Address data restructuring system incorporating semantic search technology

Abstract

The application relates to the technical field of geographic information processing and intelligent address matching, in particular to an address data reconstruction system fusing a semantic retrieval technology, which comprises a basic database, an address preprocessing module, a candidate interest point grouping module, a relationship establishing module, an address matching module, an address reconstruction output module and a closed-loop management module; the multi-source map interest point data is subjected to extraction of an organization main body name field, a hierarchical field, a subdivision type label, an alias field, an administrative division code, a spatial grid code and a text feature representation, candidate interest point groups are generated by aggregation, parent-child, containing, same-level, alias, adjacent alias and type conflict relationships are calculated and an address association table is established, candidate query and standard interest point determination are performed on a business address, and a standardized address record, a hierarchical address expression or a unified address identification is output; the application improves the accuracy and stability of the address reconstruction result.

Description

Technical Field

[0001] This invention relates to the field of geographic information processing and intelligent address matching technology, specifically to an address data reconstruction system that integrates semantic retrieval technology. Background Technology

[0002] In existing address standardization processes, it is usually necessary to normalize and match multi-source map point of interest data and business system address data in order to form a unified standard address record for retrieval, management and business use. Existing address processing solutions generally include address splitting, keyword matching, and candidate address filtering. Address data often contains organization abbreviations, historical aliases, hierarchical names, and type descriptions. By comparing fields of the address text or matching them with coordinate information, business addresses can be merged into standard addresses. However, existing solutions often rely on string similarity, simple dictionary rules, or spatial proximity for matching, which can easily lead to confusion between addresses with parent-child hierarchical relationships, addresses of different child nodes of the same subject, and aliased addresses. This results in address matching accuracy being lower than the preset qualified standard and unclear hierarchical relationships, leading to the instability of address reconstruction results that do not meet the preset standard. Summary of the Invention

[0003] The purpose of this invention is to provide an address data reconstruction system that integrates semantic retrieval technology. This system avoids confusion between addresses with parent-child hierarchical relationships, addresses of different child nodes of the same subject, and aliased addresses, which are easily confused when relying solely on string similarity, simple dictionary rules, or spatial proximity for matching. Furthermore, it can separate parent-child level addresses that are difficult to distinguish using only string similarity, reducing erroneous normalization caused by similar names or overlapping coordinates. This solves the problems of insufficient address matching accuracy, unclear hierarchical relationship identification, and poor stability of address reconstruction results in existing solutions.

[0004] The objective of this invention can be achieved through the following technical solutions: The basic database is used to store the organization's suffix thesaurus, alias thesaurus, type tree, address encoding library, thesaurus matching weight, preset matching weight, grouping weight, relation rule weight, relation judgment rules, preset aggregation threshold, matching qualification threshold and retrieval configuration parameters; The address preprocessing module is used to acquire multi-source map point of interest data and split the features, extract the organization name field, hierarchy field, subdivision type label, alias field, administrative division code and spatial grid code, and generate text feature representation and point of interest structured record based on name or alias; The candidate interest point grouping module is used to aggregate interest points with related attributes and spatial proximity based on structured records to generate candidate interest point groups; The relationship establishment module is used to analyze the parent-child and sibling relationships between interest points based on the acquired candidate interest point grouping, and guide the large language model to analyze the relationship between interest points through a preset hierarchical reasoning prompt word template, generate hierarchical relationship data, and parse and extract relationship records in which the relationship confidence is not lower than the first relationship score threshold, and write them into the address association table. The address matching module is used to perform retrieval and matching of the business address to be matched according to the first matching mode or the second matching mode, and calculate the final matching score for each candidate point of interest; when the final matching score is the highest and not lower than the preset matching qualification threshold in the basic database, the corresponding candidate point of interest is used as the standard point of interest and the standard point of interest identifier is output; when the final matching score is lower than the matching qualification threshold, the result pending manual review is output. The first matching mode is a retrieval enhancement generation matching mode based on large language model prediction and vector recall, and the second matching mode is a word matching mode based on word segmentation of professional word library and general word library. The address reconstruction output module is used to generate standardized records, hierarchical address representations, or unified identifiers based on standard points of interest and address association tables. The closed-loop management module is used to receive relationship data that is lower than the first relationship score threshold, business address records that have not returned candidate points of interest, business address records whose final matching score is lower than the matching qualification threshold, and manual review results. It updates the alias library, type tree, encoding library, word library matching weight and grouping weight, and writes relationships with a relationship confidence of not lower than the first relationship score threshold or those that have been manually confirmed into the address association table.

[0005] Preferably, the address preprocessing module includes: a main body name extraction unit, used to strip the suffixes of the point of interest names according to the organization suffix thesaurus, and extract and store the organization main body name field; and a hierarchical word extraction unit, used to extract and store the hierarchical field representing the organizational affiliation level or building functional level. The type segmentation unit is used to compare the original industry classification with the type tree and generate segmented type labels; the alias mapping unit is used to generate alias fields for points of interest based on the alias library; the administrative division positioning unit is used to generate administrative division codes and spatial grid codes corresponding to points of interest; and the semantic feature generation unit is used to generate text feature representations for semantic retrieval based on the point of interest name and alias field.

[0006] Preferably, the candidate point of interest grouping module is used to group two points of interest into the same candidate point of interest group when the following conditions are met simultaneously: the administrative division codes are the same; the character matching degree of the institution's main name field is not lower than a preset character matching threshold; the spatial distance is not greater than a preset distance threshold; and the first-level subdivision type labels are the same or belong to the same type tree upper-level node.

[0007] Preferably, the candidate point of interest grouping module is used to: calculate the comprehensive score of two points of interest based on attribute information, spatial information, type information, and administrative division information determined by the extracted institution name field, subdivision type label, administrative division code and spatial grid code, and text feature representation; and aggregate the two points of interest into the same candidate point of interest group when the comprehensive score is not lower than a preset aggregation threshold; and not aggregate when the comprehensive score is lower than the preset aggregation threshold. When the number of interest points in a single candidate interest point group exceeds a preset number, the candidate interest point group is further segmented according to a spatial grid code that is one or more levels finer than the spatial grid code.

[0008] Preferably, the specific steps of the relationship establishment module in performing hierarchical semantic graph construction are as follows: obtain the structured records of each interest point in the candidate interest point group and integrate them into a candidate cluster input to the large language model; guide it to deeply analyze the parent-child and parallel relationships between each interest point in the candidate cluster according to the preset hierarchical reasoning prompt word template; and assemble the extracted institutional entity name field, hierarchical field and subdivision type label into a formatted text sequence with sequence identifier. The formatted text sequence is input into a pre-trained large language model, and the hierarchical relationship data in a tree structure, which includes node identifiers and relationship subdivision type labels, is received from the output of the large language model. Parse the hierarchical relationship data of the tree structure, extract and receive the relationship data containing node identifiers and relationship sub-type labels, as well as the relationship confidence level; When the confidence level of a relationship is not lower than the first relationship score threshold, the corresponding parent-child relationship and sibling relationship are written into the address association table. When the confidence level of a relationship is lower than the first relationship score threshold, the corresponding relationship is sent to the manual review queue and written into the address association table after manual confirmation.

