterized in that, Based on each address label in the address label set and its corresponding candidate address fragment, a directed graph is constructed, including: Each address label is combined with its corresponding candidate address fragment to form a graph node; If the address fragment corresponding to the current graph node and the address fragment corresponding to the next graph node are sequentially adjacent in the original address text, and the address label corresponding to the current graph node and the address label corresponding to the next graph node are allowed to be adjacent in the predefined address label address structure rules, then a directed edge is established between the current graph node and the next graph node. Based on all created graph nodes and established directed edges, construct the directed graph.

7. The method according to claim 6, characterized in that, Based on the directed graph, a path search is performed to obtain at least one set of first candidate address elements, including: Search for at least one path from the graph node corresponding to the predefined start address label to the graph node corresponding to the predefined end address label, and calculate the overall score for each path; Based on the comprehensive score, at least one path is selected from all the searched paths to obtain at least one set of first candidate address elements.

8. The method according to claim 7, characterized in that, Calculate the overall score for each path, including: Obtain the node weights of all graph nodes on the path, and sum the node weights to obtain the node weight sum; wherein, the node weights represent the degree of matching between the address label and the candidate address fragment; Obtain the transition weights of all directed edges on the path, and sum the transition weights to obtain the transition weight sum; wherein, the transition weights are used to quantify the rationality and commonality of the address structure rules represented by the edges connecting two different address labels in the directed graph. The overall score is calculated based on the sum of the node weights and the sum of the transition weights.

9. The method according to claim 8, characterized in that, Obtain the transition weights of all directed edges on the path, including: Based on the edge identifiers corresponding to all directed edges on the path, query the predefined transition weight mapping table to obtain the corresponding transition weights; The transfer weight mapping table is constructed based on at least one of the following methods: Based on the address hierarchy specifications and commonality, different weights are assigned to addresses that conform to the predefined address label address structure rules. Based on the statistical probability of conforming to predefined address label address structure rules in historical address data, calculate and assign corresponding transfer weights.

10. The method according to claim 7, characterized in that, Based on the comprehensive score, at least one path is selected from all the searched paths to obtain at least one set of first candidate address elements, including: Based on the comprehensive score of each path, the top P paths with the highest comprehensive scores are selected as candidate paths; For each candidate path, the address labels and candidate address fragments corresponding to the graph node sequence contained in the candidate path are determined as a corresponding set of first candidate address elements.

11. The method according to claim 1, characterized in that, After obtaining at least one set of first candidate address elements, the method further includes: Obtain a pre-constructed first prompt word template, which includes preset slots for receiving the original address text and the multiple address fragments to guide the calculation of confidence scores; The original address text and the multiple address fragments are filled into the preset slots of the first prompt word template to obtain the confidence score calculation prompt word; Based on the confidence score, the prompt words are calculated, and the large language model is called to perform confidence score calculation to obtain the confidence score of the first candidate address element.

12. The method according to claim 1, characterized in that, Enhanced parsing is performed based on the address fragment and the original address text to obtain the target address elements corresponding to the original address text, including: Obtain a pre-constructed second prompt word template, the second prompt word template containing preset slots for receiving the original address text and the multiple address fragments, so as to guide the generation of address elements; The original address text and the multiple address fragments are filled into the preset slots of the second prompt word template to obtain the address element generation prompt word; Based on the address elements, a prompt word is generated, and a large language model is called to perform address parsing processing to obtain the target address elements.

13. The method according to claim 12, characterized in that, The large language model is invoked for address resolution processing to obtain the target address elements, including: Obtain the second candidate address element and its corresponding confidence score output by the large language model; If the confidence score meets the preset confidence score threshold, then the second candidate address element is determined as the target address element; If the confidence score meets the preset confidence score threshold, then the target address element is determined based on the first candidate address element and the second candidate address element.

14. The method according to any one of claims 1-13, characterized in that, Before segmenting the original address text, the method further includes: The original address text is formatted and normalized. Based on preset text verification rules, the text compliance of the address text after format standardization is verified. If the address text passes the text compliance check, then the subsequent segmentation steps are executed; If the address text fails the text compliance check, it is marked as an abnormal address and the processing flow ends.

15. The method according to any one of claims 1-13, characterized in that, After obtaining the target address elements corresponding to the original address text, the method further includes: Based on preset business verification rules, the target address elements are verified for business compliance. If the target address element passes the business compliance verification, then the original address text is reviewed for business processing based on the target address element. If the target address element fails the business compliance verification, a verification result containing risk markers will be generated and output.

16. The method according to claim 15, characterized in that, Based on the target address elements, the original address text is reviewed for business processing, including: The target address element and the original address text are output to a preset upper-level business system to trigger the upper-level business system to conduct business processing review on the original address text based on the target address element.

17. An address resolution device, characterized in that, include: The address fragment acquisition module is used to acquire the original address text and segment the original address text to obtain multiple address fragments; The first candidate address element acquisition module is used to perform address parsing processing on the plurality of address segments to obtain at least one set of first candidate address elements; wherein, each set of first candidate address elements includes a plurality of address tags and address segments corresponding to each address tag; The first target address element determination module is used to determine the target address element corresponding to the original address text based on the at least one set of first candidate address elements if the confidence score of any set of first candidate address elements meets the preset confidence score threshold. The second target address element determination module is used to perform enhanced parsing processing based on the address fragment and the original address text to obtain the target address element corresponding to the original address text if the confidence scores of the at least one group of first candidate address elements do not meet the preset confidence score threshold.

18. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-16.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-16.

20. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-16.

Images