POI data fusion method and system based on large language model

Abstract

The invention provides a POI data fusion method and system based on a large language model, and the method comprises the steps: carrying out the fusion detection of POI data after the first fusion, screening out a first POI candidate set, carrying out the multi-level analysis of a fusion omission mode of the first POI candidate set in the characteristics of spatial, semantic, time sequence changes, and the like, and carrying out the fusion of the first POI candidate set. According to the missing mode, secondary fusion detection is carried out to obtain a second POI candidate set, the large language model is in butt joint with an external POI database through data retrieval enhancement, the POI data recall efficiency of the large language model is improved, and finally a fusion decision of the second POI candidate set is generated according to recalled POI candidate knowledge fragments; according to the POI data fusion method, the second POI candidate set is obtained by performing omission analysis on the suspected repeated POI pairs, that is, omission mode analysis data is used as reference knowledge to be input into a large language model, and a retrieval enhancement technology is combined to replace a traditional linear weight formula, so that a self-adaptive adjustment fusion strategy is realized, and the fusion efficiency is improved. And the POI fusion accuracy and the processing efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of POI fusion, and in particular to a method and system for POI data fusion based on a large language model. Background Technology

[0002] Points of Interest (POIs) are core data resources in Geographic Information Systems (GIS) and location services, widely used in map navigation, location recommendation, smart cities, and big data analytics. With the rapid growth of multi-source data (government data, internet map platforms, user crowdsourcing, commercial organizations, etc.), the scale of multi-source POI data has expanded rapidly. The goal of POI fusion and deduplication is to remove duplicates, unify standards, and improve data usability; however, existing technologies still have significant bottlenecks, and the following problems commonly exist in the fusion process of multi-source POI data: Currently, mainstream POI data fusion methods are mainly based on traditional machine learning and rule matching techniques. Existing technologies, exemplified by patent CN117056374A, typically employ a three-stage processing architecture: "GeoHash grid recall + tree model coarse ranking + linear weight formula decision." Specifically, first, GeoHash encoding is used to divide the geographic space into grids, and POIs are recalled within the same or adjacent grids. Then, tree models such as XGBoost and Random Forest are used for coarse ranking based on features such as name similarity, address matching, and coordinate distance. Finally, a comprehensive similarity is calculated using the linear weight formula S = Q×N + A×M (where Q and A are weight coefficients, N is name similarity, and M is address similarity), and a threshold is set for fusion judgment. The main limitation of this method is its reliance on manually designed feature engineering and fixed weight configurations, making it difficult to handle complex semantic changes and dynamic business scenarios. This leads to the possibility that even after fusion, some POIs may not be successfully fused into duplicate data, requiring manual secondary screening, resulting in low efficiency and accuracy. Summary of the Invention

[0003] To address the above issues, this invention solves the problems of insufficient semantic understanding, inadequate knowledge utilization, unclean fusion, and fusion errors in existing POI fusion processing. It performs fusion detection on the POI data after the first fusion to screen out suspected duplicate POI pairs, analyzes the omission patterns of suspected duplicates, and performs a second POI recall based on the analysis results to obtain a second POI candidate set that is more suitable for the current POI data. Furthermore, it uses a large language model combined with retrieval enhancement technology to replace the traditional linear weight formula, effectively improving the accuracy and efficiency of POI fusion.

[0004] This invention is achieved through the following technical solution: Firstly, this invention provides a POI data fusion method based on a large language model, comprising: S10: Obtain the POI dataset after one fusion as the initial dataset, and perform duplicate POI detection on each POI data in the initial dataset to obtain the first POI candidate set; wherein, the first POI candidate set includes multiple first suspected duplicate POI pairs, and each POI pair includes POI pair information, suspected type identifier and confidence score; S20: Perform in-depth analysis on the spatial distribution features, semantic features, and temporal change features of the first POI candidate set, identify the omission patterns after one fusion, and obtain omission pattern analysis data. S30: Based on the fusion omission pattern analysis data, perform knowledge recall on the POI dataset after the first fusion to obtain a second POI candidate set; The second POI candidate set is input into a trained large language model for fusion decision generation. The retrieval module of the large language model is connected to a POI knowledge base. The specific steps are as follows: S401: Extract key entities and attributes from each second suspected duplicate POI pair in the second POI candidate set as retrieval conditions, perform condition retrieval in the POI knowledge base, and filter out candidate knowledge fragments related to the second suspected duplicate POI pair in the POI knowledge base. S402: Generate a query vector based on the second suspected duplicate POI pair in the second POI candidate set; calculate the semantic similarity, contextual relevance, and temporal validity of the candidate knowledge fragment based on the query vector and its corresponding key entities and attributes; and obtain the final score of the candidate knowledge by combining the POI confidence score of the second suspected duplicate POI pair; and complete the sorting and filtering of candidate knowledge. S403: Generate a fusion decision based on the matching degree between the data of another POI in the corresponding second suspected duplicate POI pair and the candidate knowledge; S50: The second POI candidate set is fused according to the fusion decision to obtain the final POI fusion result.

[0005] Further, step S20 includes: S201.a: Perform statistical analysis on the spatial distribution characteristics of the first suspected duplicate POI pairs in the first POI candidate set to obtain the distribution pattern of the first suspected duplicate POI pairs; S201.b: Compare the semantic features of the first suspected duplicate POI pairs in the first POI candidate set to obtain the category labeling differences of each POI in different data sources; S201.c: Record the temporal change characteristics of the POI information of the first suspected duplicate POI pair in the first POI candidate set as the result of temporal analysis; S202: Based on the distribution pattern, category labeling differences, and time series analysis results of the first suspected duplicate POI pair, generate fusion omission pattern analysis data.

[0006] Furthermore, the formulas for calculating semantic similarity, contextual relevance, and temporal validity in step S402 are as follows: semantic similarity :

[0007] Where q is the query vector and k is the candidate knowledge fragment vector; Contextual relevance :

[0008] in, Let i be the matching degree function for the i-th context feature. Indicates the weight of the corresponding feature; Time correlation:

[0009] Wherein, the time difference between the current time and the candidate knowledge update time, and λ is the time decay coefficient; Final score for candidate knowledge:

[0010] in, For semantic similarity weight parameters, For context-related weight parameters, For confidence weight parameters, This is the time-related weighting parameter.

