Geographic interest point retrieval method based on tree sampling and large language model

Abstract

The invention relates to a geographic interest point retrieval method based on tree sampling and a large language model. The method is suitable for location-based service scenes such as map navigation and online travel. The core process of the method comprises the following steps of: mapping a POI geographic coordinate into a character string through Geohash coding, and constructing a prefix tree index; sampling positive and negative samples on the basis of a spatial proximity strategy of a prefix tree, and training a representation model with spatial semantic expression ability in combination with unsupervised spatial contrast loss; and recalling a candidate POI set through a vector database, constructing a Prompt template by the candidate POI set and user query, inputting a large language model after LoRA fine tuning, and completing semantic rearrangement in combination with a retrieval enhancement generation (RAG) mechanism. According to the scheme, unified modeling of geographic space and text semantics is achieved, the comprehensive matching precision of retrieval results is effectively improved, dynamic changes such as POI addition and position change can be rapidly adapted, the cross-region and cross-domain generalization ability is achieved, and a more accurate and efficient retrieval solution is provided for various position-based service systems.

Description

Technical Field

[0001] This invention belongs to the field of geographic information processing and information retrieval technology, and in particular relates to a geographic point of interest retrieval method based on tree sampling and large language model. Background Technology

[0002] Points of interest (POI) retrieval is a core supporting technology based on location services (LBS), and is widely used in scenarios such as map navigation, online travel, and logistics delivery. Its core objective is to accurately match the user's query intent with the target POI, while meeting the dual requirements of semantic relevance of text and geographic proximity. It directly affects service efficiency and user experience, and has significant practical value and application prospects.

[0003] Current POI retrieval generally adopts a two-stage architecture of "recall-ranking": the recall stage uses text matching algorithms such as BM25 and Lucene to quickly filter candidate POIs, prioritizing retrieval speed; the ranking stage relies on semantic models such as BERT and CoIBERT to optimize the order of results, focusing on improving semantic matching accuracy. In recent years, some studies have attempted to introduce large language models (LLM), but most have failed to effectively integrate the structured spatial information of POIs, such as geographic coordinates and address codes, and have only remained at the textual semantic level of modeling; moreover, existing models mostly rely on static training data, using fixed parameters to cope with dynamic changes such as the addition of POIs, location changes, and name updates, lacking a flexible adaptation mechanism.

[0004] While existing technologies have made some progress in a single dimension, key bottlenecks remain: First, the integration of spatial information and textual semantics is insufficient, making it difficult for search results to simultaneously consider both "semantic relevance" and "spatial proximity," easily leading to invalid matches that are semantically similar but geographically distant. Second, there is insufficient dynamic adaptability to POI data; static models cannot respond to data updates in a timely manner, affecting the timeliness of results. Third, large language models lack adaptation solutions when processing multi-source structured POI data, failing to fully utilize their semantic understanding and reasoning capabilities. These problems result in limited accuracy of existing retrieval systems in complex query scenarios, making it difficult to meet the high requirements of scenarios such as instant services and real-time navigation. Therefore, there is an urgent need for a POI retrieval solution that can achieve unified spatial-semantic modeling and possess dynamic adaptability. Summary of the Invention

[0005] To address the problems existing in the background art, one aspect of the present invention provides a geographic point of interest retrieval method based on tree sampling and a large language model, comprising:

[0006] S1: Obtain a user-point of interest dataset, which includes: user query text information, point of interest description information, and point of interest geographic location information; wherein, the user query text information is marked with corresponding target points of interest;

[0007] S2: Map the geographic location information of the points of interest to strings using the Geohash encoding algorithm, and construct a prefix tree structure based on the commonality of the Geohash encoding prefixes of the points of interest;

[0008] S3: Based on the interest point prefix tree structure, positive and negative samples are sampled on the prefix tree according to the hierarchical relationship of the Geohash encoding prefix of the interest points, and positive and negative sample pairs of interest points are constructed to pre-train the interest point encoder.

