POI knowledge base construction method and system

Abstract

The invention provides a method and a system for constructing a POI knowledge base, which are characterized in that on the basis of an existing POI dictionary and POI data in the same field, related knowledge of a POI is acquired from a POI field dictionary by RAG retrieval to correct and complement the relationship and attribute of the POI data subjected to entity recognition, so that data nodes for constructing a POI knowledge graph are uniform in format, and the POI knowledge graph can be constructed more accurately. The method comprises the following steps of: establishing a multi-layer POI knowledge base in a multi-layer POI database, and establishing a multi-layer retrieval category on different granularity levels on the basis of multi-modal vectorization processing and HNSW, so that the final multi-layer POI knowledge base can be combined with an RAG retrieval technology, high-level features can be firstly matched on the basis of the retrieval category in a subsequent retrieval process, and layer-by-layer downward retrieval can be carried out on the basis of the retrieval category in the subsequent retrieval process. According to the method, the problems that retrieval needs to be carried out in different databases in the POI fusion technology, the retrieval range is large, and time consumption is long are solved, and the knowledge storage and retrieval efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of POI fusion and its data storage, and in particular to a method and system for constructing a POI knowledge base. Background Technology

[0002] Points of Interest (POIs) are core data resources in Geographic Information Systems (GIS) and location services, widely used in map navigation, location recommendation, smart cities, and big data analytics. With the rapid growth of multi-source data (government data, internet map platforms, user crowdsourcing, commercial organizations, etc.), the scale of multi-source POI data has expanded rapidly. The goal of POI fusion and deduplication is to remove duplicates, unify standards, and improve data usability; however, existing technologies still have significant bottlenecks, and the following problems commonly exist in the fusion process of multi-source POI data: Current POI fusion processes rely on extensive searches of existing industry databases to compare the similarity of POI data from different sources. This not only makes it difficult to standardize the format of POI data from different sources during the fusion process, but also requires searching across different databases, resulting in a large search scope and long processing times. Therefore, there is an urgent need for a new method for constructing a POI knowledge base that enables efficient knowledge storage and retrieval for diverse POI data. Summary of the Invention

[0003] To address the above issues, this invention uses existing POI domain dictionaries and POI data in the same domain as a foundation. It employs RAG retrieval to extract relevant POI knowledge from the domain dictionary, correcting and supplementing the relationships and attributes of the POI data after entity recognition. This ensures a unified data node format and complete internal relationships for constructing the POI knowledge graph. Finally, based on multimodal vectorization processing and HNSW, a graph index is constructed for the high-dimensional vector of each POI. Multi-level retrieval categories are built at different granularities, enabling the final multi-level POI knowledge base to combine with RAG retrieval technology. In subsequent retrieval processes, it first matches high-level features based on the retrieval category, then searches downwards layer by layer until a precise match is found at the lowest level. This solves the problem of POI fusion technology requiring retrieval in different databases, resulting in a large retrieval scope and long processing time, thus improving knowledge storage and retrieval efficiency.

[0004] This invention is achieved through the following technical solution: Firstly, this invention provides a method for constructing a POI knowledge base, comprising: S10: Collect multi-source heterogeneous POI data, and perform data cleaning and standardization on each heterogeneous POI data to obtain a set of POI data in a unified format; S20: Perform BERT encoding, linearized classification, and conditional random field processing on the POI data to obtain entity annotation results for each POI data, wherein the entity annotation results include entity text, start position, end position, entity category, and confidence score; S30: The entity annotation results are used to perform RAG retrieval and recall in the annotated POI dictionary to generate an entity candidate set, wherein the POI dictionary includes prior knowledge of the standard name and common variants of the entity. S40: Calculate similarity features based on the data in the entity candidate set, complete the sorting and disambiguation of entities, and generate standardized POI entity nodes. Each standardized POI entity node includes the original recognition text, start position, end position, entity category, standardized entity ID, and confidence score. S50: Based on the standardized POI entity nodes, complete the relation features and attribute features of each POI entity node from the POI dictionary, and construct a POI knowledge graph based on the completed POI entity nodes; S60: Based on the structure of the POI knowledge graph and the POI entity nodes, perform multimodal vectorization processing and construct the HNSW graph index to obtain a multi-level POI knowledge base.

[0005] Further, step S60 includes: S601.a: Based on the structure of the POI knowledge graph, the geographic coordinate information of the POI entity nodes, and the distance decay weight function, obtain the graph-structured embedding vector of each POI entity node. S601.b: Perform semantic modeling on the descriptive text of each POI entity node to obtain the semantic vector of the text of each POI entity node; S602: Construct positive sample pairs based on the graph structured embedding vector and text description of each POI entity node, and construct negative sample pairs based on the text semantic vector and text description of each POI entity node. Train the alignment network by maximizing the cosine similarity of the positive sample pairs and minimizing the similarity of the negative sample pairs to obtain the trained alignment network. S603: Align the graph structured embedding vector and the text semantic vector of each POI entity node to obtain the high-dimensional multimodal vector of each POI entity node. S604: Construct an HNSW graph index based on each high-dimensional multimodal vector to obtain a multi-level POI knowledge base.

[0006] Further, step S20 includes: S201: Sequentially perform word tokenization on the text data of each POI data. Each POI data is split into token sequence data, and each token in the token sequence data includes the corresponding token ID, attention mask and token type mask. S202: Sequentially encode each token in the token sequence data using BERT to obtain the context hidden vector of each token; S203: Perform linear classification mapping on the context hidden vector of each token to obtain the non-normalized prediction score of each token on multiple preset labels; S204: Perform conditional random field processing based on the nonnormalized prediction score of the token, and use dynamic programming to find the maximum probability tag sequence of the token; S205: Post-process the text data of the POI and the maximum probability label sequence of its corresponding token data to obtain the entity annotation result for each POI data.

[0007] Further, step S50 includes: S501.a: Identify the original text of each POI entity node according to the preset explicit rule pattern, and extract the explicit relationship of standardized POI entity nodes based on the matched data and the keyword dictionary contained in the POI dictionary. S501.b: Perform deep learning recognition on the original identified text of the POI entity node to obtain the relation label score of each token in the original identified text, and capture the dependency relationship in the same text based on the relation label score and the label transition score; S501.c: Perform attribute completion prediction for each standardized POI entity node based on the POI dictionary, and complete the POI attributes with the attribute with the highest prediction probability; S502: Construct a POI knowledge graph based on POI entity node data, explicit relationships of each POI entity node, dependencies of each POI entity node, and attributes.