[0011] Further, step S30 includes: S301: Based on the distribution pattern of the first suspected duplicate POI pair, perform dynamic search radius determination and multi-level geographic query on the POI dataset after one fusion to obtain the most relevant POI candidate data based on geographic proximity, and obtain geographic suspected duplicate POI data. S302: Based on the differences in category labeling of each POI in different data sources, semantic recall is performed on the POI knowledge in the POI dataset after one fusion. The candidate data of POIs that are synonymous with the corresponding POI are obtained by recalling keywords through the inverted index, and the TF-IDF score of each candidate POI is calculated to obtain the suspected duplicate POI data in the text. S303: Based on the time series analysis results, filter out POI data that have undergone time series changes in the fused POI dataset, and perform correlation queries on the filtered data in conjunction with a knowledge base to obtain suspected duplicate time series POI data. S304: Take the intersection of the geographic suspected duplicate POI data, text suspected duplicate POI data, and time series suspected duplicate POI data to obtain the second POI candidate set.

[0012] Furthermore, the large language model is trained through the following steps: The first training set is input into the general large language model for training until the objective function of the general large language model converges, and the converged general large language model is obtained. The first training set includes geographic information, business POIs and related question answering scenarios. For each lightweight adapter module inserted into the converged general-purpose large language model, prefix tuning is performed at the model input end. The real POI pairs, the fusion results of the annotations, and the high-quality domain data extracted from the historical question-answering logs in the knowledge base are obtained as the second training set. The second training set is input into the converged general-purpose large model for fine-tuning to obtain the trained large language model.

[0013] Furthermore, after step S50, the method further includes: S60: Add the omission pattern analysis data to the monitoring rules, monitor the final POI fusion result according to all monitoring rules, and obtain the fusion anomaly detection result. When fusion omission occurs, input the omission pattern corresponding to the monitoring rule and the suspected duplicate POI pair into the large model.

[0014] On the other hand, the present invention also provides a POI data fusion system based on a large language model, which includes: First POI candidate set generation module: used to obtain the POI dataset after one fusion as the initial dataset, and to perform duplicate POI detection on each POI data in the initial dataset to obtain the first POI candidate set; wherein, the first POI candidate set includes multiple first suspected duplicate POI pairs, and each POI pair includes POI pair information, suspected type identifier and confidence score; The fusion omission pattern analysis module is used to perform in-depth analysis on the spatial distribution features, semantic features, and temporal change features of the first POI candidate set, identify the omission patterns after one fusion, and obtain omission pattern analysis data. Second POI candidate set generation module: used to perform knowledge retrieval on the POI dataset after one fusion based on the fusion omission pattern analysis data to obtain a second POI candidate set; A large language model connected to a POI knowledge base, the large language model comprising: The retrieval enhancement submodule is used to extract key entities and attributes as retrieval conditions for each second suspected duplicate POI pair in the second POI candidate set, perform conditional retrieval in the POI knowledge base, and filter out candidate knowledge fragments related to the second suspected duplicate POI pair in the POI knowledge base. The POI data retrieval submodule is used to generate query vectors based on the second suspected duplicate POI pairs in the second POI candidate set, calculate the semantic similarity, contextual relevance, and temporal validity of the candidate knowledge fragments based on the query vectors and their corresponding key entities and attributes, and obtain the final score of the candidate knowledge by combining the POI confidence score of the second suspected duplicate POI pairs; and complete the sorting and filtering of candidate knowledge.

[0015] Decision generation submodule: Generates a fusion decision based on the matching degree between the data of another POI in the corresponding second suspected duplicate POI pair and the candidate knowledge; POI fusion module: Fusion the second POI candidate set according to the fusion decision to obtain the final POI fusion result.

[0016] Furthermore, the fusion omission pattern analysis module includes: Spatial pattern analysis submodule: used to perform statistical analysis on the spatial distribution characteristics of the first suspected duplicate POI pairs in the first POI candidate set, and obtain the distribution pattern of the first suspected duplicate POI pairs; Semantic pattern analysis submodule: used to compare the semantic features of the first suspected duplicate POI pairs in the first POI candidate set to obtain the category labeling differences of each POI in different data sources; The timeliness analysis submodule is used to record the time-series change characteristics of the POI information of the first suspected duplicate POI pair in the first POI candidate set as the time-series analysis result; Omission pattern analysis data generation submodule: used to generate fused omission pattern analysis data based on the distribution pattern of the first suspected duplicate POI pair, the difference in category labeling, and the time series analysis results.

[0017] Furthermore, the second POI candidate set generation module includes: The geographic proximity channel recall submodule is used to dynamically determine the search radius and perform multi-level geographic queries on the POI dataset after one fusion based on the distribution pattern of the first suspected duplicate POI pair, to obtain the most relevant POI candidate data based on geographic proximity relationship pairs, and to obtain the geographic suspected duplicate POI data. The text similarity recall submodule is used to perform semantic recall of POI knowledge in the POI dataset after one fusion based on the difference in category labeling of each POI in different data sources. It retrieves POI candidate data that are synonymous with the corresponding POI through inverted index keywords, and calculates the TF-IDF score of each candidate POI to obtain suspected duplicate POI data. The knowledge-enhanced recall submodule is used to filter POI data that have undergone time-series changes in the fused POI dataset based on the time-series analysis results, and perform correlation queries on the filtered data in conjunction with a knowledge base to obtain suspected duplicate POI data in the time series. Candidate set generation submodule: Take the intersection of the geographic suspected duplicate POI data, text suspected duplicate POI data, and time series suspected duplicate POI data to obtain the second POI candidate set.

[0018] Furthermore, the POI data fusion system also includes: Fusion quality monitoring module: used to add the missing pattern analysis data to the monitoring rules, monitor the final POI fusion result according to all monitoring rules, obtain fusion anomaly detection results, and when fusion omission occurs, input the missing pattern corresponding to the monitoring rule and the suspected duplicate POI pair into the large model.

