Method for multilingual geographic information data understanding and knowledge extraction based on large model

Abstract

The application discloses a multilingual geographic information data understanding and knowledge extraction method based on a large model, and comprises the following steps: constructing a geographic information exclusive fine-tuning data set, performing LoRA incremental fine-tuning on a pre-trained large language model, and performing geographic task Prompt adaptation; inputting multilingual POI text into a geographic information understanding large model to generate POI semantic features and address semantic features; constructing a name-attribute-space three-dimensional joint feature space and obtaining a POI clustering result; combining standard POI corpus to identify homonym POI and synonym POI, generating standard POI information and POI standardization records; determining address coordinate matching results; generating POI detail text and detail image-text aggregation results; and optimizing a processing strategy through a geographic information processing Reward function and a PPO reinforcement learning algorithm. The application improves the accuracy of multilingual POI understanding, address matching and knowledge extraction.

Description

Technical Field

[0001] This invention relates to the field of geographic information data processing technology, and in particular to a method for understanding and extracting knowledge from multilingual geographic information data based on large models. Background Technology

[0002] With the increasing demand for multilingual map services, cross-border travel platforms, and urban spatial data operations, technologies for automatic understanding and knowledge extraction of multilingual POI text, non-standard address text, and dynamic service information have received widespread attention. Existing geographic information processing systems primarily rely on string matching, rule parsing, traditional geocoding, and manual maintenance for POI deduplication, address coordinate matching, and detailed content generation. However, these methods commonly suffer from the following problems in practical applications: The significant differences in language, aliases, abbreviations, and colloquial address expressions make it difficult for traditional word segmentation and string similarity methods to accurately identify synonymous POIs, and they are prone to mis-merging POIs with the same name but different names. Existing POI clustering methods often treat names, attributes, and spatial coordinates as independent features, lacking dynamic constraints of spatial distance on the semantic clustering process, resulting in insufficient stability in determining distant synonyms and nearby aliases. For non-standard address text, traditional geocoding candidate recall does not adequately utilize the combined use of contextual semantics, road proximity, and administrative division consistency, and coordinate matching results are prone to deviation. In addition, POI detail text, scene images, and dynamic service information usually rely on manual editing and static rule binding, making it difficult to continuously optimize spatial attenuation coefficients and address coordinate matching parameters based on processing feedback, affecting the automation and reliability of multilingual geographic information knowledge extraction.

[0003] Therefore, how to provide methods for understanding and extracting knowledge from multilingual geographic information data based on large models is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a method for multilingual geographic information data understanding and knowledge extraction based on a large model. This invention fully utilizes pre-trained large language models, LoRA incremental fine-tuning, spatial decay contrastive clustering, geocoding matching, CLIP image and text filtering, and PPO reinforcement learning techniques. It describes in detail the processing flow of multilingual POI semantic understanding, same-names-same-names identification, non-standard address coordinate matching, and detailed image and text aggregation generation. It has the advantages of strong multilingual adaptability, accurate POI standardization, high address matching accuracy, and stable knowledge extraction quality.

[0005] The method for understanding and extracting knowledge from multilingual geographic information data based on a large model according to embodiments of the present invention includes the following steps: Step 1: Construct a geographic information-specific fine-tuning dataset based on multilingual POI text and geographic annotation data, perform LoRA incremental fine-tuning and geographic task Prompt adaptation on the pre-trained large language model, and obtain a large model for geographic information understanding. Step 2: Input the multilingual POI text into the geographic information understanding model to perform language type recognition, address level parsing, POI category determination and attribute extraction, forming POI semantic features and address semantic features; Step 3: Construct a three-dimensional joint feature space of name-attribute-space based on POI semantic features and spatial coordinate features, and obtain POI clustering results by fusing Haversine distance spatial decay weights; Step 4: Based on the POI clustering results and the standardized POI corpus, combine center vector comparison, edit distance verification and alias mapping matching to identify homonymous POIs and homonymous heteronymous POIs, and generate standard POI information and standardized POI records; Step 5: For non-standard address text and spatial reference data, determine the address coordinate matching results through contextual semantic restoration, geocoding candidate recall, road proximity constraints, and administrative division consistency verification; Step Six: Based on standard POI information and address coordinate matching results, construct a detail generation constraint template, generate POI detail text through geographic information understanding of the large model, and constrain CLIP image-text similarity filtering and diffusion image generation with POI category labels and coordinate neighborhood labels to form a detail image-text aggregation result; Step 7: Combining dynamic service information and processing feedback data, construct a geographic information processing reward function that includes cluster purity, coordinate error, map-text consistency, and attachment accuracy. Optimize the processing strategy using the PPO reinforcement learning algorithm, and adjust the spatial attenuation coefficient and address coordinate matching parameters to generate multilingual geographic information knowledge extraction results.

[0006] Optionally, step one specifically includes: Extract multilingual POI text from the map point of interest database, wherein the multilingual POI text includes language identifiers, POI names, POI aliases, and address fragments; Geographically labeled data is extracted from a standard geographic information database, the geographically labeled data including standard POI identifiers and latitude and longitude coordinates; Multilingual POI texts are grouped according to standard POI identifiers. Samples with the same standard POI identifiers but different text spellings are grouped into the homologous sample group, and samples with the same POI name but different standard POI identifiers are grouped into the spatial conflict sample group. Name normalization instruction samples are generated based on homologous sample groups, and same-name foreign object discrimination instruction samples are generated based on spatial conflict sample groups. The instruction sample association records are established according to standard POI identifiers and latitude and longitude coordinates to form a geographic information-specific fine-tuning dataset. A pre-trained large language model based on the Transformer structure was selected. A geographic information-specific fine-tuning dataset was input into the pre-trained large language model. The parameters of the original large language model were frozen by LoRA incremental fine-tuning and the low-rank matrix parameters in the attention mapping layer were updated. Construct a geographic task Prompt template that limits the input fields, output fields, and output order. Based on the geographic task Prompt template, limit the model output language type, address level, POI category, and attribute fields to generate a large geographic information understanding model.

[0007] Optionally, step two specifically includes: Multilingual POI text is input into a large geographic information understanding model, and the language identifier, POI name, POI alias and address fragment are split into fields to form text units to be parsed. The parsed text units are normalized to unify full-width and half-width characters, uppercase and lowercase characters, simplified and traditional Chinese characters, numeric characters, and punctuation characters, generating normalized text; Normalized text is input into a multilingual tokenizer, which segments it into a sequence of sub-word tokens. The sub-word token sequences are then input into a Transformer encoding layer, where contextual relationships are aggregated through self-attention weights to generate a contextual semantic representation. The contextual semantic representation is input into the instruction parsing layer, which outputs language type labels, address level fields, POI category labels, and attribute field sets according to the geographic task Prompt template. The POI name, POI alias, POI category label and attribute field set are mapped to field vectors and concatenated to generate POI semantic features; Address fragments, language type labels, and address hierarchy fields are mapped to field vectors and concatenated to generate address semantic features.