[0009] Preferably, in the first matching mode, i.e. the retrieval enhancement generation matching mode, the address matching module is used to perform the following steps: perform zero-shot type prediction and address classification on the business address text based on a large language model to obtain the target address type; Based on the target address type, a multi-way query channel is established using the combined county and type index sharding, the combined city and type index sharding, the alias index sharding, the relational adjacency index sharding, and the semantic feature index sharding. After converting the business address into address matching parameters that include at least one or more of the main word encoding, type encoding, zone encoding and text feature representation, multi-path data extraction is performed, and the extraction results are merged and deduplicated to obtain a set of candidate interest points, which are then input into a large language model to perform semantic fine screening. Based on the scores and candidate set ranking results after semantic fine screening of the large language model, a preset number of candidate interest points are selected to determine the standard interest points, and the corresponding standard interest point identifiers are output.

[0010] Preferably, the address matching module is used to compare and match each candidate point of interest with the corresponding field of the business address to obtain the encoding matching result, type matching result, administrative division matching result and alias matching result, calculate the final matching score by combining the hierarchical structure information and distance information, and sort them in descending order according to the final matching score; The address matching module is also used to compare the consistency of main words, the consistency of hierarchical fields, the consistency of subdivided type tags, and the alias hits of the candidate interest points whose final matching scores rank within the aforementioned preset number, and output the standard interest points.

[0011] Preferably, in the second matching mode, i.e., the dictionary matching mode, the address matching module is configured to perform the following operations: Based on professional and general terminology databases, business addresses are professionally segmented and address elements are split to generate a set of Chinese keywords and a set of Pinyin keywords. Initial screening is performed based on preset weighting factors and adjustable coordinate precision. The adjustable coordinate precision is configured to initially limit the search range in the coordinate space according to a set latitude and longitude tolerance. The initial matching weights are determined based on historical data according to the address number, main body words of the organization, and type words. The matching is performed step by step, including at least one or more of the following: exact match, fuzzy match, pinyin match, and hierarchical priority match. Exact match means that the business address keyword and the corresponding field of the point of interest are completely consistent. Fuzzy match means that the character matching degree of the business address keyword and the corresponding field of the point of interest is not lower than the preset fuzzy match threshold. Pinyin match means that the pinyin representation of the business address keyword and the corresponding field of the point of interest are consistent or have the same prefix. Hierarchical priority match means that when a hierarchical field exists, the hierarchical field is compared first and then the main body word is compared. The initial candidate set is screened by combining spatial distance and preset rule verification to determine the final matching result and output standard interest point identifiers. The aforementioned preset rule verification includes at least the door number consistency verification and the institution type alignment verification. At the same time, a dynamic weight adjustment mechanism based on feedback from manual review results is introduced to generate weight adjustment values, and the weight of the hit keyword categories is dynamically optimized and updated individually.

[0012] Preferably, the closed-loop management module is used to send the following data into the manual review process: relationship data in the relationship establishment module whose relationship score is lower than the second relationship score threshold, wherein the second relationship score threshold is less than the first relationship score threshold; business address records that did not return candidate points of interest in the aforementioned first matching mode; and business address records that returned candidate points of interest in the thesaurus matching mode but whose highest matching score is lower than the matching qualification threshold. After manual review and confirmation, perform at least one of the following update operations: write the correspondence between business addresses and standard points of interest into the alias database; update the keyword category weights in the thesaurus matching mode; update the grouping weights in the aforementioned candidate point of interest grouping module; and write the confirmed hierarchical relationship into and replace the original relationship record in the address association table.

[0013] Preferably, the basic database is also used to store semantic retrieval indexes corresponding to text feature representations; the text feature representations refer to vectors that can be used for semantic similarity calculation, obtained by encoding the text of point of interest names, alias fields, or business address text.

[0014] The beneficial effects of this invention are: 1. This invention establishes a basic database, an address preprocessing module, a candidate address grouping module, a relationship establishment module, an address matching module, an address reconstruction output module, and a closed-loop management module. During the address preprocessing stage, the point of interest name is split into an institution name field, a hierarchy field, a subdivision type label, an alias field, an administrative division code, a spatial grid code, and a text feature representation. This allows for the separate storage of subject information, hierarchy information, type information, and spatial information, facilitating subsequent grouping, relationship judgment, and matching calculations. This addresses addresses with multi-level nested parent-child hierarchical structures, such as those in universities, hospitals, and industrial parks. It reduces problems caused by relying solely on string similarity, simple dictionary rules, or spatial proximity relationships, such as confusion between parent institutions and child buildings, misclassification of aliased addresses, and unclear hierarchical relationship identification. This improves the accuracy and stability of the address reconstruction results. 2. This invention first filters and aggregates candidate addresses based on the following conditions: consistent administrative division codes, matching degree of the institutional entity name field characters reaching a threshold, spatial distance not exceeding a preset distance threshold, and consistent sub-type labels or belonging to the same type of tree upper-level node. Then, it combines attribute information similarity, distance score corresponding to spatial distance, type consistency, and administrative division consistency for comprehensive scoring and aggregation. When the group size exceeds a preset number, it performs secondary segmentation according to a more detailed one-level or multi-level spatial grid code, so that the global original points of interest are first aggregated into candidate groups that meet the quantity limit conditions. Then, the relationship establishment module calculates the relationship score and writes it into the address association table based on the institutional entity name field similarity rules, hierarchical field correspondence rules, sub-type label consistency rules, administrative division consistency rules, and spatial location proximity rules. This enables the identification of parent-child relationships, inclusion relationships, sibling relationships, alias relationships, and neighboring heteronym relationships within a limited and narrowed search range, effectively suppressing mis-aggregation and misjudgment of cross-type, cross-district, and same-name neighboring entities. Attached Figure Description