[0019] This invention performs fusion detection on the POI data after one fusion, filters out suspected duplicate POI pairs, and then analyzes fusion omission patterns from multiple levels of features such as spatial, semantic, and temporal changes. Combining the fusion omission pattern data with knowledge retrieval from the POI data after one fusion, a second POI candidate set is obtained. Subsequently, based on a large language model and combined with retrieval enhancement technology, candidate knowledge fragments closest to the second suspected duplicate POI pairs in the second POI candidate set are retrieved from a standard knowledge base. A fusion decision is generated based on the candidate knowledge fragments and the second POI candidate set. Finally, the suspected duplicate POI pairs are fused according to the fusion decision. To avoid the re-omission of the same omission pattern, each detected omission pattern is used as a monitoring rule to monitor the final POI fusion result. When a fusion omission occurs, the omission pattern corresponding to the monitoring rule and the suspected duplicate POI pair are input into the large model, further improving the fusion accuracy. Attached Figure Description

[0020] Figure 1 This is a structural block diagram of the POI fusion system of the present invention; Figure 2 For execution Figure 1 The flowchart of the POI data fusion method of the system shown is illustrated. Figure 3 This is a structural block diagram of the fusion omission pattern analysis module of the present invention; Figure 4 For execution Figure 3 The flowchart of the omission pattern analysis method for the module shown; Figure 5 This is a structural block diagram of the second POI candidate set generation module of the present invention; Figure 6 For execution Figure 4 The flowchart shows the method for generating the second POI candidate set in the module shown. Detailed Implementation

[0021] POI data possesses unique characteristics that significantly distinguish it from general text data: First, it exhibits multimodal features, simultaneously containing textual information (name, address, description), geographic information (latitude and longitude coordinates), classification information (business category), and temporal information (business hours, update time). Second, it exhibits dynamic evolution, as POI information changes with factors such as business status, address changes, and brand restructuring. Third, it suffers from data source heterogeneity, with different map service providers, government departments, and enterprises employing different data collection standards and encoding methods. Fourth, it faces geospatial constraints, requiring simultaneous consideration of semantic similarity and geographic proximity in POI similarity assessment. These unique characteristics make traditional text matching and general machine learning methods ineffective. Therefore, this invention attempts to address the complex linguistic phenomena in POI fusion by leveraging the semantic understanding and reasoning capabilities of large language models. This allows fusion decisions to be dynamically adjusted based on the textual and contextual information within the POI data, thereby improving the accuracy of POI data fusion.

[0022] Based on the above analysis, the present invention proposes the following... Figure 1 This is a POI fusion system. It performs fusion detection on POI data after initial fusion, filtering out suspected duplicate POI pairs. Then, it analyzes the fusion omission patterns of suspected duplicate POI pairs from multiple levels based on spatial, semantic, and temporal variation features. Subsequently, it uses the fusion omission pattern data to perform knowledge retrieval from the initial fusion data to obtain a second POI candidate set. Then, based on a large language model and combined with retrieval enhancement techniques, it retrieves candidate knowledge fragments that are closest to the second candidate set from a standard POI knowledge base. Finally, it generates a fusion decision based on the candidate knowledge fragments and the second POI candidate set, obtaining a fusion scheme. Finally, it achieves the fusion of suspected duplicate POI pairs according to the fusion scheme. The system includes: a first POI candidate set generation module 10, a fusion omission pattern analysis module 20, a second POI candidate set generation module 30, a large language model 40, a POI fusion module 50, and a preset POI knowledge base. Please refer to [link to relevant documentation]. Figure 2 The working process between the modules in this system is as follows: The first POI candidate set generation module 10 is used to perform step S10: obtain the POI dataset after one fusion as the initial dataset, and perform duplicate POI detection on each POI data in the initial dataset to obtain the first POI candidate set; wherein, the first POI candidate set includes multiple first suspected duplicate POI pairs, and each POI pair includes POI pair information, suspected type identifier and confidence score.

[0023] This initial dataset uses a standardized data structure and includes key fields: a unique POI identifier, name, address, latitude and longitude coordinates, category label, contact number, and a source list field recording the data source fusion history, providing a complete data foundation for subsequent quality inspection. This step is specifically carried out in three directions: (1) The anomaly detection mechanism based on geographic density is as follows: by calculating the number of POIs of the same category within the radius around each POI, and comparing it with the historical average density of the area type, a density anomaly score is generated. When the density score exceeds a preset threshold, the area is marked as a high-density anomaly area, indicating that there may be duplicate POIs that have not been identified by traditional methods.

[0024] (2) Suspected duplicate detection based on name similarity: A name similarity detection model is constructed using the edit distance algorithm, and the similarity of POI names is calculated within the GeoHash seven-bit precision grid. By counting the minimum number of character insertion, deletion, and replacement operations, the standardized edit distance is calculated. Combined with spatial distance constraints, when the name similarity exceeds a preset threshold and the spatial distance is less than fifty meters, the POI pair is marked as a suspected duplicate object.

[0025] (3) Address representation difference detection mechanism, specifically: establish a multi-level address parsing model to extract key geographic identifiers such as province, city, district / county, street, house number, and building from the address. Use the longest common subsequence algorithm to calculate address text similarity, and at the same time construct a house number proximity detection function. By extracting the numerical part of the house number and calculating the numerical difference, identify POI objects with different address representations but similar actual locations.

[0026] The fusion omission pattern analysis module 20 is used to execute step S20: to perform in-depth analysis on the spatial distribution features, semantic features, and temporal variation features of the suspected duplicate POI candidate set, identify the omission patterns after one fusion, and obtain omission pattern analysis data. Please refer to... Figure 3 and Figure 4 The fusion omission pattern analysis module 20 includes a parallel spatial pattern analysis submodule 201.a, a semantic pattern analysis submodule 201.b, a timeliness analysis submodule 201.c, and an omission pattern analysis data generation submodule 202. The spatial pattern analysis submodule 201.a is used to execute step S201.a: perform statistical analysis on the spatial distribution characteristics of the first suspected duplicate POI pair in the first POI candidate set to obtain the distribution pattern of the first suspected duplicate POI pair.

[0027] Statistical analysis was performed on the spatial distribution characteristics of suspected duplicate POI pairs, calculating the mean, standard deviation, and extreme values ​​of spatial distances to identify spatial clustering patterns. Density clustering algorithms were used to cluster the geographic coordinates of suspected duplicate POIs, identifying spatially concentrated areas of duplicate POIs. The impact of GeoHash grid boundary effects on traditional fusion methods was analyzed, and the distribution patterns of suspected duplicate POI pairs located near grid boundaries were statistically analyzed.

