Address resolution method, model training method, device, storage medium and program product

Abstract

The embodiment of the application provides an address resolution method, a model training method, equipment, a storage medium and a program product. The method can comprise: preprocessing original address information through a task coordination intelligent module to obtain preprocessed address information; if the preprocessed address information does not satisfy a preset condition, searching for similar address information similar to the preprocessed address information through an address retrieval intelligent module; calling a target map application through a map calling intelligent module to obtain similar geographic point of interest information matched with the preprocessed address information; and finally performing multi-source collaborative reasoning based on multi-source heterogeneous information (i.e. the original address information, the similar address information and the similar address point of interest information) through a decision-making intelligent module to obtain target address information. Since multi-source heterogeneous information including the original address information can be obtained, and the multi-source heterogeneous information is mutually verified to resolve the original address information, the resolution accuracy of the original address is improved.

Description

Technical Field

[0001] This application relates to the field of geographic information processing, and in particular to an address resolution method, a model training method, an apparatus, a storage medium, and a program product. Background Technology

[0002] In service scenarios such as real estate transactions and logistics delivery, the original address information provided by users is crucial for providing services. However, in practice, original address information often suffers from problems such as non-standard descriptions, missing information, semantic ambiguity, and misuse of alternative names, making it impossible to accurately locate the target based on the original address information. Therefore, it is necessary to parse the original address information to obtain accurate target address information.

[0003] In related technologies, a general text model can be used to generate an original address vector corresponding to the original address information. Based on the original address vector, addresses similar to the original address information can be retrieved from an address database to obtain the target address information. However, in the above method, because there are multiple duplicate place names or semantically similar place names in reality (such as Nanjing Road and Gulou District existing in multiple locations, and Jianshe Road 1 and Jianshe Road 2 being semantically similar), the locations indicated by the similar addresses retrieved through the general text model may differ significantly from the locations indicated by the original address information. Furthermore, because the original address information provided by the user is inaccurate, the accuracy of the similar addresses retrieved based on the inaccurate original address information is also very limited, resulting in poor parsing accuracy of the original address information. Summary of the Invention

[0004] This application provides an address resolution method, a model training method, a device, a storage medium, and a program product to improve the accuracy of resolving raw address information.

[0005] This application provides an address resolution method, comprising: preprocessing original address information through a task coordination intelligent module to obtain preprocessed address information; if the preprocessed address information does not meet preset conditions, performing the following operations: searching for similar address information corresponding to the preprocessed address information through an address retrieval intelligent module; invoking a target map application through a map invocation intelligent module to obtain similar geographic interest point information corresponding to the preprocessed address information; and performing multi-source collaborative reasoning through a decision intelligent module based on the original address information, the similar address information, and the similar geographic interest point information to obtain target address information.

[0006] This application provides a model training method, comprising: acquiring training samples, wherein the training samples include original address information, real latitude and longitude coordinates, and real grid index identifiers; generating a predicted address vector with a nested structure based on the original address information using a first initial model, and predicting the corresponding predicted latitude and longitude coordinates and predicted grid index identifiers based on the multi-level prefix sub-vectors in the predicted address vectors; calculating the loss value of the first initial model based on the real latitude and longitude coordinates, the real grid index identifiers, the predicted latitude and longitude coordinates, and the predicted grid index identifiers using a hierarchical joint loss function, wherein the hierarchical joint loss function includes loss terms corresponding to multiple levels, and the loss terms corresponding to any level include Haversian distance loss, planar distance loss, and hexagonal hierarchical spatial index classification loss; updating the model parameters of the first initial model based on the loss value until the model training termination condition is met, thereby obtaining a geographic embedding model.

[0007] This application provides a model training method, comprising: acquiring training samples, wherein the training samples include first original address information, similar address information, similar geographic points of interest information, and labeled addresses; performing supervised fine-tuning on a second initial model based on the training samples to obtain an intermediate model; acquiring triple samples, wherein the triple samples include prompt words, correct examples, and incorrect examples, wherein the prompt words include the first original address information, the similar address information, and the similar geographic points of interest information; the correct examples are semantically and format-consistent with the labeled addresses, and the incorrect examples are semantically or format-inconsistent with the labeled addresses; performing preference training on the intermediate model using the triple samples to obtain a decision model; and performing knowledge distillation on the decision model to obtain a lightweight decision model.

[0008] This application provides an address resolution system, including: a task coordination intelligent module, an address retrieval intelligent module, a map invocation intelligent module, a decision-making intelligent module, and a fast resolution intelligent module. The task coordination intelligent module is used to preprocess the original address information to obtain preprocessed address information. The task coordination intelligent module is also used to, when the preprocessed address information meets preset conditions, call the fast resolution intelligent module to resolve the preprocessed address information to obtain target address information. The task coordination intelligent module is also used to, when the preprocessed address information does not meet the preset conditions, send the preprocessed address information to the address retrieval intelligent module and the map invocation intelligent module. The address retrieval intelligent module is used to find similar address information corresponding to the preprocessed address information and return the similar address information to the task coordination intelligent module. The map invocation intelligent module is used to call a target map application to obtain similar geographic interest point (PI) information corresponding to the preprocessed address information and return the similar PI information to the task coordination intelligent module. The decision-making intelligent module is used to perform multi-source collaborative reasoning based on the original address information, the similar address information, and the similar PI information sent by the task coordination intelligent module to obtain the target address information.

[0009] This application also provides an electronic device, including: a memory and a processor; the memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program in the memory to implement the steps in the address resolution method or the model training method.

[0010] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the address resolution method or the model training method.

[0011] This application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, enables the processor to implement the steps in the address resolution method or the model training method.

[0012] In this embodiment, the original address information can be preprocessed using a task coordination intelligent module to obtain preprocessed address information. If the preprocessed address information does not meet preset conditions, an address retrieval intelligent module searches for similar address information that is semantically and spatially similar to the preprocessed address information. A map invocation intelligent module invokes the target map application to obtain similar geographic points of interest information matching the preprocessed address information. Finally, a decision intelligent module performs multi-source collaborative reasoning based on multi-source heterogeneous information (i.e., original address information, similar address information, and similar address point of interest information) to obtain the target address information. Thus, because multi-source heterogeneous information, including the original address information, can be obtained and cross-verified to parse the original address information, the accuracy of original address parsing is improved. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating an address resolution method provided for an exemplary embodiment of this application; Figure 2 A schematic diagram of an address index provided for an exemplary embodiment of this application; Figure 3 A schematic diagram illustrating the process of an address resolution method provided for an exemplary embodiment of this application; Figure 4 A flowchart illustrating a model training method provided for an exemplary embodiment of this application. Figure 1 ; Figure 5 A flowchart illustrating a model training method provided for an exemplary embodiment of this application. Figure 2 ; Figure 6 A schematic diagram of an address resolution system provided for an exemplary embodiment of this application; Figure 7 A schematic diagram of an address resolution device provided for an exemplary embodiment of this application; Figure 8 A schematic diagram of the structure of a model training device provided as an exemplary embodiment of this application. Figure 1 ; Figure 9 A schematic diagram of the structure of a model training device provided as an exemplary embodiment of this application. Figure 2 ; Figure 10 This is a schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] It should be noted that, in the cases involving user information in the embodiments of this application, 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, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0016] In service scenarios such as real estate transactions, information recommendations, house auctions, and logistics delivery, the original address information provided by users is a crucial basis for providing services. However, in practical applications, original address information often suffers from problems such as non-standard descriptions, missing information, semantic ambiguity, and misuse of alternative names, making it impossible to accurately locate users based on the original address information. Therefore, it is necessary to parse the original address information to obtain accurate target address information.

[0017] In related technologies, there is a solution 1: based on predefined regular expressions, matching is performed in the address dictionary, the address elements corresponding to the preset fields are extracted from the original address information, and the extracted address elements are standardized to obtain the target address information. However, solution 1 has the following disadvantages: (1) every time a new naming pattern appears, a new regular expression needs to be written, the preset address dictionary is always lagging behind, and the rules need to be continuously iterated manually, resulting in high maintenance costs; (2) different regions have different ways of saying addresses (such as "lane", "alley", "hutong", etc.), requiring a large number of localized rules; (3) for complex semantic structures (such as intersections, north of xxx road, etc.), the rule matching lacks contextual understanding capabilities.

[0018] There is also a second option: directly calling existing map applications to parse the original address information to obtain the target address information. However, the disadvantages of the second option are: (1) Since existing map applications are designed for public navigation, the target address information obtained by parsing can usually only be located at the center of a road or point of interest, and cannot be accurate to a specific building; (2) If the original address information provided by the user includes non-address noise information (such as Mr. Zhang's phone number 1381234, No. 159, Third Road, District C, City B) and address aliases (such as "Sky City" being aliased as "Cloud Sky City"), the existing map applications will be unable to identify the valid address information.

[0019] Furthermore, there is a third option: using a general text model to generate the original address vector corresponding to the original address information, and based on the original address vector, searching the address database for addresses similar to the original address information to obtain the target address information. However, the disadvantages of the third option are: (1) because there are multiple duplicate place names or semantically similar place names in reality (such as Nanjing Road, Gulou District, etc., and Jianshe Road 1 and Jianshe Road 2 being semantically similar in many places), the location indicated by the similar address retrieved by the general text model may be very different from the location indicated by the original address information; (2) because the original address information provided by the user is inaccurate, the accuracy of the similar address retrieved based on the inaccurate original address information is also very limited.

[0020] In summary, all three schemes suffer from poor address resolution accuracy.

[0021] To address the aforementioned technical problems, in this embodiment, the address resolution task is decomposed into multiple intelligent modules working collaboratively. First, the task coordination intelligent module preprocesses the original address information to obtain preprocessed address information. If the preprocessed address information does not meet preset conditions, the address retrieval intelligent module searches for similar address information that is semantically and spatially similar to the preprocessed address information. The map invocation intelligent module invokes the target map application to obtain similar geographic POI (Point of Interest) information matching the preprocessed address information. Finally, the decision intelligent module performs multi-source collaborative reasoning based on multi-source information (i.e., original address information, similar address information, and similar address point of interest information) to obtain the target address information. In this way, because multi-source information corresponding to the original address information can be obtained and cross-verified based on this information for multi-source collaborative reasoning, the accuracy of original address resolution is improved.

[0022] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart illustrating an address resolution method provided for an exemplary embodiment of this application. Figure 1 As shown, the method includes: S11. The original address information is preprocessed through the task coordination intelligent module to obtain preprocessed address information.

[0024] S12. If the preprocessed address information does not meet the preset conditions, then perform the following operations: S13. Use the address retrieval intelligent module to find similar address information corresponding to the preprocessed address information.

[0025] S14. Use the map to call the intelligent module to call the target map application to obtain similar geographic points of interest information corresponding to the preprocessed address information.

[0026] S15. Based on the original address information, similar address information, and similar geographic points of interest information, the target address information is obtained through multi-source collaborative reasoning via the decision intelligence module.

[0027] In this embodiment, the executing entity of the address resolution method is not limited; it can be an electronic device or an address resolution device installed in the electronic device. The address resolution device can be implemented in software or a combination of software and hardware. The address resolution device can be a processor in the electronic device. For ease of understanding, the following description uses an electronic device as the executing entity.

[0028] In this embodiment of the application, the original address information may have problems such as non-standard expression, missing information, semantic ambiguity, and misuse of aliases. For example, the original address information may be "Unit F, Building X, No. 31, Street E, District C, Telephone 1234". This original address information contains the special symbol "#" and non-address noise information "Telephone 1234", and lacks the city name and specific community name.

[0029] In the embodiments of this application, the task coordination intelligent module can also be called the task coordination intelligent agent. The task coordination intelligent module can be used to preprocess the original address information to obtain preprocessed address information, and perform adaptive routing on the preprocessed address information to achieve task coordination.

[0030] The task coordination intelligent module can be associated with the address preprocessing model to preprocess the original address information and obtain preprocessed address information.

[0031] Optionally, preprocessing may include at least one of the following: anomaly detection (i.e., identification of special symbols, typos, abbreviations, missing fields, word order anomalies, etc.), sequence correction, address format conversion, etc.

[0032] For example, if the original address information is as shown in the example above, after the task coordination intelligent module calls the address preprocessing model, the preprocessed address information can be "Room F, Unit X, Building 31, Street E, District C, City B".

[0033] It should be noted that the address preprocessing model has limited processing capabilities for the original address information. It may not be able to remove all non-address noise information or fill in all missing information. Therefore, the obtained preprocessed address information may require further parsing and processing.

[0034] In step S12, the task coordination intelligent module can evaluate the address complexity of the preprocessed address information based on preset conditions, and trigger a simple parsing channel or a complex parsing channel according to the address complexity.

[0035] Optionally, the preset condition can be: the structure of the preprocessed address information conforms to any preset address template.

[0036] If the preprocessed address information meets the preset conditions, that is, the structure of the preprocessed address information conforms to the preset address template, then the address complexity of the preprocessed address information can be determined to be simple, and the simple parsing channel can be triggered; if the preprocessed address information does not meet the preset conditions, that is, the structure of the preprocessed address information does not conform to the preset address template, then the address complexity of the preprocessed address information can be determined to be complex, and the complex parsing channel can be triggered, that is, steps S13~S15 are executed.