[0009] S4: Use the user query text information and its corresponding target interest points as positive sample pairs, sample the remaining interest points other than the target interest points and form negative sample pairs with the user query text information, and train the query encoder and interest point encoder.

[0010] S5: Extract interest point vectors using the trained interest point encoder and add the interest point vectors to the database;

[0011] S6: Input the user's query text information into the trained query encoder to extract the user's query vector. Based on the user's query vector, query the database for the K most similar interest point vectors to the query vector and construct a candidate interest point set.

[0012] S7: Construct based on the candidate interest point set and the user's query text information Template; to be built The large language model, after fine-tuning the template input, sorts the interest points in the candidate interest point set to obtain the sorted interest point retrieval results.

[0013] Another aspect of the present invention provides a geographic point of interest retrieval system based on tree sampling and a large language model. The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the geographic point of interest retrieval method based on tree sampling and a large language model.

[0014] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for retrieving geographic points of interest based on tree sampling and a large language model.

[0015] The present invention has at least the following beneficial effects

[0016] This invention achieves unified modeling of geospatial and textual semantics by constructing a Geohash encoding and prefix tree index, sampling positive and negative samples using a spatial proximity strategy, and training a representation model using unsupervised spatial contrast loss. This effectively improves the overall performance of retrieval results in terms of spatial relevance and semantic matching. Simultaneously, by leveraging LoRA fine-tuning technology and the Retrieval Augmentation (RAG) mechanism, the large language model efficiently adapts to POI retrieval tasks, enabling rapid response to dynamic changes such as new POIs and location changes. This significantly improves the timeliness and accuracy of the system in scenarios such as instant services and real-time navigation. Furthermore, its flexible encoding and fine-tuning process can adapt to POI data from different regions and categories, providing a more accurate, efficient, and generalizable retrieval solution for various location-based service systems. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of an embodiment of the present invention;

[0018] Figure 2 This is an example diagram of the prefix tree and positive and negative samples in an embodiment of the present invention;

[0019] Figure 3 This is a comparison chart of the test performance of the present invention and various existing models on the GEOTest dataset. Detailed Implementation

[0020] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0021] Please see Figure 1 One aspect of the present invention provides a geographic point of interest retrieval method based on tree sampling and a large language model, comprising:

[0022] S1: Obtain a user-point of interest dataset, which includes: user query text information, point of interest description information, and point of interest geographic location information; wherein, the user query text information is marked with corresponding target points of interest;

[0023] In this embodiment, the core of the user-point of interest dataset comes from publicly available standardized datasets in the field of geographic information retrieval (such as GEOGLUE), historical search logs from real-world scenarios such as map navigation or online travel, and structured POI base databases provided by map service providers or local life service platforms. The publicly available datasets already contain the corresponding annotations of queries and target POIs, the search logs extract the association between users' natural queries and the POIs they clicked and selected, and the POI base database provides complete geographic coordinates and descriptive information.

[0024] Please see Figure 2 S2: Map the geographic location information of the points of interest to strings using the Geohash encoding algorithm, and construct a prefix tree structure based on the commonalities of the Geohash encoding prefixes of the points of interest;

[0025] In this embodiment, the Geohash encoding algorithm is first used to map the geographic location information (latitude and longitude) of each point of interest into a one-dimensional string encoding. Then, prefix commonality analysis is performed on the Geohash encodings of all points of interest, and a prefix tree (Trie) index structure is constructed based on the encoded character sequence. Each level of the tree corresponds to a character node of the Geohash encoding. The path from the root node to the leaf node uniquely identifies a continuous spatial region. Each leaf node is associated with the basic information of all points of interest in that region. Thus, the spatial proximity of points of interest is intuitively reflected through the prefix hierarchy relationship, providing efficient index support for subsequent positive and negative sample sampling.

[0026] S3: Based on the interest point prefix tree structure, positive and negative samples are sampled on the prefix tree according to the hierarchical relationship of the Geohash encoding prefix of the interest points, and positive and negative sample pairs of interest points are constructed to pre-train the interest point encoder.

