POI address type identification method based on knowledge graph and multiple neural networks

Abstract

The invention provides a POI address type identification method based on a knowledge graph and multiple neural networks, and the method comprises the steps: constructing the knowledge graph containing AOI nodes, POI nodes and address nodes through multi-source data, and building a position, proximity and inclusion relation; a hierarchical recognition mechanism is adopted, specifically, categories are directly returned through graph query, if no nodes exist, a keyword library is matched through a rule model, and a deep learning model (BERT is used for extracting text / geographic features, and dynamic weight fusion classification is carried out after CNN and BiLSTM-Attention combined extraction) fused with a knowledge graph is adopted for the fuzzy POI; and meanwhile, the problem of graph association query of the implied POI is solved, and the new POI automatically updates the graph. According to the method, the fuzzy / implicit POI recognition accuracy is remarkably improved, and the POIs with the same coordinates are effectively distinguished.

Description

Technical Field

[0001] The present invention belongs to the field of geographic information processing and artificial intelligence technology, and specifically relates to a POI address type recognition method based on knowledge graph and multi-neural network. Background Art

[0002] In today's rapidly developing digitalization and intelligentization, improving logistics and distribution efficiency, as well as making informed decisions about regional planning and commercial layout, rely heavily on accurate identification of POI address types. POIs can be generally categorized into large and small POIs. Small POIs are basic point-like entities of interest with specific geographic coordinates on a map, such as convenience stores and gas stations. They are small in scale and have relatively simple functions. Large POIs, on the other hand, are geographical entities of a certain scale and complexity, known as Areas of Interest (AOIs). These are large geographic spatial units formed by the aggregation of multiple small POIs, such as large business districts, industrial parks, and commercial complexes. However, the existence of ambiguous and implicit POI addresses has become a major obstacle to optimizing logistics and distribution, as well as scientifically planning regional areas. Ambiguous POIs are POIs whose names make it difficult to determine their type. For example, "Sunshine World" could represent a residential complex, a leisure complex, or even an office building. An implicit POI refers to an address that lacks a clear POI name, such as "No. 2, Cuihu North Road, Wuhua District, Kunming City", which only contains the city, district, road and road number. It is difficult to see that its corresponding POI is "Yunnan University".

[0003] Existing POI address type recognition methods can be roughly divided into the following three categories: keyword matching, statistical models and neural networks.

[0004] (1) The advantages of keyword matching are high efficiency and strong interpretability. It can process large-scale data sets quickly and return classification results in a short time. It has good classification effects on POIs with obvious classification attributes, such as POIs containing "university", "beef noodles", "hair salon", etc. The disadvantages are poor semantic understanding ability. It relies only on keyword matching and has difficulty understanding the semantic information and contextual relationships behind the keywords. It has weak generalization ability and cannot effectively classify unseen keywords or newly appeared place names. It also has poor fault tolerance. If there are typos, synonyms, polysemous words, etc. in the address data, it will lead to inaccurate matching and affect the classification effect.

[0005] (2) The advantages of statistical models are good stability and robustness, and they can effectively deal with noise and interference information in address data. The model is highly interpretable, and the degree of influence of each feature on the classification result can be clearly seen. The model has a certain learning ability and can automatically establish the connection between address features and classification labels. The disadvantages are that feature engineering is complex, requires a lot of time and experience, and has a certain reliance on empirical knowledge. It is easy to overfit or underfit, and has high requirements for training data.

[0006] (3) The advantage of the neural network model is that it has a strong feature learning ability and can automatically learn complex, nonlinear feature representations from large amounts of data without manually designing a complex feature extraction process. It is also highly robust to noise and outliers in address data. However, its disadvantage is that it is prone to overfitting and is not ideal for complex address classification scenarios, such as fuzzy POIs and implicit POIs.

[0007] Fuzzy POI addresses and implicit POI addresses account for a large proportion in real address data, and the above three POI address classification methods cannot effectively process these addresses. Therefore, it is urgent to develop a new POI address classification method to meet the growing intelligent needs in logistics distribution, regional planning and other fields. Summary of the Invention

[0008] In order to overcome the problems in the background technology, the present invention provides a POI address type recognition method based on knowledge graph and multi-neural network.