[0008] Furthermore, before constructing the HNSW graph index based on each high-dimensional multimodal vector in step S604, the method further includes: The high-dimensional multimodal vector is mapped to a low-dimensional multimodal vector according to the mapping rule obtained from training; wherein, the mapping rule is calculated in the following manner: Where x represents the input high-dimensional multimodal vector; z represents the mapped low-dimensional latent representation; q_ (z|x) represents the encoder distribution; p_θ(x|z) represents the decoder reconstruction distribution; KL() represents the Kullback–Leibler divergence; Ltask represents the task performance loss; β and γ are weight coefficients; The low-dimensional multimodal vector is subjected to product quantization to obtain a compressed multimodal vector.

[0009] On the other hand, the present invention also provides a POI knowledge base construction system, which includes: a data format standardization unit: used to collect multi-source heterogeneous POI data, and to perform data cleaning and standardization on each heterogeneous POI data to obtain a set of POI data in a unified format; POI Entity Recognition Unit: Used to perform BERT encoding, linearization classification and conditional random field processing on the POI data to obtain entity annotation results for each POI data, wherein the entity annotation results include entity text, start position, end position, entity category and confidence score; Candidate set recall unit: used to perform RAG retrieval on the entity annotation results in the annotated POI dictionary to generate entity candidate sets, wherein the POI dictionary includes prior knowledge of the standard names and common variants of the entities; POI Entity Ranking and Disambiguation Unit: Used to calculate similarity features based on the data in the entity candidate set, complete the ranking and disambiguation of entities, and generate standardized POI entity nodes. Each standardized POI entity node includes the original recognition text, start position, end position, entity category, standardized entity ID, and confidence score. Knowledge graph generation unit: used to complete the relation features and attribute features of each POI entity node from the POI dictionary based on the standardized POI entity nodes, and to construct a POI knowledge graph based on the completed POI entity nodes; Vectorized storage and index building unit: used to perform multimodal vectorization processing and HNSW graph index building based on the structure of the POI knowledge graph and POI entity nodes to obtain a multi-level POI knowledge base.

[0010] Furthermore, the vectorized storage and indexing construction unit includes: Graph structure embedding vectorization subunit: used to obtain the graph structure embedding vector of each POI entity node based on the structure of the POI knowledge graph, the geographic coordinate information of the POI entity node, and the distance decay weight function. Alignment network training subunit: used to construct positive sample pairs based on the graph structured embedding vector and text description of each POI entity node, and construct negative sample pairs based on the text semantic vector and text description of each POI entity node. The alignment network is trained by maximizing the cosine similarity of positive sample pairs and minimizing the similarity of negative sample pairs, and the trained alignment network is obtained. Alignment Network: Used to align the graph structured embedding vector and the text semantic vector of each POI entity node to obtain a high-dimensional multimodal vector of each POI entity node. HNSW Index Building Subunit: Used to build HNSW graph indexes based on each high-dimensional multimodal vector, resulting in a multi-level POI knowledge base.

[0011] Furthermore, the POI entity recognition unit includes: Lexicalization subunit: used to sequentially lexicalize the text data of each POI data. Each POI data is split into token sequence data, and each token in the token sequence data includes the corresponding token ID, attention mask and token type mask. Bidirectional encoding subunit: used to sequentially encode each token in the token sequence data using BERT to obtain the context hidden vector of each token; Linear classification subunit: used to perform linear classification mapping on the context hidden vector of each token to obtain the non-normalized prediction score of each token on multiple preset labels; Sequence prediction subunit: Used to perform conditional random field processing based on the nonnormalized prediction score of the token, and to find the maximum probability tag sequence of the token through dynamic programming. Entity annotation sub-unit: Used to post-process the text data of POI and the maximum probability label sequence of each corresponding token data to obtain the entity annotation result of each POI data.

[0012] Furthermore, the knowledge graph generation unit includes: Explicit Relationship Extraction Subunit: Used to identify the original recognition text of each POI entity node according to the preset explicit rule pattern, and to complete the explicit relationship extraction of standardized POI entity nodes based on the matched data and the keyword dictionary contained in the POI dictionary. Dependency extraction subunit: used to perform deep learning recognition on the original recognition text of the POI entity node to obtain the relation label score of each token in the original recognition text, and capture the dependency relationship in the same text based on the relation label score and the label transition score; Attribute completion subunit: Used to predict attribute completion for each standardized POI entity node based on the POI dictionary, and complete the POI attribute with the attribute with the highest prediction probability; Knowledge Graph Construction Subunit: Used to construct a POI knowledge graph based on POI entity node data, explicit relationships of each POI entity node, dependency relationships of each POI entity node, and attributes.

[0013] Furthermore, the vectorized storage and indexing construction unit also includes: Variational autoencoder subunit: used to perform low-dimensional mapping on the high-dimensional multimodal vector according to the mapping rule obtained through training, to obtain a low-dimensional multimodal vector; wherein, the mapping rule is calculated in the following manner: Where x represents the input high-dimensional multimodal vector; z represents the mapped low-dimensional latent representation; q_ (z|x) represents the encoder distribution; p_θ(x|z) represents the decoder reconstruction distribution; KL() represents the Kullback–Leibler divergence; Ltask represents the task performance loss; β and γ are weight coefficients; Vector compression subunit: used to perform product quantization on the low-dimensional multimodal vector to obtain the compressed multimodal vector.