[0008] Optionally, step three specifically includes: A POI sample set is established based on POI semantic features and spatial coordinate features. The POI semantic features are linearly mapped to generate name-attribute embedding vectors. The spatial coordinate features are normalized and linearly mapped to generate spatial embedding vectors. The name-attribute embedding vectors and spatial embedding vectors are concatenated and input into a fully connected mapping layer to generate joint embedding vectors. A three-dimensional joint feature space of name-attribute-space is constructed. For any two POI samples, convert latitude and longitude coordinates into radians, and calculate the spherical spatial distance based on the square of half the latitude difference, the square of half the longitude difference, the cosine product of the latitude radian value, and the Earth's radius. The spatial attenuation weight is generated by multiplying the spherical spatial distance by the spatial attenuation coefficient, taking the negative value, and then performing a natural exponential operation. The semantic cosine similarity is obtained by dividing the joint embedding vector dot product of the two POI samples by the product of the L2 norms. The semantic cosine similarity is then multiplied by the spatial decay weight to generate the spatial decay corrected similarity. POI samples with the same standard POI identifier and spatial decay correction similarity not less than the preset positive sample threshold are constructed as positive sample pairs. POI samples with normalized edit distance not greater than the preset name similarity threshold and different standard POI identifiers, or POI samples with spatial decay correction similarity less than the preset conflict threshold, are constructed as conflict negative sample pairs. For each anchor point sample, the spatial decay correction similarity of positive sample pairs is summed after temperature coefficient scaling and natural exponent operation to generate a positive sample aggregation term. The semantic cosine similarity of conflicting negative sample pairs is multiplied by the spatial repulsion weight, and summed after temperature coefficient scaling and natural exponent operation to generate a conflicting negative sample aggregation term. The spatial repulsion weight is a weight minus the corresponding spatial decay weight. Divide the positive sample aggregation term by the sum of the positive sample aggregation term and the conflicting negative sample aggregation term to obtain the normalized probability of the positive sample. Take the natural logarithm of the positive sample aggregation term and then take the opposite number to generate the spatial decay-aware InfoNCE loss value. The spatial decay perception InfoNCE loss value of all anchor point samples is averaged to generate batch contrast loss. The parameters of the linear mapping layer and the fully connected mapping layer are then backpropagated and updated based on the batch contrast loss. Based on the updated joint embedding vector, density clustering is performed according to the preset neighborhood radius and preset minimum number of samples to generate POI clustering results.

[0009] Optionally, step four specifically includes: Based on the POI clustering results, the set of POIs within each cluster and the cluster center embedding vector are determined. The cluster center embedding vector is then compared with the standard POI vector. The standard POI vector is obtained by encoding the standard POI name, standard category, and standard address in the standard POI corpus using the geographic information understanding big model. The standard POIs with a central semantic proximity of not less than the preset central comparison threshold are used as candidate standard POIs, and a clustering decision matrix is ​​established for any two POIs in the same cluster. The clustering decision matrix includes name relationship markers, standard anchoring markers, alias hit markers, administrative division consistency markers, and spatial proximity markers. When two POIs have different names and the same candidate POI, and satisfy any two of the alias hit flag, administrative division consistency flag, and spatial proximity flag, write the same object with different names association edge; When two POIs have the same name or the normalized edit distance is not greater than the preset name similarity threshold, and there are any of the following: different candidate standard POIs, inconsistent administrative divisions, or standard coordinate distance not less than the preset foreign object distance threshold, write the same-name foreign object conflict edge. Based on the connection edges of the same object with different names, a connected subgraph is constructed. POIs in the same connected subgraph are merged into the same POI entity unit. When there are conflicting edges of the same object with different names in the connected subgraph, it is re-divided according to the candidate standard POI, administrative division level and standard coordinate neighborhood. Generate standard POI information for a set of POIs with the same name but different names, write conflict edges, splitting criteria, cluster numbers and original POI text into a set of POIs with the same name but different names, and generate a normalized POI record.

[0010] Optionally, step five specifically includes: For non-standard address text and spatial reference data, the non-standard address text is input into the geographic information understanding model. According to the geographic task Prompt template, administrative division fragments, road fragments, house number fragments, building fragments and surrounding POI fragments are extracted to generate address semantic slots. The spatial reference data includes administrative division boundaries, road centerlines, building outlines, house numbers and surrounding POI spatial indexes. The system performs administrative division completion, road name standard mapping, and address hierarchy reorganization on the address semantic slots. Missing administrative divisions are completed based on higher-level administrative divisions, colloquial road fragments are mapped to standard road names, and standardized address expressions are generated. Candidate coordinates are retrieved from spatial reference data based on administrative division fragments, road fragments, house number fragments, and surrounding POI fragments. A set of candidate coordinates is generated, and each candidate coordinate is associated with a candidate road, candidate building, candidate house number, and candidate administrative division. Calculate the vertical distance from the candidate coordinates to the centerline of the candidate road, generate a road proximity ranking, and compare the candidate administrative division with the administrative division fragment in the address semantic slot to generate an administrative division consistency verification result. The address semantic slot is compared with the candidate address number to generate a address consistency verification result. The surrounding POI fragment in the address semantic slot is compared with the surrounding POI name in the candidate coordinate neighborhood to generate a surrounding POI consistency verification result. A comprehensive score for candidate coordinates is generated based on the road proximity ranking, administrative division consistency verification results, house number consistency verification results, and surrounding POI consistency verification results. The candidate coordinates with the highest score and not less than the preset coordinate matching threshold are used as the address coordinate matching results, and non-standard address texts that do not meet the preset coordinate matching threshold are written into the address set to be verified.

[0011] Optionally, step six specifically includes: Extract POI category, standard POI name, standard address and standard attributes based on standard POI information, generate POI category label, determine target coordinates and its neighborhood range based on address coordinate matching results, and generate coordinate neighborhood label; A detail generation constraint template is constructed based on the POI category label and coordinate neighborhood label. The detail generation constraint template includes a name field, a category field, an address field, a service field, a surrounding field, and a display field. By understanding the geographic information, the large model fills in the details to generate constraint templates, and generates POI detail text in the order of standard POI name, POI category, standard address, service content and surrounding relationship; Image retrieval text is generated based on POI detail text. The image retrieval text and candidate image set are input into CLIP image encoder to generate text vector and image vector. The dot product of the text vector and image vector is divided by the product of their L2 norms to obtain the text-image consistency. The candidate image set is a set of images with the same POI category label or whose coordinate neighborhood label falls within the target coordinate neighborhood range. When there is an image in the candidate image set with a text-image consistency score not less than the preset text-image consistency threshold, the image with the highest text-image consistency score is selected as the POI scene image. When the corresponding image does not exist in the candidate image set, the constraint words for generating the diffusion image are concatenated in the order of category field, neighborhood field and detail text field. The POI scene image is generated by the diffusion image generation model and the detail image and text aggregation result is formed.