[0009] To achieve the above object, the present invention is implemented through the following technical solutions:

[0010] The POI address type recognition method based on knowledge graph and multi-neural network includes the following steps:

[0011] S1: Build a POI address knowledge graph:

[0012] S1a: Obtain POI address data from multiple source data;

[0013] S1b: Preprocess the data, including cleaning duplicate and erroneous data, and aligning entities through name similarity, address, and latitude and longitude matching;

[0014] S1c: Use the BiLSTM+CRF sequence labeling model for address parsing, identify entity boundaries and types in addresses, and deduce spatial inclusion relationships;

[0015] S1d: Construct a knowledge graph containing three types of entities and three types of relationships, where proximity is determined by the distance between POIs being less than 500 meters;

[0016] S2: POI address type identification:

[0017] S2a: Perform address parsing on the classified address text and extract POI entities;

[0018] S2b: If the POI node exists in the knowledge graph, directly return its category attribute;

[0019] S2c: If it does not exist, the type is identified through the rule model: the POI name is matched based on the keyword library and the three-level classification result is output;

[0020] S2d: If the rule model cannot be classified, the fuzzy POI recognition model integrated with the knowledge graph is used for processing.

[0021] S3: Handling implicit POI addresses:

[0022] When the address parsing fails to extract a POI, the address text is used as the query condition to find the associated POI node in the knowledge graph and return the category;

[0023] If there is no associated node, classification is performed according to the S2 trigger rule model or the fuzzy POI recognition model;

[0024] S4: Trigger the graph update mechanism: add category attributes to the newly added POI nodes and store them in the knowledge graph.

[0025] Furthermore, the entity alignment in S1 specifically includes merging entities by calculating name similarity and comparing addresses and longitude and latitude based on name differences of the same POI in different data sources.

[0026] Furthermore, in the knowledge graph of S1:

[0027] The three types of entities include AOI nodes, POI nodes and address nodes, and the three types of relationships include positional relationships, proximity relationships and inclusion relationships;

[0028] Among them, AOI nodes and POI nodes contain six attributes: alias, abbreviation, category, latitude and longitude, creation time, and abandonment time.

[0029] Furthermore, in S2, when the rule model cannot classify, the fuzzy POI recognition model is used for processing. First, BERT is used to encode the address text to generate text features, then the related information is queried from the knowledge graph and converted into text descriptions, and the geographic location features are generated by BERT. Then, the text features and geographic location features are fused, and features are jointly extracted through CNN and BiLSTM-Attention. The dynamic weight fusion mechanism is used to generate the final feature vector, and finally the classification result is output through the fully connected layer and Softmax.

[0030] Furthermore, the rule model of S2 is based on a three-level classification system, including 3 major categories and 371 sub-categories, and the keyword library covers POI aliases, abbreviations, and dialect expressions.

[0031] Furthermore, the dynamic weight fusion mechanism in the fuzzy POI recognition model of S2 includes:

[0032] Perform linear transformation and splicing on the output feature I of CNN cnn and the output feature I of BiLSTM-Attention lstm respectively;

[0033] Generate a fusion coefficient β through a learnable weight matrix W β and the Sigmoid function, where the calculation formula is β = Sigmoid(W β [W1I cnn ; W2I lstm + b β );

[0034] According to the fusion coefficient β, perform weighted summation on I cnn and I lstm to obtain the final feature I, and its calculation formula is I = β · I cnn + (1 - β) · I lstm ;

[0035] Finally, send the final feature I into a multi-classifier composed of a fully connected layer and a Softmax function to convert it into a final classification decision: y = Softmax(W f I + b f ).

[0036] Furthermore, the CNN of S2 uses one-dimensional convolutional kernels with 3 different window sizes, extracts multi-granularity features, and then fuses them through max pooling.

[0037] Advantages of the present invention:

[0038] The present invention realizes three major breakthrough improvements through the fusion of a knowledge graph and multiple neural networks:

[0039] (1) Break through the bottleneck of fuzzy or implicit POI recognition:

[0040] Establish a hierarchical processing mechanism: The rule model preferentially matches the dialect and alias library (such as "Jiayuan" → residence), covering 371 types of detailed scenarios; for the remaining fuzzy POIs, associate the attributes of surrounding POIs through the knowledge graph (such as querying the density of restaurants near "Yiran Garden"), and comprehensively determine the category by combining deep learning.

[0041] Implicit address resolution: Implicit addresses are associated with their corresponding AOI nodes through a graph (e.g., mapped to the "Seventh Block" business district), solving the positioning problem without POI names and significantly improving classification accuracy.