[0027] Preferably, the step of sampling positive and negative samples according to the hierarchical relationship of the Geohash encoding prefixes of the points of interest includes:

[0028] Set points of interest and The Geohash codes are respectively and The prefix matching length of the two Defined as:

[0029]

[0030] in, Indicates the first part of the Geohash encoding One character; This represents the maximum value function; if the matching length... , If the threshold is set to determine positive samples, then... Determined as Positive samples; if , The threshold for negative sample determination. ),but Determined as Negative samples.

[0031] Preferably, the pre-training of the interest point encoder includes:

[0032] The constructed loss function:

[0033]

[0034] in, Indicate points of interest The set of positive samples; Indicate points of interest Geographic vector; Indicate points of interest The geographic vector of the positive sample; Indicates hyperparameters; Indicate points of interest The set of negative samples; Indicates an indicator function; Indicate points of interest The geographic vector of the negative sample; Represents the cosine similarity function; Represents the loss function;

[0035] Based on the constructed loss function Backpropagation gradient updates the address encoding branch of the interest point encoder.

[0036] In this embodiment, precise sampling guided by prefix trees allows the model to establish an association between similar encoding prefixes and closer spatial distances during the pre-training stage, thereby improving its ability to capture geospatial relationships. The address encoding branch parameters are updated in a targeted manner so that the geographic vectors of points of interest can effectively reflect spatial location differences. Spatial correlations are mined in advance through unsupervised learning, reducing the complexity of subsequent joint training tasks and helping the model converge to the optimal state more quickly.

[0037] S4: Use the user query text information and its corresponding target interest points as positive sample pairs, sample the remaining interest points other than the target interest points and form negative sample pairs with the user query text information, and train the query encoder and interest point encoder.

[0038] Preferably, the interest point encoder includes: a text encoding branch and an address encoding branch; the text encoding branch includes: a cascaded linear layer, a LayerNorm layer, and a ReLU activation function; the address encoding branch includes: a cascaded Geo_Embedding layer (customized embedding layer), a Pos_Embedding layer (location embedding layer), an Attention_Pooling layer (attention pooling layer), and a Transformer encoding layer; the text encoding branch is used to encode the interest point description information to generate the interest point text vector; the address encoding branch is used to encode the interest point geographic location information to generate the interest point geographic vector; the interest point geographic vector and the interest point text vector are concatenated to obtain the interest point vector.

[0039] Preferably, the query encoder includes: a cascaded linear layer, a LayerNorm layer, and a ReLU activation function.

[0040] Preferably, the loss function used to train the query encoder and the interest encoder includes:

[0041]

[0042] in, Represents the loss function; Interest point vectors representing target interest points; Represents the user's query vector; A set of interest point vectors representing negative samples of user query text information; A vector representing the interest points of negative samples of user query text information; Indicates hyperparameters; Indicates an indicator function; This represents the cosine similarity function.

[0043] In this embodiment, user query text information and corresponding target POIs are used as positive sample pairs to ensure accurate association between query intent and target POI. Other POIs besides the target POI are randomly sampled from the dataset and paired with the user query text information to form negative sample pairs, creating a training sample structure of one positive and multiple negative pairs. The query encoder and the POI encoder are trained simultaneously, with the query encoder responsible for learning the semantic features of the user query. The POI encoder continues to optimize the fusion effect of text vectors and geographic vectors, and the two share the same vector space parameters. A contrastive loss function is used to calculate the matching degree between the query vector and the POI vector through cosine similarity, maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. The gradient is backpropagated to update all parameters of the two encoders. The query encoder and the POI encoder learn in the same vector space, so that the query intent-POI semantics-geographic space form a strong association, solving the feature fragmentation problem caused by separate training. Through supervised training with direct association between query and POI, the model can accurately capture the deep needs of user queries, reduce invalid matches caused by semantic ambiguity, and the optimized two encoders can generate more discriminative vector representations.