[0014] This invention formats and unifies multi-source, multi-modal POI data, abstracting it into a triple structure of entity, attribute, and relation. A pre-trained BERT+CRF model and a corresponding domain POI dictionary are used to identify entities for each POI, yielding candidate entities for knowledge base construction. After sorting, disambiguation, relation extraction, and attribute completion, each POI node has complete entity, attribute, and relation information. A POI knowledge graph is constructed based on the POI entity nodes. The POI node data and edge information in the obtained knowledge graph are then processed into graph structure vectorization and text vectorization. This unifies the multi-source, multi-modal information (including graph structure, text description, and geographic features) in the POI knowledge graph into a computable and searchable high-dimensional vector representation. A designed alignment network is used to integrate the structured information (graph) and unstructured information (text) modalities of the POI knowledge graph into a shared semantic space. In Space, a unified mapping of different modal vectors of the same POI entity further improves the subsequent retrieval and data fusion effect, realizing the unification of POI data. Finally, HNSW (Hierarchical Navigable Small World) is used to construct a graph index for the high-dimensional vector of each POI, building multi-level retrieval categories at different granularities. This allows the final multi-level POI knowledge base to combine with RAG retrieval technology. In the subsequent retrieval process, it can first match high-level features based on the retrieval category, and then search downwards layer by layer until the most accurate match is found at the bottom level. This solves the problem that POI fusion technology needs to search in different databases, resulting in a large retrieval range and long processing time, thus improving the efficiency of knowledge storage and retrieval. Attached Figure Description

[0015] Figure 1 This is a structural block diagram of the POI knowledge base construction system of the present invention; Figure 2 For execution Figure 1 The flowchart shows the POI knowledge base construction method of the system shown. Figure 3 This is a block diagram of the POI entity recognition unit; Figure 4 For execution Figure 3 The flowchart of the unit shown; Figure 5 A structural diagram of the knowledge graph generation unit; Figure 6 For execution Figure 5 The flowchart of the unit shown; Figure 7 A structural block diagram of the vectorized storage and indexing building unit; Figure 8For execution Figure 7 The flowchart of the unit shown. Detailed Implementation

[0016] For multi-source POI fusion technology, the key lies in the inconsistent storage formats and methods of POI data from different sources. Because of this, existing POI fusion methods match each POI data point using traditional text matching. However, multi-source POI data often contains aliases, name changes, and abbreviations, leading to problems such as insufficient semantic understanding, inadequate knowledge utilization, unclean fusion, and fusion errors. Research has shown that knowledge graphs can extract and complete different POI data; however, table-based and graph-based storage methods suffer from low retrieval efficiency and a large number of null values ​​in the data tables, respectively. Therefore, it is necessary to reconstruct the data in the knowledge graph to enable rapid retrieval and matching in subsequent POI fusion. This invention attempts to unify the multi-source, multi-modal information (including graph structure, text description, and geographic features) in the POI knowledge graph into a computable and searchable high-dimensional vector representation, providing a foundation for downstream similarity retrieval, clustering, and knowledge fusion.

[0017] The following is a description of some of the technologies used in this invention, for better understanding of this invention.

[0018] Knowledge graphs are data models that represent real-world entities (such as people, places, and concepts) and their relationships as a network structure, consisting of "nodes (entities)" and "edges (relationships)" to form a structured semantic knowledge base. They help search engines shift from string-based retrieval to concept-based retrieval, thereby providing more accurate, structured search results and in-depth knowledge, and are applied in fields such as search, recommendation, and intelligent question answering. BERT: Bidirectional Encoder Representations from Transformers (BERT) is a natural language processing model developed by Google that learns bidirectional representations of text, significantly improving the ability to understand unlabeled text in context across many different tasks.

[0019] RAG (Retrieval-Augmented Generation): A technique that combines the capabilities of retrieval systems and generative models to enhance the quality of generated data by retrieving relevant knowledge.

[0020] BIO annotation system: An annotation method in named entity recognition. B (Begin) indicates the start of an entity, I (Inside) indicates the inside of an entity, and O (Outside) indicates a non-entity.

[0021] Conditional Random Fields (CRFs): Probabilistic graphical models used for sequence labeling that can model the transition relationships between labels and ensure the global consistency of the output label sequence.

[0022] HNSW (Hierarchical Navigable Small World) is a data structure for quickly finding similar vectors. It organizes vector data into a multi-layered graph structure: the upper layer is sparse but covers a wide area, while the lower layer is dense but exhibits strong locality. During a search, the approximate region is quickly located in the upper layer, and then a precise search is performed in the lower layer—similar to first finding a city on a map and then finding a specific address on a street map, significantly improving the speed of large-scale vector retrieval.

[0023] Based on the above analysis, the present invention proposes the following... Figure 1 A POI knowledge base construction system is disclosed. It can construct a standardized knowledge graph from multimodal POI data, and perform text semantic vectorization on the descriptive text of POI entity nodes based on the knowledge graph. Simultaneously, an improved Geo-GNN is used to embed the graph structure of POI entities in the knowledge graph. A cross-modal semantic alignment module, also known as a cross-modal semantic alignment module, performs unified mapping between the text semantic vectors and graph-structured embedding vectors of the same POI entity node in a shared semantic space, resulting in high-dimensional multimodal vectors. Finally, HNSW (Hierarchical Navigable Small World) graph indexing is applied to the high-dimensional multimodal vectors of each POI entity node to obtain the final multi-layered POI knowledge base. The system includes: a data format standardization unit 10, a POI entity recognition unit 20, a candidate set recall unit 30, a POI entity ranking and disambiguation unit 40, a knowledge graph generation unit 50, and a vectorized storage and index construction unit 60. (See also...) Figure 2 The working process between the various units in this system is as follows: The data format standardization unit 10 is used to perform step S10: collect multi-source heterogeneous POI data, and perform data cleaning and standardization on each heterogeneous POI data to obtain a set of POI data in a unified format.

[0024] The collected heterogeneous data (such as CSV, xlsx, txt, JSON, etc.) is converted into a unified format. Generally, the data is unified into JSON format, and in the case of large amounts of data, it can be stored in Parquet format.

[0025] The POI entity recognition unit 20 is used to perform step S20: perform BERT encoding, linearization classification and conditional random field processing on the POI data to obtain the entity annotation result for each POI data, wherein the entity annotation result includes entity text, start position, end position, entity category and confidence score.

[0026] Please see Figure 3 and Figure 4 The POI entity recognition unit 20 includes: a lexicalization subunit 201 for performing step S201: sequentially lexicalizing the text data of each POI data, each POI data being split into token sequence data, and each token in the token sequence data including the corresponding token ID, attention mask and token typemask; Input processing employs a unified tokenization strategy, splitting text into tokens and generating token IDs, attention masks, and token type masks to ensure input alignment with tags. In this invention, a BIO annotation system is pre-built, defining entity category tags B-MERCHANT / I-MERCHANT (merchant name), B-BRAND / I-BRAND (brand name), B-GEO / I-GEO (geographic location), B-STREET / I-STREET (street information), B-BUILDING / I-BUILDING (building), and O (non-entity).