[0042] (2) Overcoming the difficulty of classifying POIs with the same coordinates: In large AOIs, even if the longitude and latitude are the same, the store types can be accurately distinguished by address hierarchy information (floor / area).

[0043] (3) Build an evolvable geographic intelligence system: New POIs automatically trigger the entry of node and category attributes into the database, continuously expanding the coverage of the map; integrate logistics addresses, social platforms, and government public data, eliminate data ambiguity through entity alignment technology, support real-time recognition of tens of millions of POIs, and provide decision-making support for logistics route optimization and urban commercial layout. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a flow chart of POI address type identification of the present invention;

[0046] Figure 2 This is a diagram of the fuzzy POI recognition model architecture of the present invention;

[0047] Figure 3 This is the POI address knowledge graph construction diagram of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0049] Example 1: Identification of POI Address Types in Large Commercial Districts

[0050] 1. Scenario Description

[0051] A logistics company needs to identify the type of "No. 302, 3rd Floor, Building A, Shuncheng Shopping Center, Wuhua District, Kunming City". This address has an implicit POI (no clear store name) and is located in a large AOI (stores at the same longitude and latitude in the shopping center are densely populated).

[0052] 2. Knowledge Graph Construction

[0053] 1. Data collection and preprocessing

[0054] (1) Extract the original record of “No. 302, 3rd Floor, Building A, Shuncheng Shopping Center” from SF Express data;

[0055] (2) Obtain related data from the POI of Amap: "Shuncheng Shopping Center (AOI) → Contains → Haidilao (POI, latitude and longitude: 102.71°E, 25.04°N)";

[0056] (3) Through entity alignment: merge "Shuncheng Shopping Center" and "Shuncheng MALL" into the same AOI node (name similarity 92%, longitude and latitude deviation <10 meters).

[0057] 2. Address resolution and graph construction

[0058] Input the address text into the BiLSTM+CRF model (BIOE annotation system), and the output is:

[0059] B-province:yun, I-province:nan, E-province:province;

[0060] B-city:Kunming, E-city:city;

[0061] B-district: Wuhua, I-district: District, E-district: District;

[0062] B-road: Shuncheng, I-road: Shopping, I-road: Center, E-road: Center;

[0063] B-building: Tower A, E-building: Tower;

[0064] B-floor: 3rd floor, E-floor: 1st floor;

[0065] B-room: No. 302, E-room: No.

[0066] 3. Construct knowledge graph nodes:

[0067] (1) AOI node: Shuncheng Shopping Center (category: commercial complex, latitude and longitude: 102.71°E, 25.04°N);

[0068] (2) Address node: No. 302, 3rd Floor, Building A;

[0069] (3) Relationship: (Shuncheng Shopping Center)-[Include]->(No. 302, 3rd Floor, Building A).

[0070] III. Implicit POI Recognition

[0071] 1. Address Resolution and Atlas Query:

[0072] Input the address "Room 302, 3rd Floor, Block A, Shuncheng Shopping Center, Wuhua District, Kunming City" into the BiLSTM+CRF model;

[0073] Analysis Result: No POI name is extracted → Determine as an implicit POI

[0074] 2. Query Associated Nodes in the Atlas:

[0075] Locate the address node "Room 302, 3rd Floor, Block A";

[0076] [[ID=--]]Find that it is associated with the POI node "Haidilao" (Category: Catering - Hot Pot) through the [Belonging Region] relationship.

[0077] 3. Output Result: Directly return the category "Physical Venue - Catering - Hot Pot".

[0078] IV. Fuzzy POI Recognition (assuming the address is "Yiran Garden in Shuncheng Shopping Center" and there is no keyword match)

[0079] 1. Rule Model Failure: The name "Yiran Garden" does not hit the keyword library;

[0080] 2. Fuzzy POI Model Processing:

[0081] (1) BERT Text Encoding: Generate text feature vectors for "Yiran Garden";

[0082] (2) Knowledge Graph Query: The neighboring POIs are (Haidilao, Category: Catering), (Uniqlo, Category: Clothing Retail), with a distance < 500 meters;

[0083] (3) Belonging AOI: Shuncheng Shopping Center (Category: Commercial Complex);

[0084] (4) Triple to Text: "Shuncheng Shopping Center contains restaurants and clothing stores" → BERT generates geographical features;

[0085] 3. Feature Fusion and Classification:

[0086] (1) Concatenate features;

[0087] (2) CNN Layer: Use 3 types of convolutional kernels (h = 2, 3, 4) to extract local features → Max pooling → Output;