[0015] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the address data reconstruction system that integrates semantic retrieval technology provided in the embodiments of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 The address data reconstruction system integrating semantic retrieval technology includes: a basic database for storing an organization suffix thesaurus, alias thesaurus, type tree, address encoding library, thesaurus matching weights, preset matching weights, grouping weights, relation rule weights, relation judgment rules, preset aggregation thresholds, matching qualification thresholds, and retrieval configuration parameters; The address preprocessing module is used to acquire multi-source map point of interest data and split the features, extract the organization name field, hierarchy field, subdivision type label, alias field, administrative division code and spatial grid code, and generate text feature representation and point of interest structured record based on name or alias; The candidate interest point grouping module is used to aggregate interest points with related attributes and spatial proximity based on structured records to generate candidate interest point groups; The relationship establishment module is used to group the acquired candidate points of interest, guide the large language model to analyze the parent-child and sibling relationships between points of interest through a preset hierarchical reasoning prompt word template, generate hierarchical relationship data, and parse and extract relationship records whose relationship confidence is not lower than the first relationship score threshold, and write them into the address association table. The address matching module is used to perform retrieval and matching of the business address to be matched according to the first matching mode or the second matching mode, and calculate the final matching score for each candidate point of interest. When the final matching score is the highest and not lower than the preset matching qualification threshold in the basic database, the corresponding candidate point of interest is used as the standard point of interest and the standard point of interest identifier is output. When the final matching score is lower than the matching qualification threshold, the result pending manual review is output. Among them, the first matching mode is a retrieval enhancement generation matching mode based on large language model prediction and vector recall, and the second matching mode is a word matching mode based on word segmentation of professional word library and general word library; The address reconstruction output module generates standardized records, hierarchical address representations, or unified identifiers based on standard points of interest and association tables. Closed-loop management module: Used to receive relation data below the first relation score threshold, business address records that did not return candidate points of interest, business address records whose final matching score is below the matching qualification threshold, and manual review results; update the alias library, type tree, encoding library, word library matching weight and grouping weight; and write relations with a relation confidence of not less than the first relation score threshold or those confirmed by manual review into the association table. The address preprocessing module includes: a main name extraction unit, which is used to strip the suffixes of the point of interest names based on the organization suffix thesaurus, and extract and store the organization main name field; The hierarchical term extraction unit is used to extract and store hierarchical fields that represent organizational hierarchy or building functional hierarchy; The type segmentation unit is used to compare the original industry classification with the type tree and generate segmented type labels; the alias mapping unit is used to generate alias fields for points of interest based on the alias library; the administrative division positioning unit is used to generate administrative division codes and spatial grid codes corresponding to points of interest; and the semantic feature generation unit is used to generate text feature representations for semantic retrieval based on the point of interest name and alias field. The candidate point of interest grouping module is used to group two points of interest into the same candidate point of interest group when the following conditions are met simultaneously: the administrative division codes are the same; The character matching degree of the organization name field is not lower than the preset character matching threshold; the spatial distance is not greater than the preset distance threshold; the first-level sub-type labels are consistent or belong to the same type of tree parent node; The candidate point of interest grouping module is used to: calculate the comprehensive score of two points of interest based on attribute information, spatial information, type information, and administrative division information determined by the extracted institution name field, sub-type label, administrative division code, spatial grid code, and text feature representation; and aggregate the two points of interest into the same candidate point of interest group when the comprehensive score is not lower than the preset aggregation threshold; and not aggregate when the comprehensive score is lower than the preset aggregation threshold. When the number of interest points in a single candidate interest point group exceeds the preset number, the candidate interest point group is further segmented according to a spatial grid code that is one or more levels finer than the spatial grid code. The specific steps of the relationship establishment module to construct the hierarchical semantic graph are as follows: obtain the structured records of each interest point in the candidate interest point group and integrate them into the candidate cluster input into the large language model. Guide it to deeply analyze the parent-child and parallel relationships between each interest point in the candidate cluster according to the preset hierarchical reasoning prompt word template. Assemble the organization main name field, hierarchical field and subdivision type label into a formatted text sequence with sequence identifier. The formatted text sequence is input into a pre-trained large language model, and the hierarchical relation data in a tree structure, which includes node identifiers and relation sub-type labels, is received from the output of the large language model. Parse the hierarchical relationship data of the tree structure, extract and receive the relationship data containing node identifiers, relationship sub-type labels, and relationship confidence scores; When the relationship confidence is not lower than the first relationship score threshold, the corresponding parent-child relationship and sibling relationship are written into the address association table.

[0018] The basic database is also used to store semantic retrieval indexes corresponding to text feature representations; text feature representations refer to vectors that can be used for semantic similarity calculations, obtained by encoding the text of point of interest names, alias fields, or business addresses.