[0012] Optionally, step seven specifically includes: Based on dynamic service information and processing feedback data, the entity name, time field and location field in the dynamic service information are matched with the target POI to generate dynamic information entity pointing relationship. The target POI is the POI that has been standardized and matched with coordinates. Based on the POI merging results, coordinate verification results, image and text verification results, and dynamic information linking results, the cluster purity score, coordinate error score, image and text consistency score, and linking accuracy score are calculated respectively. Count the number of clustering errors, coordinate errors, image / text errors, and attachment errors; calculate and normalize the proportion of each error type; and generate corresponding reward weights. The cluster purity score, coordinate error score, image-text consistency score, and attachment accuracy score are multiplied by their respective reward weights and then summed to obtain the comprehensive reward value of the geographic information processing reward function. The PPO reinforcement learning algorithm receives the comprehensive reward value, uses the four scores as the policy state, and uses the spatial decay coefficient adjustment and address coordinate matching parameter adjustment as the policy action. The spatial decay coefficient and address coordinate matching parameter are then adjusted according to the policy action. Increase the spatial attenuation coefficient when the number of conflicting edges with the same name increases; decrease the spatial attenuation coefficient when the number of disconnected edges with the same name but different names increases. When the coordinate error score is lower than the preset coordinate feedback threshold, increase the weight of each item in the address coordinate matching parameters and the coordinate matching threshold. When the accuracy score of the connection is lower than the preset connection feedback threshold, the dynamic service information candidates are reordered according to the name semantic matching score, coordinate neighborhood matching score and time validity score. The updated spatial attenuation coefficient and address coordinate matching parameters are written back to the next processing flow to generate multilingual geographic information knowledge extraction results.

[0013] The beneficial effects of this invention are: This invention constructs a dedicated fine-tuning dataset for geographic information using multilingual POI text and geo-annotated data. It then employs LoRA incremental fine-tuning and geographic task Prompt adaptation to generate a large-scale geographic information understanding model. This addresses the instability of general models in understanding multilingual place names, aliases, and colloquial addresses, improving the consistency of language type recognition, address hierarchy resolution, and attribute extraction. Furthermore, it utilizes a name-attribute-space three-dimensional joint feature space and the InfoNCE contrastive clustering algorithm, which incorporates Haversine distance space decay weights, to overcome the problems of traditional POI clustering relying solely on text similarity, leading to the mismerging of distant homonyms and the omission of nearby heteronyms. This improves the understanding of homonymous POIs and homo-attribute POIs. The system improves the accuracy of identifying POIs (Points of Interest) with different names; it addresses the issue of non-standard address text coordinate matching offsets through contextual semantic restoration, geocoding candidate recall, road proximity constraints, and administrative division consistency verification, thereby enhancing the stability of address coordinate matching; it constrains POI detail text and scene images with POI category labels and coordinate neighborhood labels through detail generation constraint templates, CLIP image-text similarity filtering, and diffusion image generation, reducing manual editing and inconsistencies between images and text; and it feeds clustering purity, coordinate error, image-text consistency, and attachment accuracy back to spatial attenuation coefficients and address coordinate matching parameters through geographic information processing reward functions and PPO reinforcement learning algorithms, achieving closed-loop optimization of processing strategies. Its significance lies in its ability to support multilingual map data cleaning, cross-border POI standardization, urban spatial data operation, and automatic attachment of dynamic service information, improving the automation, accuracy, and continuous adaptation capabilities of geographic information knowledge extraction. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows the multilingual geographic information data understanding and knowledge extraction method based on a large model proposed in this invention. Figure 2 This is a schematic diagram of the multilingual geographic information data understanding and knowledge extraction method based on a large model proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1-2 A method for understanding and extracting knowledge from multilingual geographic information data based on large models includes the following steps: Step 1: Construct a geographic information-specific fine-tuning dataset based on multilingual POI text and geographic annotation data, perform LoRA incremental fine-tuning and geographic task Prompt adaptation on the pre-trained large language model, and obtain a large model for geographic information understanding. Step 2: Input the multilingual POI text into the geographic information understanding model to perform language type recognition, address level parsing, POI category determination and attribute extraction, forming POI semantic features and address semantic features; Step 3: Construct a three-dimensional joint feature space of name-attribute-space based on POI semantic features and spatial coordinate features, and obtain POI clustering results by fusing Haversine distance spatial decay weights; Step 4: Based on the POI clustering results and the standardized POI corpus, combine center vector comparison, edit distance verification and alias mapping matching to identify homonymous POIs and homonymous heteronymous POIs, and generate standard POI information and standardized POI records; Step 5: For non-standard address text and spatial reference data, determine the address coordinate matching results through contextual semantic restoration, geocoding candidate recall, road proximity constraints, and administrative division consistency verification; Step Six: Based on standard POI information and address coordinate matching results, construct a detail generation constraint template, generate POI detail text through geographic information understanding of the large model, and constrain CLIP image-text similarity filtering and diffusion image generation with POI category labels and coordinate neighborhood labels to form a detail image-text aggregation result; Step 7: Combining dynamic service information and processing feedback data, construct a geographic information processing reward function that includes cluster purity, coordinate error, map-text consistency, and attachment accuracy. Optimize the processing strategy using the PPO reinforcement learning algorithm, and adjust the spatial attenuation coefficient and address coordinate matching parameters to generate multilingual geographic information knowledge extraction results.

[0017] In this embodiment, step one specifically includes: Retrieve multilingual POI text from the map point of interest library. A POI is a geospatial point of interest object with a name, category, and coordinates. The map point of interest library is a data set that stores POI names, POI aliases, address fragments, and language identifiers. Multilingual POI text is POI names, POI aliases, and address fragments with language identifiers. Obtain geographic annotation data from the standard geographic information database. The standard geographic information database is a collection of data that stores standard POI identifiers and latitude and longitude coordinates. The geographic annotation data consists of standard POI identifiers and latitude and longitude coordinates corresponding to multilingual POI text. Multilingual POI texts are grouped according to standard POI identifiers. Samples with the same standard POI identifiers but different text spellings are grouped into the homologous sample group, and samples with the same POI name but different standard POI identifiers are grouped into the spatial conflict sample group. Name normalization instruction samples are generated based on the homologous sample group, and same-name foreign object discrimination instruction samples are generated based on the spatial conflict sample group. The name normalization instruction samples are training samples with the original POI text as input and the standard POI name as output. The same-name foreign object discrimination instruction samples are training samples with the same-name POI text and the corresponding latitude and longitude coordinates as input and different POI discrimination results as output. The name normalization instruction samples, the same-name foreign object discrimination instruction samples, and the geographic annotation data are bound together to form a geographic information-specific fine-tuning dataset. The geographic information-specific fine-tuning dataset is a set of training samples constructed for multilingual POI understanding, name normalization, and same-name foreign object discrimination tasks. A pre-trained large language model based on the Transformer structure was selected. The pre-trained large language model is a generative language model that has been trained on a general corpus and has multilingual semantic encoding and instruction following capabilities. A geographic information-specific fine-tuning dataset was input into the pre-trained large language model. The low-rank adaptation parameters were updated by LoRA incremental fine-tuning. LoRA incremental fine-tuning is a training method that freezes the original parameters of the pre-trained large language model and updates the low-rank matrix parameters in the attention mapping layer. The low-rank adaptation parameters are the low-rank matrix parameters inserted into the attention mapping layer. Construct a geographic task prompt template, which is an instruction text template that limits the input fields, output fields, and output order. Constrain the model output language type, address level, POI category, and attribute fields through the geographic task prompt template to generate a large geographic information understanding model. This implementation constructs a geographic information-specific fine-tuning dataset by using homologous sample groups and spatial conflict sample groups. This enables the pre-trained large language model to simultaneously learn multilingual name normalization and same-name-different-object discrimination rules during LoRA incremental fine-tuning. This solves the problem of insufficient understanding of POI aliases, cross-lingual expressions, and spatial conflict samples by general large models, and improves the scene adaptability of geographic information understanding large models.