[0044] S5: Extract interest point vectors using the trained interest point encoder and add the interest point vectors to the database;

[0045] S6: Input the user's query text information into the trained query encoder to extract the user's query vector. Based on the user's query vector, query the database for the K most similar interest point vectors to the query vector and construct a candidate interest point set.

[0046] S7: Construct based on the candidate interest point set and the user's query text information Template; to be built The large language model, after fine-tuning the template input, sorts the interest points in the candidate interest point set to obtain the sorted interest point retrieval results.

[0047] In this embodiment, the Prompt template is constructed according to the structure of query text + Geohash + candidate POI list + task instructions as follows:

[0048] System command: Sort the list of candidate points of interest based on address semantic relevance and Geohash spatial proximity, and return the sorted numeric index (index starts from 0).

[0049] User query: Building 59, Taiziyuan, Lakeside Garden, East of Wenxin Kindergarten

[0050] Query Geohash: wtm79bjgvzsr

[0051] Candidate POI list: [

[0053] id:0|address:Building 59, Taiziyuan, Lakeside Garden (50 meters east of Wenxin Kindergarten)|geohash:wtm79bjgvz89,

[0054] id:1|address:Building 61, Taiziyuan, Lakeside Garden (80 meters east of Wenxin Kindergarten)|geohash:wtm79bjgvz9a,

[0055] id:2|address: No. 125 Wen'er Road, Hangzhou Preschool Teachers College, Zhejiang Normal University (South Gate)|geohash:wtmkm9r8xbbc,

[0056] id:3|address:Building 20, Shuangwan International City, No. 239 Shixiang Road|geohash:wtmkw1x500t7,

[0057] id:4|address: No. 95, Shizhuang Street, intersection of Zaogou Highway|geohash:wtms6j27reef ]

[0059] Please return the sorted indices:

[0060] The above Prompt template is input into the LoRA-tuned GLM4-9B model. The model incorporates a Retrieval Augmentation (RAG) mechanism and finally sorts the results by semantic accuracy and spatial proximity, outputting a sorted index. The model returns the sorted index: [1,2,3,4,5], corresponding to the following search result list:

[0061] 1. Lakeside Garden Prince Garden Building 59 (50 meters east of Wenxin Kindergarten);

[0062] 2. Lakeside Garden Prince Garden Building 61 (80 meters east of Wenxin Kindergarten);

[0063] 3. No. 125 Wen'er Road, Hangzhou Preschool Teachers College of Zhejiang Normal University (South Gate);

[0064] 4. Building 20, Shuangwan International City, No. 239 Shixiang Road;

[0065] 5. No. 95, Shizhuang Subdistrict, at the intersection of Zaogou Highway;

[0066] The model activates semantic understanding and spatial association reasoning capabilities through Prompt, accurately selects target POIs and completes reasonable ranking.

[0067] Preferably, the large language model includes a GLM4-9B model, LLaMA, ChatGLM, or a Transformer architecture large language model. Fine-tuning the large language model includes: randomly sampling M interest points and adding them to a candidate interest point set; and constructing a large language model based on the candidate interest point set and the user's query text information. Templates will be built The template is input into a large language model. The cross-entropy loss function is constructed with the ranking of the target interest points corresponding to the user's query text information in the large language model as the first constraint. The large language model is fine-tuned using LoRa to obtain the fine-tuned large language model.

[0068] In this embodiment, M points of interest are randomly sampled and added to the candidate point of interest set, forming a training sample structure of query + target POI + candidate POI set; a prompt template is constructed using the structure of query text + Geohash + candidate POI list + task instruction, and the predicted candidate POI ranking index sequence is generated; with the large language model ranking the target POI first as a constraint, a cross-entropy loss function is constructed: one-hot encoding for target POI index 1 (one-hot encoding is 1 if the target point of interest is ranked 1, and encoding is 0 for other index positions); the cross-entropy loss measures the difference between the model's predicted sequence and the true label, forcing the model to learn the rule of prioritizing target POI ranking; a low-rank matrix A is inserted into the self-attention layer of the large language model, only updating the parameters of matrices A and B without modifying the original model weights, and the constructed training samples are input into the model, iterating at a learning rate of 1e-4, minimizing the cross-entropy loss through backpropagation, so that the model is adapted to the POI ranking task.