[0028] The semantic pattern analysis submodule 201.b is used in S201.b to compare the semantic features of the first suspected duplicate POI pair in the first POI candidate set to obtain the difference in category labeling of each POI in different data sources.

[0029] This study systematically analyzes typical patterns of POI name differences, including semantic variations such as homophones, abbreviations versus full names, and Chinese-English translations. It delves into address representation differences, identifying issues such as missing administrative levels, use of geographical aliases, and coexistence of old and new place names. Furthermore, it analyzes category labeling differences for the same POI across different data sources, identifying inconsistencies in category mapping.

[0030] The timeliness analysis submodule 201.c is used to execute step S201.c: record the time-series change characteristics of the POI information of the first suspected duplicate POI pair in the first POI candidate set as the time-series analysis result.

[0031] This analysis addresses issues of POI information loss due to changes over time, including dynamic factors such as updated business information, changes in contact details, and adjustments to business scope. It also identifies the impact of business changes on POI integration, such as the effects of brand restructuring, store changes, and changes in operating entities on data consistency.

[0032] The omission pattern analysis data generation submodule 202 is used to execute step S202: generate fused omission pattern analysis data based on the distribution pattern, category labeling differences, and time series analysis results of the first suspected duplicate POI pair.

[0033] By analyzing the omission patterns and reasons of traditional fusion methods at multiple levels, data support is provided for the optimization of subsequent intelligent decision-making strategies.

[0034] The second POI candidate set generation module 30 is used to perform knowledge retrieval from the POI data after one fusion based on the fusion omission pattern data to obtain the second POI candidate set. Please refer to [link to module 30]. Figure 5 , Figure 6 The second POI candidate set generation module 30 includes: a geographic proximity channel recall submodule 301, a text similarity recall submodule 302, a knowledge-enhanced recall submodule 303, and a candidate set generation submodule 304.

[0035] The geographic proximity channel recall submodule 301 is used to execute step S301: Based on the distribution pattern of the first suspected duplicate POI pair, the POI dataset after one fusion is dynamically searched and multi-level geographic queries are performed to obtain the most relevant POI candidate data based on geographic proximity relationship pairs, and thus obtain the geographic suspected duplicate POI data.

[0036] The system determines POI data and POI categories based on the query text. Based on the distribution patterns of suspected duplicate POI pairs, it calculates the density of similar POIs within a 1-kilometer radius of the target POI. A GeoHash index is then used to quickly count the number of similar POIs within adjacent grid cells. The basic search radius is then adjusted according to the density level. In high-density areas (number of similar POIs > 50), the search radius is adjusted to 60% of the base value to avoid recalling too many irrelevant candidates; in medium-density areas (number of similar POIs 20-50), the search radius is adjusted to 80% of the base value; and in low-density areas (number of similar POIs < 20), the search radius is expanded to 120% of the base value to ensure no potential duplicate POIs are missed.

[0037] Subsequently, a multi-level geographic query is performed, employing a two-stage strategy of coarse recall followed by precise filtering to balance retrieval efficiency and accuracy. The first stage uses a GeoHash index for rapid coarse-grained recall, while the second stage uses an R-tree index for precise distance calculation and filtering. The GeoHash coarse recall stage leverages the spatial locality of GeoHash encoding. The system first calculates the 7-bit GeoHash code for the target POI, then determines the number of neighboring GeoHash grids to be retrieved based on the search radius. The correspondence between search radius and GeoHash precision is determined through a pre-calculation table: for radii less than 200 meters, the target grid and its 8 directly adjacent grids are retrieved; for radii between 200 and 500 meters, the target grid and its 24 second-order adjacent grids are retrieved; and for radii greater than 500 meters, the target grid and its larger range of adjacent grids are retrieved. The R-tree precise distance calculation stage performs precise geographic distance calculations on the candidate POIs from the coarse recall. The system uses the Haversine formula to calculate spherical distances, which considers the influence of the Earth's curvature on distance calculations and is suitable for precise measurements over longer distances. For close-range scenarios with distances less than 1 kilometer, the system uses the Euclidean distance formula in a Cartesian coordinate system, resulting in higher computational efficiency. The specific implementation of distance calculation considers a balance between accuracy and efficiency. The system pre-calculates trigonometric function values ​​within commonly used distance ranges and establishes a lookup table to reduce real-time computation overhead. For batch distance calculation tasks, vectorized computation technology is employed, utilizing the CPU's SIMD instruction set to calculate the distances between multiple POI pairs in parallel. Finally, the most relevant POI candidate data is sorted based on the distances between POI pairs to obtain geographically suspected duplicate POI data.

[0038] The text similarity recall submodule 302 is used to perform step S302: based on the difference in category labeling of each POI in different data sources, semantic recall of POI knowledge in the POI dataset after one fusion is performed, POI candidate data that are synonymous with the corresponding POI are obtained by recalling keywords through inverted index, and the TF-IDF score of each candidate POI is calculated to obtain suspected duplicate POI data.

[0039] Text similarity recall performs candidate retrieval based on text information such as the name and address of the POI. When constructing text similarity features, the semantic pattern analysis results of step S103 are additionally introduced, and semantic variants such as common abbreviations / full names, homophones, and Chinese-English translations are incorporated into the recall rules. For example, when it is detected that "McDonald's" and "Golden Arches" appear in the same GeoHash neighborhood, the system will force them into the candidate pair to reduce omissions caused by semantic differences. Keyword recall based on the inverted index is the basic link of text similarity retrieval. The system first extracts query keywords from the target POI, and then quickly locates candidate POIs containing these keywords through the inverted index. Moreover, the system has established a large-scale synonym dictionary covering the POI field, including Chinese-English translations of brands, mappings between local common names and standard names, and synonymous expressions of industry terms. For example, the synonym expansion of "McDonald's" includes "McDonald's", "Golden Arches", "M's"; the synonym expansion of "KFC" includes "Kentucky Fried Chicken", "Kaifeng Cuisine", "K's", etc. During the synonym expansion process, the system sets different weights for different synonyms, with the official name having the highest weight, common abbreviations having the second highest weight, and local common names having the lowest weight. The inverted index retrieval uses the TF-IDF scoring model to calculate text similarity. For each term in the query term set, the system obtains the list of POIs containing that term and its term frequency information from the inverted index. Then, the TF-IDF score of each candidate POI is calculated: tf_idf_score=(tf / max_tf_in_document)×log(total_documents / document_frequency) where tf is the occurrence frequency of the term in the POI, max_tf_in_document is the highest term frequency in the POI, total_documents is the total number of POIs, and document_frequency is the number of POIs containing that term. The multi-term score fusion adopts an accumulation strategy, and the total score of the candidate POI is the sum of the TF-IDF scores of all query terms it contains. To avoid long text POIs obtaining unfairly high scores, the system performs length normalization on the total score. Sorting is completed according to the TF-IDF scores of the candidate POIs to obtain text suspected duplicate POI data.