[0018] In this embodiment, step two specifically includes: Multilingual POI text is input into a large geographic information understanding model. The language identifiers, POI names, POI aliases and address fragments in the multilingual POI text are split into fields to form text units to be parsed. The text units to be parsed are processed by character normalization. Character normalization is a text preprocessing method that uniformly encodes full-width and half-width characters, uppercase and lowercase characters, simplified and traditional Chinese characters, numeric characters and punctuation characters, and generates normalized text. Normalized text is input into a multilingual word segmenter, which is a word segmentation component that supports character segmentation and sub-word encoding in different languages. The normalized text is segmented into a sequence of sub-word tokens, and the sub-word token is the smallest encoding unit that preserves semantic fragments. The sequence of sub-word tokens is input into the Transformer encoding layer, which is a semantic encoding structure composed of a self-attention layer and a feedforward mapping layer. The contextual relationship between sub-word tokens is aggregated through self-attention weights to generate a contextual semantic representation. The contextual semantic representation is input into the instruction parsing layer, which is a parsing structure of the output fields constrained by the geographic task Prompt template. The language type is identified for the text unit to be parsed, and language type labels are generated. Based on contextual semantic representation and language type tags, address fragments are parsed for address hierarchy to generate address hierarchy fields. The address hierarchy is the hierarchical position of administrative division, road, house number, building and floor in the address text. Based on contextual semantic representation and address level fields, POI names and POI aliases are used to determine POI categories and generate POI category labels. The POI categories are geographical points of interest business categories in catering, supermarkets, transportation, accommodation, medical care, education and public services. Based on contextual semantic representation, address level fields, and POI category labels, POI attribute content is extracted to generate a set of attribute fields. The POI attribute content consists of text fields that can represent the POI entity, including the business entity name, service type, location description, and alias expression. After mapping the POI name, POI alias, POI category label and attribute field set to corresponding field vectors, the vectors are concatenated to generate POI semantic features. After mapping the address fragment, language type label, and address level field to their respective field vectors, the vectors are concatenated to generate address semantic features. This implementation method uses character normalization, a multilingual word segmenter, a Transformer encoding layer, and an instruction parsing layer to decompose multilingual POI text into language type, address level, POI category, and attribute fields. It further generates POI semantic features and address semantic features, solving the problem that traditional string matching is difficult to understand the multilingual semantics and address level relationships.

[0019] In this embodiment, step three specifically includes: A POI sample set is established based on POI semantic features and spatial coordinate features. The spatial coordinate features are two-dimensional spatial location vectors obtained by transforming POI latitude and longitude coordinates. Linear mapping is performed on the semantic features of POIs to generate name-attribute embedding vectors, and coordinate normalization and linear mapping are performed on the spatial coordinate features to generate spatial embedding vectors. The name-attribute embedding vector and the spatial embedding vector are concatenated, and a joint embedding vector is generated through a fully connected mapping layer. Based on the joint embedding vector of all POI samples, a three-dimensional joint feature space of name-attribute-space is constructed. For any two POI samples, first convert their latitude and longitude coordinates to radians. Then calculate the square of half the latitude difference, the square of half the longitude difference, and the cosine product of their latitude radian values. Add the square of half the latitude difference, the cosine product of the latitude radian value, and the square of half the longitude difference. Take the square root of the sum, take the arcsine, and multiply by twice the Earth's radius to obtain the spherical spatial distance between the two POI samples. The Haversine distance is the spatial distance calculated based on the latitude and longitude radian difference and the Earth's radius, which is the shortest distance on the Earth's surface. Spatial decay weights are generated based on spherical spatial distances. The spatial decay weights are obtained by multiplying the spherical spatial distance by the spatial decay coefficient, taking the negative value, and then performing natural exponential calculation. The spatial decay weights decrease as the spherical spatial distance increases. Calculate the semantic cosine similarity between any two POI samples. The semantic cosine similarity is obtained by performing a dot product operation on the joint embedding vectors of the two POI samples and dividing it by the product of the two joint embedding vectors' L2 norm. The semantic cosine similarity is multiplied by the spatial decay weight to generate the spatial decay corrected similarity, which is used to simultaneously characterize the semantic closeness and spatial proximity of two POI samples. Positive sample pairs are constructed from POI samples with the same standard POI identifier and spatial decay correction similarity not less than the preset positive sample threshold. The standard POI identifier is derived from geographic annotation data. POI samples with normalized edit distance not greater than the preset name similarity threshold and different standard POI identifiers, or POI samples with spatial decay correction similarity less than the preset conflict threshold, are constructed as conflict negative sample pairs. For each anchor sample, the spatial decay correction similarity of the positive sample pair corresponding to the anchor sample is divided by the temperature coefficient and then the natural exponent is calculated. The results are summed to generate a positive sample aggregation term. The temperature coefficient is a parameter that controls the smoothness of the contrastive learning similarity distribution. Multiply the semantic cosine similarity of the conflicting negative sample pair corresponding to the anchor sample by the spatial repulsion weight. The spatial repulsion weight is the weight obtained by subtracting the spatial decay weight of the corresponding conflicting negative sample pair from one. Then, divide the multiplication result by the temperature coefficient and perform natural exponentiation. Finally, sum the results to generate the conflicting negative sample aggregation term. Divide the positive sample aggregation term by the sum of the positive sample aggregation term and the conflicting negative sample aggregation term to obtain the positive sample normalization probability of the anchor sample. Take the natural logarithm of the positive sample normalization probability and then take the opposite number to generate the spatial attenuation perception InfoNCE loss value of the anchor sample. The spatial decay-aware InfoNCE loss values ​​of all anchor point samples are averaged to generate batch contrast loss, and the mapping parameters of the joint embedding vector are updated based on the batch contrast loss. Based on the updated joint embedding vector, density clustering is performed according to the preset neighborhood radius and preset minimum number of samples to generate POI clustering results. The POI clustering results include cluster numbers, POI sets within clusters, and cluster center embedding vectors. This implementation constructs a three-dimensional joint feature space of name, attribute, and space through name-attribute embedding vectors and spatial embedding vectors, enabling POI name, category attribute, and latitude and longitude location to participate in clustering determination. Spatial attenuation weights are generated through Haversine distance, and the InfoNCE comparative clustering process is introduced to enhance the aggregation of semantically similar and spatially adjacent POIs, while separating semantically similar but spatially distant POIs as conflicting negative samples, thereby reducing the false merging of homonymous objects and the omission of homonymous object identification.