[0069] By employing LoRA fine-tuning combined with cross-entropy loss function constraints and Prompt template construction, the model achieves efficient adaptation to POI retrieval and ranking tasks. When applied to the semantic re-ranking of candidate points of interest, the synergistic effect of the two methods not only accurately integrates the semantic needs of user queries with the spatial proximity features of POIs, significantly improving the ranking accuracy of retrieval results, but also quickly responds to dynamic changes such as the addition of new POIs and changes in their location, effectively reducing model illusions and providing users with retrieval results that better meet their real needs.

[0070] Please see Figure 3This application utilizes the Python and PyTorch deep learning framework to conduct simulation experiments on the publicly available GEOGLUE dataset, which covers diverse user query text, point-of-interest descriptions, geographic coordinate data, and annotations linking queries to target points of interest. Recall@10 / 50 / 100 and MRR@100 are used as core evaluation metrics. The application compares this solution with traditional text retrieval models such as BM25 and Lucene, semantic retrieval models such as BERT and CoIBERT, recent large language models such as BGE-M3 and Gte-Qwen2, and MGeo (retrival) geographic retrieval models. Comparative experiments show that the proposed scheme achieves Recall@10, 50, and 100 values ​​of 0.6287, 0.7594, and 0.8085, respectively, and an MRR@100 of 0.4447. Compared to the best baseline model Gte-Qwen2, the proposed scheme improves Recall@100 by approximately 9.6% and MRR@100 by approximately 19.1%, significantly outperforming all comparison models. This fully validates the effectiveness of the proposed scheme in achieving unified spatial-semantic modeling through Geohash encoding, spatial contrastive learning, and LoRA fine-tuning of the large language model, greatly enhancing the recall capability and ranking accuracy of geographic point of interest retrieval.

[0071] Another aspect of the present invention provides a geographic point of interest retrieval system based on tree sampling and a large language model. The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the geographic point of interest retrieval method based on tree sampling and a large language model.

[0072] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for retrieving geographic points of interest based on tree sampling and a large language model.

[0073] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0074] In summary, this invention achieves unified modeling of geospatial and textual semantics through Geohash encoding and prefix tree indexing, spatial proximity strategy for sampling positive and negative samples, and unsupervised spatial contrastive loss for training the representation model. This effectively improves the overall performance of retrieval results in terms of spatial relevance and semantic matching. Simultaneously, by leveraging LoRA fine-tuning technology and the Retrieval Augmentation (RAG) mechanism, the large language model efficiently adapts to POI retrieval tasks, enabling rapid response to dynamic changes such as new POIs and location changes. This significantly improves the timeliness and accuracy of the system in scenarios such as instant services and real-time navigation. Furthermore, its flexible encoding and fine-tuning process can adapt to POI data from different regions and categories, providing a more accurate, efficient, and generalizable retrieval solution for various location-based service systems.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for retrieving geographic points of interest based on tree sampling and a large language model, characterized in that, include: S1: Obtain a user-point of interest dataset, which includes: user query text information, point of interest description information, and point of interest geographic location information; wherein, the user query text information is marked with corresponding target points of interest; S2: Map the geographic location information of the points of interest to strings using the Geohash encoding algorithm, and construct a prefix tree structure based on the commonality of the Geohash encoding prefixes of the points of interest; S3: Based on the interest point prefix tree structure, positive and negative samples are sampled on the prefix tree according to the hierarchical relationship of the Geohash encoding prefix of the interest points, and positive and negative sample pairs of interest points are constructed to pre-train the interest point encoder. S4: Use the user query text information and its corresponding target interest points as positive sample pairs, sample the remaining interest points other than the target interest points and form negative sample pairs with the user query text information, and train the query encoder and interest point encoder. S5: Extract interest point vectors using the trained interest point encoder and add the interest point vectors to the database; S6: Input the user's query text information into the trained query encoder to extract the user's query vector. Based on the user's query vector, query the database for the K most similar interest point vectors to the query vector and construct a candidate interest point set. S7: Construct based on the candidate interest point set and the user's query text information Template; to be built The large language model, after fine-tuning the template input, sorts the interest points in the candidate interest point set to obtain the sorted interest point retrieval results.