[0019] This implementation takes the normalization of multi-source map points of interest and business system addresses as the main business processing objective, and the system is deployed on at least one application server, at least one database server, and at least one retrieval server. A basic database is established in the database server to store the organization suffix thesaurus, alias thesaurus, type tree, address encoding thesaurus, thesaurus matching weights, preset matching weights, grouping weights, relation rule weights, relation judgment rules, preset aggregation thresholds, matching qualification thresholds, and semantic retrieval index configuration parameters. The application server respectively hosts the address preprocessing module, the candidate interest point grouping module, the relationship establishment module, the address matching module, the address reconstruction output module, and the closed-loop management module; The retrieval server is used to store the semantic retrieval index corresponding to the text feature representation; the text feature representation refers to the vector that can be used for semantic similarity calculation, which is obtained by encoding the text of the point of interest name, alias field or business address; the address association table refers to the data table that records the parent-child, containment, sibling, alias, neighboring synonym and type conflict relationships between standard points of interest. In this embodiment, the address preprocessing module receives raw records of points of interest from the map data interface; each raw record of point of interest includes at least the point of interest name, address text, latitude and longitude, and original industry classification; after the application server reads the organization suffix thesaurus from the basic database, it performs suffix stripping on the point of interest name; The suffix stripping unit is configured to match and strip segments according to the preset suffix priority in the organization suffix dictionary: if the end of the point of interest name matches an organization suffix, the matching suffix will be separated from the main candidate string. If the name also contains a phrase indicating the organizational hierarchy or building function level, it will be retained in the hierarchy field and not merged into the organization main name field. For example, for the outpatient building of the First Affiliated Hospital of Guangxi Medical University, the main name extraction unit outputs Guangxi Medical University as the main name field of the institution, and the hierarchical word extraction unit outputs the First Affiliated Hospital and the outpatient building as hierarchical fields; In the subsequent grouping stage, entity aggregation needs to be performed using the entity name field, and in the relationship establishment stage, the hierarchical field needs to be used to determine the parent-child relationship. Through the field splitting operation performed in the address preprocessing stage, corresponding dimension input data is provided for the subsequent grouping, relationship judgment and matching calculation. The type subdivision unit reads the type tree and compares the original industry classification with the type tree nodes level by level. If the original classification can be directly mapped to a fine-grained node, the fine-grained node is written into the subdivision type label. If the original classification can only be mapped to a higher-level node, the higher-level node code is retained and processed according to the higher-level node in subsequent groupings. The alias mapping unit queries the alias database for synonyms, abbreviations, or common spellings to generate an alias field; the division positioning unit looks up the administrative division code based on the coordinates of the point of interest and generates a spatial grid code according to the preset grid precision; the semantic feature generation unit performs text encoding on the point of interest name and alias field, generates a text feature representation, and writes it into the vector index of the retrieval server. The semantic feature generation unit calls a pre-trained natural language processing model, such as a bidirectional encoder representation model or an enhanced bidirectional encoder representation model, as the underlying encoder. It concatenates the interest point name and alias fields into an input character sequence in sequence and adds specific delimiters at the beginning and end of the sequence. Extract the hidden layer vectors from the model's output layer that represent sentence-level semantics, such as the vector of the first position, or perform mean pooling on the hidden layer vectors of all characters. Before deployment, the pre-trained natural language processing model needs to be fine-tuned using positive and negative samples of historical standard addresses. The fine-tuning process uses the triplet loss function. in, The value of the triplet loss function. As the base address vector, For positive sample address vectors, For negative sample address vectors, It is a metric for the Euclidean or cosine distance between vectors. The preset boundary threshold; After the above processing is completed, each point of interest forms a structured record of the point of interest. This record includes at least the following fields: institution name, hierarchy, sub-type label, alias, administrative division code, spatial grid code, spatial location, and text feature representation. After receiving the structured records of interest points, the candidate interest point grouping module prioritizes conditional filtering and aggregation. The character matching degree uses edit distance, longest common subsequence, or a combination of both to obtain normalized similarity. When using edit distance, normalized similarity is calculated by subtracting the ratio of edit distance to the maximum length of the two strings from 1; when using longest common subsequence, normalized similarity is calculated by dividing twice the length of the longest common subsequence by the sum of the lengths of the two strings. This normalization process ensures that the calculation results of names of different lengths are strictly mapped to the 0 to 1 range, so as to make an equivalent comparison with the uniformly set preset threshold; under a set of implementation parameters, the preset character matching threshold value can be set to 0.85, and the preset distance threshold value can be set to 2000m. This distance is used to cover physical areas such as university campuses and hospital parks whose spatial boundary dimensions exceed the preset threshold. When two points of interest simultaneously meet the following conditions: consistent administrative division codes, character matching degree of the main body name field not less than 0.85, spatial distance not greater than 2000m, consistent first-level subdivision type labels, or being in the same type tree upper-level node, the candidate point of interest grouping module will classify them into the same candidate point of interest group; this method is suitable for map data whose coordinate fluctuations are within the preset deviation range and whose naming conforms to the structured standard. For points of interest whose coordinate drift exceeds the preset deviation range or whose names are abbreviated or commonly known, the candidate point of interest grouping module performs a comprehensive scoring aggregation method. The application server calculates the similarity of attribute information, the distance score corresponding to spatial distance, the consistency of type, and the consistency of administrative division for any two points of interest, and sums them according to the group weights to obtain the comprehensive score; The attribute information similarity is represented by the cosine similarity of the text features of the organization's main name field. Character matching degree According to the preset weighting coefficients The formula is obtained by linear weighted summation: in, For attribute information similarity, Cosine similarity for the text features of the organization's main name field. For character matching degree, weight coefficient The distance score is determined according to the mapping relationship between spatial distance and preset distance range. When the distance exceeds the spatial tolerance limit, the distance score is zero. Under a set of implementation parameters, the spatial tolerance limit can be set to 5000m. Type consistency is calculated using an exponential decay function, and the formula is as follows: in, For type consistency, This represents the shortest path distance between two subcategories in the type tree. The preset attenuation base and its value range is [value range missing]. The consistency of administrative divisions is determined by the corresponding preset quantitative score based on the matching level of the administrative division codes: the first preset division value is assigned when the administrative division codes are completely consistent; the second preset division value is assigned when the municipal-level administrative division codes are consistent but the district / county-level administrative division codes are inconsistent; and the third preset division value is assigned in other cases. The first preset division value is greater than the second preset division value, and the second preset division value is greater than the third preset division value. Under a set of implementation parameters, the grouping weights of semantics, space, type, and administrative division can be initialized to 0.40, 0.30, 0.20, and 0.10 respectively, and the preset aggregation threshold value can be set to 0.78. If the overall score is not lower than 0.78, the two points of interest will be aggregated into the same candidate point of interest group; for non-standard subject name and alias fields, by setting the group weight of semantic similarity feature items greater than the preset ratio coefficient, and combining the spatial distance threshold value and the consistency judgment condition of subdivided type labels, entity objects whose subject name characters match but whose spatial location does not meet the threshold value or whose business type does not match are filtered. When the number of interest points in a single candidate interest point group exceeds the preset number, the system performs a secondary segmentation; under a set of implementation parameters, the preset number can be set to 50. The application server reads the spatial grid code of each point of interest in the group and re-divides the data set according to a spatial grid code that is one or more levels finer than the current spatial grid code, thereby dividing the candidate point of interest group with a number exceeding a preset threshold into multiple subgroups. This secondary segmentation step is used to limit the data processing scale of a single relationship comparison in the subsequent relationship establishment stage, and to avoid continuously aggregating objects whose spatial distance is greater than the preset distance threshold but whose name matching degree meets the conditions into the same group. After receiving the candidate interest point group, the relationship establishment module first organizes the interest points in the group into unified structured data in the application server. Each data point includes at least the organization name field, alias field, sub-type label, level field, administrative division code and spatial location. The relationship establishment module constructs a hierarchical semantic graph based on the acquired candidate interest points. Specifically, it integrates the structured data of each interest point within a group into a candidate cluster, and assembles the extracted institution name field, hierarchy field, and sub-type labels into a formatted text sequence. The formatted text sequence of the candidate cluster is then input into a pre-trained large language model. The large language model is guided by a preset hierarchical reasoning prompt template to perform deep reasoning analysis on the parent-child and sibling relationships between complex entities such as affiliated hospitals, outpatient buildings, and industrial parks, based on the context. The system receives hierarchical relationship data in a tree structure, containing node identifiers and relationship sub-type labels, output by the large language model to construct a hierarchical semantic address knowledge graph. Furthermore, it extracts parent-child and sibling relationship records with a relationship confidence level not lower than a preset threshold and writes them into an address association table. In one specific embodiment, to fully highlight the core capabilities of the large language model in structured reasoning of address hierarchy relationships, the system constructs the following preset hierarchical reasoning prompt template and inputs it into the pre-trained large language model: "You are an address semantic parsing expert; the following is a list (JSON format) of candidate points of interest (POIs) belonging to the same entity; please use deep semantic reasoning capabilities to analyze the physical inclusion relationship (parent-child) and functional parallel relationship (sibling) between them." Constraints: First, the spatial scope of a child entity must be within its parent or functionally subordinate to its parent; second, only JSON-formatted node-edge relationships should be output, and no additional explanations are allowed. Input data: [A formatted text sequence containing the organization name field, hierarchy field, and sub-category label]. The system receives the JSON array output by the model and performs format validation. At the model inference end, it filters edge relationships with confidence scores below a preset initial threshold, extracts relationship data with confidence scores above the preset initial threshold for subsequent judgment, and persistently stores hierarchy relationship records with confidence scores above the first relationship score threshold in the address association table. Under a set of implementation parameters, the first relationship score threshold can be set to 0.90; if the relationship score is not lower than 0.90, the application server directly writes the relationship into the address association table; if the relationship score is lower than 0.90, the corresponding relationship is written into the manual review queue. The manual review queue can be read by the review terminal. After the review result is returned, the application server writes the confirmed relationship into the address association table. Under another set of preferred parameters, the second relationship score threshold can also be set to 0.75. Relationships below this value are given priority to be sent to the manual review queue, while relationships above this value but below the first relationship score threshold can also enter the manual confirmation process. Based on the dual threshold mechanism, relationship data with scores between the first relationship score threshold and the second relationship score threshold are judged to have differences in individual fields, while relationship data with scores below the second relationship threshold are judged to have conflicts in multiple feature elements. Both are sent to the manual review queue for confirmation. After the above modules are completed, the address reconstruction output module can use standard points of interest and address association tables to generate standardized address records, hierarchical address expressions, or unified address identifiers. Standardized address records must include at least the standard point of interest name, administrative division code, standard sub-type label, and standard point of interest identifier; hierarchical address representation is used to show the hierarchical link from parent entity to child entity; The unified address identifier is used for external business systems to call it in a fixed number manner; the address reconstruction output module generates and outputs standardized address records aligned to the last-level entity node based on the separated and processed organization entity name field, hierarchy field, subdivision type label and spatial grid code, as well as the hierarchical mapping relationship recorded in the address association table. In this embodiment, when the above system is applied to the hierarchical address data processing of universities, hospitals, and industrial parks, the address preprocessing module, the candidate interest point grouping module, and the relationship establishment module perform dimensionality reduction and separation calculations on entities with hierarchical relationships, thereby improving the accuracy of entity identification and matching at each level.