[0037] In this embodiment, since the complexity of preprocessed address information can be evaluated in the task coordination intelligent module, and adaptive routing can be performed based on the address complexity, simple preprocessed address information is distributed to the simple parsing channel (i.e., the preprocessed address information is parsed by the parsing intelligent module), and complex preprocessed address information is distributed to the complex parsing channel (i.e., the preprocessed address information is parsed by the address retrieval intelligent module, the map calling intelligent module, and the decision intelligent module working together). Therefore, not only is the address parsing accuracy high, but the average response latency and parsing cost of address parsing are also reduced.

[0038] In this embodiment of the application, the address retrieval intelligent module can also be called the address retrieval intelligent agent. The address retrieval intelligent module can be used to find similar address information corresponding to the preprocessed address information.

[0039] The address retrieval intelligent module can be associated with a geographic embedding model. This model generates preprocessed address vectors corresponding to preprocessed address information. The module then employs an approximate nearest neighbor search algorithm to search for similar address vectors in the address index based on the preprocessed address vectors, thereby determining similar address information.

[0040] Optionally, the preprocessed address vector can be a multi-dimensional elastic vector, supporting the elastic representation of preprocessed address information in different preset dimensions. For example, the preprocessed address vector can support elastic representation in 128, 256, 512, and 1024 dimensions.

[0041] The address index can be constructed using the HNSW (Hierarchical Navigable Small World) algorithm. The address index is a multi-level graph structure, including a top level, at least one middle level, and a bottom level. Each level contains multiple nodes.

[0042] Below, in conjunction with Figure 2 The address index is explained.

[0043] Figure 2 This is a schematic diagram of an address index provided for an exemplary embodiment of this application. Please refer to [link / reference]. Figure 2 The address index can include a bottom layer, a middle layer, and a top layer. In the address index, the number of nodes decreases layer by layer from bottom to top. For example, the bottom layer includes all nodes, i.e., nodes N1 to N8. The middle layer includes some nodes extracted from the bottom layer, such as nodes N2, N5, and N7. The top layer includes some nodes extracted from the middle layer, such as nodes N2 and N7.

[0044] For any node in any layer, the node can include an address vector and an address identifier corresponding to historical address information. The address identifier is used to associate the historical address information with the address vector corresponding to that historical address information.

[0045] Optionally, the historical address information may include the historical address and its corresponding latitude and longitude coordinates, and may also include information such as postal code.

[0046] For example, node N2 may include address vector 2 and address identifier ID002 corresponding to historical address information 2. Address identifier ID002 can associate historical address information 2 and address vector 2.

[0047] In any layer, a node can be connected to multiple neighboring nodes based on the similarity between the address vectors in each node.

[0048] It's important to note that the maximum number of connections a node can have differs across different layers. The maximum number of connections refers to the maximum number of adjacent nodes a node can connect to. Nodes at the bottom layer typically have the highest maximum number of connections, while those at the top layer typically have the lowest, meaning the maximum number of connections decreases from bottom to top.

[0049] For example, such as Figure 2In the middle layer, if the maximum number of connections for a node is 2, then node N2 can connect to nodes N1 and N3, and node N2's actual number of connections is 2, which equals the maximum number of connections 2; node N1 can connect to node N2, and node N1's actual number of connections is 1, which is less than the maximum number of connections 2. If the maximum number of connections for a node is 1 in the top layer, then node N2 can connect to node N7, and its actual number of connections is 1, which equals the maximum number of connections 1.

[0050] Since each node in the address index stores the address vector corresponding to historical address information, the address retrieval intelligent module can use an approximate nearest neighbor search algorithm to find an address vector in the address index that is similar to the preprocessed address vector based on the preprocessed address vector, and use it as the similar address vector corresponding to the preprocessed address vector.

[0051] In this embodiment of the application, the search for similar address vectors in the address index based on the approximate nearest neighbor search algorithm can reduce the retrieval complexity and retrieval time, and improve retrieval efficiency.

[0052] For example, if in Figure 2 In the address index shown, based on the preprocessed address vector, address vector 1 in node N1, address vector 2 in node N2, and address vector 3 in node N3 are determined to be similar address vectors. Then, the historical address information 1 corresponding to address vector 1 can be used as similar address information 1, the historical address information 2 corresponding to address vector 2 can be used as similar address information 1, and the historical address information 3 corresponding to address vector 3 can be used as similar address information 3.

[0053] In this embodiment of the application, the map calling intelligent module can also be called the map calling intelligent agent. The map calling intelligent module can call the target map application to obtain similar geographic interest point information corresponding to the preprocessed address information.

[0054] Alternatively, the target map application can be an existing map application.

[0055] The map invocation intelligent module can generate geographic point of interest (POI) query requests and send them to the target map application. The POI query request may include preprocessed address information.

[0056] After receiving a geographic point of interest (POI) query request, the target map application can respond to the POI query request by retrieving similar POI information from the POI database based on preprocessed address information.

[0057] Optionally, the target map application can retrieve at least one candidate geographic point of interest (POI) from the geographic POI database based on the preprocessed address information, calculate at least one similarity between the candidate POI and the preprocessed address information, and then determine initial similar POIs from the candidate POIs based on the at least one similarity. The number of initial similar POIs can be at least one.

[0058] Optionally, determining initial similar point of interest information from at least one candidate geographic point of interest information based on at least one similarity can include the following two methods: Method 1: Select candidate geographic points of interest with a similarity greater than a preset similarity threshold as initial similar points of interest.

[0059] Method 2: Sort at least one candidate geographic point of interest (POI) information according to the similarity from high to low, and determine the top M candidate POI information in the similarity ranking as the initial similar POI information. Here, M is a positive integer. M can be preset. For example, M can be 3.

[0060] The target map application can package the initial similar point of interest (POI) information to generate geographic point of interest response data, and then send the geographic point of interest response data to the map invocation intelligence module. The geographic point of interest response data includes the initial similar geographic point of interest information.

[0061] It should be noted that the initial similar geographic interest point (PIP) information may include multiple initial pieces of information about similar PIPs (such as PIP type, PIP name, latitude and longitude coordinates, corresponding geographic region, detailed address, tags, postal code, service type, service rating, etc.). However, not all of this initial information may be useful for address resolution in this solution. Therefore, after receiving the PIP response data, the map invocation intelligent module can extract multiple field values ​​corresponding to multiple preset fields from any initial similar PIP information in the response data, and combine these multiple field values ​​according to a preset template to obtain the similar PIP information. The amount of information in the similar PIP information is less than or equal to the amount of information in the initial similar PIP information.

[0062] Optionally, multiple preset fields may include type, name, location (latitude and longitude coordinates), region (geographic region), and address (detailed address).

[0063] Optionally, the similar geographic point of interest (POI) information may include information such as the type, name, latitude and longitude coordinates, corresponding geographic region, and detailed address of the similar POIs. The similar POI information can be structured; for example, it can be in JSON format.

[0064] For example, if the geographic point of interest (POI) response data includes three initial similar POIs, the map calling intelligent module can extract the following from the initial similar POI 1 in the response data: the field value corresponding to type is "place name and address information; transportation place name; road name"; the field value corresponding to name is "E Road"; the field value corresponding to location is "119.941134, 30.267310"; the field value corresponding to region is "City C District, Road E"; and the field value corresponding to address is "City B, District C, Road E". Then, it combines these five field values ​​according to a preset template to obtain similar POI 1. For example, similar POI 1 could be {"type": "place name and address information; transportation place name; road name", "name": "Road E"; "location": "119.941134, 30.267310", "region": "City C District", "address": City B, District C, Road E}.

[0065] It should be noted that the execution order of steps S13 and S14 is not important; steps S13 and S14 can be executed in parallel to reduce the processing latency of address resolution.

[0066] In this embodiment of the application, the decision intelligence module can also be referred to as the decision intelligence agent. The decision intelligence module can perform multi-source collaborative reasoning based on the original address information, similar address information and similar geographic points of interest information to obtain the target address information.

[0067] In step S11, the task coordination intelligent module can send the original address information to the decision intelligent module; in step S13, after the address retrieval intelligent module obtains the similar address information corresponding to the preprocessed address information, it can send the similar address information to the decision intelligent module; in step S14, after the map retrieval intelligent module obtains the similar geographic point of interest (POI) information, it can send the similar POI information to the decision intelligent module. After receiving the original address information, similar address information, and similar POI information, the decision intelligent module can perform multi-source collaborative reasoning based on the original address information, similar address information, and similar POI information to obtain the target address information.

[0068] It should be noted that the task coordination intelligence module sends the raw address information to the decision intelligence module, rather than the preprocessed address information. This is because the raw address information can provide the original semantics, express the user's true intentions, avoid the bias of the preprocessed address, increase the heterogeneity of multi-source information, and improve the accuracy of multi-source collaborative reasoning.

[0069] The target address information may include the target address and target latitude and longitude coordinates, and may also include at least one of the following: confidence level, address identifier, address alias, address feature label, and structured address field.

[0070] The target address can be any address that conforms to any preset address template. The target address can be a complete, standard, human-readable natural language address.

[0071] A structured address field can include address values ​​corresponding to multiple preset address fields, such as city, district, street, and house number. The structured address field can be stored as key-value pairs or in JSON format.

[0072] In this embodiment, the decision intelligence module can perform multi-source collaborative reasoning on the original address information, similar address information, and similar geographic points of interest information to obtain the target address information. For example, if the original address information is as illustrated above, the target address information may include: the target address is Room F, Unit X, Building 31, H Community, E Road, C District, City B, with latitude and longitude coordinates of (119.93964, 30.267382), a confidence level of 0.95, an address feature tag of telephone number 1234, and an address identifier of ID11012.

[0073] In this embodiment, the original address information can be preprocessed using a task coordination intelligent module to obtain preprocessed address information. If the preprocessed address information does not meet preset conditions, an address retrieval intelligent module searches for similar address information that is semantically and spatially similar to the preprocessed address information. A map invocation intelligent module invokes the target map application to obtain similar geographic points of interest information matching the preprocessed address information. Finally, a decision intelligent module performs multi-source collaborative reasoning based on multi-source heterogeneous information (i.e., original address information, similar address information, and similar address point of interest information) to obtain the target address information. Thus, because multi-source heterogeneous information, including the original address information, can be obtained and cross-verified to parse the original address information, the accuracy of original address parsing is improved.

[0074] In this embodiment of the application, there are no restrictions on the specific implementation of the above step S11 "preprocessing the original address information through the task coordination intelligent module to obtain preprocessed address information".

[0075] In an exemplary embodiment, the task coordination intelligence module can be associated with an address preprocessing model, which may include a noise identification layer, a sequence correction layer, and an address output layer. Optionally, the preprocessed address information can be obtained by preprocessing the original address information through the task coordination intelligence module via the following steps S111-S113: S111. Through the noise identification layer, anomalies are identified in the original address information to obtain address noise characteristics.

[0076] S112. Through the sequence correction layer, based on the address noise characteristics, the original address information is corrected at the sequence level to generate an address correction sequence.

[0077] S113. Through the address output layer, the address correction sequence is converted into preprocessed address information that conforms to the preset address format.

[0078] Optionally, anomaly identification includes at least one of the following: special symbols, misspellings, abbreviations, missing fields, and abnormal word order.

[0079] Address noise features can be used to indicate non-address noise information in the original address information. Address noise features can include anomaly location and anomaly type. Optionally, address noise features can be structured information or feature vectors, etc.

[0080] For example, if the original address information is as shown in the example above, the original address information can be anomaly identified by the noise identification layer. If it is identified that "#" is a special symbol, "telephone 1234" is a distractor word, and the city name and community name are missing, then the address noise features can include: "#" is a special symbol, "telephone 1234" is a distractor word, and the city name and community name are missing.

[0081] Optionally, the sequence correction layer can dynamically edit the original address information at the character or word sequence level based on the address noise characteristics to generate address correction sequences with more complete address structure and more standardized semantics.

[0082] Optionally, the address correction sequence can be a noise-corrected address text sequence in natural language form.

[0083] For example, if the address noise characteristics are as shown above, the sequence correction layer can correct "#" to "building" in the original address information and remove "telephone 1234", resulting in the address correction sequence as Room F, Unit X, Building 31, Street E, District C.

[0084] Optionally, if the address correction sequence is missing city names or the hierarchical order is disordered, the address output layer can fill in the city names in the address correction sequence and sort the address elements according to the standard hierarchy to obtain preprocessed address information that conforms to the preset address format.

[0085] For example, if the address correction sequence is as shown in the example above, then through the address output layer, the city name can be inferred to be city B based on "area C" in the original address information, and "city B" can be added before area C in the address correction sequence to obtain preprocessed address information. The preprocessed address information can be: City B, area C, street D, road E, building 31, unit X, room F.

[0086] In this embodiment of the application, the original address information is preprocessed by the task coordination intelligent module to obtain preprocessed address information. Since the preprocessed address information has removed some non-address noise and filled in some missing information, the preprocessed address information is more accurate and standardized than the original address information. Subsequent parsing based on such preprocessed address improves the accuracy of address parsing.