[0040] The knowledge-enhanced recall sub-module 303 is used to execute step S303: screen the POI data with temporal changes in the once-fused POI dataset according to the temporal analysis results, and perform an association query on the screened data in combination with a knowledge base to obtain temporally suspected duplicate POI data.

[0041] Knowledge-enhanced recall uses a knowledge base to perform POI association queries, enabling the discovery of duplicate relationships that are difficult to identify based on surface features. This recall process leverages timeliness analysis results to identify dynamic evolutions such as company name changes, brand restructuring, and store changes. For POIs that have undergone temporal evolution, the system includes their historical version information in the recall scope. For example, when a chain enterprise changes its name, the system automatically recalls POI stores under the old brand name to reduce omissions. The association between POIs and enterprise entities employs a cascading approach using multiple matching strategies. In the exact matching stage, the system establishes an association through a perfect match between the POI name and the enterprise name; in the fuzzy matching stage, the system uses algorithms such as edit distance and Jaro-Winkler similarity to identify enterprises with similar names; in the address matching stage, the system establishes an association through a structured match between the POI address and the enterprise's registered address; and in the legal representative matching stage, the system establishes an association through matching the POI operator information with the enterprise's legal representative information. Same-enterprise store queries are the most direct method for discovering duplicate POIs. Once a target POI is associated with a specific enterprise entity, the system queries all associated store POIs of that enterprise. These store POIs have a high probability of duplication, especially for chain enterprises. The system assigns a high confidence score of 0.9 to store relationships with the same enterprise. Related enterprise queries reveal more complex enterprise relationships. The enterprise knowledge graph records investment relationships, holding relationships, parent-subsidiary relationships, etc., between enterprises. The system uses a graph traversal algorithm to find related enterprises of the target enterprise, including parent companies, subsidiaries, and sister companies. Store POIs of related enterprises may also have duplicate relationships, but the confidence level is relatively low, and the system assigns a medium confidence score of 0.7. Enterprise change history queries identify change events such as enterprise renaming, mergers, and spin-offs. When an enterprise changes its name, POIs under the old and new enterprise names may actually be the same entity; when an enterprise merges, the POIs of the merged enterprise may be integrated into the merging party; when an enterprise spins off, the POIs of the original enterprise may be assigned to the new spun-off enterprise. The system tracks these change events, identifies potential duplicate POI relationships, and ultimately obtains time-series suspected duplicate POI data.

[0042] The candidate set generation submodule 304 is used to perform step S302: take the intersection of the geographic suspected duplicate POI data, text suspected duplicate POI data, and time series suspected duplicate POI data to obtain the second POI candidate set.

[0043] The first POI candidate set is used to initially analyze the omission patterns. Then, based on the omission patterns, knowledge retrieval is performed from the three levels of the fused POI dataset. Finally, the intersection of the three is taken to make the final POI candidate set more complete and accurate.

[0044] The large language model 40 is used to quickly retrieve the closest candidate knowledge from the POI knowledge base based on the second POI candidate set using knowledge enhancement technology, and to generate a fusion decision based on the similarity between the second POI candidate set and the candidate knowledge. Specifically, it includes: a retrieval enhancement submodule 401, a POI data retrieval submodule 402, and a decision generation submodule 403. It should be understood that a standard POI knowledge base is based on an existing POI domain dictionary and POI data in the same domain. It uses RAG retrieval to obtain relevant POI knowledge from the POI domain dictionary to correct and complete the relationships and attributes of the POI data after entity recognition. Through graph indexing construction technology, it constructs multi-level POI knowledge bases with multi-level retrieval categories at different granularities.

[0045] The retrieval enhancement submodule 401 is used to perform step S401: extract key entities and attributes as retrieval conditions for each second suspected duplicate POI pair in the second POI candidate set, perform condition retrieval in the POI knowledge base, and filter out candidate knowledge fragments related to the second suspected duplicate POI pair in the POI knowledge base.

[0046] The retrieval enhancement submodule 402 is used to perform step S402: generate a query vector based on the second suspected duplicate POI pair in the second POI candidate set, calculate the semantic similarity, contextual relevance and temporal validity with the candidate knowledge fragment based on the query vector and its corresponding key entities and attributes, and obtain the final score of the candidate knowledge by combining the POI confidence score of the second suspected duplicate POI pair; and complete the sorting and screening of candidate knowledge.

[0047] The system generates a semantic vector from the user query q, and calculates the following four types of scores for each candidate knowledge fragment k recalled from the knowledge base: 1) Semantic Similarity Semantic similarity is calculated based on the cosine similarity between the vectorized query vector q and the candidate fragment vector k, as shown in the following formula:

[0048] Where q is the query vector and k is the candidate knowledge fragment vector, this method is used to measure semantic proximity.

[0049] 2) Context Relevance This is used to measure the degree of matching between candidate knowledge and the query context. Contextual features may include category consistency, geographical proximity, entity structure consistency, industry category, address structure, brand, etc. Its general weighted form is:

[0050] in: Let i be the matching degree function for the i-th context feature. Indicates the weight of the corresponding feature; : indicates summation over all context features (the number can be expanded).

[0051] The number of features N can be dynamically expanded according to business needs, including various contextual elements such as category consistency, geographical proximity, entity matching degree, industry consistency, brand similarity, address structure matching degree, and data source consistency.

[0052] 3) Confidence Score The confidence level of candidate knowledge is calculated based on its source quality, entity recognition confidence level, relation extraction confidence level, and attribute completion model output probability, etc.