[0020] In this embodiment, step four specifically includes: Based on the POI clustering results, the set of POIs within each cluster and the cluster center embedding vector are determined. The cluster center embedding vector is compared with the standard POI vector in the standard POI corpus. The standard POI vector is obtained by encoding the standard POI name, standard category and standard address in the standard POI corpus through the geographic information understanding big model. The dot product of the cluster center embedding vector and the standard POI vector is divided by the product of their L2 norms to obtain the center semantic proximity. The standard POIs whose center semantic proximity reaches the preset center comparison threshold are used as candidate standard POIs. For any two POIs within the same cluster, a clustering decision matrix is ​​established. The clustering decision matrix includes name relationship markers, standard anchor markers, alias hit markers, administrative division consistency markers, and spatial proximity markers. The name relationship marker is generated based on the normalized edit distance. The standard anchor marker is generated based on whether the candidate standard POIs are the same. The alias hit marker is generated based on the standard alias set hit result. The administrative division consistency marker is generated based on the address level field comparison result. The spatial proximity marker is generated based on whether the standard coordinate distance is less than a preset proximity distance threshold. When two POIs have different names and the same candidate standard POI, and both satisfy any two of the alias hit marker, administrative division consistency marker, and spatial proximity marker, a homonymous association edge is written between them. When two POIs have the same name or the normalized edit distance is not greater than a preset name similarity threshold, and both have any one of the following: different candidate standard POIs, inconsistent administrative divisions, or standard coordinate distance reaching a preset heteronym distance threshold, a homonymous conflict edge is written between them. Based on the edges of the same object with different names, a connected subgraph of the same object with different names is constructed. POIs within the same connected subgraph are merged into the same POI entity unit, and this POI entity unit is marked as a set of POIs with the same object but different names. When there are conflicting edges of the same object with different names within the same connected subgraph, the connected subgraph is split a second time according to the candidate standard POI, administrative division level, and standard coordinate neighborhood to generate mutually independent POI entity units. The split POI entity units are then marked as a set of POIs with the same object but different names. For a set of POIs with the same name but different names, the name with the highest frequency that matches the standardized POI corpus is selected as the standard POI name. The remaining names are written into the standard alias set, and the standard category, standard address, and standard coordinates are merged to generate standard POI information. For a set of POIs with the same name but different objects, the standard POI name, standard address, and standard coordinates corresponding to each POI entity unit are retained, and the conflict edges of the same-named but different objects, the splitting basis, the cluster number, and the original POI text are written into the POI normalization record. This implementation method uses a clustering decision matrix, homonymous association edges, and homonymous conflict edges to perform intra-cluster secondary determination and connected subgraph splitting on the POI clustering results. This solves the problem that single clustering can easily mis-merge homonymous objects or miss homonymous objects, and makes the standard POI information and POI normalized records have clear entity boundaries.

[0021] In this embodiment, step five specifically includes: For non-standard address text and spatial reference data, the non-standard address text is input into the geographic information understanding model. According to the geographic task Prompt template, administrative division fragments, road fragments, house number fragments, building fragments and surrounding POI fragments are extracted to generate address semantic slots. Spatial reference data includes administrative division boundaries, road centerlines, building outlines and surrounding POI spatial indexes. The address semantic slots are a set of structured fields used to carry address level fields and spatial pointing fields. The address semantic slots are restored to contextual semantics, missing administrative division fragments are filled in according to the superior administrative divisions in the spatial reference data, colloquial road fragments are mapped to standard road names, and building fragments and house number fragments are reorganized in the order of administrative division, road, house number, and building to generate a standardized address expression. Based on the standardized address representation, geocoding candidate recall is performed. Geocoding candidate recall is a process of retrieving candidate coordinates from spatial reference data according to administrative division fragments, road fragments, house number fragments and surrounding POI fragments, generating a set of candidate coordinates, and binding each candidate coordinate to a candidate road, candidate building and candidate administrative division. The candidate coordinate set is subjected to road proximity constraints and administrative division consistency checks. The road proximity constraint is to calculate the vertical distance from the candidate coordinate to the centerline of the candidate road and generate a road proximity sort from the nearest to the farthest. The administrative division consistency check is to compare whether the candidate administrative division is consistent with the administrative division fragment in the address semantic slot. Based on the road proximity ranking, administrative division consistency verification results, house number fragment matching results, and surrounding POI fragment matching results, road proximity scores, administrative division consistency scores, house number matching scores, and surrounding POI matching scores are generated respectively. The road proximity scores, administrative division consistency scores, house number matching scores, and surrounding POI matching scores are weighted and summed to obtain the candidate coordinate comprehensive score. The candidate coordinate comprehensive ranking is generated according to the candidate coordinate comprehensive score from high to low. The candidate coordinates with the highest comprehensive score in the candidate coordinate comprehensive ranking and whose comprehensive score meets the preset coordinate matching threshold are taken as the address coordinate matching results. Non-standard address texts that do not meet the preset coordinate matching threshold are written into the address set to be verified. This implementation extracts administrative division fragments, road fragments, house number fragments, building fragments, and surrounding POI fragments through address semantic slots. It also combines road proximity sorting, administrative division consistency verification, house number consistency verification, and surrounding POI consistency verification to solve the coordinate matching offset problem caused by missing non-standard address text, colloquial language, and unclear spatial orientation.

[0022] In this embodiment, step six specifically includes: Based on standard POI information, POI categories, standard POI names, standard addresses, and standard attributes are extracted to generate POI category tags. POI category tags are category markers that limit the scope of the POI details text topic and the scope of the scene image content. Based on the address coordinate matching results, the target coordinates and their neighborhood range are determined to generate coordinate neighborhood tags. Coordinate neighborhood tags are geographic markers that limit the spatial orientation of the scene image and the surrounding descriptive content. A detail generation constraint template is constructed based on POI category labels and coordinate neighborhood labels. The detail generation constraint template includes name field, category field, address field, service field, surrounding field, and display field. The fields in the detail generation constraint template are populated by the geographic information understanding model, and the text is organized in the order of standard POI name, POI category, standard address, service content, and surrounding relationship to generate POI detail text. The image retrieval text is generated based on the POI details text. The image retrieval text and the candidate image set are respectively input into the CLIP image encoder to generate text vectors and image vectors. The dot product of the text vector and the image vector is divided by the product of their L2 norms to obtain the text-image consistency. The candidate image set is the set of images that have a matching relationship with the POI category label or coordinate neighborhood label. When there is an image in the candidate image set whose text-image consistency reaches the preset text-image consistency threshold, the image with the highest text-image consistency is selected as the POI scene image; when there is no image in the candidate image set that reaches the preset text-image consistency threshold, the POI category label, coordinate neighborhood label, and POI detail text are combined into diffusion image generation constraint words, and the POI scene image is generated through the diffusion image generation model. The POI detail text, POI scene image, text-image consistency, and corresponding constraint label are bound together to form a detail text-image aggregation result. This implementation constructs a detail generation constraint template by using POI category labels and coordinate neighborhood labels, and uses CLIP image encoder to filter candidate images or uses diffusion image to generate constraint words to generate POI scene images, thus solving the problem of the disconnect between POI detail text and scene images and improving the spatial and content consistency of the detail image aggregation results.