[0027] Taking the text "Starbucks Sanlitun Store, Sanlitun Taikoo Li, Chaoyang District, Beijing, Coffee Shop" as an example, the POI entity recognition unit splits it into the following token sequence: [CLS] Starbucks Sanlitun Store, Sanlitun Taikoo Li, Chaoyang District, Beijing, Coffee Shop [SEP] Its corresponding BIO tag is: [B-BRAND I-BRAND I-BRAND I-BRAND I-BRAND O B-GEO I-GEO I-GEO O B-GEOI-GEO I-GEO I-GEO O B-STREET I-STREET I-STREET O B-BUILDING O B-INDUSTRY I-INDUSTRY O] Each token is mapped to a token ID, and an attention mask and a token type mask are generated simultaneously. The attention mask is used to identify the positions of the actual input token and the padding tokens to ensure that the model only calculates attention for valid tokens; the token type mask is used to distinguish different segments in a text pair.

[0028] The bidirectional encoding subunit 202 is used to perform step S202: sequentially encode each token in the token sequence data using BERT to obtain the context hidden vector of each token; This bidirectional encoding subunit consists of 12 Transformer layers, and generates a 768-dimensional contextual hidden vector after each batch of token sequence input.

[0029] The linear classification subunit 203 is used to perform step S203: perform linear classification mapping on the context hidden vector of each token to obtain the non-normalized prediction score of each token on multiple preset labels; This subunit can map hidden vectors to a multi-dimensional label space according to preset mapping rules, and output the logits of each token in each category (logits represent the non-normalized prediction scores of each token in different label categories), which serve as the Emission score of the linear classification subunit and are used to model the label dependencies of the entire sequence. The linear classification subunit 203 defines a label transition matrix A, where A[i][j] represents the probability of transitioning from label i to label j. For example, the transition probability from B-BRAND to I-BRAND is usually higher, while the probability from B-BRAND to B-GEO is lower. Given an input sequence x and a label sequence y, the formula for calculating the sequence probability using CRF is: P(y|x) = exp(Σ(score(x,y))) / Z(x) score(x,y) = Σ(Emission[i][y[i]]+ Transition[y[i-1]][y[i]]) Where Emission[i][y[i]] represents the score when token i is predicted to be label y[i], Transition[y[i-1]][y[i]] represents the score when label y[i-1] is transferred to y[i], and Z(x) is a normalization factor used to ensure that the sum of the sequence probabilities is 1.

[0030] Sequence prediction subunit 204 is used to perform step S204: perform conditional random field processing based on the nonnormalized prediction score of the token, and find the maximum probability tag sequence of the token through dynamic programming. The specific calculation process is as follows: The forward probability α[t][i] of each time step t and label i is: α[t][i]= log(Σ exp(α[t-1][j]+ A[j][i]+ E[i][t])) Where α[t][i] represents the maximum logarithmic probability of the sequence ending at position t with label i, A[j][i] is the label transition score, E[i][t] is the Emission score, and j iterates through all labels in the previous time step. Finally, by backtracking backward, the label sequence with the highest forward probability is found.

[0031] The entity annotation subunit 205 is used to perform step S205: post-processing based on the text data of POI and the maximum probability label sequence of each corresponding token data to obtain the entity annotation result of each POI data.

[0032] Post-processing includes entity boundary correction (ensuring that a B-label is followed by a corresponding I-label), conflict resolution (selecting the label with the highest confidence as the final label when there are multiple labels at the same location), and entity merging (combining consecutive BI labels into a complete entity). Taking the example text "Starbucks Sanlitun Store, Sanlitun Taikoo Li, Chaoyang District, Beijing, coffee shop" as an example, the output after decoding and post-processing is shown in the table below:

[0033] Table 1 The entity annotation results are shown in Table 1.

[0034] The candidate set recall unit 30 is used to perform step S30: to perform RAG retrieval and recall on the entity annotation results in the annotated POI dictionary to generate an entity candidate set, wherein the POI dictionary includes the standard name of the entity and prior knowledge of common variants; The POI dictionary, as a subset and prerequisite resource of the knowledge base, primarily provides prior knowledge of standard names and common variations. Its construction process includes: collecting candidate words from business registration information, map open platform data, mainstream brand directories, and industry standard documents; standardizing the collected words, including merging synonyms, converting between simplified and traditional Chinese characters, and mapping Chinese and English; organizing the words into brand dictionaries, geographical name dictionaries, industry terminology dictionaries, and building type dictionaries according to different categories; and establishing a correspondence between standard names and common variations for each entry. This dictionary supports dynamic updates during system operation; that is, when new, uncollected entries are discovered during the recognition process, they can be added to the dictionary, thus achieving bidirectional iteration between the dictionary and the knowledge base.

[0035] After entity recognition and decoding are completed, the system outputs structured entity results, including entity text, start position, end position, entity category, and confidence score. These results serve as the input candidate set recall unit 30 for candidate text entities. This unit is responsible for matching and recalling the recognized text entities with standardized entity nodes in the POI dictionary or previously identified standardized entity nodes, thus providing basic data for ranking and disambiguation. First, the system performs fast string matching on text entities using an inverted index, using entities with the same or highly similar names as the candidate set. Simultaneously, it uses vectorized semantic retrieval technology to recall semantically similar entities. Semantic retrieval calculates the cosine similarity between the text embedding vector of the recognized entity and the vector representation of entities in standardized entity nodes in the POI dictionary or previously identified standardized entity nodes, thereby recalling candidate entities with different or synonymous names.

[0036] The POI entity ranking and disambiguation unit 40 is used to perform step S40: calculate similarity features based on the data in the entity candidate set, complete the ranking and disambiguation of entities, and generate standardized POI entity nodes. Each standardized POI entity node includes the original recognition text, start position, end position, entity category, standardized entity ID, and confidence score.