[0053] The following applies to filtering: when >0.9: High-confidence candidate When 0.7≤ ≤0.9: Medium confidence candidate when <0.7: Remove or reduce the ranking weight 4) Temporal Validity Candidate knowledge decays over time, therefore a time decay function is introduced:

[0054] in: This is the time difference between the current time and the candidate knowledge update time. This is the time decay coefficient.

[0055] Based on the above four categories of indicators, the final scoring formula for candidate knowledge in this invention is defined as follows:

[0056] in, For semantic similarity weight parameters, For context-related weight parameters, For confidence weight parameters, This is the time-related weighting parameter.

[0057] The decision generation submodule 403 is used to perform step S304: generate a fusion decision based on the matching degree between the data of another POI in the corresponding second suspected duplicate POI pair and the candidate knowledge.

[0058] The matching degree between the suspected duplicate pair and another POI in the second suspected duplicate pair and the candidate knowledge is calculated. If the match is completed with the candidate knowledge within the threshold range, it is indicated that the suspected duplicate pair is a duplicate pair.

[0059] Furthermore, the large language model is trained through the following steps: The first training set is input into the general large language model for training until the objective function of the general large language model converges, and the converged general large language model is obtained. The first training set includes geographic information, business POIs and related question answering scenarios.

[0060] The system first selects a suitable pre-trained large language model based on the POI business requirements and application scenarios. The selected model can be a general-purpose large model, such as GPT, Tongyi Qianwen, Doubao, or the Deepseek series, or a vertically adapted large model whose training corpus includes geographic information, commercial POIs, and related question-and-answer scenarios. During the selection process, the system comprehensively considers factors such as model parameter size, inference efficiency, knowledge coverage, knowledge retention capability, and deployment cost to ensure efficient and accurate answer generation during subsequent fine-tuning and RAG retrieval enhancement stages. After model selection, the system prepares the environment for model deployment and subsequent interface integration, including allocating computing resources, defining API interfaces, and specifying input / output protocol specifications.

[0061] For each lightweight adapter module inserted into the converged general-purpose large language model, prefix tuning is performed at the model input end. The real POI pairs, the fusion results of the annotations, and the high-quality domain data extracted from the historical question-answering logs in the knowledge base are obtained as the second training set. The second training set is input into the converged general-purpose large model for fine-tuning to obtain the trained large language model.

[0062] The system fine-tunes a pre-trained large language model using high-quality domain data extracted from a POI knowledge base, real POI pairs, fused annotation results, and historical question-and-answer logs. Adapter fine-tuning involves inserting lightweight adapter modules into each layer of the model, enabling it to learn domain-specific knowledge without modifying the original parameters. Prefix tuning adds trainable prefix vectors to the model's input, allowing it to dynamically reference external knowledge during generation, thus improving the professionalism and consistency of the answers. During fine-tuning, the system ensures the model accurately identifies entities and attribute relationships in the knowledge base, prioritizing external knowledge while maintaining its original general knowledge capabilities. After fine-tuning, the model can more efficiently integrate external knowledge to generate answers that meet professional requirements.

[0063] The POI fusion module 50 is used to perform step S50: fusion the second POI candidate set according to the fusion decision to obtain the final POI fusion result.

[0064] Driven by the POI fusion module 50, physical merging of the candidate POI set is performed. When processing each fusion candidate, the POI fusion module 50 first reads its confidence value and follows different execution paths according to predefined thresholds: when the confidence value is >0.90, the system automatically enters the "high confidence automatic fusion" process, directly performs merging in the target table and records fusion metadata; when the confidence value is 0.70 ≤ confidence value ≤ 0.90, the system performs the fusion operation and marks the record as "requiring subsequent manual review" for subsequent manual or expert sampling; when the confidence value is <0.70, the system does not perform automatic merging, but writes the candidate into the "pending review queue", triggering a manual or semi-automatic decision-making process and delaying fusion.

[0065] Attribute merging employs different priorities and merging rules based on field type: For name fields, the longest and most standardized string conforming to naming conventions is prioritized as the primary name, while other strings are stored as a set of aliases; for address fields, entries with the most complete hierarchical level (province / city / district / street / house number / building) are selected as the standard address, while address variations are preserved for reverse lookup; coordinate fields are merged using a weighted average, with the weights determined by the data source trust score (source_trust_score), coordinate precision (coordinate_precision, provided in the source metadata), and recency factor (recency_factor), calculated using the following formula: weight_i = α * source_trust_score_i + β * coordinate_precision_i + γ* recency_factor_i merged_coordinate = Σ(coordinate_i × weight_i) / Σ(weight_i) Where α, β, and γ are normalization coefficients. Category merging selects the most specific subcategory as the primary category (e.g., "restaurants > hot pot" takes precedence over "restaurants"), while related but broader categories are written into the auxiliary label field. Contact information merging first performs standardization (phone number formatting, removal of spaces / special characters, and unification of international dialing codes), then deduplication, retaining all valid contact information and recording the source and credibility of each contact method. This yields the final high-quality POI fusion result.

[0066] Preferably, to avoid the recurrence of the same omission pattern, the POI fusion system of the present invention further includes: Fusion quality monitoring module: used to add the missing pattern analysis data to the monitoring rules, monitor the final POI fusion result according to all monitoring rules, obtain fusion anomaly detection results, and when fusion omission occurs, input the missing pattern corresponding to the monitoring rule and the suspected duplicate POI pair into the large model.

[0067] The fusion quality monitoring module continuously utilizes the anomaly identification results from S10 as a "monitoring baseline" to ensure that the same anomaly patterns do not reappear after fusion. The spatial and temporal pattern analysis results from S20 are mapped to monitoring rules. For example, within boundary areas, the system pays special attention to whether the merged result crosses unreasonable grids; in brand renaming or store change scenarios, the system monitors whether the merged result retains both old and new identifiers. Fusion quality monitoring includes routine statistical indicators and sampling verification. Routine statistical indicators include: fusion rate (fusion_rate = (initial_count - final_count) / initial_count), daily new fusion count, and fusion distribution with stratified confidence levels. Sampling verification uses statistical methods to determine the sample size (e.g., calculating the required sample size under a 95% confidence level and a 5% error tolerance), manually reviews samples, and writes back the review results for calculating precision (precision = correct_fusions / total_fusions) and recall estimation (through comparison with known standard datasets or historical labeled sets). The system runs automatic sampling scripts periodically (e.g., daily / weekly) and automatically issues manual review tasks. When an anomaly is detected, the omission patterns corresponding to the monitoring rules and the suspected duplicate POIs are input into the large model for re-fusion. As the system processes more and more POI data, its fusion accuracy becomes higher.