[0020] In a preferred embodiment of the present invention, the address matching module, in the first matching mode, i.e., the retrieval enhancement generation matching mode, is used to perform the following steps: Zero-shot type prediction and address classification based on a large language model are performed on the business address text to obtain the target address type; Based on the target address type, multiple query channels are established using county and type combined index sharding, city and type combined index sharding, alias index sharding, relational adjacency index sharding, and semantic feature index sharding. After converting the business address into address matching parameters that include at least one or more of the main word encoding, type encoding, zone encoding and text feature representation, multi-path data extraction is performed, and the extraction results are merged and deduplicated to obtain a set of candidate interest points, which are then input into the large language model to perform semantic fine screening. Based on the scores and candidate set ranking results after semantic refinement of the large language model, a preset number of candidate interest points are selected to determine the standard interest points, and the corresponding standard interest point identifiers are output.

[0021] This implementation describes the execution process of the address matching module in the associated query matching mode; the module is deployed on the application server, calls the multi-way index in the retrieval server, and calls the address association table, alias library and address encoding library in the database server; The associated query matching method refers to retrieving candidate points of interest in parallel across multiple index shards based on the target address type and administrative division information of the business address, and then reordering the candidate points of interest using the hierarchical relationship in the address association table. When a business system submits a business address to be matched, the application server first splits the business address text into address elements, extracting main words, candidate sub-type labels, administrative division keywords, alias keywords, and text feature representations; the target address type is the address category given by the address classification step, such as school, hospital, commercial building, or building, etc. The target address type is used as a filtering condition to limit the recall search range of subsequent multi-channel retrieval; if the business address contains multiple type prompt words at the same time, the address matching module retains one or more type codes with a predicted score not lower than the preset type score threshold as candidate types, and uses them respectively in subsequent query channels; After determining the target address type, the retrieval server establishes multiple query channels according to index shards; the district / county combined index shard is used to prioritize the retrieval of standard points of interest with complete administrative division codes and matching sub-type labels; the city-level combined index shard is used when district-level information is missing but city-level information exists; Alias ​​index fragments are used to recall abbreviations, common names, or historical names; relational adjacency index fragments are used to recall points of interest that are adjacent to or related to known entities based on the address association table. Semantic feature index sharding is used to recall semantically similar points of interest by representing text features; for missing information features in business address text records, multiple query channels are established to avoid the data recall volume being lower than the preset requirements due to the use of a single retrieval channel; The address matching module converts the business address into address matching parameters; the address matching parameters include at least one or more of the following: subject word encoding, type encoding, zone encoding, and text feature representation; The subject word encoding uses a non-cryptographic hash algorithm to map the subject words stored in the basic database into fixed-length globally unique integer identifiers and store them in the address encoding library. When a hash collision occurs, the administrative division code to which the subject word belongs is introduced as a secondary salt value for secondary hashing to ensure the uniqueness of the integer identifier in the global address encoding library. The type encoding comes from the type tree mapping, the division encoding comes from the administrative division parsing result, and the text feature representation comes from the text encoding model consistent with the preprocessing stage. The application server sends these parameters to each query channel to perform multi-channel data extraction. The retrieved results are then merged by primary key and deduplicated within the application server to obtain a set of candidate points of interest. After obtaining the set of candidate interest points, the address matching module calculates the first round of matching scores; for each candidate interest point, the encoding matching result, type matching result, administrative division matching result, alias matching result, relational adjacency matching result, and semantic similarity result are judged respectively. The encoding matching results reflect whether the encoding of the main words of the business address is consistent with the standard encoding of the candidate points of interest; the type matching results reflect whether the target type of the business address is consistent with the sub-type label of the candidate points of interest. The administrative division matching result reflects the consistency of district / county or city-level codes; the alias matching result reflects whether the business address keyword appears in the candidate point of interest alias field; the relational adjacency matching result reflects whether the candidate has an adjacency relationship with other known hierarchical nodes; The aforementioned five results all adopt discrete quantization mapping rules: when they are completely consistent or hit, the corresponding result is 1.0, otherwise it is 0.0; the semantic similarity result reflects the degree of closeness between the business address text feature representation and the candidate interest point text feature representation. The specific quantization calculation rule is: calculate the cosine similarity between the business address text feature representation vector and the candidate interest point text feature representation vector, that is, the dot product of the two vectors divided by the product of the magnitudes of the two vectors; When performing this calculation, if the magnitude of any vector is detected to be zero, the semantic similarity result is directly reset to 0.0 to avoid division by zero anomalies. When the similarity is less than 0, it is forcibly truncated and reset to 0.0, thereby strictly mapping the semantic similarity result to the normalized range of 0.0 to 1.0; the above six normalized results are weighted and summed according to the preset matching weight to obtain the matching score, and a preset number of candidate interest points are selected and sorted according to the score; under a set of implementation parameters, the preset number can be set to 30, which is configured to limit the processing scale of subsequent data re-sorting under the preset data recall rate index; For the sorted candidate points of interest, the address matching module continues to calculate the final matching score; the final matching score is a weighted sum of the encoding score, type score, area division, level score, alias score and distance score; the distance score is only included in the calculation when the business address has coordinates attached or the coordinates can be obtained from the business context; If there are no coordinates, the distance score can be set to zero. At the same time, in order to ensure the consistency of the full score benchmark and avoid the evaluation deviation caused by the decrease in the total score, resulting in the value being lower than the fixed matching qualification threshold, the system dynamically normalizes and amplifies the initial matching weights of the remaining five items according to their original relative proportions, so that the sum of the weights of the remaining items involved in the calculation is equal to 1.0 again. To ensure consistency in quantification, the mapping rule for distance scores is as follows: within the spatial distance range of 0 to 500m, the distance score is always 1.0; within the range of 500 to 2000m, it linearly decays to 0.5 with distance; and beyond 2000m, an exponential decay function based on distance is used to smoothly transition to 0. Under a set of implementation parameters, the initial matching weights for each item can be set to 0.35, 0.20, 0.15, 0.15, 0.10, and 0.05, respectively. In the above weight configuration, the sum of the matching weights corresponding to the subject word matching and text encoding related item scores is set to 0.55, configured to prioritize the uniqueness of the institution subject name field; the hierarchical matching weight is set to 0.15, configured to distinguish between parent and child entities by adjusting the scores when there are hierarchical attribute differences between the business address to be matched and the candidate point of interest. The calculation of hierarchical scores depends on the address association table; the application server first queries whether the current candidate point of interest has child nodes and whether the business address contains child node feature words; If a candidate point of interest has a child node associated record, and the business address contains words that match the preset child node feature word library, then when the current candidate point of interest is the corresponding parent node, the system will generate a negative score adjustment for the candidate point of interest. If the candidate point of interest corresponds to the same level as the business address, and the business address contains information on the level of affiliated institutions, buildings, or departments, and it completely matches the level field to which the candidate point of interest belongs, then a positive score is awarded. In all other cases, a score of zero is recorded. In specific quantitative examples, the aforementioned negative score can be set to -1, and the positive score can be set to 1. To avoid the introduction of negative scores causing the final weighted sum to drop to an unreasonable negative number or causing abnormalities in the underlying sorting logic, the system sets up a score truncation protection mechanism when calculating the final matching score: if the final result of the weighted sum is lower than 0, it will be forcibly truncated and reset to 0. Through the aforementioned positive and negative scoring adjustment and zero-score cutoff mechanism, the matching score ranking priority of parent interest points that only match the institutional entity name field but do not match the hierarchical field is reduced, so that the comprehensive score is within the preset effective value range. After the final matching score is completed, the address matching module sorts the candidate points of interest in descending order of the final matching score, and further compares the consistency of the main words, the consistency of the hierarchical fields, the consistency of the subdivision type tags, and the alias hit rate of the top-ranked candidates. The fine-grained verification configuration performs a secondary element comparison on the candidate points of interest that rank high in the final matching score: if the subject words and hierarchical fields of the candidate point of interest are completely consistent with the business address, then the candidate point of interest is given the highest ranking priority; if the subject words are consistent but the hierarchical fields are inconsistent, then the alias match and subdivided type tags are combined for further judgment; if the subject word character matching degree reaches the preset matching threshold but the administrative division code is inconsistent, then its ranking priority is reduced. After the optimal standard point of interest is determined through careful verification, the address matching module also needs to perform threshold verification, that is, to determine whether the final matching score corresponding to the optimal standard point of interest is not lower than the matching qualification threshold. The matching qualification threshold is determined based on the retrospective calibration of a manual confirmation test set containing at least 10,000 historical business addresses. Specifically, it is constrained to ensure that the system's automatic recall accuracy reaches 95%, and the score corresponding to the inflection point of the accuracy-recall curve is selected as the benchmark threshold. Under a set of preferred implementation parameters, the qualified matching threshold after calibration is set to 0.65; if it is not lower than the qualified matching threshold, the best standard interest point is determined as the final standard interest point and passed to the address reconstruction output module to generate standardized address results. When the final matching score is lower than the matching qualification threshold, the system will output the current candidate point of interest as a result pending manual review and pass it to the closed-loop management module; For example, the business address is written as "Outpatient Building of the First Affiliated Hospital of Guangxi Medical University". The main keyword "Guangxi Medical University" is consistent with the main name field of the standard point of interest. The alias keyword "First Affiliated Hospital" hits the alias field of the candidate point of interest. The business address also includes the sub-node feature word "Outpatient Building". If Guangxi Medical University, the First Affiliated Hospital of Guangxi Medical University, and the outpatient building of the First Affiliated Hospital of Guangxi Medical University all exist in the candidate set, all three may be recalled in the first round of matching. However, in the final matching stage, the parent Guangxi Medical University is penalized because it has child nodes and its business address contains child node feature words. The First Affiliated Hospital also scores lower than the outpatient building node because its level is not completely consistent. Therefore, the outpatient building node is selected as the best standard point of interest. When the sum of the final matching scores of the last-level candidate interest points in multiple dimensions is not lower than the aforementioned matching qualification threshold, the last-level candidate interest point is determined as the standard interest point. When the final matching score of the business address to be matched is lower than the matching qualification threshold, the closed-loop management module terminates the automatic output of standardized address records and sends the corresponding record to the review queue to generate a result pending manual review.