[0087] The task coordination intelligent module can determine whether the preprocessed address information meets preset conditions. If it does, in an exemplary embodiment, the target address information can also be determined based on the following steps S16-S17: S16. If the preprocessed address information meets the preset conditions, a geocoding request is sent to the target map application through the fast parsing intelligent module.

[0088] S17. Receive the target address information returned by the target map application in response to the geocoding request.

[0089] In this embodiment of the application, the fast parsing intelligent module can also be called the fast parsing intelligent agent. The fast parsing intelligent module can be used to call the target map application to obtain target address information.

[0090] If the preprocessed address information meets preset conditions, the task coordination intelligent module can send the preprocessed address information to the fast parsing intelligent module. After receiving the preprocessed address information, the fast parsing intelligent module can generate a geocoding request and send it to the target map application. The geocoding request may include the preprocessed address information.

[0091] After receiving a geocoding request, the target map application can parse the preprocessed address information in the geocoding request according to multiple preset address fields to obtain the target address information, and then send the target address information to the fast parsing intelligent module.

[0092] Optionally, multiple preset address fields may include city, district, street, and house number, etc.

[0093] Optionally, the target address information includes a target address and / or a structured address field, wherein the structured address field includes address values ​​corresponding to multiple preset address fields. Each preset address field and its corresponding address value is a key-value pair.

[0094] For example, if the preprocessed address information is: Room F, Unit X, Building 31, E Road, Street D, District C, City B, the fast parsing intelligent module can send a geocoding request to the target map application. The geocoding request can include this preprocessed address information. The target map application can extract the address value corresponding to "City" from the preprocessed address information according to multiple preset fields, such as "City B," "District C," and so on, to obtain the target address "Room F, Unit X, Building 31, E Road, Street D, District C, City B" and a structured address field. The structured address field can include multiple key-value pairs such as city-City B and district-District C. The target map application can determine that target address information 1 includes the target address and the structured address field, and return target address information 1 to the fast parsing intelligent module.

[0095] In this embodiment of the application, when the preprocessed address information meets the preset conditions, the target map application is called by the fast parsing intelligent module to parse the preprocessed address, thus constructing a simple parsing channel. This eliminates the need to perform complex multi-source collaborative reasoning, significantly reducing parsing latency and computational overhead, and improving the processing efficiency of address parsing. This forms an intelligent parsing architecture in which simple addresses use a simple parsing channel and complex addresses use a complex parsing channel, effectively balancing address parsing efficiency and parsing accuracy.

[0096] In this embodiment of the application, there are no restrictions on the specific implementation of step S13, "finding similar address information corresponding to the preprocessed address information through the address retrieval intelligent module".

[0097] In an exemplary embodiment, the address retrieval intelligent module can be associated with a geographic embedding model. The following steps S131-S133 can be used to find similar address information corresponding to the preprocessed address information: S131. Generate a preprocessed address vector corresponding to the preprocessed address information through a geographic embedding model.

[0098] S132. Based on the preprocessed address vector, retrieve similar address vectors in the address index.

[0099] S133. Based on the address identifiers corresponding to the similar address vectors, obtain the similar address information corresponding to the similar address vectors from the address database.

[0100] Optionally, the geographic embedding model may include an input encoding layer, a context encoding layer, a feature aggregation layer, and a nested projection layer.

[0101] Optionally, in step S131, the preprocessed address vector corresponding to the preprocessed address information can be generated through a geographic embedding model in the following manner: the preprocessed address information is segmented and embedded through the input encoding layer to obtain a corresponding word sequence; the word sequence is semantically modeled through the context encoding layer to obtain a first feature vector; the first feature vector is pooled through the feature aggregation layer to obtain a second feature vector; the second feature vector is projected onto a high-dimensional embedding space through a nested projection layer to obtain an address embedding vector with a nested structure; the first N dimensions are truncated from the address embedding vector as the preprocessed address vector.

[0102] Optionally, the input encoding layer may include an address-specific word segmenter and an embedded encoder. The address-specific word segmenter can segment the preprocessed address information according to a preset dictionary to obtain multiple tokens.

[0103] For any given lexical unit, an embedding encoder can be used to map the lexical unit into a vector of fixed dimensions, resulting in the lexical vector corresponding to that lexical unit. The lexical vector can then be used to represent the semantics of the corresponding lexical unit.

[0104] Following the order of the tokens in the preprocessed address information, the token vectors corresponding to each token are combined and processed to obtain a token sequence corresponding to the processed address information. The token sequence can be in matrix form and can be used to represent the semantics of the preprocessed address information.

[0105] For example, if the preprocessed address information is: "Building 31, Unit X, Room F, Street D, District C, City B, Road E", then a dedicated address segmenter can be used to segment this preprocessed address information, resulting in 7 tokens: "City B", "District C", "Street D", "Road E", "Building 31", "Unit X", and "Room F". An embedded encoder can be used to map these 7 tokens to fixed-dimensional token vectors. For example, token vector 1 for "City B" could be [0.56, 0.12, -0.34, ..., 0.77], token vector 2 for "District C" could be [0.91, -0.22, 0.45, ..., -0.12], ..., and token vector 7 for "Room F" could be [0.85, 0.16, 0.37, ..., 0.30]. Furthermore, the token vectors corresponding to these 7 tokens can be combined according to their order in the preprocessed address information to obtain a token sequence. The first row of the lexical sequence is lexical vector 1, the second row is lexical vector 2, the third row is lexical vector 3, ..., the seventh row is lexical vector 7.

[0106] In one alternative embodiment, the context encoding layer can employ a Transformer encoder. The Transformer is a deep learning model architecture based on the attention mechanism for processing sequential data (such as natural language), comprising an encoder and a decoder. The encoder consists of multiple stacked layers of the same structure, each containing a self-attention mechanism and a feedforward neural network. Its main function is to transform the input sequence (such as a sentence of source language) into a series of continuous, semantically rich representation vectors. The decoder also consists of multiple layers, which not only include a self-attention mechanism but also introduce an "encoder-decoder attention" mechanism to focus on the input representations produced by the encoder when generating the output sequence (such as a sentence of target language). The decoder operates in an autoregressive manner, generating output tokens one by one, with the prediction of each new token depending on previously generated tokens. In practical applications, there are also variations that use only an encoder or only a decoder, each suitable for different downstream tasks.

[0107] Since the word vector corresponding to each word in the word sequence is a static embedding, that is, the contextual semantics of the word in the preprocessed address information are not considered, the word sequence can be processed by context encoding layer to perform context modeling, that is, based on the context information of each word, the word vector corresponding to the word is modified to obtain a new word vector corresponding to each word, and the combination is used to obtain the first feature vector.

[0108] The first feature vector can be in matrix form. The first feature vector can include the contextual semantics corresponding to any word in the word sequence.

[0109] For example, for “C area”, the context coding layer can determine that C area is in city B, rather than other areas with the same name, based on city B. Therefore, the word vector 2 corresponding to “C area” can be modified to obtain the new word vector 2 as [0.10, -0.20, 0.62, ..., -0.18].

[0110] In an optional embodiment, the first feature vector can be subjected to average pooling through a feature aggregation layer, that is, the first feature vector is compressed into a second feature vector of fixed length. The second feature vector can be a row vector, and the second feature vector can be used to characterize the overall semantics of the preprocessed address information.

[0111] Optionally, in the nested projection layer, a learnable projection matrix W can be used to map the low-dimensional second feature vector to a higher-dimensional space to obtain the initial projection vector. Optionally, the following projection formula can be used: Z = W * v + b Where Z represents the address embedding vector, v represents the second feature vector, W represents the projection matrix, and b represents the bias matrix.

[0112] Optionally, nested constraints are set in the nested projection layer. These nested constraints are learned by the loss function during the training of the geographic embedding model. Therefore, when generating the address embedding vector Z, the nested projection layer has already encoded the geographic hierarchical information into the prefix of the address embedding vector Z. The address embedding vector Z has a nested structure, that is, the prefix sub-vectors of multiple preset dimensions in the address embedding vector can respectively represent the semantics of the preprocessed address information at different geographic granularities.

[0113] After obtaining the address embedding vector, the first N dimensions can be extracted from it as a preprocessed address vector. Here, N is a preset dimension. The preprocessed address vector can be a multi-dimensional flexible vector. For example, if N is 1024, the first 1024 dimensions can be extracted from the address embedding vector as the preprocessed address vector.

[0114] It should be noted that the larger N is, the more accurate the representation of preprocessed address information by the preprocessed address vector. For example, a 1024-dimensional preprocessed address vector represents preprocessed address information more accurately than a 128-dimensional preprocessed address vector.

[0115] Optionally, in step S132, similar address vectors can be retrieved in the address index based on the preprocessed address vector as follows: calculate the similarity between the preprocessed address vector and the address vectors in each node of the top layer, and determine the node corresponding to the address vector with a similarity greater than a third preset threshold as the first matching node; determine at least one candidate node in the next layer based on the first matching node, and calculate the similarity between the preprocessed address vector and the address vectors in at least one candidate node, and determine the candidate node corresponding to the address vector with a similarity greater than a fourth preset threshold as the second matching node; repeat the above process until the bottom layer of the address index is reached, and determine multiple matching nodes in the bottom layer; determine similar address vectors according to the similarity between the preprocessed address vector and the address vectors in the multiple matching nodes.

[0116] Optionally, the third preset threshold can be preset. For example, the third preset threshold can be 70%.

[0117] Optionally, at least one candidate node may include a first matching node, or it may include a node that has a connection relationship with the first matching node. That is, the first matching node can be a candidate node, and a node that has a connection relationship with the first matching node can also be a candidate node.

[0118] For example, if the address index is as follows Figure 2As shown, if node N2 is the first matching node, since there is a connection between node N2 and node N5 in the middle layer, but no connection between node N7 and node N2, two candidate nodes are determined based on node N2: node N2 and node N5.

[0119] It should be noted that since the similarity between the first matching node and the preprocessed address vector has already been calculated at the top level, in an optional embodiment, when calculating the similarity between the preprocessed address vector and the address vector in at least one candidate node, if the candidate node is the first matching node, there is no need to calculate the similarity between the preprocessed address vector and the first matching node again, and the previously calculated similarity can be reused.

[0120] Optionally, the fourth preset threshold can be preset. For example, the fourth preset threshold can be 80%.

[0121] In an optional embodiment, when determining similar address vectors based on the similarity between the preprocessed address vector and the address vectors in multiple matching nodes, address vectors with a similarity greater than a fifth preset threshold can be determined as similar address vectors, or the top Q address vectors in the similarity sort from high to low can be determined as similar address vectors, where Q is a positive integer.

[0122] The address database may include multiple address identifiers and historical address information corresponding to each address identifier.

[0123] For example, if the address index is as follows Figure 2As shown, the similarity 2 between the preprocessed address vector and address vector 2 in node N2 of the top layer, and the similarity 3 between address vector 3 in node N3, can be calculated. If the third preset threshold is 70%, similarity 2 is 90%, and similarity 3 is 60%, since similarity 2 is greater than the third preset threshold of 70% and similarity 3 is less than the third preset threshold of 70%, node N2 can be determined as the first matching node. In the next layer (i.e., the middle layer) corresponding to the top layer, two candidate nodes can be determined based on node N2: node N2 and node N5. The similarity 5 between the preprocessed address vector and address vector 5 in node N5 can be calculated. If similarity 5 is 75% less than the fourth preset threshold of 80%, and similarity 2 is 90% greater than the fourth preset threshold of 80%, then node N2 can be determined as the second matching node. In the next layer (bottom layer) corresponding to the middle layer, i.e., in the bottom layer, three matching nodes can be determined based on node N2: node N1, node N2, and node N3. The similarity 1 between the preprocessed address vector and address vector 1 in node N1, and the similarity 3 between the preprocessed address vector and address vector 3 in node N3 can be calculated. If the fifth preset threshold is 90%, the similarity 1 is 95%, the similarity 2 is 90%, and the similarity 3 is 91%, then address vector 1 in node N1, address vector 2 in node N2, and address vector 3 in node N3 can be determined to be similar address vectors.

[0124] If the address identifier corresponding to address vector 1 in node N1 is ID001, the address identifier corresponding to address vector 2 in node N2 is ID002, and the address identifier corresponding to address vector 3 in node N3 is ID003, then in the address database, the historical address information 1 corresponding to address vector 1 can be determined based on address identifier ID001, the historical address information 2 corresponding to address vector 2 can be determined based on address identifier ID002, and the historical address information 3 corresponding to address vector 3 can be determined based on address identifier ID003. In this case, the three historical address information are similar address information.

[0125] In this embodiment, preprocessed address information is transformed into preprocessed address vectors through a geographic embedding model. Similar address vectors are then efficiently retrieved from the address index based on the preprocessed address vectors. The corresponding similar address information is then obtained through the address identifiers corresponding to the similar address vectors. This achieves high-precision semantic matching for non-standard, ambiguous, or incomplete addresses. Compared to traditional keyword matching and rule parsing, preprocessed address vectors are better able to capture the deep geographic semantics in preprocessed address information, effectively solving problems such as homonyms, misspellings, and address format differences. At the same time, with the help of an efficient approximate nearest neighbor index structure (i.e., address index), similar address vectors can be retrieved quickly, improving retrieval efficiency and accuracy, increasing the accuracy of determining similar address information, and thus improving the accuracy of determining target address information.