[0037] After candidate entities are generated, the system enters the ranking and disambiguation stage. The entity ranking and disambiguation unit prioritizes candidate entities using a machine learning ranking model, comprehensively considering the following features: string similarity between the entity name and the identified text, semantic matching between the entity and the context text, the entity's popularity or weight in the knowledge base (e.g., number of stores, brand influence), geographic information matching, and consistency between the entity's category and the context. By weighting and scoring each feature, the model ranks candidate entities, placing the most likely matching entities at the top. After ranking, the system performs disambiguation processing. Disambiguation not only considers the matching probability of a single entity but also introduces a collective disambiguation mechanism, which integrates the co-occurrence relationships of multiple entities in the same text, the consistency of the context theme, and the logical constraints between entities. For example, for the text "Starbucks Sanlitun Store," if entities such as "Taikoo Li" and "coffee shop" also appear, the system will determine that the standardized brand entity ID corresponding to "Starbucks" is M001 based on geographic proximity and category consistency, while "Sanlitun" and "Taikoo Li" are merged and mapped to the same geographic entity GEO001, avoiding incorrect mappings caused by name duplication or ambiguity. In addition, the system combines the existing relationship information of entities in the POI dictionary, such as the hierarchical relationship between brand headquarters and stores, and the affiliation relationship between business districts and streets, to perform global consistency verification on candidate entities, ensuring that the final mapping results are reasonable in both semantics and logic.

[0038] Ultimately, the entity information output from the sorting and disambiguation stages still includes the original recognized text, start position, end position, entity category, standardized entity ID, and confidence score. These structured and standardized entity nodes will serve as standardized POI entities in the knowledge base, supporting subsequent tasks such as relation extraction, attribute completion, and cross-source POI data fusion. Through this process, the system can effectively solve problems such as diverse entity names, complex addresses, and semantic ambiguity in multi-source data, achieving high-precision and high-coverage entity mapping and providing a reliable data foundation for building a modular RAG knowledge base.

[0039] The knowledge graph generation unit 50 is used to perform S50: supplement the relation features and attribute features of each POI entity node from the POI dictionary according to the standardized POI entity nodes, and construct a POI knowledge graph based on the supplemented POI entity nodes; Please see Figure 5 and Figure 6 The knowledge graph generation unit 50 includes: The explicit relation extraction subunit 501.a is used to perform step S501.a: identify the original recognition text of each POI entity node according to the preset explicit rule pattern, and complete the explicit relation extraction of standardized POI entity nodes according to the matched data and the keyword dictionary contained in the POI dictionary. First, explicit relationships are extracted using predefined pattern rules. These rules are constructed by combining POI domain standards, business registration information, address and geographic structure from map open platforms, industry standard documents, and brand directories. By analyzing the co-occurrence patterns of multi-source POI text, the system predefines explicit relationship types and constructs extraction keyword categories in the POI dictionary. Specific keywords, such as "located in," "belonging to," "opened in," "belonging to a chain," etc., are used, and these rules are encoded into pattern matching templates. This allows for the accurate extraction of direct semantic relationships between entities from input text. For example, when the input text is "Starbucks Sanlitun store is located in Sanlitun Taikoo Li," the rule pattern can directly generate the relation triple: (Starbucks Sanlitun store, located in, Sanlitun Taikoo Li). The dependency extraction subunit 501.b is used to perform step S501.b: perform deep learning recognition on the original recognition text of the POI entity node to obtain the relation label score of each token in the original recognition text, and capture the dependency relationship in the same text based on the relation label score and the label transition score.

[0040] A relation extraction model based on a deep neural network is invoked. This model's design references the general approach of open-source relation extraction frameworks (OpenNRE, spaCy Relation Extraction module). The model adopts the same 12-layer Transformer structure and 768-dimensional hidden states as the BERT-Base-Chinese encoder to achieve architectural consistency and module commonality. On this basis, a linear classification layer is added to map each token hidden vector to the relation label space, and dropout is used to prevent overfitting. For global relation modeling, a CRF or relation classifier can be combined, integrating the emission score with the label transition score to ensure that dependencies between multiple relations in the same text are captured. For example, given the input text "Starbucks Sanlitun store belongs to the brand Starbucks, headquartered in Seattle, USA", the model can output the following latent relations: Latent relations: (Starbucks Sanlitun store, belongs to the brand, Starbucks), (Starbucks, headquartered in, Seattle, USA).

[0041] Furthermore, for relation results from multiple sources, the knowledge graph generation unit 50 introduces consistency verification and confidence calculation mechanisms. Consistency verification ensures that multi-source relations do not conflict in terms of geographical location, industry category, and contextual dependencies by logically and semantically checking the relations of the same entity pair extracted from different data sources. For example, for "Starbucks Sanlitun store is located in Sanlitun Taikoo Li" and "Starbucks Sanlitun store is located in Sanlitun, Chaoyang District," the system uses administrative division and geographical hierarchy rules to verify the consistency of the two relations and merges or adjusts the relations to make them logically coherent. Logical constraints between attributes (such as the requirement that store industry category must be compatible with brand category) are also verified at this stage. Confidence calculation assigns a probability or score to each relation or attribute, representing the system's reliability. The probability predicted by the deep learning model, the template matching strength of rule matching, the weighted score of multi-source fusion, and global logical constraints all participate in the confidence calculation. For example, the relation (Starbucks Sanlitun store, located in Sanlitun Taikoo Li) might have a confidence score of 0.97 from model prediction, 0.99 from rule matching, and a final confidence score of 0.985 after multi-source fusion. Furthermore, if the relation is logically consistent with existing information in the knowledge base, the confidence score can be further increased; if a conflict exists, the relation can be reduced or discarded.

[0042] The attribute completion subunit 501.c is used to perform step S501.c: perform attribute completion prediction for each standardized POI entity node based on the POI dictionary, and complete the POI attribute with the attribute with the highest prediction probability.