[0022] In a preferred embodiment of the present invention, the address matching module is configured to perform the following operations in the second matching mode, i.e., the dictionary matching mode: Based on professional and general terminology databases, business addresses are professionally segmented and address elements are split to generate a set of Chinese keywords and a set of Pinyin keywords. Initial screening is performed based on preset weighting factors and adjustable coordinate precision. The adjustable coordinate precision is configured to initially limit the search range in the coordinate space according to the set latitude and longitude tolerance. The initial matching weights are determined based on historical data according to the address number, main body words of the organization, and type words. The matching is performed step by step, including at least one or more of the following: exact match, fuzzy match, pinyin match, and hierarchical priority match. Exact match means that the business address keyword and the corresponding field of the point of interest are completely consistent. Fuzzy match means that the character matching degree of the business address keyword and the corresponding field of the point of interest is not lower than the preset fuzzy match threshold. Pinyin match means that the pinyin representation of the business address keyword and the corresponding field of the point of interest are consistent or have the same prefix. Hierarchical priority match means that when a hierarchical field exists, the hierarchical field is compared first and then the main body word is compared. The initial candidate set is screened by combining spatial distance and preset rule verification to determine the final matching result and output standard interest point identifiers. The aforementioned preset rule verification includes at least the door number consistency verification and the institution type alignment verification. At the same time, a dynamic weight adjustment mechanism based on feedback from manual review results is introduced to generate weight adjustment values, and the weight of the hit keyword categories is dynamically optimized and updated separately. The closed-loop management module is used to send the following data into the manual review process: relationship data in the relationship establishment module whose relationship score is lower than the second relationship score threshold, wherein the second relationship score threshold is less than the first relationship score threshold; business address records in the aforementioned first matching mode that did not return candidate points of interest; And business address records that return candidate points of interest in the thesaurus matching mode but whose highest matching score is lower than the matching qualification threshold; and after manual review and confirmation, perform at least one of the following update operations: write the correspondence between business addresses and standard points of interest into the alias database; update the keyword category weights in the thesaurus matching mode; Update the grouping weights in the aforementioned candidate point of interest grouping module; write the confirmed hierarchical relationship into and replace the original relationship record in the address association table.