[0023] In this embodiment, step seven specifically includes: A dynamic information entity pointing relationship is established based on dynamic service information and processing feedback data. The dynamic service information includes business hours, service status, activity content, and temporary control text related to the target POI. The target POI is the POI that has been standardized and matched with coordinates. The processing feedback data includes the POI merging results, coordinate verification results, image and text verification results, and dynamic information linking results obtained through manual verification or historical feedback. The dynamic information entity pointing relationship is the correspondence between dynamic service information and target POI. The cluster purity score is calculated based on the POI merging results. The number of samples in the same cluster that are verified as the same standard POI is divided by the total number of samples in that cluster to obtain the cluster-level purity. The cluster-level purity of all clusters is then averaged to obtain the cluster purity score. The coordinate error score is calculated based on the coordinate verification results. The spherical space distance between the address coordinate matching result and the verification coordinate is divided by the preset maximum allowable coordinate error to obtain the coordinate error normalization value. The coordinate error normalization value is limited to the range of zero to one as the deviation penalty value, and the result after subtracting the deviation penalty value from one is used as the coordinate error score. The consistency score between text and image is calculated based on the text and image verification results. The consistency score between text and image in the aggregated text and image results is obtained by multiplying the text and image consistency scores in the aggregated text and image results by the preset text and image weights and the preset verification weights, respectively. The accuracy score is calculated based on the dynamic information connection results. The number of dynamic service information correctly connected to the target POI is divided by the total number of connected dynamic service information to obtain the accuracy score. The number of clustering errors, coordinate errors, image-text errors, and linking errors in the statistical processing feedback data is calculated. The number of each type of error is divided by the total number of errors to obtain the corresponding error type percentage. The percentage of each error type is then normalized to generate reward weights corresponding to the clustering purity score, coordinate error score, image-text consistency score, and linking accuracy score. The cluster purity score, coordinate error score, image-text consistency score, and attachment accuracy score are multiplied by their respective reward weights and then summed to obtain the comprehensive reward value of the geographic information processing reward function. The PPO reinforcement learning algorithm receives a comprehensive reward value, and uses cluster purity score, coordinate error score, image-text consistency score and attachment accuracy score as policy states, and spatial attenuation coefficient adjustment amount and address coordinate matching parameter adjustment amount as policy actions to update the processing strategy of the large geographic information understanding model. The address coordinate matching parameters include road proximity constraint weight, administrative division consistency verification weight, house number matching weight, surrounding POI matching weight and coordinate matching threshold. When the cluster purity score is lower than the preset cluster feedback threshold and the number of conflicting edges between objects with the same name increases, the spatial decay coefficient is increased; when the cluster purity score is lower than the preset cluster feedback threshold and the number of disconnected edges between objects with the same name increases, the spatial decay coefficient is decreased. When the coordinate error score is lower than the preset coordinate feedback threshold, increase the weight of road proximity constraint, administrative division consistency verification, house number matching and surrounding POI matching, and tighten the coordinate matching threshold; when the splicing accuracy score is lower than the preset splicing feedback threshold, update the dynamic information candidate filtering order in the processing strategy. The updated spatial attenuation coefficient and address coordinate matching parameters are written back to the next round of POI clustering and address coordinate matching process, and dynamic information attachment is performed in combination with the updated processing strategy to generate multilingual geographic information knowledge extraction results. This implementation constructs a geographic information processing reward function based on cluster purity score, coordinate error score, image-text consistency score, and attachment accuracy score, and generates reward weights based on the proportions of clustering errors, coordinate errors, image-text errors, and attachment errors. The PPO reinforcement learning algorithm uses these four scores as policy states and the spatial decay coefficient adjustment and address coordinate matching parameter adjustment as policy actions, achieving closed-loop updates of the spatial decay coefficient and address coordinate matching parameters. This improves the system's adaptability to multilingual expressions, regional address differences, and dynamic service information changes.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a multilingual POI cleaning and detail generation scenario on a cross-border map service platform. The test data included 120,000 multilingual POI texts, 86,000 geographic annotation data entries, 21,000 non-standard address texts, and 15,000 dynamic service information entries. The text languages ​​included Chinese, Japanese, English, and Thai. Common problems in the scenario included the same store having different names such as "Wang's Convenience Store," "Old Wang's Convenience Store," and "WangConvenience Store," as well as POIs with the same name but different objects in different areas called "Central Cafe." Traditional string matching easily resulted in incorrect or missed merging.

[0025] In this embodiment, the geographic information-specific fine-tuning dataset consists of a homologous sample group and a spatial conflict sample group, with 52,000 homologous sample groups and 14,000 spatial conflict sample groups. The pre-trained large language model adopts a generative language model based on the Transformer structure, with LoRA rank set to 8, learning rate set to 2e-4, and fine-tuning rounds of 3. The joint embedding vector dimension of the name-attribute-spatial three-dimensional joint feature space is set to 512 dimensions, the spatial decay coefficient is set to 0.0045 / km, the temperature coefficient is set to 0.07, the preset positive sample threshold is set to 0.72, the preset conflict threshold is set to 0.38, the preset neighborhood radius of density clustering is set to 0.42, and the preset minimum sample size is set to 4. In address coordinate matching, the preset coordinate matching threshold is set to 0.76, and the weights for road proximity constraints, administrative division consistency verification, house number matching, and surrounding POI matching are set to 0.30, 0.25, 0.25, and 0.20, respectively. The image-text consistency threshold is set to 0.68, and the preset clustering feedback threshold, preset coordinate feedback threshold, and preset attachment feedback threshold in PPO reinforcement learning are set to 0.90, 0.86, and 0.88, respectively.

[0026] Comparison Method A is a traditional rule-based matching method, using string similarity, administrative division rules, and ordinary geocoding to complete POI merging and address matching; Comparison Method B is a general large-scale model method, using a pre-trained large language model without geo-information-specific fine-tuning for text parsing, combined with ordinary density clustering; Comparison Method C is an unimproved model, which uses a geo-information understanding large-scale model and ordinary InfoNCE clustering, but does not introduce Haversine distance spatial decay weights, nor does it use the geo-information processing reward function for parameter write-back. This invention, however, fully utilizes a geo-information understanding large-scale model, a spatial decay-aware InfoNCE contrastive clustering algorithm, a details generation constraint template, and a PPO reinforcement learning algorithm.

[0027] Table 1. Comparison of the effectiveness of the present invention and the comparative method in multilingual geographic information data processing.

[0028] As shown in Table 1, traditional rule-based matching methods are unstable in scenarios involving multilingual aliases, colloquial addresses, and objects with the same name but different characteristics. The F1 score for POI duplicate recognition is only 81.5%, and the false merging rate for objects with the same name reaches 9.8%, indicating that relying solely on string similarity and static rules is insufficient to handle cross-linguistic expressions in real-world map data. While general large-scale model methods improve language understanding capabilities, they lack geographically specific fine-tuning datasets and spatial constraints. The address-level resolution accuracy remains at 83.5%, and the false merging rate for objects with the same name is 7.2%, demonstrating that general semantic understanding cannot directly replace geospatial determination.

[0029] The unimproved model shows a significant improvement over the general large model, indicating that LoRA incremental fine-tuning and geographic task prompt adaptation can improve the stability of POI semantic features and address semantic features. However, it does not introduce Haversine distance spatial decay weights, resulting in mismerging of semantically similar but spatially distant POIs. The POI duplicate identification F1 score is 91.8%, lower than the 96.4% of the present invention. The present invention uses the spatial decay-aware InfoNCE comparative clustering algorithm, directly applying spherical spatial distance to clustering training, reducing the mismerging rate of homonymous objects to 1.7%, a decrease of approximately 63.0% compared to the unimproved model.