[0068] Simultaneously, anomaly detection and performance monitoring can be introduced to monitor system operation. The anomaly detection mechanism covers two categories: semantic / logical anomalies and behavioral / performance anomalies. Semantic and logical anomalies include geographical distance anomalies (e.g., the distance between merged pairs > 2km, or extremely inconsistent with the geographical distribution of POI categories), category conflicts (mutually exclusive situations between primary and secondary categories after merging), and missing or inconsistent fields (e.g., complete conflict between primary business hours and historical records). Row-level anomaly detection can use statistical rules (mean ± nσ) or model-based anomaly detection (based on Isolation Forest or time-series-based anomaly detection such as EWMA, CUSUM) to identify sudden batch anomalies. Performance monitoring includes processing latency, CPU / memory usage, disk I / O, and queue length. If preset thresholds are exceeded, automatic expansion or alarms are triggered. Once an anomaly is detected, the system should handle it according to severity level: minor anomalies are subject to automatic rate limiting and the operations and maintenance team is notified; moderate anomalies suspend related tasks and are added to the manual review queue; severe anomalies trigger system rollback or manual intervention. The quality scoring system includes, but is not limited to, the following dimensions: completeness score (completeness_score = filled_fields / total_fields), accuracy score (derived by comparing with authoritative data sources), consistency score (evaluation of logical consistency within fields), and timeliness score (based on update intervals). These scores are weighted to form a single quality score for sorting and reporting. The system should display key metrics in a monitoring platform (such as Prometheus + Grafana) and configure threshold alerts and a daily quality snapshot report sending mechanism.

[0069] This invention performs fusion detection on the POI data after one fusion, filters out suspected duplicate POI pairs, and then analyzes fusion omission patterns from multiple levels of features such as spatial, semantic, and temporal changes. Combining the fusion omission pattern data with knowledge retrieval from the POI data after one fusion, a second POI candidate set is obtained. Subsequently, based on a large language model and combined with retrieval enhancement technology, candidate knowledge fragments closest to the second suspected duplicate POI pairs in the second POI candidate set are retrieved from a standard knowledge base. A fusion decision is generated based on the candidate knowledge fragments and the second POI candidate set. Finally, the suspected duplicate POI pairs are fused according to the fusion decision. To avoid the re-omission of the same omission pattern, each detected omission pattern is used as a monitoring rule to monitor the final POI fusion result. When a fusion omission occurs, the omission pattern corresponding to the monitoring rule and the suspected duplicate POI pair are input into the large model, further improving the fusion accuracy.

[0070] Compared to existing technologies, this invention has the following advantages: This invention is the first to introduce RAG technology into the field of POI data fusion, designing a multi-level intelligent recall mechanism and developing a RAG-enhanced fusion decision module, achieving a technological leap from traditional rule-driven to intelligent knowledge-driven approaches. This technical solution not only achieves significant breakthroughs in POI fusion accuracy and processing efficiency, but more importantly, it establishes a scalable and replicable intelligent data fusion technology architecture, which plays a crucial role in promoting data governance and application services in related fields such as geographic information systems, location services, and smart cities.

[0071] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A POI data fusion method based on a large language model, characterized in that, include: S10: Obtain the POI dataset after one fusion as the initial dataset, and perform duplicate POI detection on each POI data in the initial dataset to obtain the first POI candidate set; wherein, the first POI candidate set includes multiple first suspected duplicate POI pairs, and each POI pair includes POI pair information, suspected type identifier and confidence score; S20: Perform in-depth analysis on the spatial distribution features, semantic features, and temporal change features of the first POI candidate set, identify the omission patterns after one fusion, and obtain omission pattern analysis data. S30: Based on the fusion omission pattern analysis data, perform knowledge recall on the POI dataset after the first fusion to obtain a second POI candidate set; The second POI candidate set is input into a trained large language model for fusion decision generation. The retrieval module of the large language model is connected to a POI knowledge base. The specific steps are as follows: S401: Extract key entities and attributes as search conditions for each second suspected duplicate POI pair in the second POI candidate set, perform conditional search in the POI knowledge base, and filter out candidate knowledge fragments related to the second suspected duplicate POI pair in the POI knowledge base. S402: Generate a query vector based on the second suspected duplicate POI pair in the second POI candidate set; calculate the semantic similarity, contextual relevance, and temporal validity of the candidate knowledge fragment based on the query vector and its corresponding key entities and attributes; and obtain the final score of the candidate knowledge by combining the POI confidence score of the second suspected duplicate POI pair; and complete the sorting and screening of candidate knowledge. S403: Generate a fusion decision based on the matching degree between the data of another POI in the corresponding second suspected duplicate POI pair and the candidate knowledge; S50: The second POI candidate set is fused according to the fusion decision to obtain the final POI fusion result.

2. The POI data fusion method according to claim 1, characterized in that, Step S20 includes: S201.a: Perform statistical analysis on the spatial distribution characteristics of the first suspected duplicate POI pairs in the first POI candidate set to obtain the distribution pattern of the first suspected duplicate POI pairs; S201.b: Compare the semantic features of the first suspected duplicate POI pairs in the first POI candidate set to obtain the category labeling differences of each POI in different data sources; S201.c: Record the temporal change characteristics of the POI information of the first suspected duplicate POI pair in the first POI candidate set as the result of temporal analysis; S202: Based on the distribution pattern, category labeling differences, and time series analysis results of the first suspected duplicate POI pair, generate fusion omission pattern analysis data.

3. The POI data fusion method according to claim 2, characterized in that, The formulas for calculating semantic similarity, contextual relevance, and temporal validity in step S402 are as follows: semantic similarity : Where q is the query vector and k is the candidate knowledge fragment vector; Contextual relevance : in, Let i be the matching degree function for the i-th context feature. Indicates the weight of the corresponding feature; Time correlation : in, λ represents the time difference between the current time and the candidate knowledge update time, where λ is the time decay coefficient. Final score for candidate knowledge: in, For semantic similarity weight parameters, For context-related weight parameters, For confidence weight parameters, This is the time-related weighting parameter.