[0023] In the second matching mode, the address matching module uses the thesaurus matching mode to perform retrieval and matching. The closed-loop management module receives sample data with matching scores below the matching qualification threshold and sends them to the manual review process. Based on the manual review results, it writes back and updates the database parameters. This method also runs on the application server and the database server, and can complete address standardization without relying on full semantic reordering. Once the address matching module receives the business address, it performs address element splitting to generate a set of Chinese keywords and a set of Pinyin keywords. The set of Chinese keywords is used to directly compare with the main name field, alias field, address number field, and type field of the point of interest organization. The set of pinyin keywords is used to handle homophones, near-homophones, or cases where pinyin is used during input; the determination of the initial matching weight is based on the historical distribution statistics in the basic database, specifically calculated based on the inverse document frequency of various words in the historical address database, and quantitative intervals are set according to address number, main body words of organization, and type words respectively; Since the repetition rate of house numbers in different addresses is lower than the preset repetition probability threshold and the distinguishing ability is higher than the preset distinguishing feature threshold, the initial matching weight parameter of the house number is configured in the range of 0.4 to 0.5. The main body keywords are used to identify the entity itself, and their weight range is set to 0.3 to 0.4; the category keywords are used to indicate the industry category attribute of the point of interest, and their initial matching weight range is set to 0.1 to 0.2. During the step-by-step matching process, the address matching module first performs a complete match; if the business address keywords and the corresponding fields of the points of interest are completely consistent, then that field is set as the first preset matching weight score. If they are not completely identical, then fuzzy matching is entered; the fuzzy matching threshold is set to be no less than the preset fuzzy matching threshold for the main word, which can be set to no less than 0.85 under a set of implementation parameters; the character matching degree can still be calculated by edit distance, longest common subsequence or a combination thereof; When fuzzy matching fails to distinguish multiple candidates, the module performs pinyin matching to determine whether the pinyin keywords of the business address are consistent with or have the same prefix as the pinyin representation of the corresponding field of the point of interest. If there is a clear hierarchy field in the business address, then hierarchy priority matching will be started, and the hierarchy field will be compared first. If the hierarchy field matches, the main words will be compared further. For entity records with multi-level hierarchical structures and repeated main words, the comparison and judgment based on the hierarchical field is performed first, and the hierarchical consistency judgment result is used as a limiting condition for the ranking priority of candidate points of interest. The score for the keyword matching mode is the sum of the hit results of each type of keyword and their weights; when a house number is hit, the weight of the house number contributes to the score; when a main word is hit, the weight of the main word contributes to the score. When a type word is hit, the type word weight is added to the score; when there is a hierarchical priority hit, candidate interest points with the same level are selected first under the condition of equal subject words. The system sorts the returned candidate points of interest by total score and selects the one with the highest score as the candidate standard point of interest. If the highest matching score is not lower than the matching qualification threshold, the standard point of interest is directly output. When there are candidate points of interest and the highest matching score is lower than the matching qualification threshold, the system outputs the result pending manual review. In this embodiment, the closed-loop management module receives three types of data; the first type is relationship data from the relationship establishment module whose relationship score is lower than the second relationship score threshold; The second category is business address records that did not return candidate points of interest in the first matching mode mentioned above; the third category is business address records that returned candidate points of interest in the thesaurus matching mode but whose highest matching score was lower than the matching qualification threshold. The second relationship score threshold is less than the first relationship score threshold; when the first relationship score threshold is set to 0.90 and the second relationship score threshold is set to 0.75, relationship data with a relationship confidence score lower than 0.75 are sent to the manual review queue. After manual review is completed, the closed-loop management module performs at least one update operation based on the review results; if it is manually confirmed that a certain business address is a commonly used abbreviation or historical name of a certain standard point of interest, the correspondence between the business address and the standard point of interest is written into the alias database. The next time the same or similar spelling is encountered, the new alias field can be used directly for retrieval and scoring in both the thesaurus matching mode and the associated query matching method; if the thesaurus matching is confirmed to be correct by manual verification, a positive weight adjustment value will be added to the keyword category of this hit. If a mismatch is manually confirmed, a negative weight adjustment value is added to the keyword category that was hit this time; the keyword category that was not hit does not participate in the update; combined with a set of optional implementation parameters, the update logic of the keyword category weight is: add the current period weight to the product of the preset learning rate, the review feedback value and the hit feature value to obtain the updated weight for the next period. Among them, the review feedback value takes a positive value, such as +1, when the match is correct, and a negative value, such as -1, when the match is incorrect; the hit feature value takes 1 when the corresponding category is hit, and 0 when the category is not hit. The learning rate can be set in the range of 0.01 to 0.05. To avoid continuous negative feedback causing excessive decay of the weight of a certain type of keyword or even turning it into a negative value, thus causing abnormal score calculation, the system introduces a lower limit protection mechanism for the underlying weight. After each formula calculation, if the resulting weight for the next cycle is less than a specific percentage of the initial weight lower limit for that keyword type, for example... If so, its callback will be forcibly corrected to this dynamic lower limit value; Meanwhile, to prevent the unlimited growth of a certain type of weight due to long-term continuous positive feedback from disrupting the relative balance of the preset initial weight range, the system simultaneously introduces a weight upper limit protection mechanism: if the weight obtained in the next cycle exceeds the preset maximum weight cutoff value, such as 1.2 times the upper limit of the initial weight range of this type of keyword, it will be forcibly cut off and corrected to the maximum weight cutoff value. The aforementioned weight update and two-way truncation protection mechanism is configured to dynamically reduce the weight parameters of general organization suffix words, and increase the weight values ​​of main words, hierarchical words and address number matching categories when the weight parameters are between the initial dynamic lower limit of the weight and the maximum weight truncation value. In addition to updating the keyword category weights in the thesaurus matching mode, the closed-loop management module can also update the grouping weights in the aforementioned candidate interest point grouping module; When the manual review results indicate that the spatial distance is within the preset range and the grouping error rate of inconsistent type characteristics exceeds the preset proportion, the weight of type consistency will be dynamically increased or the weight of distance will be decreased. When the proportion of misgrouped records in the abbreviation and full name matching fields exceeds the preset proportion threshold, the weight coefficient of attribute information similarity in the comprehensive score aggregation method is increased, and the calculation parameters of the candidate interest point grouping module and the address matching module are updated synchronously. For the correction of hierarchical relationships, when the closed-loop management module receives the relationship results after manual confirmation, it writes the confirmed hierarchical relationship into and replaces the original relationship record in the address association table. Replacement operations are applied first to relationship records that were originally written automatically with low confidence or are pending review, so that subsequent related queries and hierarchical priority matching use the latest manually confirmed results. In this implementation, even without enabling complex semantic reordering, the system can still achieve continuous improvement by relying on dictionary matching, pinyin matching, hierarchical priority matching, and manual review and write-back. The closed-loop management module is configured to write non-standard address text representations, spelling variants, and hierarchical verification results into the alias library and address association table, respectively, and update the corresponding weight parameters. The updated alias library, address association table, and weight parameters are then used to perform normalization processing on subsequent address records.

[0024] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are merely for explaining the principles of the invention and should not be construed as limiting the scope of protection of the invention in any way. Based on this explanation, those skilled in the art can readily conceive of other specific embodiments of the invention without inventive effort, and these equivalent variations or substitutions are all included within the scope defined by the claims of this application.

Claims

1. An address data reconstruction system integrating semantic retrieval technology, characterized in that, include: The basic database is used to store the organization's suffix thesaurus, alias thesaurus, type tree, address encoding library, thesaurus matching weight, preset matching weight, grouping weight, relation rule weight, relation judgment rules, preset aggregation threshold, matching qualification threshold and retrieval configuration parameters; The address preprocessing module is used to acquire multi-source map point of interest data and split the features, extract the organization name field, hierarchy field, subdivision type label, alias field, administrative division code and spatial grid code, and generate text feature representation and point of interest structured record based on name or alias; The candidate interest point grouping module is used to aggregate interest points with related attributes and spatial proximity based on structured records to generate candidate interest point groups; The relationship establishment module is used to analyze the parent-child and sibling relationships between interest points based on the acquired candidate interest point grouping, and guide the large language model to analyze the relationship between interest points through a preset hierarchical reasoning prompt word template, generate hierarchical relationship data, and parse and extract relationship records in which the relationship confidence is not lower than the first relationship score threshold, and write them into the address association table. The address matching module is used to perform retrieval and matching of the business address to be matched according to the first matching mode or the second matching mode, and calculate the final matching score for each candidate point of interest. When the final matching score is the highest and not lower than the preset matching qualification threshold in the basic database, the corresponding candidate interest point is used as the standard interest point and the standard interest point identifier is output. When the final matching score is lower than the matching qualification threshold, the result pending manual review is output. The first matching mode is a retrieval enhancement generation matching mode based on large language model prediction and vector recall, and the second matching mode is a word matching mode based on word segmentation of professional word library and general word library. The address reconstruction output module is used to generate standardized records, hierarchical address representations, or unified identifiers based on standard points of interest and address association tables. The closed-loop management module is used to receive relationship data that is lower than the first relationship score threshold, business address records that have not returned candidate points of interest, business address records whose final matching score is lower than the matching qualification threshold, and manual review results. It updates the alias library, type tree, encoding library, word library matching weight and grouping weight, and writes relationships with a relationship confidence of not lower than the first relationship score threshold or those that have been manually confirmed into the address association table.