[0126] Since an address index is needed to find similar address information corresponding to preprocessed address information, in an exemplary embodiment, an address index can also be pre-constructed based on the following steps S18~S20: S18. Generate address vectors corresponding to multiple historical address information through a geographic embedding model.

[0127] S19. Retrieve the address identifiers corresponding to multiple historical address information from the address database.

[0128] S20. Based on the address vector and address identifier corresponding to each historical address information, construct an address index using the HNSW algorithm.

[0129] Optionally, multiple historical address information can be obtained from the address database, and address vectors corresponding to multiple historical address information can be generated through a geographic embedding model.

[0130] Optionally, when constructing the address index based on the address vector and address identifier corresponding to each historical address information, the maximum number of layers L in the address index can be preset (e.g., L=3). A node corresponding to a historical address information is used as the starting node, and this starting node is added to each layer simultaneously (i.e., the bottom layer, middle layer, and top layer). When inserting subsequent nodes based on this starting node, a top-down search for candidate nearest neighbors in each layer can be initiated from the starting node, and the distance between each subsequent node and the inserted node can be calculated to determine whether nodes are connected, until all nodes are inserted, resulting in the address index constructed by the HNSW algorithm.

[0131] Subsequently, the nodes corresponding to the remaining historical address information are inserted sequentially: For each node to be inserted, starting from the initial node, a top-down search strategy is used to find its nearest neighbor candidate node in each layer, and several nearest neighbors are selected and connected based on vector distance, thereby gradually constructing a multi-layer graph structure. This process is repeated until all nodes are inserted, resulting in an address index constructed based on the HNSW algorithm.

[0132] It should be noted that inserting a node means storing the address vector and address identifier corresponding to the historical address information into the address index, so that when performing address retrieval, the address identifier can be used to trace back to the historical address information.

[0133] For example, address vector 1 corresponding to historical address information 1, address vector 2 corresponding to historical address information 2, ..., address vector 8 corresponding to historical address information 8 can be generated through a geographic embedding model. Address identifier ID001 corresponding to historical address information 1, address identifier ID002 corresponding to historical address information 2, ..., address identifier ID008 corresponding to historical address information 8 can be obtained from the address database. Using the HNSW algorithm, address vector 2 and address identifier ID002 corresponding to historical address information 2 can be stored in the address index, forming node N2 in the address index; address vector 1 and address identifier ID001 corresponding to historical address information 1 can be inserted into the address index, forming node N1 in the address index, and a connection is established between node N1 and node N2; ...; until address vector 8 and address identifier ID008 corresponding to historical address information 8 are inserted into the address index, forming node N8 in the address index, resulting in... Figure 2 The address index is shown.

[0134] In this embodiment of the application, there are no restrictions on the specific implementation of the above step S15, "to obtain target address information by performing multi-source collaborative reasoning through a decision intelligence module based on original address information, similar address information, and similar geographic point of interest information".

[0135] In an exemplary embodiment, the decision intelligence module can be associated with a lightweight decision model, which may include a multi-source input layer, a credibility judgment layer, an evidence fusion layer, and a decision output layer. The target address information can be obtained through multi-source collaborative reasoning based on original address information, similar address information, and similar geographic points of interest information via the decision intelligence module through the following steps S151-S154: S151. Receive original address information, similar address information, and similar geographic points of interest information through the multi-source input layer.

[0136] S152. Through the trusted judgment layer, based on the original address information, trusted judgment is made on similar address information and similar geographic points of interest information.

[0137] S153. If either the similar address information or the similar geographic point of interest information is reliable, then the target address information is determined based on the reliable similar information through the decision output layer.

[0138] S154. If both the similar address information and the similar geographic point of interest information are reliable, then the evidence fusion layer performs conflict detection on the similar address information and the similar geographic point of interest information, and the decision output layer determines the target address information based on the conflict detection results.

[0139] In one optional embodiment, the multi-source input layer may include a first interface corresponding to the task coordination intelligent module, through which the original address information, similar address information, and similar geographic points of interest information sent by the task coordination intelligent module are received.

[0140] In this optional embodiment, after the address retrieval intelligent module obtains similar address information, it needs to return the similar address information to the task coordination intelligent module; after the map call intelligent module obtains similar geographic point of interest information, it needs to return the similar geographic point of interest information to the task coordination intelligent module.

[0141] In another optional embodiment, the multi-source input layer may include a first interface corresponding to the task coordination intelligence module, a second interface corresponding to the address retrieval intelligence module, and a third interface corresponding to the map invocation intelligence module. The multi-source output layer can receive raw address information sent by the task coordination intelligence module through the first interface, receive similar address information sent by the address retrieval intelligence module through the second interface, and receive similar geographic points of interest information sent by the map invocation intelligence module through the third interface.

[0142] In this optional embodiment, after the address retrieval intelligent module obtains similar address information, it does not need to return similar address information to the task coordination intelligent module; after the map calling intelligent module obtains similar geographic point of interest information, it does not need to return similar geographic point of interest information to the task coordination intelligent module.

[0143] In this embodiment of the application, the trust judgment layer in the lightweight decision model may include a first trust submodule and a second trust submodule. The first trust submodule can be used to make trust judgments on similar address information; the second trust submodule can be used to make trust judgments on similar geographic points of interest information.

[0144] In an optional embodiment, in the trust judgment layer, the first trust submodule can perform trust judgment on similar address information in the following manner: based on the latitude and longitude coordinates in the similar address information, obtain the grid index identifier corresponding to the similar address information in the hexagonal hierarchical spatial index database to obtain at least one grid index identifier; perform deduplication processing on the at least one grid index identifier to obtain K grid index identifiers, where K is a positive integer; if K is less than or equal to a preset number, the similar address information is determined to be trustworthy; if K is greater than the preset number, the similar address information is determined to be untrustworthy.

[0145] A hexagonal hierarchical spatial index database can be constructed based on a hexagonal grid system. In a hexagonal grid system, the Earth's surface can be divided into multiple nested hexagonal grids, each with a different granularity. Each hexagon has a unique grid index identifier. Therefore, a hexagonal hierarchical spatial index database can include multiple hexagonal grids at different levels, as well as the latitude and longitude range and grid index identifier corresponding to each hexagonal grid.

[0146] Optionally, for any similar address information, the first trusted submodule can call the interface of the hexagonal hierarchical spatial index database, input the latitude and longitude coordinates in the similar address information into the hexagonal hierarchical spatial index database, determine the hexagonal grid at the specified granularity to which the latitude and longitude coordinates belong, and determine the grid index identifier corresponding to the hexagonal grid as the grid index identifier corresponding to the similar address information.

[0147] The first trusted submodule can perform deduplication on at least one grid index identifier to obtain K grid index identifiers. K can be preset.

[0148] If K is less than or equal to the preset number, it indicates that the similar address information is clustered. In other words, the addresses indicated by the multiple similar address information are very close to each other and are very likely to point to the same location. Therefore, the similar address information can be determined to be reliable.

[0149] If K is greater than the preset number, it indicates that the similar addresses are discrete, meaning that the addresses indicated by the multiple similar address information are far apart and may point to different locations. It is uncertain which similar address information is accurate, so the similar address information can be judged as unreliable.

[0150] For example, if there are three similar address information entries, where similar address information 1 includes latitude and longitude coordinates A (119.93964, 30.267382), the first trusted submodule can input this latitude and longitude coordinate A into the hexagonal hierarchical spatial index database. If this latitude and longitude coordinate A belongs to hexagonal grid 10001, and the grid index identifier corresponding to hexagonal grid 10001 is 894a6652c7fffff, then the grid index identifier 1 corresponding to similar address information 1 can be determined to be 894a6652c7fffff. Similarly, the grid index identifier 2 corresponding to similar address information 2 can be determined to be 894a6652c7fffff, and the grid index identifier 3 corresponding to similar address information 3 can be determined to be 894a665283fffff, thus obtaining three grid index identifiers. The first trusted module can then deduplicate these three grid index identifiers, obtaining two grid index identifiers, i.e., K=2. If the preset quantity is 2, since K equals the preset quantity of 2, the three similar address information can be determined to be reliable.

[0151] It should be noted that the credibility assessment of similar address information involves treating all similar address information as a whole. If it is considered credible, then all similar address information is trustworthy; if it is considered untrustworthy, then all similar address information is untrustworthy.

[0152] Optionally, in the trust judgment layer, the second trust submodule can perform trust judgment on similar geographic point of interest information based on the original address information in the following way: semantic extraction is performed on the original address information to obtain keywords in the original address information; for any similar geographic point of interest information, if the similar geographic point of interest information includes keywords, then the similar geographic point of interest information is determined to be trustworthy; if the similar geographic point of interest information does not include keywords, then the similar geographic point of interest information is determined to be untrustworthy.

[0153] For example, if the original address information is as shown in the example above, the second trusted submodule can extract the keyword "E Road" from the original address information. If similar geographic point of interest information 1 is as shown in the example above, the second trusted submodule can determine that similar geographic point of interest information 1 includes the keyword "E Road", and thus determine that similar geographic point of interest information 1 is trusted; if similar geographic point of interest information 2 includes the keyword, then similar geographic point of interest information 2 is trusted; if similar geographic point of interest information 3 does not include the keyword "E Road", then similar geographic point of interest information 3 is untrustworthy.

[0154] In this embodiment of the application, if either the similar address information or the similar geographic point of interest information is reliable, the target address information is determined based on the reliable similarity information, including the following two cases: Case 11: Similar address information is reliable, but similar geographic points of interest information is not reliable.

[0155] In this case, the target address information can be determined based on similar address information.

[0156] Since there is at least one similar address information, and the similar address information is reliable, meaning that all of the at least one similar address information is reliable, the decision output layer can determine the target similar address information from the at least one similar address information, and determine the target address information based on the target similar address information.

[0157] The target similar address information can be the similar address information among the at least one similar address information whose similarity with the preprocessed address information is greater than a second preset threshold.

[0158] Optionally, the second preset threshold can be preset. For example, the second preset threshold can be 90%.

[0159] For example, if there are 3 similar address information, and the similarity between similar address information 1, similar address information 2, and similar address information 3 and the preprocessed address information is 93%, 89%, and 88% respectively, and if the second preset threshold is 90%, then similar address information 1 can be determined as the target similar address information, and the target address information can be determined based on similar address information 1.

[0160] Case 12: The similar address information is unreliable, but there is reliable similar geographic point of interest information.

[0161] In this case, the target address information can be determined based on credible similar geographic points of interest information.

[0162] Since there is at least one credible similar geographic point of interest information, the decision output layer can determine the target geographic point of interest information from at least one credible similar geographic point of interest information, and determine the target address information based on the target geographic point of interest information.

[0163] The target geographic point of interest information can be the trusted similar geographic point of interest information among the at least one trusted similar geographic point of interest information, whose similarity with the preprocessed address information is greater than a first preset threshold.

[0164] Optionally, the first preset threshold can be preset. For example, the first preset threshold can be 90%.

[0165] For example, if there are two credible similar geographic point of interest information, namely similar geographic point of interest information 1 and similar geographic point of interest information 2, and if the similarity between similar geographic point of interest information 1 and similar geographic point of interest information 2 and the preprocessed address information is 91% and 85% respectively, and if the first preset threshold is 90%, then similar geographic point of interest information 1 can be determined as the target similar address information, and thus similar geographic point of interest information 1 can be determined as the target address information.

[0166] Optionally, if both the similar address information and the similar geographic point of interest information are reliable, then in step S154, conflict detection can be performed on the similar address information and the similar geographic point of interest information through the evidence fusion layer via steps A1 to A4 as follows: A1. Determine the target similar address information from the similar address information, and extract the first latitude and longitude coordinates and the first geographical region from the target similar address information.

[0167] A2. Identify the target geographic point of interest information from similar geographic point of interest information, and extract the second latitude and longitude coordinates and the second geographic region from the target geographic point of interest information.

[0168] A3. Calculate the target distance based on the first and second latitude and longitude coordinates.

[0169] A4. Based on the target distance, the first geographic region, and the second geographic region, determine whether there is a conflict between similar address information and similar geographic point of interest information.

[0170] Optionally, the evidence fusion layer can determine target similar address information from at least one similar address information. The target similar address information can be the similar address information among the at least one similar address information whose similarity to the preprocessed address information is greater than a second preset threshold.

[0171] The evidence fusion layer can extract the first latitude and longitude coordinates and the first geographic region from the target similar address information. For example, the evidence fusion layer can determine that similar address information 1 is the target similar address information, and extract the first latitude and longitude coordinates as latitude and longitude coordinates A (119.93964, 30.267382) and the first geographic region as region C from similar address information 1.

[0172] Optionally, the evidence fusion layer can determine the target geographic point of interest (OPI) information from at least one similar OPI information. The target OPI information can be the OPI information among the at least one similar OPI information whose similarity to the preprocessed address information is greater than a first preset threshold.