[0030] In terms of address coordinate matching, this invention reduces the median error of address coordinates to 13.9m through contextual semantic restoration, geocoding candidate recall, road proximity constraints, administrative division consistency verification, house number consistency verification, and surrounding POI consistency verification, which is approximately 70.8% lower than traditional rule-based matching methods. Regarding detailed text and image aggregation, this invention utilizes POI category labels and coordinate neighborhood labels to constrain CLIP text and image similarity filtering and diffusion image generation, achieving a text and image consistency score of 0.86, indicating that the generated POI detail text and POI scene images are more consistent in category, location, and displayed content.

[0031] Regarding dynamic information attachment and continuous optimization, this invention constructs a geographic information processing reward function by using cluster purity scoring, coordinate error scoring, image-text consistency scoring, and attachment accuracy scoring. It then adjusts the spatial attenuation coefficient and address coordinate matching parameters using a PPO reinforcement learning algorithm, achieving a dynamic information attachment accuracy of 92.5% and reducing the manual review ratio to 6.4%. This demonstrates that this invention can not only complete one-time POI cleaning and address matching but also continuously optimize subsequent processing strategies based on processing feedback data.

[0032] In summary, this embodiment demonstrates that the present invention can effectively solve the problems of unstable multilingual POI text understanding, difficulty in distinguishing between homonyms and homonyms, non-standard address coordinate matching offset, inconsistency between detailed images and text, and inaccurate dynamic service information attachment. While ensuring processing accuracy, it reduces the proportion of manual review and the processing time per 10,000 records. It is applicable to scenarios such as cross-border map services, urban spatial data operation, construction of cultural tourism POI knowledge base, and automatic updating of dynamic service information.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for multilingual geographic information data understanding and knowledge extraction based on large models, characterized in that: Includes the following steps: Step 1: Construct a geographic information-specific fine-tuning dataset based on multilingual POI text and geographic annotation data, perform LoRA incremental fine-tuning and geographic task Prompt adaptation on the pre-trained large language model, and obtain a large model for geographic information understanding. Step 2: Input the multilingual POI text into the geographic information understanding model to perform language type recognition, address level parsing, POI category determination and attribute extraction, forming POI semantic features and address semantic features; Step 3: Construct a three-dimensional joint feature space of name-attribute-space based on POI semantic features and spatial coordinate features, and obtain POI clustering results by fusing Haversine distance spatial decay weights; Step 4: Based on the POI clustering results and the standardized POI corpus, combine center vector comparison, edit distance verification and alias mapping matching to identify homonymous POIs and homonymous heteronymous POIs, and generate standard POI information and standardized POI records; Step 5: For non-standard address text and spatial reference data, determine the address coordinate matching results through contextual semantic restoration, geocoding candidate recall, road proximity constraints, and administrative division consistency verification; Step Six: Based on standard POI information and address coordinate matching results, construct a detail generation constraint template, generate POI detail text through geographic information understanding of the large model, and constrain CLIP image-text similarity filtering and diffusion image generation with POI category labels and coordinate neighborhood labels to form a detail image-text aggregation result; Step 7: Combining dynamic service information and processing feedback data, construct a geographic information processing reward function that includes cluster purity, coordinate error, map-text consistency, and attachment accuracy. Optimize the processing strategy using the PPO reinforcement learning algorithm, and adjust the spatial attenuation coefficient and address coordinate matching parameters to generate multilingual geographic information knowledge extraction results.

2. The method for multilingual geographic information data understanding and knowledge extraction based on a large model according to claim 1, characterized in that, Step one specifically includes: Extract multilingual POI text from the map point of interest database, wherein the multilingual POI text includes language identifiers, POI names, POI aliases, and address fragments; Geographically labeled data is extracted from a standard geographic information database, the geographically labeled data including standard POI identifiers and latitude and longitude coordinates; Multilingual POI texts are grouped according to standard POI identifiers. Samples with the same standard POI identifiers but different text spellings are grouped into the homologous sample group, and samples with the same POI name but different standard POI identifiers are grouped into the spatial conflict sample group. Name normalization instruction samples are generated based on homologous sample groups, and same-name foreign object discrimination instruction samples are generated based on spatial conflict sample groups. The instruction sample association records are established according to standard POI identifiers and latitude and longitude coordinates to form a geographic information-specific fine-tuning dataset. A pre-trained large language model based on the Transformer structure was selected. A geographic information-specific fine-tuning dataset was input into the pre-trained large language model. The parameters of the original large language model were frozen by LoRA incremental fine-tuning and the low-rank matrix parameters in the attention mapping layer were updated. Construct a geographic task Prompt template that limits the input fields, output fields, and output order. Based on the geographic task Prompt template, limit the model output language type, address level, POI category, and attribute fields to generate a large geographic information understanding model.

3. The method for multilingual geographic information data understanding and knowledge extraction based on a large model according to claim 1, characterized in that, Step two specifically includes: Multilingual POI text is input into a large geographic information understanding model, and the language identifier, POI name, POI alias and address fragment are split into fields to form text units to be parsed. The parsed text units are normalized to unify full-width and half-width characters, uppercase and lowercase characters, simplified and traditional Chinese characters, numeric characters, and punctuation characters, generating normalized text; Normalized text is input into a multilingual tokenizer, which segments it into a sequence of sub-word tokens. The sub-word token sequences are then input into a Transformer encoding layer, where contextual relationships are aggregated through self-attention weights to generate a contextual semantic representation. The contextual semantic representation is input into the instruction parsing layer, which outputs language type labels, address level fields, POI category labels, and attribute field sets according to the geographic task Prompt template. The POI name, POI alias, POI category label and attribute field set are mapped to field vectors and concatenated to generate POI semantic features; Address fragments, language type labels, and address hierarchy fields are mapped to field vectors and concatenated to generate address semantic features.

4. The method for multilingual geographic information data understanding and knowledge extraction based on a large model according to claim 1, characterized in that, Step three specifically includes: A POI sample set is established based on POI semantic features and spatial coordinate features. The POI semantic features are linearly mapped to generate name-attribute embedding vectors. The spatial coordinate features are normalized and linearly mapped to generate spatial embedding vectors. The name-attribute embedding vectors and spatial embedding vectors are concatenated and input into a fully connected mapping layer to generate joint embedding vectors. A three-dimensional joint feature space of name-attribute-space is constructed. For any two POI samples, convert latitude and longitude coordinates into radians, and calculate the spherical spatial distance based on the square of half the latitude difference, the square of half the longitude difference, the cosine product of the latitude radian value, and the Earth's radius. The spatial attenuation weight is generated by multiplying the spherical spatial distance by the spatial attenuation coefficient, taking the negative value, and then performing a natural exponential operation. The semantic cosine similarity is obtained by dividing the joint embedding vector dot product of the two POI samples by the product of the L2 norms. The semantic cosine similarity is then multiplied by the spatial decay weight to generate the spatial decay corrected similarity. POI samples with the same standard POI identifier and spatial decay correction similarity not less than the preset positive sample threshold are constructed as positive sample pairs. POI samples with normalized edit distance not greater than the preset name similarity threshold and different standard POI identifiers, or POI samples with spatial decay correction similarity less than the preset conflict threshold, are constructed as conflict negative sample pairs. For each anchor point sample, the spatial decay correction similarity of positive sample pairs is summed after temperature coefficient scaling and natural exponent operation to generate a positive sample aggregation term. The semantic cosine similarity of conflicting negative sample pairs is multiplied by the spatial repulsion weight, and summed after temperature coefficient scaling and natural exponent operation to generate a conflicting negative sample aggregation term. The spatial repulsion weight is a weight minus the corresponding spatial decay weight. Divide the positive sample aggregation term by the sum of the positive sample aggregation term and the conflicting negative sample aggregation term to obtain the normalized probability of the positive sample. Take the natural logarithm of the positive sample aggregation term and then take the opposite number to generate the spatial decay-aware InfoNCE loss value. The spatial decay perception InfoNCE loss value of all anchor point samples is averaged to generate batch contrast loss. The parameters of the linear mapping layer and the fully connected mapping layer are then backpropagated and updated based on the batch contrast loss. Based on the updated joint embedding vector, density clustering is performed according to the preset neighborhood radius and preset minimum number of samples to generate POI clustering results.