[0043] In the attribute completion section, the knowledge graph generation unit uses rule-based reasoning to infer administrative divisions based on geographic coordinates and industry categories based on business scope. First, the entity text "Starbucks Sanlitun Store" and its context "located in Sanlitun Taikoo Li, coffee shop" are used as input sequences. The BERT encoder generates a 768-dimensional context representation for each token. Simultaneously, the identified standardized entity IDs (such as brand M001 and geographic entity GEO001) are used as additional features input to the attribute completion prediction model. Several linear layers are added to the output to predict different attribute categories, such as industry category, chain status, store size, and administrative division. A self-attention mechanism is used to learn the correlation between text features and historical attributes, thereby accurately predicting missing attributes. For example, during pre-training, the system learns that the industry category label for "Starbucks Sanlitun Store" is "coffee and beverages," its chain status is "chain," and its administrative division is "Chaoyang District, Beijing." This unit, trained on numerous similar samples, has learned that when the text contains "coffee shop" and the brand is "Starbucks," it highly likely corresponds to the F&B industry category and has a chain status of "chain." Therefore, during the inference phase, "Starbucks Sanlitun Store" and its context are input into the trained knowledge graph generation unit. The knowledge graph generation unit outputs the raw scores (logits) for each attribute category. These logits are unnormalized values ​​reflecting the model's relative confidence in each candidate category. To convert these raw scores into an interpretable probability distribution, the system applies the softmax function to the logits. Softmax is a commonly used normalization function that maps a set of real numbers to probability values ​​between 0 and 1, ensuring that the sum of the probabilities for all categories is 1. Its mathematical formula is:

[0044] in, This represents the predicted probability of category i. The original score for category i is given, with the denominator being the exponential sum of the scores from all categories, ensuring the output represents an effective probability distribution. In attribute prediction, softmax enables the model to generate a confidence score for each attribute category, reflecting the model's probability that the category is the correct attribute. For example, for industry category prediction, the model calculates logits for candidate categories such as "coffee and beverage," "comprehensive catering," and "tea shop," obtains the corresponding probabilities through softmax, and finally selects the category with the highest probability as the prediction result, while retaining its confidence value; for example, "coffee and beverage" has a confidence value of 0.943. Similarly, for chain status prediction, the model outputs logits for the categories "chain" and "non-chain," obtains the predicted probability after softmax normalization, and selects the category with the highest probability, "chain," with a confidence value of 0.981. The same applies to administrative division attributes; after softmax normalization, the model outputs a probability of 0.972 for "Chaoyang District, Beijing." Through softmax normalization, not only can the model output be mapped to intuitive probabilities, but it also facilitates subsequent consistency verification and confidence calculation with rule-based inference results, ensuring the accuracy and credibility of the predicted attributes.

[0045] Ultimately, the "Starbucks Sanlitun Store" was completed with full attribute information, including the industry category "coffee and beverages", the chain status "chain", and the administrative division "Chaoyang District, Beijing".

[0046] The knowledge graph construction subunit 502 is used to perform step S502: constructing a POI knowledge graph based on POI entity node data, explicit relationships of each POI entity node, dependency relationships of each POI entity node, and attributes.

[0047] Once all POI entity nodes have completed relationship extraction and attribute completion, a POI knowledge graph is constructed based on the relationships between the entity nodes.

[0048] The vectorized storage and index building unit 60 is used to perform step S60: based on the structure of the POI knowledge graph and the POI entity nodes, perform multimodal vectorization processing and HNSW graph index building to obtain a multi-level POI knowledge base.

[0049] The vectorized storage and indexing construction unit 60 includes a graph structure embedding vectorization subunit 601.a for performing step S601.a: obtaining the graph structure embedding vector of each POI entity node based on the structure of the POI knowledge graph, the geographic coordinate information of the POI entity node, and the distance decay weight function; The text semantic vectorization subunit 601.b is used to perform step S601.b: semantic modeling of the descriptive text of each POI entity node to obtain the text semantic vector of each POI entity node; The alignment network training subunit 602 is used to perform step S602: construct positive sample pairs based on the graph structured embedding vector and text description of each POI entity node, construct negative sample pairs based on the text semantic vector and text description of each POI entity node, and train the alignment network by maximizing the cosine similarity of the positive sample pairs and minimizing the similarity of the negative sample pairs to obtain the trained alignment network. The alignment network 603 is used to perform step S603: align the graph structured embedding vector and the text semantic vector of each POI entity node to obtain a high-dimensional multimodal vector of each POI entity node. HNSW index construction subunit 604 is used to perform step S604: construct an HNSW graph index based on each high-dimensional multimodal vector to obtain a multi-level POI knowledge base.

[0050] A Hierarchical Navigable Small World (HNSW) graph index is built for the compressed vector of each POI. This index achieves near-nearest neighbor search with logarithmic time complexity by constructing a multi-layer graph structure. An early stopping mechanism is introduced during graph construction, terminating the connection process prematurely when the number of connected neighbor nodes reaches a preset threshold or the similarity falls below a threshold, thereby controlling the graph density and query latency. In the intermediate semantic clustering layer, the system performs density-based semantic clustering within each geographic region. The clustering algorithm employs an improved DBSCAN method, combining vector similarity and POI functional category information to define neighborhood relationships. During clustering, density parameters are dynamically adjusted to group functionally similar and geographically proximate POIs into the same cluster. Each cluster maintains a central vector as a representative for fast matching during coarse-grained retrieval. In the top-level spatial partitioning layer, the system builds a geospatial index based on an improved R-tree data structure. Unlike traditional R-trees, this spatial index considers both geographic boundaries and semantic density during node splitting, ensuring that POIs within the same spatial region have a certain semantic relevance. Spatial indexes support multiple retrieval modes, including range queries, k-nearest neighbor queries, and composite queries.

[0051] Preferably, the vectorized storage and indexing construction unit further includes a variational autoencoder subunit and a vector compression subunit. Before constructing the HNSW graph index based on each high-dimensional multimodal vector, the variational autoencoder subunit performs low-dimensional mapping on the high-dimensional multimodal vector according to the training-derived mapping rules to obtain low-dimensional multimodal vectors; wherein, the mapping rules are calculated in the following manner: Where x represents the input high-dimensional multimodal vector; z represents the mapped low-dimensional latent representation; q_ (z|x) represents the encoder distribution; p_θ(x|z) represents the decoder reconstruction distribution; KL() represents the Kullback–Leibler divergence; Ltask represents the task performance loss; β and γ are weight coefficients.