[0173] The evidence fusion layer can extract second latitude and longitude coordinates and a second geographic region from the target geographic point of interest information. For example, the evidence fusion layer can identify similar geographic point of interest information 1 as the target geographic point of interest information, and extract the second latitude and longitude coordinates as latitude and longitude coordinates B (119.941134, 30.267310) and the second geographic region as region C from this similar geographic point of interest information 1. The evidence fusion layer can calculate that the target distance between latitude and longitude coordinates A (119.93964, 30.267382) and latitude and longitude coordinates B (119.941134, 30.267310) is approximately 129 meters.

[0174] In step A4, based on the target distance, the first geographic region, and the second geographic region, it is determined whether there is a conflict between similar address information and similar geographic point of interest information, including the following two cases: Case 21: If the target distance is less than or equal to the distance threshold, and the first geographic region is the same as the second geographic region.

[0175] In this case, it is determined that there is no conflict between similar address information and similar geographic point of interest information.

[0176] Optionally, the distance threshold can be preset. For example, the distance threshold can be 500 meters.

[0177] For example, if the target distance is 129 meters, which is less than the distance threshold of 500 meters, and the first geographic region and the second geographic region are both in area C, then the conflict result between similar address information and similar geographic point of interest information can be determined as no conflict.

[0178] Case 22: If the target distance is greater than the distance threshold, and / or the first geographic region is inconsistent with the second geographic region.

[0179] In this case, it is determined that there is a conflict between similar address information and similar geographic point of interest information.

[0180] For example, if the target distance is 600 meters, which is greater than the distance threshold of 500 meters, it can be determined that there is a conflict between similar address information and similar geographic point of interest information; if the first geographic region is area C and the second geographic region is area R, it can also be determined that there is a conflict between similar address information and similar geographic point of interest information.

[0181] Optionally, if there is no conflict between the similar address information and the similar geographic point of interest (POI) information, the decision output layer can determine the target address information based on the target POI information in the similar POI information. If there is a conflict between the similar address information and the similar POI information, the target address information is determined based on the target similar address information in the similar address information.

[0182] If there is no conflict between similar address information and similar geographic point of interest (POI) information, meaning they both point to the vicinity of the same address, then since similar POI information is more accurate, structured, and authoritative than similar address information, and naturally supports semantic understanding (users often use POI names to describe locations), while similar address information is determined based on historical address information and may contain noise from user input, outdated information, or non-standard addresses, the target address information can be determined based on the more accurate and authoritative target similar POI information when there is no conflict between them.

[0183] If there is a conflict between similar address information and similar geographic points of interest information, the similar address information determined based on historical address information is more representative of the real address (e.g., similar address information has been successfully delivered). Therefore, the target address information can be determined based on the target similar address information in the similar address information.

[0184] It should be noted that in step S153 or step S154 above, when determining the target address information based on the target geographic point of interest information or the target similar address information, the backbone address elements (such as address elements refined to the community level) in the target address information can be determined based on the target geographic point of interest information or the target similar address information. Fine-grained address elements such as building, unit, and house number can be extracted from the original address information. Then, the backbone address elements and fine-grained address elements are combined to obtain the target address information.

[0185] For example, if similar address information 1 is the target similar address information, and similar address information 1 includes: Unit G, Building 31, Road E, District C, City B (New H Community), with latitude and longitude coordinates A (119.93964, 30.267382), then based on similar address information 1, the main address element can be determined to be H Community, Road E, District C, City B. Fine-grained address elements can be extracted from the original address information, including: Building 31, Unit X, Room F. Then, the main address element and fine-grained address elements can be combined to obtain the target address: Unit F, Building X, H Community, Road E, District C, City B. Alternatively, the latitude and longitude coordinates A in similar address information 1 can be determined as the latitude and longitude coordinates (119.93964, 30.267382) corresponding to the target address. Then, the target address information can include Unit F, Building X, H Community, Road E, District C, City B, as well as the latitude and longitude coordinates (119.93964, 30.267382).

[0186] Optionally, the decision output layer can also identify "telephone 1234" as the address feature label corresponding to the target address information in the original address information.

[0187] In this embodiment, a lightweight and robust multi-source collaborative address resolution mechanism is constructed, significantly improving the accuracy and practicality of address resolution. The lightweight decision model first integrates multi-source information such as address databases and geographic POIs to expand the coverage of address representation. Then, a trust judgment layer performs trust judgment on similar address information and similar geographic points of interest based on keywords or structural features in the original address information. When a single source is trustworthy, the target address information can be directly determined based on the trustworthy source to ensure efficiency. When multiple sources are all trustworthy, conflict detection (such as latitude and longitude consistency, geographic region consistency, etc.) can be further performed through an evidence fusion layer, and the decision output layer intelligently determines the target address information based on the conflict detection results. This design not only avoids the limitations of a single data source but also maintains high stability in complex scenarios such as identical names in different locations, incomplete addresses, and ambiguous descriptions. It also balances lightweight design with real-time performance, efficiently outputting complete target address information containing standard text and latitude and longitude, directly supporting precise positioning and automated processing in downstream applications such as logistics and map services, significantly improving the accuracy, robustness, and engineering feasibility of address resolution.

[0188] Below, based on the above embodiments, and in conjunction with Figure 3 The above address resolution method will be further explained through specific examples.

[0189] Figure 3 This is a schematic diagram illustrating an address resolution method provided for an exemplary embodiment of this application. Please refer to [link / reference]. Figure 3 The task coordination intelligent module can be associated with the address preprocessing model.

[0190] If the original address information is: Room F, Unit X, Building 31, E Road, Street D, District C, with phone number 1234, then the task coordination intelligent module can preprocess the original address information through the address preprocessing model to obtain the preprocessed address information as: Room F, Unit X, Building 31, E Road, Street D, District C, City B.

[0191] The task coordination intelligent module can determine whether the preprocessed address information meets preset conditions. If the preprocessed address information meets the preset conditions, the task coordination intelligent module can send the preprocessed address information to the fast parsing intelligent module. The fast parsing intelligent module can call the target map application to parse the preprocessed address information to obtain target address information 1. Target address information 1 may include a target address and / or a structured address field. The structured address field may include address values ​​corresponding to multiple preset address fields, such as... Figure 3 As shown, the target address information 1 may include address values ​​corresponding to multiple preset address fields, namely: city - city B, district - district C, street - street D, road - road E, building - building 31, unit - unit X, and door number - room F.

[0192] If the preprocessed address information does not meet the preset conditions, the task coordination intelligent module can send the preprocessed address information to the address retrieval intelligent module and the map calling intelligent module.

[0193] The address retrieval intelligent module can be associated with a geographic embedding model. Through the geographic embedding model, the address retrieval intelligent module generates preprocessed address vectors corresponding to preprocessed address information. Based on these preprocessed address vectors, it retrieves three similar address vectors from the address index. Then, based on the address identifiers corresponding to these three similar address vectors, it obtains the similar address information corresponding to these three similar address vectors from the address database, namely, similar address information 1, similar address information 2, and similar address information 3, as shown in Table 1 below: Table 1 The address retrieval intelligent module can return the information of these three similar addresses to the task coordination intelligent module.

[0194] The map invocation intelligent module can invoke the target map application to obtain three similar geographic interest points (PIs) corresponding to the preprocessed address information, namely similar PI 1, similar PI 2, and similar PI 3, as shown in Table 2 below: Table 2 The map calling intelligent module can return information on the three similar geographic points of interest to the task coordination intelligent module.

[0195] After receiving the three similar address information as shown in Table 1 above and the three similar geographic points of interest information as shown in Table 2 above, the task coordination intelligent module can send the original address information, the three similar address information and the three similar geographic points of interest information to the decision intelligent module.

[0196] The decision intelligence module can be linked to a lightweight decision model. Based on the original address information, the decision intelligence module can make a reliable judgment on three similar address information and three similar geographic points of interest using the lightweight decision model.

[0197] If the three similar address information sets are reliable, the target address information can be determined based on the target similar address information among these three sets. Similarly, if the three similar geographic points of interest (GPIs) are reliable, the target address information can be determined based on the target GPI information among these three sets. If both the three similar address information sets and the three similar GPIs are reliable, it can be further determined whether the three similar address information sets conflict with the three similar GPIs. If there is no conflict, the target address information can be determined based on the target GPI information among these three sets; if there is a conflict, the target address information can be determined based on the target similar address information among these three sets.

[0198] When determining the target address information, the decision intelligence module can also perform semantic understanding of the original address information and determine that "telephone 1234" in the original address information can be used as an address feature label.

[0199] For example, if similar address information 1 out of three similar address information is the target similar address information, then target address information 2 can be determined based on similar address information 1, such as... Figure 3 As shown, target address information 2 may include target address 2 as: Room F, Unit X, Building 31, H Community, E Road, C District, City B, with a confidence level of 0.95 and address feature label as telephone number 1234.

[0200] In this embodiment, a multi-intelligent module collaborative framework is constructed based on a task coordination intelligent module, a fast parsing intelligent module, an address retrieval intelligent module, a map retrieval intelligent module, and a decision-making intelligent module. This framework decomposes the address parsing task into a multi-intelligent module collaborative processing mechanism, significantly improving address parsing accuracy. In real-world service scenarios, after applying this embodiment, the address parsing accuracy increased to 93%, with a 61.24% hit rate within 100 meters and a 93.91% hit rate within 500 meters. This effectively meets the demand for precise positioning in relevant services, reduces the proportion of abnormal address locations, decreases the workload of manual address verification, and saves significant labor costs.

[0201] Furthermore, in this embodiment, an adaptive routing system is designed in the task coordination intelligence module. This system can dynamically assess address complexity, offloading simple tasks to low-cost, high-speed channels, while complex tasks are handled by the expert intelligence module. This design significantly optimizes the average response time and computational cost of address resolution while ensuring high accuracy.

[0202] Furthermore, compared to the above-mentioned Scheme 1, the decision intelligence module and address retrieval intelligence module in this application embodiment have powerful semantic understanding capabilities, can automatically handle problems such as aliases, word order reversal, and noise, and do not require manual maintenance of fragile rules, realizing a leap from "hard coding" to "soft learning"; and the introduction of the address retrieval intelligence module and map calling intelligence module makes this application embodiment inherently possess geocoding capabilities, and can directly output high-precision latitude and longitude coordinates.

[0203] Compared to the second scheme mentioned above, this embodiment achieves complementary data sources. When existing map applications fail to resolve the addresses of newly built buildings or self-built houses, the address retrieval intelligent module can obtain real historical address information from the address database as similar address information, thereby improving effective location clues and greatly enhancing the resolution accuracy of non-standard addresses. Furthermore, the decision intelligent module can critically view the similar geographic points of interest information returned by the map calling intelligent module calling the target map application. It can combine similar address information for cross-validation and conflict detection, avoiding blindly accepting the low-precision results returned by the target map application, and improving the controllability and accuracy of address resolution.

[0204] Compared to the above-mentioned scheme three, the embodiment of this application does not rely on a single data source, but integrates similar address information filtered from massive historical address information in the address database and similar geographic points of interest information returned by the target map application. The information sources are wider and the robustness is stronger. Moreover, the decision intelligence module can not only perform simple information aggregation, but also perform complex logical reasoning based on the mutual verification of multi-source heterogeneous information, thereby improving the accuracy of address parsing.

[0205] In addition, compared with the end-to-end model, the embodiments of this application have stronger geographic perception capabilities and interpretability: by integrating geospatial supervision signals such as real latitude and longitude and grid index identifiers, and adopting nested address vector representation, it can not only output structured address fields, but also directly support high-precision coordinate positioning; at the same time, multi-source input (original address information, similar address information, similar geographic interest point information) and hierarchical decision-making mechanism make the address resolution process transparent and controllable, which facilitates error attribution and system iteration, effectively overcomes the opacity of the black box generation process of the end-to-end model, and improves the stability of address resolution.

[0206] The embodiments of this application are also superior to traditional multi-task learning models: they do not rely on a large amount of fine-grained manual annotation (such as segment labels), but automatically generate training signals using massive amounts of real historical address and map data, reducing data costs; through a unified vector retrieval and preference alignment mechanism, address element identification, conflict detection and final decision are internalized in the collaborative reasoning process, avoiding dependencies and conflicts between multiple tasks, and realizing a data-driven, end-to-end, lightweight and efficient intelligent address resolution closed loop.

[0207] In any of the above embodiments, the address retrieval intelligent module can be associated with a geographic embedding model, which is obtained by training a first initial model based on MRL (Matryoshka Representation Learning).

[0208] Below, in conjunction with Figure 4 The training process of the geographic embedding model is explained.

[0209] Figure 4 A flowchart illustrating a model training method provided for an exemplary embodiment of this application. Figure 1 Please see. Figure 4 The method includes: S41. Obtain training samples.

[0210] S42. Using the first initial model, a predicted address vector with a nested structure is generated based on the original address information. Based on the multi-level prefix sub-vectors in the predicted address vector, the corresponding predicted latitude and longitude coordinates and predicted grid index identifiers are predicted respectively.

[0211] S43. Based on the real latitude and longitude coordinates, the real grid index identifier, the predicted latitude and longitude coordinates, and the predicted grid index identifier, calculate the loss value of the first initial model through a hierarchical joint loss function.