5. The method for multilingual geographic information data understanding and knowledge extraction based on a large model according to claim 1, characterized in that, Step four specifically includes: Based on the POI clustering results, the set of POIs within each cluster and the cluster center embedding vector are determined. The cluster center embedding vector is then compared with the standard POI vector. The standard POI vector is obtained by encoding the standard POI name, standard category, and standard address in the standard POI corpus using the geographic information understanding big model. The standard POIs with a central semantic proximity of not less than the preset central comparison threshold are used as candidate standard POIs, and a clustering decision matrix is ​​established for any two POIs in the same cluster. The clustering decision matrix includes name relationship markers, standard anchoring markers, alias hit markers, administrative division consistency markers, and spatial proximity markers. When two POIs have different names and the same candidate POI, and satisfy any two of the alias hit flag, administrative division consistency flag, and spatial proximity flag, write the same object with different names association edge; When two POIs have the same name or the normalized edit distance is not greater than the preset name similarity threshold, and there are any of the following: different candidate standard POIs, inconsistent administrative divisions, or standard coordinate distance not less than the preset foreign object distance threshold, write the same-name foreign object conflict edge. Based on the connection edges of the same object with different names, a connected subgraph is constructed. POIs in the same connected subgraph are merged into the same POI entity unit. When there are conflicting edges of the same object with different names in the connected subgraph, it is re-divided according to the candidate standard POI, administrative division level and standard coordinate neighborhood. Generate standard POI information for a set of POIs with the same name but different names, write conflict edges, splitting criteria, cluster numbers and original POI text into a set of POIs with the same name but different names, and generate a normalized POI record.

6. The method for multilingual geographic information data understanding and knowledge extraction based on a large model according to claim 1, characterized in that, Step five specifically includes: For non-standard address text and spatial reference data, the non-standard address text is input into the geographic information understanding model. According to the geographic task Prompt template, administrative division fragments, road fragments, house number fragments, building fragments and surrounding POI fragments are extracted to generate address semantic slots. The spatial reference data includes administrative division boundaries, road centerlines, building outlines, house numbers and surrounding POI spatial indexes. The system performs administrative division completion, road name standard mapping, and address hierarchy reorganization on the address semantic slots. Missing administrative divisions are completed based on higher-level administrative divisions, colloquial road fragments are mapped to standard road names, and standardized address expressions are generated. Candidate coordinates are retrieved from spatial reference data based on administrative division fragments, road fragments, house number fragments, and surrounding POI fragments. A set of candidate coordinates is generated, and each candidate coordinate is associated with a candidate road, candidate building, candidate house number, and candidate administrative division. Calculate the vertical distance from the candidate coordinates to the centerline of the candidate road, generate a road proximity ranking, and compare the candidate administrative division with the administrative division fragment in the address semantic slot to generate an administrative division consistency verification result. The address semantic slot is compared with the candidate address number to generate a address consistency verification result. The surrounding POI fragment in the address semantic slot is compared with the surrounding POI name in the candidate coordinate neighborhood to generate a surrounding POI consistency verification result. A comprehensive score for candidate coordinates is generated based on the road proximity ranking, administrative division consistency verification results, house number consistency verification results, and surrounding POI consistency verification results. The candidate coordinates with the highest score and not less than the preset coordinate matching threshold are used as the address coordinate matching results, and non-standard address texts that do not meet the preset coordinate matching threshold are written into the address set to be verified.

7. The method for multilingual geographic information data understanding and knowledge extraction based on a large model according to claim 1, characterized in that, Step six specifically includes: Extract POI category, standard POI name, standard address and standard attributes based on standard POI information, generate POI category label, determine target coordinates and its neighborhood range based on address coordinate matching results, and generate coordinate neighborhood label; A detail generation constraint template is constructed based on the POI category label and coordinate neighborhood label. The detail generation constraint template includes a name field, a category field, an address field, a service field, a surrounding field, and a display field. By understanding the geographic information, the large model fills in the details to generate constraint templates, and generates POI detail text in the order of standard POI name, POI category, standard address, service content and surrounding relationship; Image retrieval text is generated based on POI detail text. The image retrieval text and candidate image set are input into CLIP image encoder to generate text vector and image vector. The dot product of the text vector and image vector is divided by the product of their L2 norms to obtain the text-image consistency. The candidate image set is a set of images with the same POI category label or whose coordinate neighborhood label falls within the target coordinate neighborhood range. When there is an image in the candidate image set with a text-image consistency score not less than the preset text-image consistency threshold, the image with the highest text-image consistency score is selected as the POI scene image. When the corresponding image does not exist in the candidate image set, the constraint words for generating the diffusion image are concatenated in the order of category field, neighborhood field and detail text field. The POI scene image is generated by the diffusion image generation model and the detail image and text aggregation result is formed.

8. The method for multilingual geographic information data understanding and knowledge extraction based on a large model according to claim 1, characterized in that, Step seven specifically includes: Based on dynamic service information and processing feedback data, the entity name, time field and location field in the dynamic service information are matched with the target POI to generate dynamic information entity pointing relationship. The target POI is the POI that has been standardized and matched with coordinates. Based on the POI merging results, coordinate verification results, image and text verification results, and dynamic information linking results, the cluster purity score, coordinate error score, image and text consistency score, and linking accuracy score are calculated respectively. Count the number of clustering errors, coordinate errors, image / text errors, and attachment errors; calculate and normalize the proportion of each error type; and generate corresponding reward weights. The cluster purity score, coordinate error score, image-text consistency score, and attachment accuracy score are multiplied by their respective reward weights and then summed to obtain the comprehensive reward value of the geographic information processing reward function. The PPO reinforcement learning algorithm receives the comprehensive reward value, uses the four scores as the policy state, and uses the spatial decay coefficient adjustment and address coordinate matching parameter adjustment as the policy action. The spatial decay coefficient and address coordinate matching parameter are then adjusted according to the policy action. Increase the spatial attenuation coefficient when the number of conflicting edges with the same name increases; decrease the spatial attenuation coefficient when the number of disconnected edges with the same name but different names increases. When the coordinate error score is lower than the preset coordinate feedback threshold, increase the weight of each item in the address coordinate matching parameters and the coordinate matching threshold. When the accuracy score of the connection is lower than the preset connection feedback threshold, the dynamic service information candidates are reordered according to the name semantic matching score, coordinate neighborhood matching score and time validity score. The updated spatial attenuation coefficient and address coordinate matching parameters are written back to the next processing flow to generate multilingual geographic information knowledge extraction results.

Images