[0052] The task performance loss is used to maintain the discriminative ability of the dimensionality-reduced vector in downstream retrieval tasks; the weight coefficients are used to balance reconstruction accuracy, distribution regularization, and task performance. After training, the mapping rule can be expressed as: Z=μ (x) That is, using the pre-trained encoder parameters The input vector x is projected into the latent semantic space to obtain a low-dimensional representation z.

[0053] This system optimizes storage efficiency and computational performance by performing task-oriented dimensionality reduction. The dimensionality reduction process employs a variational autoencoder (VAE) architecture, where the encoder maps the original high-dimensional vector *x* to a low-dimensional latent representation *z*, and the decoder attempts to reconstruct the original vector from *z*. Unlike traditional VAEs, this dimensionality reduction network introduces a task-related regularization term, ensuring that the reduced vector maintains high discriminative power in POI retrieval tasks. During training, the system uses a multi-objective optimization strategy. The loss function includes the reconstruction error *L_recon = ||x - x||²*, the KL divergence regularization term *L_KL = KL(q(z|x)||p(z))*, and the task performance loss *L_task*. The task performance loss is calculated by performing a POI similarity ranking task on the reduced vector, ensuring that the dimensionality reduction process does not significantly impair downstream retrieval performance. The system dynamically adjusts the target dimension size based on storage resource constraints and query latency requirements, achieving a balance between storage efficiency and retrieval accuracy.

[0054] Subsequently, the vector compression subunit performs product quantization on the low-dimensional multimodal vector to obtain the compressed multimodal vector.

[0055] After dimensionality reduction, the system further applies vector quantization technology to compress the vectors. Using product quantization, the d-dimensional vector is decomposed into m sub-vectors, each of which undergoes independent k-means clustering quantization. The quantization codebook is updated through an iterative optimization algorithm, minimizing quantization error while maintaining the relative distance relationships between vectors.

[0056] This invention formats and unifies multi-source, multi-modal POI data, abstracting it into a triple structure of entity, attribute, and relation. A pre-trained BERT+CRF model and a corresponding domain POI dictionary are used to identify entities for each POI, yielding candidate entities for knowledge base construction. After sorting, disambiguation, relation extraction, and attribute completion, each POI node has complete entity, attribute, and relation information. A POI knowledge graph is constructed based on the POI entity nodes. The POI node data and edge information in the obtained knowledge graph are then processed into graph structure vectorization and text vectorization. This unifies the multi-source, multi-modal information (including graph structure, text description, and geographic features) in the POI knowledge graph into a computable and searchable high-dimensional vector representation. A designed alignment network is used to integrate the structured information (graph) and unstructured information (text) modalities of the POI knowledge graph into a shared semantic space. In Space, a unified mapping of different modal vectors of the same POI entity further improves the subsequent retrieval and data fusion effect, realizing the unification of POI data. Finally, HNSW (Hierarchical Navigable Small World) is used to construct a graph index for the high-dimensional vector of each POI, building multi-level retrieval categories at different granularities. This allows the final multi-level POI knowledge base to combine with RAG retrieval technology. In the subsequent retrieval process, it can first match high-level features based on the retrieval category, and then search downwards layer by layer until the most accurate match is found at the bottom level. This solves the problem that POI fusion technology needs to search in different databases, resulting in a large retrieval range and long processing time, thus improving the efficiency of knowledge storage and retrieval.

[0057] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A POI knowledge base construction method, characterized by, The method comprises the following steps: S10: collecting multi-source heterogeneous POI data, and performing data cleaning and standardization on each heterogeneous POI data to obtain a set of POI data in a unified format; S20: performing BERT encoding, linear classification and conditional random field processing on the POI data to obtain an entity annotation result of each POI data, wherein the entity annotation result comprises entity text, a start position, an end position, an entity category and a confidence score; S30: performing RAG retrieval recall on the entity annotation result in a labeled POI dictionary to generate an entity candidate set, wherein the POI dictionary comprises prior knowledge of standard names and common variants of entities; S40: calculating similarity features according to data in the entity candidate set, completing entity sorting and disambiguation, and generating a standardized POI entity node, wherein each standardized POI entity node comprises original recognized text, a start position, an end position, an entity category, a standardized entity ID and a confidence score; S50: completing relationship features and attribute features of each POI entity node from the POI dictionary according to the standardized POI entity node, and constructing a POI knowledge graph according to the completed POI entity node; S60: performing multi-modal vectorization processing and HNSW graph index construction according to the structure of the POI knowledge graph and the POI entity node to obtain a multi-level POI knowledge base.

2. The POI knowledge base construction method of claim 1, wherein, The step S60 comprises: S601.a: obtaining a graph-structured embedding vector of each POI entity node according to the structure of the POI knowledge graph, geographic coordinate information of the POI entity node and a distance decay weight function; S601.b: performing semantic modeling on descriptive text of each POI entity node to obtain a text semantic vector of each POI entity node; S602: constructing a positive sample pair according to the graph-structured embedding vector of each POI entity node and the text description thereof, constructing a negative sample pair according to each POI entity node text semantic vector and the text description thereof, training an alignment network by maximizing the cosine similarity of the positive sample pair and minimizing the similarity of the negative sample pair, and obtaining the trained alignment network; S603: aligning the graph-structured embedding vector corresponding to each POI entity node and the POI entity node text semantic vector to obtain a high-dimensional multi-modal vector of each POI entity node; S604: constructing an HNSW graph index according to each high-dimensional multi-modal vector to obtain a multi-level POI knowledge base.

3. The POI knowledge base construction method of claim 2, wherein, The step S20 comprises: S201: tokenizing text data of each POI data in sequence, each POI data being split into token sequence data, and each token in the token sequence data comprising a corresponding token ID, an attention mask and a token type mask; S202: performing BERT encoding on each token in the token sequence data in sequence to obtain a context hidden vector of each token; S203: Linear classification mapping is performed on the context hidden vector of each token to obtain non-normalized prediction scores of each token on multiple preset labels; S204: Conditional random field processing is performed according to the non-normalized prediction scores of the token, and the maximum probability label sequence of the token is obtained by a dynamic programming method; S205: Post-processing is performed according to the text data of the POI and the maximum probability label sequence of each corresponding token data to obtain the entity annotation result of each POI data.