[0212] S44. Based on the loss value, update the model parameters of the first initial model until the model training termination condition is met, and obtain the geographic embedding model.

[0213] For any training sample, the training sample may include original address information, real latitude and longitude coordinates, and real grid index identifiers. The original address information can be unstructured address text from user input or historical records. The real latitude and longitude coordinates are the actual geographic coordinates corresponding to the original address information, serving as precise supervisory labels for continuous spatial locations. The real grid index identifiers are discrete region identifiers generated at different geographic granularities in a hexagonal hierarchical spatial indexing system based on the latitude and longitude coordinates corresponding to the original address information, used to characterize the regional affiliation of the original address at different geographic granularities.

[0214] Optionally, the training samples may also include labeled address vectors, which are high-dimensional vectors with nested structures.

[0215] Training samples can be derived from historically successfully parsed address records, authoritative geographic databases, or high-precision POI data to ensure authenticity and reliability.

[0216] During the training of the first initial model, the introduction of dual supervision signals, namely latitude and longitude coordinates and grid index identifiers, not only preserves the location information of continuous space, but also introduces discretized hierarchical regional semantics, providing structured priors for subsequent nested representation learning and improving the first initial model's ability to perceive geographical layers with different coverage areas. The nested design of labeled address vectors can guide the first initial model to learn hierarchical address representations, laying the foundation for subsequent hierarchical loss calculation and multi-granularity retrieval.

[0217] Optionally, the first initial model can be a pre-trained language model. For example, the first initial model can be a model based on the Transformer architecture.

[0218] In step S42, the original address information is input into the first initial model. After text encoding and context encoding, a high-dimensional semantic vector is output. Subsequently, this vector is mapped to the maximum target dimension (e.g., 1024 dimensions) through a shared underlying adaptation network to obtain the predicted address vector. This predicted address vector has a nested structure, and its arbitrary prefix sub-vectors (e.g., the first 64 dimensions, the first 128 dimensions, and the first 256 dimensions) constitute sub-vectors of the corresponding geographic granularity. That is, the predicted address vector can include multiple layers of prefix sub-vectors.

[0219] For each level of prefix subvector, the first N dimensions can be extracted from the predicted address vector and input into a georegression head. The georegression head then predicts the corresponding latitude and longitude coordinates, which are then constrained to a legal geographic range through post-processing. This prefix subvector can be input into multiple grid index identifier classification heads to predict the corresponding predicted grid index identifiers. The predicted grid index identifiers can include discrete region identifiers at different geographic granularities.

[0220] In one optional embodiment, the hierarchical joint loss function can be a weighted sum of the loss terms corresponding to each level (i.e., each target dimension), and the loss terms corresponding to each level include: Haversian distance loss, planar distance loss, and hexagonal hierarchical spatial index classification loss.

[0221] Haversine distance loss can be obtained by calculating the great circle distance between predicted latitude and longitude coordinates and actual latitude and longitude coordinates based on spherical geometry, and is applicable to global-scale positioning.

[0222] Planar distance loss refers to the Euclidean distance calculated by approximating the Earth as a plane in a local region. Using planar distance loss can accelerate convergence and improve accuracy in small areas.

[0223] The hexagonal hierarchical spatial index classification loss refers to the cross-entropy loss between predicted grid index labels and true grid index labels. The hexagonal hierarchical spatial index classification loss can enhance the initial model's ability to discriminate the boundaries of discrete regions.

[0224] The Havelsein distance loss, planar distance loss, and hexagonal hierarchical spatial index classification loss can be weighted and summed according to geographical level or task importance to form an end-to-end hierarchical joint loss function.

[0225] In step S44, based on the loss value calculated in step S43, all trainable parameters of the first initial model (including the parameters of the adaptation network) can be updated by backpropagation until the training termination condition is met, thus obtaining the geographic embedding model. The geographic embedding model can generate a high-dimensional address vector with a nested structure based on the input address text.

[0226] The training termination condition may include at least one of the following: reaching the maximum number of training rounds, convergence of the validation set loss, or geolocation error being less than a preset threshold.

[0227] The trained geographic embedding model can output a high-dimensional address vector with a nested structure based on the input address text. Any prefix of this high-dimensional address vector (such as the first 64 / 128 / 256 dimensions) can be used independently for geographic retrieval, naturally supporting multi-granularity, truncated approximate nearest neighbor search; and the vector space embeds continuous coordinate semantics and discrete regional hierarchical structure, effectively solving address ambiguity problems such as "same name, different location" and "cross-level model".

[0228] In this embodiment, during the model training phase, the original address information is used as input. Latitude and longitude coordinates and hexagonal grid index identifiers are introduced as dual supervision signals, and an MRL prefix nesting mechanism is used to train the first initial model. This allows the first initial model to learn multi-scale geographic representations in a single forward pass. The resulting geographic embedding model can map address text into structured, measurable, and truncated geographic semantic vectors without manual rules or post-processing. Essentially, it constructs a dynamic, fine-grained, and highly synchronized private geographic knowledge base with the real world, effectively compensating for the shortcomings of public map data in terms of coverage breadth, update timeliness, and end-point details. This is significantly superior to traditional address parsing schemes based on string matching or single vector retrieval.

[0229] In any of the above embodiments, the decision intelligence module can be associated with a lightweight decision model, which can be obtained by knowledge distillation of the decision model, and the decision model can be obtained by supervised fine-tuning and preference training.

[0230] Below, based on any of the above embodiments, combined with Figure 5 The training process of the lightweight decision-making model is explained.

[0231] Figure 5 A flowchart illustrating a model training method provided for an exemplary embodiment of this application. Figure 2 Please see. Figure 5 The method includes: S51. Obtain training samples, which include the first original address information, similar address information, similar geographic points of interest information, and labeled addresses.

[0232] S52. Based on the training samples, the second initial model is fine-tuned under supervision to obtain the intermediate model.

[0233] S53. Obtain triplet samples, which include prompt words, correct examples, and incorrect examples.

[0234] S54. The intermediate model is trained with preference using triplet samples to obtain the decision model.

[0235] S55. Perform knowledge distillation on the decision model to obtain a lightweight decision model.

[0236] In any training sample, the initial raw address information can be unstructured address text input by the user. Similar address information can be semantically similar historical address information retrieved by the address retrieval intelligent module. Similar geographic points of interest (POI) information can be POI results returned by the target map application invoked by the map invocation intelligent module. The labeled address can be a standard target address labeled by humans, which can serve as the "correct answer" for supervised learning.

[0237] Optionally, the second initial model can be a large language model using the Transformer architecture.

[0238] In step S52, any training sample is input into the second initial model, with the labeled address as the target output. Standard sequence-to-sequence or autoregressive language modeling loss is used to perform SFT (Supervised Fine-Tuning) on ​​the second initial model. The second initial model can extract key information from multi-source data (i.e., the first original address information, similar address information, and similar geographic points of interest information) and generate standard addresses with standardized format and accurate semantics. After fine-tuning, an intermediate model is obtained.

[0239] By supervising and fine-tuning the second initial model, an intermediate model is obtained. The intermediate model can initially generate multi-source collaborative understanding and address standardization capabilities, laying the foundation for more refined preference alignment in the future; at the same time, it avoids training from scratch, significantly reducing data and computing costs.

[0240] In any triplet sample, the prompt word includes the first original address information, similar address information, and similar geographic points of interest information. The prompt word integrates the first original address information, similar address information, and similar geographic points of interest information to form a complete contextual input.

[0241] A correct example is a correct address that is highly consistent with the labeled address in terms of semantics (consistent location) and format (complete city area and street number, no noise).

[0242] Error examples: Error addresses generated by intermediate models or other strategies, such as semantic errors (locating to other cities), format errors (missing district names), or the inclusion of raw noise (retaining special symbols such as "#").

[0243] Based on triplet samples, providing correct and incorrect examples can guide intermediate models to learn human-preferred address representation conventions, effectively suppressing address noise and format confusion, and improving the reliability and usability of generating target address information.

[0244] In step S54, a preference alignment algorithm can be used to train the intermediate model to generate correct examples rather than incorrect examples with a higher probability.

[0245] Alternatively, the preference alignment algorithm can be DPO (Direct Preference Optimization), Pairwise Ranking Loss, or RLHF (Reinforcement Learning from Human Feedback).

[0246] Preference training enables intermediate models to generate correct target address information, improving decision robustness in ambiguous, conflicting, or noisy scenarios.

[0247] In step S55, a decision model can be used as the teacher model, and a smaller neural network (such as a small Transformer) can be used as the student model, trained using knowledge distillation techniques. The student model learns to mimic the output distribution of the teacher model under the same input (such as the final target address information), thereby significantly reducing the number of parameters and inference latency while retaining most of the performance, ultimately resulting in a lightweight decision model suitable for high-concurrency, low-latency online service scenarios.

[0248] By performing knowledge distillation on the decision model, the model size can be compressed and computational overhead reduced with almost no loss of address resolution accuracy, resulting in a lightweight decision model. This lightweight model can support real-time processing of hundreds of millions of address resolution requests, improving the overall throughput and resource efficiency of the multi-intelligent module collaborative framework.

[0249] In the embodiments of this application, there are no restrictions on the specific implementation of the above step S52 "supervised fine-tuning of the second initial model based on training samples to obtain an intermediate model".

[0250] Optionally, the second initial model includes a multi-head self-attention layer, which includes a query matrix, a key matrix, a value matrix, and an output matrix. In an exemplary embodiment, supervised fine-tuning of the second initial model based on training samples can be performed through the following steps S521-S524 to obtain an intermediate model: S521. Inject low-rank adapters into the query matrix, key matrix, value matrix, and output matrix, and freeze the original weight parameters of the second initial model.

[0251] S522. Using the second initial model, a predicted address is generated based on the first original address information, similar address information, and similar geographic points of interest information.

[0252] S523. Construct a supervised loss function based on the predicted address and the labeled address.

[0253] S524. Update the parameters of the low-rank adapter according to the supervised loss function to obtain the intermediate model.

[0254] To avoid the high computational cost and forgetting risk caused by fine-tuning all parameters of the second initial model, a low-rank adapter can be injected into the key weight matrices of the multi-head attention layer of the initial model, namely the query matrix (Q), key matrix (K), value matrix (V), and output matrix (O). The low-rank adapter is the product of two low-rank matrices.

[0255] For example, if the low-rank adapter is the product of low-rank matrices E1 and E2, then injecting the low-rank adapter into the query matrix Q will yield an updated query matrix Q1, i.e., Q1 = Q + E1 * E2. Here, the low-rank matrices E1 and E2 are trainable matrices, while the original query matrix Q remains frozen during training and does not participate in the training process.

[0256] By injecting a low-rank adapter and training its parameters to fine-tune the second initial model, the number of trainable parameters is significantly reduced, training costs are lowered, and the address resolution task is learned with only a small number of adaptation parameters. This enables efficient and stable domain specialization of the second initial model.

[0257] In step S522, the first original address information, similar address information, and similar geographic interest point information from the training samples can be concatenated into a structured prompt word according to a preset template and input into the second initial model injected with a low-rank adapter. The second initial model can generate formatted predicted addresses based on the frozen backbone network and the trainable low-rank adapter.

[0258] In step S523, the second initial model adopts a word-by-word alignment supervised training method. This model generates a predicted address sequence based on the prompt word. The word probability distribution output at each step in the sequence generation process is matched word-by-word with the target word at the corresponding position in the labeled address sequence. The cumulative value of word-by-word loss is calculated by the sequence-level cross-entropy loss function and used as the training loss value of the second initial model.

[0259] This loss value can comprehensively measure the deviation between the predicted address and the labeled address in terms of semantic accuracy and format regularity, and serve as the basis for optimization of the second initial model in the supervised fine-tuning stage.

[0260] In step S524, a gradient descent-type optimizer can be used to backpropagate and update the parameters (i.e., matrices E1 and E2) in the low-rank adapter, while all the original weights of the second initial model remain frozen until the training termination condition is met (such as reaching the maximum number of training iterations or the loss value is below a threshold), thus obtaining an intermediate model, which is a lightweight fine-tuning model that has achieved efficient adaptation on the address resolution task.

[0261] In this embodiment, a high-efficiency, economical, and scalable address decision model training paradigm is constructed by introducing a low-rank adapter into the multi-head self-attention layer of the second initial model and combining multi-source data input with supervised fine-tuning. This training method achieves accurate adaptation of large models in the address resolution domain with minimal training overhead, without increasing address resolution latency, effectively balancing performance, cost, and deployment feasibility. It provides key technical support for building a lightweight, high-accuracy multi-intelligent module collaborative framework.

[0262] In this embodiment of the application, there are no restrictions on the specific implementation of the above step S54 "training the intermediate model with preferences using triplet samples to obtain the decision model".

[0263] In an exemplary embodiment, the decision model can be obtained by training the intermediate model with preferences using triplet samples through the following steps S541-S544: S541. For any triplet sample, based on the prompt word, calculate the first generation probability of the correct example and the second generation probability of the incorrect example through the intermediate model.

[0264] S542. Construct a preference loss function based on the first generation probability and the second generation probability.