2. The address data reconstruction system integrating semantic retrieval technology according to claim 1, characterized in that, The address preprocessing module includes: The entity name extraction unit is used to strip the suffixes of the point of interest names based on the organization suffix thesaurus, and extract and store the organization entity name field; The hierarchical term extraction unit is used to extract and store hierarchical fields that represent organizational hierarchy or building functional hierarchy; Type subdivision unit, used to compare the original industry classification with the type tree and generate subdivision type labels; Alias ​​mapping unit, used to generate alias fields for points of interest based on the alias library; The zoning location unit is used to generate administrative division codes and spatial grid codes corresponding to points of interest; The semantic feature generation unit is used to generate text feature representations for semantic retrieval based on the point of interest name and alias fields.

3. The address data reconstruction system integrating semantic retrieval technology according to claim 1, characterized in that, The candidate interest point grouping module is used to group two interest points into the same candidate interest point group when the following conditions are met simultaneously: The administrative division codes are consistent; The character matching degree of the organization name field is not lower than the preset character matching threshold value; The spatial distance is no greater than a preset distance threshold. The first-level sub-type labels are consistent or belong to the same type of tree parent node.

4. The address data reconstruction system integrating semantic retrieval technology according to claim 1, characterized in that, The candidate interest point grouping module is used to: calculate the comprehensive score of two interest points based on attribute information, spatial information, type information and administrative division information determined by the extracted institution name field, sub-type label, administrative division code and spatial grid code and text feature representation, and aggregate the two interest points into the same candidate interest point group when the comprehensive score is not lower than the preset aggregation threshold value; When the overall score is lower than the preset aggregation threshold, aggregation is not performed; When the number of interest points in a single candidate interest point group exceeds a preset number, the candidate interest point group is further segmented according to a spatial grid code that is one or more levels finer than the spatial grid code.

5. The address data reconstruction system integrating semantic retrieval technology according to claim 1, characterized in that, The specific steps for constructing the hierarchical semantic graph by the relationship establishment module are as follows: The structured records of each interest point in the candidate interest point group are obtained and integrated into a candidate cluster input into the large language model. The model is guided to deeply analyze the parent-child and parallel relationships between each interest point in the candidate cluster according to the preset hierarchical reasoning prompt word template. The extracted institutional main name field, hierarchical field and subdivision type label are assembled into a formatted text sequence with sequence identifier. The formatted text sequence is input into a pre-trained large language model, and the hierarchical relationship data in a tree structure, which includes node identifiers and relationship subdivision type labels, is received from the output of the large language model. Parse the hierarchical relationship data of the tree structure, extract and receive the relationship data containing node identifiers and relationship sub-type labels, as well as the relationship confidence level; When the confidence level of a relationship is not lower than the first relationship score threshold, the corresponding parent-child relationship and sibling relationship are written into the address association table. When the confidence level of a relationship is lower than the first relationship score threshold, the corresponding relationship is sent to the manual review queue and written into the address association table after manual confirmation.

6. The address data reconstruction system integrating semantic retrieval technology according to claim 1, characterized in that, In the first matching mode, i.e., the retrieval enhancement generation matching mode, the address matching module performs the following steps: Zero-shot type prediction and address classification based on a large language model are performed on the business address text to obtain the target address type; Based on the target address type, a multi-way query channel is established using the combined county and type index sharding, the combined city and type index sharding, the alias index sharding, the relational adjacency index sharding, and the semantic feature index sharding. After converting the business address into address matching parameters that include at least one or more of the main word encoding, type encoding, zone encoding and text feature representation, multi-path data extraction is performed, and the extraction results are merged and deduplicated to obtain a set of candidate interest points, which are then input into a large language model to perform semantic fine screening. Based on the scores and candidate set ranking results after semantic fine screening of the large language model, a preset number of candidate interest points are selected to determine the standard interest points, and the corresponding standard interest point identifiers are output.

7. The address data reconstruction system integrating semantic retrieval technology according to claim 6, characterized in that, The address matching module is used to compare and match each candidate point of interest with the corresponding field of the business address to obtain the encoding matching result, type matching result, administrative division matching result and alias matching result. The module then calculates the final matching score by combining the hierarchical structure information and distance information, and sorts the results in descending order according to the final matching score. The address matching module is also used to compare the consistency of main words, the consistency of hierarchical fields, the consistency of subdivided type tags, and the alias hits of the candidate interest points whose final matching scores rank within the aforementioned preset number, and output the standard interest points.

8. The address data reconstruction system integrating semantic retrieval technology according to claim 1, characterized in that, In the second matching mode, i.e., the dictionary matching mode, the address matching module is configured to perform the following operations: Based on professional and general terminology databases, business addresses are professionally segmented and address elements are split to generate a set of Chinese keywords and a set of Pinyin keywords. Initial screening is performed based on preset weighting factors and adjustable coordinate precision. The adjustable coordinate precision is configured to initially limit the search range in the coordinate space according to a set latitude and longitude tolerance. The initial matching weights are determined based on historical data according to the address number, main body words of the organization, and type words. The matching is performed step by step, including at least one or more of the following: exact match, fuzzy match, pinyin match, and hierarchical priority match. Exact match means that the business address keyword and the corresponding field of the point of interest are completely consistent. Fuzzy match means that the character matching degree of the business address keyword and the corresponding field of the point of interest is not lower than the preset fuzzy match threshold. Pinyin match means that the pinyin representation of the business address keyword and the corresponding field of the point of interest are consistent or have the same prefix. Hierarchical priority match means that when a hierarchical field exists, the hierarchical field is compared first and then the main body word is compared. The initial candidate set is screened by combining spatial distance and preset rule verification to determine the final matching result and output standard interest point identifiers. The aforementioned preset rule verification includes at least the door number consistency verification and the institution type alignment verification. At the same time, a dynamic weight adjustment mechanism based on feedback from manual review results is introduced to generate weight adjustment values, and the weight of the hit keyword categories is dynamically optimized and updated individually.

9. The address data reconstruction system integrating semantic retrieval technology according to claim 8, characterized in that, The closed-loop management module is used to send the following data into the manual review process: Relationship data in the relationship establishment module whose relationship score is lower than the second relationship score threshold, wherein the second relationship score threshold is less than the first relationship score threshold; The aforementioned first matching mode did not return any candidate points of interest in the business address records; And business address records that return candidate points of interest in the thesaurus matching mode but whose highest matching score is lower than the matching qualification threshold; After manual review and confirmation, perform at least one of the following update operations: Write the mapping between business addresses and standard points of interest into the alias database; Update the keyword category weights in the thesaurus matching pattern; Update the grouping weights in the aforementioned candidate interest point grouping module; Write the confirmed hierarchical relationship into the address association table and replace the original relationship record.

10. The address data reconstruction system integrating semantic retrieval technology according to claim 1, characterized in that, The basic database is also used to store semantic retrieval indexes corresponding to text feature representations; the text feature representations refer to vectors that can be used for semantic similarity calculation, obtained by encoding the text of point of interest names, alias fields, or business addresses.

Images