2. The geographic point of interest retrieval method based on tree sampling and a large language model according to claim 1, wherein the positive and negative sample sampling according to the hierarchical relationship of the Geohash encoding prefix of the point of interest includes: Set points of interest and The Geohash codes are respectively and The prefix matching length of the two Defined as: in, Indicates the first part of the Geohash encoding One character; This represents the maximum value function; if the matching length... , If the threshold is set to determine positive samples, then... Determined as Positive samples; if , The threshold for negative sample determination. ),but Determined as Negative samples.

3. The geographic point of interest retrieval method based on tree sampling and a large language model according to claim 1, wherein the point of interest encoder comprises: Text encoding branch and address encoding branch; The text encoding branch includes a cascaded linear layer, a LayerNorm layer, and a ReLU activation function; the address encoding branch includes a cascaded Geo_Embedding layer, a Pos_Embedding layer, an Attention_Pooling layer, and a Transformer encoding layer; the text encoding branch is used to encode the description information of the interest point to generate a text vector of the interest point; the address encoding branch is used to encode the geographical location information of the interest point to generate a geographical vector of the interest point; the geographical vector and the text vector of the interest point are concatenated to obtain the interest point vector of the interest point.

4. The geographic interest point retrieval method based on tree sampling and a large language model according to claim 3, wherein the pre-training of the interest point encoder includes: The constructed loss function: in, Indicate points of interest The set of positive samples; Indicate points of interest Geographic vector; Indicate points of interest The geographic vector of the positive sample; Indicates hyperparameters; Indicate points of interest The set of negative samples; Indicates an indicator function; Indicate points of interest The geographic vector of the negative sample; Represents the cosine similarity function; Represents the loss function; Based on the constructed loss function Backpropagation gradient updates the address encoding branch of the interest point encoder.

5. The geographic point of interest retrieval method based on tree sampling and a large language model according to claim 3, wherein the query encoder comprises: Cascaded linear layers, LayerNorm layers, and ReLU activation functions.

6. The geographic interest point retrieval method based on tree sampling and a large language model according to claim 3 or 5, wherein the loss function used for training the query encoder and the interest point encoder includes: in, Represents the loss function; Interest point vectors representing target interest points; Represents the user's query vector; A set of interest point vectors representing negative samples of user query text information; A vector representing the interest points of negative samples of user query text information; Indicates hyperparameters; Indicates an indicator function; This represents the cosine similarity function.

7. The geographic point of interest retrieval method based on tree sampling and a large language model according to claim 1, wherein the large language model includes a GLM4-9B model, LLaMA, ChatGLM, or a large language model with a Transformer architecture, and fine-tuning the large language model includes: Randomly sample M points of interest and add them to the candidate point of interest set. Construct a database based on the candidate point of interest set and the user's query text information. Templates will be built The template is input into a large language model. The cross-entropy loss function is constructed with the ranking of the target interest points corresponding to the user's query text information in the large language model as the first constraint. The large language model is fine-tuned using LoRa to obtain the fine-tuned large language model.

8. A geographic point of interest retrieval system based on tree sampling and a large language model, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute a geographic point of interest retrieval method based on tree sampling and a large language model as described in any one of claims 1 to 7.

9. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by a processor, implements a geographic point of interest retrieval method based on tree sampling and a large language model as described in any one of claims 1 to 7.

Images