4. The POI knowledge base construction method of claim 3, wherein, The step S50 comprises: S501.a: According to the preset explicit rule, the original recognition text of each POI entity node is identified, and the explicit relationship extraction of the standardized POI entity node is completed according to the matched data and the keyword dictionary contained in the POI dictionary; S501.b: The original recognition text of the POI entity node is identified by deep learning to obtain the relationship label score of each token of the original recognition text, and the dependency relationship in the same text is captured according to the relationship label score and the label transition score; S501.c: According to the POI dictionary, attribute completion prediction is performed on each standardized POI entity node to predict the POI attribute with the highest prediction probability; S502: The POI knowledge graph is constructed according to the POI entity node data, the explicit relationship of each POI entity node, the dependency relationship of each POI entity node and the attribute.

5. The POI knowledge base construction method according to any one of claims 1-4, characterized in that, Before the step S604 of constructing the HNSW graph index according to each high-dimensional multi-modal vector, further comprising: The high-dimensional multi-modal vector is mapped to a low-dimensional multi-modal vector according to the mapping rule obtained by training; wherein the mapping rule is calculated by the following method: where x denotes the input high-dimensional multi-modal vector; z denotes the mapped low-dimensional latent representation; q (z | x) is the encoder distribution; p_ (x | z) is the decoder reconstruction distribution; KL() denotes the Kullback-Leibler divergence; Ltaskis the task performance loss; β, γ are the weight coefficients; The low-dimensional multi-modal vector is subjected to product quantization to obtain a compressed multi-modal vector.

6. A POI knowledge base construction system characterized by comprising: Comprise: A data format standardization unit is configured to collect multi-source heterogeneous POI data, and perform data cleaning and standardization on each heterogeneous POI data to obtain a set of POI data in a unified format; A POI entity recognition unit is configured to perform BERT encoding, linear classification and conditional random field processing on the POI data to obtain an entity annotation result of each POI data, wherein the entity annotation result comprises entity text, start position, end position, entity category and confidence score; A candidate set recall unit is configured to perform RAG retrieval recall on the entity annotation result in the annotated POI dictionary to generate an entity candidate set, wherein the POI dictionary comprises prior knowledge of standard names and common variants of entities; A POI entity sorting and disambiguation unit is configured to calculate similarity features according to data in the entity candidate set to complete entity sorting and disambiguation, and generate a standardized POI entity node, wherein each standardized POI entity node comprises original recognition text, start position, end position, entity category, standardized entity ID and confidence score; The knowledge graph generation unit is configured to complete the relationship features and attribute features of each POI entity node from the POI dictionary according to the standardized POI entity node, and construct a POI knowledge graph according to the completed POI entity node; The vector storage and index construction unit is configured to perform multi-modal vectorization processing and HNSW graph index construction according to the structure of the POI knowledge graph and the POI entity node, and obtain a multi-level POI knowledge base.

7. The POI knowledge base construction system of claim 6, wherein, The vector storage and index construction unit comprises: The graph structure embedding vectorization subunit is configured to obtain a graph structure embedding vector of each POI entity node according to the structure of the POI knowledge graph, the geographic coordinate information of the POI entity node, and a distance decay weight function; The alignment network training subunit is configured to construct a positive sample pair according to the graph structure embedding vector of each POI entity node and the text description thereof, construct a negative sample pair according to each POI entity node text semantic vector and the text description, train an alignment network by maximizing the cosine similarity of the positive sample pair and minimizing the similarity of the negative sample pair, and obtain a trained alignment network; The alignment network is configured to align the graph structure embedding vector corresponding to each POI entity node and the POI entity node text semantic vector, and obtain a high-dimensional multi-modal vector of each POI entity node. The HNSW index construction subunit is configured to construct an HNSW graph index according to each high-dimensional multi-modal vector, and obtain a multi-level POI knowledge base.

8. The POI knowledge base construction system of claim 7, wherein, The POI entity recognition unit comprises: The tokenization subunit is configured to sequentially tokenize the text data of each POI data, and each POI data is split into token sequence data, and each token in the token sequence data comprises a corresponding token ID, an attention mask, and a token type mask; The bidirectional encoding subunit is configured to sequentially perform BERT encoding on each token in the token sequence data, and obtain a context hidden vector of each token; The linear classification subunit is configured to perform linear classification mapping on the context hidden vector of each token, and obtain a non-normalized prediction score of each token on a plurality of preset labels; The sequence prediction subunit is configured to perform conditional random field processing according to the non-normalized prediction score of the token, and obtain a maximum probability label sequence of the token by a dynamic programming method; The entity annotation subunit is configured to perform post-processing according to the text data of the POI and the maximum probability label sequence of each corresponding token data of the POI, and obtain an entity annotation result of each POI data.

9. The POI knowledge base construction system of claim 8, wherein, The knowledge graph generation unit comprises: The explicit relation extraction subunit is configured to identify the original recognition text of each POI entity node according to a preset explicit rule mode, and complete explicit relation extraction of the standardized POI entity node according to the matched data and a keyword dictionary contained in the POI dictionary. The dependency relation extraction subunit is configured to perform deep learning recognition on the original recognized text of the POI entity node to obtain a relation label score of each token of the original recognized text, and capture the dependency relation in the same text according to the relation label score and a label transition score. The attribute completion subunit is configured to perform attribute completion prediction on each standardized POI entity node according to a POI dictionary to obtain a POI attribute with a highest prediction probability. The knowledge graph construction subunit is configured to construct a POI knowledge graph according to the POI entity node data, the explicit relation of each POI entity node, the dependency relation of each POI entity node and the attribute.

10. The POI knowledge base building system according to any one of claims 6-9, characterized in that, The vectorization storage and index construction unit further comprises: The variational auto-encoding subunit is configured to perform low-dimensional mapping on the high-dimensional multi-modal vector according to a mapping rule obtained through training to obtain a low-dimensional multi-modal vector. where x denotes the input high-dimensional multi-modal vector; z denotes the mapped low-dimensional latent representation; q (z | x) is the encoder distribution; p_ (x | z) is the decoder reconstruction distribution; KL() denotes the Kullback-Leibler divergence; Ltaskis the task performance loss; β, γ are the weight coefficients; The vector compression subunit is configured to perform product quantization on the low-dimensional multi-modal vector to obtain a compressed multi-modal vector.

Images