[0265] S543. Update the model parameters of the intermediate model according to the preference loss function until the model training termination condition is met, and obtain the decision model.

[0266] In step S541, for any triplet sample, the prompt words in the triplet sample can be input into the intermediate model, and the first generation probability of the intermediate model generating a correct example and the second generation probability of generating an incorrect example are calculated, with the correct example and the incorrect example as the target output sequence respectively.

[0267] The first generation probability represents the likelihood that the intermediate model will generate a correct example under the same input; the second generation probability represents the likelihood that the intermediate model will generate an incorrect example under the same input.

[0268] Based on the first and second generation probabilities, a preference loss function is constructed to encourage intermediate models to generate more correct examples and suppress the generation of incorrect examples. The loss value of the intermediate model can be calculated using the preference loss function.

[0269] In step S543, a gradient descent-type optimization algorithm can be used to backpropagate and iteratively update the trainable parameters of the intermediate model according to the preference loss function until the training termination condition is met (such as reaching the maximum number of training iterations or the loss value being lower than a threshold), thus obtaining the decision model. This decision model can generate target addresses that conform to standards, are semantically accurate, and have a unified format based on multi-source address data.

[0270] By training the intermediate model with preferences, we can guide its refined behavior, thereby improving the accuracy of address resolution in the final decision model.

[0271] In this embodiment of the application, there are no restrictions on the specific implementation of step S55, "performing knowledge distillation on the decision model to obtain a lightweight decision model".

[0272] In an exemplary embodiment, knowledge distillation of the decision model can be performed through the following steps S551-S554 to obtain a lightweight decision model: S551. The second original address information is processed by the decision model to obtain the first score vector, which includes the unnormalized score of each candidate address element in the structured address field by the decision model.

[0273] S552. The second original address information is processed by the third initial model to obtain the second score vector; the second score vector includes the unnormalized score of each candidate address element in the structured address field by the third initial model.

[0274] S553. Construct a distillation loss function based on the relative entropy between the first score vector and the second score vector.

[0275] S554. Based on the distillation loss function, update the model parameters of the third initial model until the model training termination condition is met, and obtain the lightweight decision model.

[0276] The second source address information can be non-standard address text entered by the user.

[0277] In S551, the second raw address information to be processed can be input into the decision model (as a teacher model). The decision model outputs multiple candidate elements (such as possible district-level options like "City B", "District C1", "District C2", etc.) for the structured address fields (such as city, district, street, house number, etc.) in the second raw address information, and calculates an unnormalized score for each candidate element. The unnormalized scores corresponding to multiple candidate elements can form a first score vector, reflecting the confidence distribution of the decision model for each candidate address element.

[0278] It should be noted that unnormalized scores retain the relative strength and uncertainty information of the decision model output, which can convey fine-grained knowledge better than using only hard labels (one-hot) or the final generated text, and provide richer supervision signals for lightweight decision models.

[0279] The third initial model can be a model with a smaller structure and fewer parameters compared to the decision model. For example, the third initial model can be a small Transformer model.

[0280] In step S552, the same second original address information is input into the third initial model. The third initial model can also output multiple candidate elements for the structured address field in the second original address information, and calculate the corresponding unnormalized score for each candidate element to obtain the second score vector.

[0281] In step S553, relative entropy (Kullback-Leibler Divergence, KL divergence) can be used to measure the difference between the first score vector and the second score vector, and a distillation loss function can be constructed to calculate the loss value of the third initial model.

[0282] In step S554, an optimization algorithm can be used to calculate the loss value based on the distillation loss function, and backpropagation and iterative updates can be performed on all trainable parameters of the third initial model. During training, the parameters of the decision model remain frozen, while the parameters of the third initial model are optimized until a preset termination condition is met (e.g., the accuracy of the third initial model meets a preset analytical threshold), resulting in a lightweight decision model. The lightweight decision model can significantly reduce computational overhead while maintaining high accuracy.

[0283] In this embodiment, by combining prompt words, supervised fine-tuning, and direct preference optimization, supplemented by knowledge distillation, a lightweight decision-making model is created that can both understand complex application logic and operate efficiently. While significantly reducing the number of parameters, memory usage, and inference latency, the lightweight decision-making model still inherits the high performance of decision-making models in handling multi-source address data, conflict detection, and format standardization. It is suitable for high-concurrency, real-time online address resolution scenarios (such as e-commerce ordering and instant delivery), improving the throughput and resource utilization efficiency of the multi-intelligent module collaborative framework.

[0284] It should be noted that the implementation of any of the above models is not limited in this application embodiment, and can be various deep learning-based neural network models. For example, the geographic embedding model can be a deep learning model with a relatively small number of model parameters, or it can be a deep learning model with a relatively large number of model parameters. The large model is just one example, and this application embodiment does not limit the number of model parameters supported by the deep learning model used, with the goal of meeting actual needs. The deep learning model involved in this application embodiment can be an artificial intelligence-based language model (LM) or a multimodal model (MM), and there is no limitation on this.

[0285] Figure 6 This is a schematic diagram of an address resolution system provided for an exemplary embodiment of this application. Please refer to [link / reference]. Figure 6 The address resolution system may include a task coordination intelligent module, an address retrieval intelligent module, a map calling intelligent module, a decision-making intelligent module, a fast resolution intelligent module, and an address database.

[0286] The task coordination intelligent module can be used to preprocess the original address information to obtain preprocessed address information.

[0287] The task coordination intelligent module can also be used to call the fast parsing intelligent module to parse the preprocessed address information and obtain the target address information when the preprocessed address information meets the preset conditions.

[0288] The task coordination intelligent module can also be used to send preprocessed address information to the address retrieval intelligent module and the map calling intelligent module when the preprocessed address information does not meet the preset conditions.

[0289] The address retrieval intelligent module can be used to find similar address information corresponding to preprocessed address information and return the similar address information to the task coordination intelligent module.

[0290] The map invocation intelligent module can be used to invoke the target map application to obtain similar geographic interest points information corresponding to the preprocessed address information, and return the similar geographic interest point information to the task coordination intelligent module.

[0291] The decision intelligence module can be used to perform multi-source collaborative reasoning based on the original address information, similar address information, and similar geographic points of interest information sent by the task coordination intelligence module to obtain the target address information.

[0292] The detailed implementation methods and beneficial effects of each step in this embodiment have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0293] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps S11 to S20 can be device A; or the execution subject of steps S11 to S17 can be device A, and the execution subject of steps S18 to S20 can be device B; and so on.

[0294] Furthermore, in some of the processes described in the above embodiments and accompanying drawings, multiple operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The sequence numbers are merely used to distinguish different operations, and the sequence numbers themselves do not represent any execution order. Additionally, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0295] Figure 7 This is a schematic diagram of an address resolution device provided for an exemplary embodiment of this application. Figure 7 As shown, the address resolution device 70 may include: a coordination module 71, a retrieval module 72, a calling module 73, and a decision module 74. Among them, The coordination module 71 is used to preprocess the original address information through the task coordination intelligent module to obtain preprocessed address information; The retrieval module 72 is used to, if the preprocessed address information does not meet the preset conditions, search for similar address information corresponding to the preprocessed address information through the address retrieval intelligent module; The calling module 73 is used to call the intelligent module through the map. If the preprocessed address information does not meet the preset conditions, the target map application is called to obtain the similar geographic interest point information corresponding to the preprocessed address information. The decision module 74 is used to obtain target address information by performing multi-source collaborative reasoning through the decision intelligence module based on the original address information, the similar address information, and the similar geographic points of interest information.

[0296] The address resolution device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0297] Figure 8 A schematic diagram of the structure of a model training device provided as an exemplary embodiment of this application. Figure 1 .like Figure 8 As shown, the model training device 80 may include: an acquisition module 81, a generation module 82, a loss calculation module 83, and an update module 84, wherein, The acquisition module 81 is used to acquire training samples, which include original address information, real latitude and longitude coordinates, and real grid index identifiers. The generation module 82 is used to generate a predicted address vector with a nested structure based on the original address information through the first initial model, and to predict the corresponding predicted latitude and longitude coordinates and predicted grid index identifier based on the multi-level prefix sub-vectors in the predicted address vector. The loss calculation module 83 is used to calculate the loss value of the first initial model based on the real latitude and longitude coordinates, the real grid index identifier, the predicted latitude and longitude coordinates, and the predicted grid index identifier, through a hierarchical joint loss function. The hierarchical joint loss function includes loss terms corresponding to multiple layers, and the loss terms corresponding to any layer include Haversian distance loss, planar distance loss, and hexagonal hierarchical spatial index classification loss. The update module 84 is used to update the model parameters of the first initial model based on the loss value until the model training termination condition is met, thereby obtaining a geographic embedding model.

[0298] The model training device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0299] Figure 9 A schematic diagram of the structure of a model training device provided as an exemplary embodiment of this application. Figure 2 .like Figure 9 As shown, the model training device 90 may include: a first acquisition module 91, a fine-tuning module 92, a second acquisition module 93, a preference training module 94, and a knowledge distillation module 95, wherein... The first acquisition module 91 is used to acquire training samples, wherein the training samples include first original address information, similar address information, similar geographic points of interest information, and labeled addresses; The fine-tuning module 92 is used to perform supervised fine-tuning of the second initial model based on the training samples to obtain an intermediate model. The second acquisition module 93 is used to acquire triplet samples, wherein the triplet samples include prompt words, correct examples, and incorrect examples, wherein the prompt words include the first original address information, the similar address information, and the similar geographic point of interest information; the correct examples are consistent with the semantics and format of the labeled address, and the incorrect examples are inconsistent with the semantics or format of the labeled address; The preference training module 94 is used to train the intermediate model with preferences using the triplet samples to obtain a decision model. The knowledge distillation module 95 is used to perform knowledge distillation on the decision model to obtain a lightweight decision model.

[0300] The model training device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0301] Figure 10 This is a schematic diagram of the structure of an electronic device provided for an exemplary embodiment of this application. Please refer to... Figure 10 The electronic device 1000 may include a memory 1001 and a processor 1002.

[0302] Memory 1001 is used to store computer programs and can be configured to store various other data to support operation on the computing platform. Examples of this data include instructions for any application or method operating on the computing platform, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0303] The processor 1002, coupled to the memory 1001, is used to execute the computer program in the memory 1001 for: preprocessing the original address information through a task coordination intelligent module to obtain preprocessed address information; if the preprocessed address information does not meet preset conditions, then performing the following operations: searching for similar address information corresponding to the preprocessed address information through an address retrieval intelligent module; calling a target map application through a map invocation intelligent module to obtain similar geographic interest point information corresponding to the preprocessed address information; and performing multi-source collaborative reasoning through a decision intelligent module based on the original address information, the similar address information, and the similar geographic interest point information to obtain the target address information.

[0304] The processor 1002 can also be used to: acquire training samples, the training samples including original address information, real latitude and longitude coordinates, and real grid index identifiers; generate a predicted address vector with a nested structure based on the original address information using a first initial model, and predict the corresponding predicted latitude and longitude coordinates and predicted grid index identifiers based on the multi-level prefix sub-vectors in the predicted address vectors; calculate the loss value of the first initial model based on the real latitude and longitude coordinates, the real grid index identifiers, the predicted latitude and longitude coordinates, and the predicted grid index identifiers using a hierarchical joint loss function, the hierarchical joint loss function including loss terms corresponding to multiple levels, and any loss term corresponding to a level including Haversian distance loss, planar distance loss, and hexagonal hierarchical spatial index classification loss; update the model parameters of the first initial model based on the loss value until the model training termination condition is met, thereby obtaining a geographic embedding model.

[0305] The processor 1002 can also be used to: acquire training samples, the training samples including first original address information, similar address information, similar geographic interest point information, and labeled addresses; perform supervised fine-tuning of a second initial model based on the training samples to obtain an intermediate model; acquire triple samples, the triple samples including prompt words, correct examples, and incorrect examples, the prompt words including the first original address information, the similar address information, and the similar geographic interest point information; the correct examples are semantically and format-consistent with the labeled addresses, and the incorrect examples are semantically or format-inconsistent with the labeled addresses; perform preference training on the intermediate model using the triple samples to obtain a decision model; and perform knowledge distillation on the decision model to obtain a lightweight decision model.

[0306] Furthermore, such as Figure 10 As shown, the electronic device also includes other components such as a communication component 1003, a display 1004, and a power supply component 1005.

[0307] Figure 10 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 10 The components shown. Additionally... Figure 10 The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the electronic device. The electronic device in this embodiment can be a terminal device such as a desktop computer, laptop computer, smartphone, or IoT device, or a server-side device such as a conventional server, cloud server, or server array. If the electronic device in this embodiment is a terminal device such as a desktop computer, laptop computer, or smartphone, it may include... Figure 10 The components within the dashed box; if the electronic device in this embodiment is implemented as a conventional server, cloud server, or server array, etc., it may be omitted. Figure 10 The component within the dashed box.

[0308] The aforementioned memory 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.