4. The POI data fusion method according to claim 3, characterized in that, Step S30 includes: S301: Based on the distribution pattern of the first suspected duplicate POI pair, perform dynamic search radius determination and multi-level geographic query on the POI dataset after one fusion to obtain the most relevant POI candidate data based on geographic proximity, and obtain geographic suspected duplicate POI data. S302: Based on the differences in category labeling of each POI in different data sources, semantic recall is performed on the POI knowledge in the POI dataset after one fusion. The candidate data of POIs that are synonymous with the corresponding POI are obtained by recalling keywords through the inverted index, and the TF-IDF score of each candidate POI is calculated to obtain the suspected duplicate POI data in the text. S303: Based on the time series analysis results, filter out POI data that have undergone time series changes in the fused POI dataset, and perform correlation queries on the filtered data in conjunction with a knowledge base to obtain suspected duplicate time series POI data. S304: Take the intersection of the geographic suspected duplicate POI data, text suspected duplicate POI data, and time series suspected duplicate POI data to obtain the second POI candidate set.

5. The POI data fusion method according to any one of claims 1-4, characterized in that, The large language model was trained through the following steps: The first training set is input into the general large language model for training until the objective function of the general large language model converges, and the converged general large language model is obtained. The first training set includes geographic information, business POIs and related question answering scenarios. Lightweight adapter modules are inserted into each layer of the converged general-purpose large language model. Prefix tuning is performed on the model input. The real POI pairs, the fusion results of the annotations, and the high-quality domain data extracted from the historical question-answering logs in the knowledge base are obtained as the second training set. The second training set is input into the converged general-purpose large model for fine-tuning to obtain the trained large language model.

6. The POI data fusion method according to claim 5, characterized in that, The method further includes the following after step S50: S60: Add the omission pattern analysis data to the monitoring rules, monitor the final POI fusion result according to all monitoring rules, and obtain the fusion anomaly detection result. When fusion omission occurs, input the omission pattern corresponding to the monitoring rule and the suspected duplicate POI pair into the large model.

7. A POI data fusion system based on a large language model, characterized in that, include: First POI candidate set generation module: used to obtain the POI dataset after one fusion as the initial dataset, and to perform duplicate POI detection on each POI data in the initial dataset to obtain the first POI candidate set; wherein, the first POI candidate set includes multiple first suspected duplicate POI pairs, and each POI pair includes POI pair information, suspected type identifier and confidence score; The fusion omission pattern analysis module is used to perform in-depth analysis on the spatial distribution features, semantic features, and temporal change features of the first POI candidate set, identify the omission patterns after one fusion, and obtain omission pattern analysis data. Second POI candidate set generation module: used to perform knowledge retrieval on the POI dataset after one fusion based on the fusion omission pattern analysis data to obtain a second POI candidate set; A large language model connected to a POI knowledge base, the large language model comprising: The retrieval enhancement submodule is used to extract key entities and attributes as retrieval conditions for each second suspected duplicate POI pair in the second POI candidate set, perform conditional retrieval in the POI knowledge base, and filter out candidate knowledge fragments related to the second suspected duplicate POI pair in the POI knowledge base. The POI data retrieval submodule is used to generate query vectors based on the second suspected duplicate POI pairs in the second POI candidate set, calculate the semantic similarity, contextual relevance, and temporal validity of the candidate knowledge fragments based on the query vectors and their corresponding key entities and attributes, and obtain the final score of the candidate knowledge by combining the POI confidence score of the second suspected duplicate POI pairs; and complete the sorting and filtering of candidate knowledge. Decision generation submodule: Generates a fusion decision based on the matching degree between the data of another POI in the corresponding second suspected duplicate POI pair and the candidate knowledge; POI fusion module: Fusion the second POI candidate set according to the fusion decision to obtain the final POI fusion result.

8. The POI data fusion system according to claim 7, characterized in that, The fusion omission pattern analysis module includes: Spatial pattern analysis submodule: used to perform statistical analysis on the spatial distribution characteristics of the first suspected duplicate POI pairs in the first POI candidate set, and obtain the distribution pattern of the first suspected duplicate POI pairs; Semantic pattern analysis submodule: used to compare the semantic features of the first suspected duplicate POI pairs in the first POI candidate set to obtain the category labeling differences of each POI in different data sources; The timeliness analysis submodule is used to record the time-series change characteristics of the POI information of the first suspected duplicate POI pair in the first POI candidate set as the time-series analysis result; Omission pattern analysis data generation submodule: used to generate fused omission pattern analysis data based on the distribution pattern of the first suspected duplicate POI pair, the difference in category labeling, and the time series analysis results.

9. The POI data fusion system according to claim 8, characterized in that, The second POI candidate set generation module includes: The geographic proximity channel recall submodule is used to dynamically determine the search radius and perform multi-level geographic queries on the POI dataset after one fusion based on the distribution pattern of the first suspected duplicate POI pair, to obtain the most relevant POI candidate data based on geographic proximity relationship pairs, and to obtain the geographic suspected duplicate POI data. The text similarity recall submodule is used to perform semantic recall of POI knowledge in the POI dataset after one fusion based on the difference in category labeling of each POI in different data sources. It retrieves POI candidate data that are synonymous with the corresponding POI through inverted index keywords, and calculates the TF-IDF score of each candidate POI to obtain suspected duplicate POI data. Knowledge Enhancement Recall Submodule: Used to filter POI data that have undergone time-series changes in the fused POI dataset based on the time-series analysis results, and perform correlation queries on the filtered data in conjunction with a knowledge base to obtain suspected duplicate time-series POI data. Candidate set generation submodule: Take the intersection of the geographic suspected duplicate POI data, text suspected duplicate POI data, and time series suspected duplicate POI data to obtain the second POI candidate set.

10. The POI data fusion system according to claim 9, characterized in that, Also includes: Fusion quality monitoring module: used to add the missing pattern analysis data to the monitoring rules, monitor the final POI fusion result according to all monitoring rules, obtain fusion anomaly detection results, and when fusion omission occurs, input the missing pattern corresponding to the monitoring rule and the suspected duplicate POI pair into the large model.

Images