[0309] The aforementioned communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0310] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can 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 can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0311] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0312] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0313] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0314] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0315] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An address resolution method, characterized in that, include: The task coordination intelligent module preprocesses the original address information to obtain preprocessed address information. If the preprocessed address information does not meet the preset conditions, the following operations are performed: The address retrieval intelligent module is used to find similar address information corresponding to the preprocessed address information. The map-invoking intelligent module calls the target map application to obtain similar geographic points of interest information corresponding to the preprocessed address information; Based on the original address information, the similar address information, and the similar geographic points of interest information, the target address information is obtained through multi-source collaborative reasoning by the decision intelligence module.

2. The method according to claim 1, characterized in that, The decision intelligence module is associated with a lightweight decision model, which is obtained by knowledge distillation of the decision model. The lightweight decision model includes a multi-source input layer, a credible judgment layer, an evidence fusion layer, and a decision output layer. Based on the original address information, the similar address information, and the similar geographic points of interest information, the target address information is obtained through multi-source collaborative reasoning via a decision intelligence module, including: The original address information, the similar address information, and the similar geographic points of interest information are received through the multi-source input layer. Through the trust judgment layer, based on the original address information, a trust judgment is made on the similar address information and the similar geographic point of interest information; If either the similar address information or the similar geographic point of interest information is reliable, then the target address information is determined based on the reliable similar information through the decision output layer. If both the similar address information and the similar geographic point of interest information are reliable, then the evidence fusion layer performs conflict detection on the similar address information and the similar geographic point of interest information, and the decision output layer determines the target address information based on the conflict detection results.

3. The method according to claim 2, characterized in that, The evidence fusion layer performs conflict detection on the similar address information and the similar geographic point of interest information, including: Target similar address information is determined from the similar address information, and first latitude and longitude coordinates and first geographical region are extracted from the target similar address information. The similarity between the target similar address information and the preprocessed address information is greater than a second preset threshold. Target geographic point of interest information is determined from the similar geographic point of interest information, and second latitude and longitude coordinates and second geographic region are extracted from the target geographic point of interest information. The similarity between the target geographic point of interest information and the preprocessed address information is greater than a first preset threshold. Calculate the target distance based on the first latitude and longitude coordinates and the second latitude and longitude coordinates; If the target distance is less than or equal to the distance threshold, and the first geographic region is consistent with the second geographic region, then it is determined that there is no conflict between the similar address information and the similar geographic point of interest information; If the target distance is greater than a distance threshold, and / or the first geographic region is inconsistent with the second geographic region, then it is determined that there is a conflict between the similar address information and the similar geographic point of interest information.

4. The method according to claim 2 or 3, characterized in that, The decision output layer determines the target address information based on the conflict detection results, including: If there is no conflict between the similar address information and the similar geographic point of interest information, then the target address information is determined based on the target geographic point of interest information in the similar geographic point of interest information; If there is a conflict between the similar address information and the similar geographic point of interest information, then the target address information is determined based on the target similar address information in the similar address information.

5. The method according to claim 2 or 3, characterized in that, The trustworthiness assessment layer performs a trustworthiness assessment on the similar address information, including: Based on the latitude and longitude coordinates in the similar address information, the grid index identifier corresponding to the similar address information is obtained in the hexagonal hierarchical spatial index database, resulting in at least one grid index identifier; The at least one grid index identifier is deduplicated to obtain K grid index identifiers, where K is a positive integer; If K is less than or equal to a preset number, then the similar address information is determined to be reliable. If K is greater than the preset number, then the similar address information is determined to be unreliable.

6. The method according to claim 2 or 3, characterized in that, Through the aforementioned trust assessment layer, based on the original address information, a trust assessment is performed on the similar geographic point of interest information, including: Semantic extraction is performed on the original address information to obtain the keywords in the original address information; For any similar geographic point of interest (POI) information, if the similar POI information includes the keyword, then the similar POI information is determined to be reliable; if the similar POI information does not include the keyword, then the similar POI information is determined to be unreliable.

7. The method according to any one of claims 1-3, characterized in that, The address retrieval intelligent module is associated with a geographic embedding model, which is obtained by training a first initial model based on a nested representation learning method. The address retrieval intelligent module searches for similar address information corresponding to the preprocessed address information, including: The geographic embedding model generates a preprocessed address vector corresponding to the preprocessed address information; the preprocessed address vector is a multi-dimensional elastic vector. Based on the preprocessed address vector, the similar address vector is retrieved in the address index, which is a multi-layer graph structure constructed based on the hierarchical navigable small-world algorithm; Based on the address identifier corresponding to the similar address vector, obtain the similar address information corresponding to the similar address vector from the address database.

8. The method according to claim 7, characterized in that, The geographic embedding model includes an input encoding layer, a context encoding layer, a feature aggregation layer, and a nested projection layer; through the geographic embedding model, a preprocessed address vector corresponding to the preprocessed address information is generated, including: The preprocessed address information is segmented and embedded through the input encoding layer to obtain the corresponding word sequence. The context encoding layer is used to perform context semantic modeling on the word sequence to obtain a first feature vector; the first feature vector includes the context semantics corresponding to any word in the word sequence. The first feature vector is pooled through the feature aggregation layer to obtain a second feature vector, which is used to characterize the overall semantics of the preprocessed address information. By using nested projection layers, the second feature vector is projected onto a high-dimensional embedding space to obtain an address embedding vector with a nested structure; the multiple prefix sub-vectors of preset dimensions in the address embedding vector respectively represent the semantics of the preprocessed address information at different geographic granularities; The first N dimensions of the address embedding vector are extracted as the preprocessed address vector, where N is a preset dimension.

9. The method according to claim 8, characterized in that, The address index includes a top layer, at least one middle layer, and a bottom layer. Each layer includes multiple nodes, and each node includes an address vector and an address identifier corresponding to historical address information. Based on the preprocessed address vector, retrieving the similar address vector from the address index includes: Calculate the similarity between the preprocessed address vector and the address vectors in each of the top-level nodes, and determine the node corresponding to the address vector with a similarity greater than a third preset threshold as the first matching node; Based on the first matching node, at least one candidate node is determined in the next layer, and the similarity between the preprocessed address vector and the address vector in the at least one candidate node is calculated. The candidate node corresponding to the address vector with a similarity greater than a fourth preset threshold is determined as the second matching node. Repeat the above process until the bottom layer of the address index is reached, where multiple matching nodes are identified. The similar address vector is determined based on the similarity between the preprocessed address vector and the address vectors in the plurality of matching nodes.

10. The method according to claim 8, characterized in that, The method further includes: The geographic embedding model generates address vectors corresponding to multiple historical address information. Retrieve the address identifiers corresponding to the multiple historical address information from the address database; The address index is constructed based on the address vector and address identifier corresponding to each historical address information through a hierarchical navigable small-world algorithm.

11. The method according to any one of claims 1-3 or 8-10, characterized in that, The map-based intelligent module invokes the target map application to obtain similar geographic points of interest information corresponding to the preprocessed address information, including: The map calling intelligent module generates a geographic point of interest query request including preprocessed address information, and sends the geographic point of interest query request to the target map application. Receive the geographic point of interest response data returned by the target map application in response to the geographic point of interest query request; In the geographic point of interest response data, multiple field values ​​corresponding to multiple preset fields are extracted; the multiple preset fields include geographic point of interest type, geographic point of interest name, and latitude and longitude coordinates. The multiple field values ​​are combined according to a preset template to obtain similar geographic point of interest information. The similar geographic point of interest information is obtained by the target map application by searching and filtering in the geographic point of interest database based on the preprocessed address information.

12. The method according to any one of claims 1-3 or 8-10, characterized in that, The method further includes: If the preprocessed address information meets the preset conditions, the geocoding request is sent to the target map application through the fast parsing intelligent module, and the geocoding request includes the preprocessed address information. The system receives target address information returned by the target map application in response to the geocoding request. The target address information includes a target address and / or a structured address field, and the structured address field includes address values ​​corresponding to multiple preset address fields.

13. The method according to any one of claims 1-3 or 8-10, characterized in that, The task coordination intelligent module is associated with the address preprocessing model, which includes a raw input layer, a noise identification layer, a sequence correction layer, and an address output layer. The task coordination intelligent module preprocesses the original address information to obtain preprocessed address information, including: The original address information is received through the original input layer; The noise identification layer is used to identify anomalies in the original address information to obtain address noise features; the anomaly identification includes at least one of the following: special symbols, typos, abbreviations, missing fields, and abnormal word order; Through the sequence correction layer, based on the address noise characteristics, the original address information is corrected at the sequence level to generate an address correction sequence; The address output layer converts the address correction sequence into preprocessed address information that conforms to a preset address format.

14. A model training method, characterized in that, The method includes: Obtain training samples, which include original address information, real latitude and longitude coordinates, and real grid index identifiers; Using the first initial model, a predicted address vector with a nested structure is generated based on the original address information. Based on the multi-level prefix sub-vectors in the predicted address vector, the corresponding predicted latitude and longitude coordinates and predicted grid index identifiers are predicted respectively. Based on the true latitude and longitude coordinates, the true grid index identifier, the predicted latitude and longitude coordinates, and the predicted grid index identifier, the loss value of the first initial model is calculated through a hierarchical joint loss function. The hierarchical joint loss function includes loss terms corresponding to multiple layers, and the loss terms corresponding to any layer include Haversian distance loss, planar distance loss, and hexagonal hierarchical spatial index classification loss. Based on the loss value, the model parameters of the first initial model are updated until the model training termination condition is met, thus obtaining the geographic embedding model.

15. A model training method, characterized in that, The method includes: Obtain training samples, which include first original address information, similar address information, similar geographic points of interest information, and labeled addresses; Based on the training samples, the second initial model is subjected to supervised fine-tuning to obtain an intermediate model; Obtain triplet samples, wherein the triplet samples include prompt words, correct examples, and incorrect examples, wherein the prompt words include the first original address information, the similar address information, and the similar geographic points of interest information; the correct examples are consistent with the semantics and format of the labeled address, and the incorrect examples are inconsistent with the semantics or format of the labeled address; The intermediate model is trained using the triplet samples to obtain the decision model; Knowledge distillation is performed on the decision model to obtain a lightweight decision model.

16. The method according to claim 15, characterized in that, The second initial model includes a multi-head self-attention layer, which includes a query matrix, a key matrix, a value matrix, and an output matrix; Based on the training samples, the second initial model is subjected to supervised fine-tuning to obtain an intermediate model, including: A low-rank adapter is injected into the query matrix, the key matrix, the value matrix, and the output matrix, and the original weight parameters of the second initial model are frozen. The low-rank adapter is the product of two low-rank matrices. Based on the first original address information, the similar address information, and the similar geographic points of interest information, a predicted address is generated using the second initial model. Based on the predicted address and the labeled address, a supervised loss function is constructed; The parameters of the low-rank adapter are updated according to the supervised loss function to obtain an intermediate model.

17. The method according to claim 15 or 16, characterized in that, The intermediate model is trained using the triplet samples to obtain a decision model, including: For any triplet sample, based on the prompt word, the intermediate model is used to calculate the first generation probability of the correct example and the second generation probability of the incorrect example. Based on the first generation probability and the second generation probability, a preference loss function is constructed; The model parameters of the intermediate model are updated according to the preference loss function until the model training termination condition is met, thus obtaining the decision model.

18. The method according to claim 15 or 16, characterized in that, Knowledge distillation is performed on the decision model to obtain a lightweight decision model, including: The second original address information is processed by a decision model to obtain a first score vector, which includes the unnormalized score of each candidate address element in the structured address field by the decision model. The second original address information is processed by the third initial model to obtain a second score vector; the second score vector includes the unnormalized score of each candidate address element in the structured address field by the third initial model; Based on the relative entropy between the first score vector and the second score vector, a distillation loss function is constructed; Based on the distillation loss function, the model parameters of the third initial model are updated until the model training termination condition is met, thus obtaining the lightweight decision model.

19. An address resolution system, characterized in that, include: Intelligent module for task coordination, intelligent module for address retrieval, intelligent module for map retrieval, intelligent module for decision-making, and intelligent module for rapid analysis; The task coordination intelligent module is used to preprocess the original address information to obtain preprocessed address information; The task coordination intelligent module is also used to, when the preprocessed address information meets the preset conditions, call the fast parsing intelligent module to parse the preprocessed address information to obtain the target address information; The task coordination intelligent module is also used to send the preprocessed address information to the address retrieval intelligent module and the map calling intelligent module when the preprocessed address information does not meet the preset conditions. The address retrieval intelligent module is used to find similar address information corresponding to the preprocessed address information and return the similar address information to the task coordination intelligent module. The map invocation intelligent module is used to invoke the target map application to obtain similar geographic interest point information corresponding to the preprocessed address information, and return the similar geographic interest point information to the task coordination intelligent module. The decision intelligence module is used to perform multi-source collaborative reasoning based on the original address information, the similar address information, and the similar geographic point of interest information sent by the task coordination intelligence module to obtain the target address information.

20. An electronic device, characterized in that, include: Memory and processor; The memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program in the memory to implement the steps of the method according to any one of claims 1-18.

21. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method according to any one of claims 1-18.

22. A computer program product, characterized in that, Includes a computer program / instruction that, when executed by a processor, causes the processor to perform the steps of the method according to any one of claims 1-18.

Images