[0088] (3) BiLSTM - Attention Layer: Learn global dependencies → Output;

[0089] (4) Dynamic Fusion;

[0090] (5) Softmax classification: The output with the highest probability is "physical place - catering - Chinese restaurant". V. Technical effect verification (as shown in the following table)

[0091]

[0092]

[0093] In this example, implicit POI location is addressed through knowledge graph association reasoning, such as "Most stores in Shuncheng Shopping Center are restaurants / retailers." Dynamic feature fusion enhances BiLSTM's ability to capture contextual dependencies, preventing misidentification of stores with the same coordinates. Furthermore, a dialect keyword library combined with a graph update mechanism continuously improves the coverage of the rule model. This example fully demonstrates a closed-loop process from data collection → knowledge construction → hierarchical identification → dynamic update, demonstrating the solution's effectiveness and innovation in complex scenarios.

[0094] Example 2: Industrial Park POI Address Type Identification

[0095] 1. Scenario Description

[0096] An energy group needed to classify the ambiguous address of an industrial park, "Nengchuang Base, No. 105, Yunda West Road, Kunming Economic Development Zone," and encountered the following issues:

[0097] Ambiguous name: "Energy Innovation Base" may be a factory, R&D center or logistics warehouse;

[0098] Complexity of the industrial environment: POIs with the same coordinates exist within the same industrial park (for example, the longitude and latitude of an office building and a workshop overlap).

[0099] 2. Knowledge Graph Query and Rule Matching

[0100] 1. Address resolution:

[0101] BiLSTM+CRF model output:

[0102] B-province:Yunnan, E-province:Nanjing, B-city:Kunming, E-city:city;

[0103] B-district: Economic Development Zone, I-district: District, E-district: District;

[0104] B-road:Yunnan University, I-road:West, E-road:Road;

[0105] B-roadno:105,E-roadno:No.

[0106] B-poi: Nengchuang, I-poi: base, E-poi: base;

[0107] POIs were extracted successfully.

[0108] 2. Graph query and rule model:

[0109] (1) There is no "Energy Creation Base" node in the knowledge graph → trigger rule model;

[0110] (2) Keyword database matching: "base" is associated with industrial manufacturing in the dialect database;

[0111] (3) But no specific subcategory is found (such as keywords such as "photovoltaic" and "battery") → the rule model returns "other".

[0112] 3. Fuzzy POI Model Deep Reasoning

[0113] 1. BERT dual-path feature extraction, extracting text features and geographic features respectively.

[0114] 2. Joint analysis of multiple neural networks:

[0115] (1) Splicing features;

[0116] (2) CNN local feature extraction: The convolution kernel with a window size of h = 3 captures keywords such as "photovoltaic" and "battery";

[0117] (3) BiLSTM-Attention global modeling: learning the industrial association of the sequence "nearby photovoltaic module factory-battery warehouse-distribution room";

[0118] (4) Dynamic weight fusion.

[0119] 4. Industrial Scenario Specificity Verification

[0120] 1. Field verification:

[0121] (1) The "Energy Innovation Base" is actually a new energy technology research and development center, including laboratories and small-scale pilot production lines;

[0122] (2) Misjudgment avoidance analysis: The traditional method misjudged the keyword "base" as a factory (with an accuracy rate of only 47%); the present invention infers the R&D attributes through the association of neighboring POIs (photovoltaic plant + battery warehouse).

[0123] 2. Example of distinguishing POIs with the same coordinates:

[0124] Address: Room 301, Building A, No. 105 Yunda West Road (same latitude and longitude as the Energy Innovation Base): CNN captures "Room" → office unit, BiLSTM models "Building A" → administrative building; dynamic fusion (β = 0.63) → output: "Company Enterprise - Service Industry - Information Technology" (IT Office).

[0125] 3. Comparison of technical effects (as shown in the table below)

[0126]

[0127] This example overcomes the ambiguity of industrial POI names by leveraging the industrial chain relationships of neighboring POIs (photovoltaic plant + battery warehouse → new energy R&D). A β value of 0.72 strengthens BiLSTM's ability to capture spatial semantic links ("neighborhood - power distribution room" implies R&D rather than production). By adding "energy innovation base → new energy R&D" to the rule base, similar scenarios in the future will no longer require deep learning. This example highlights the innovative nature of this technology in complex industrial scenarios. By mining industrial chain associations within a knowledge graph and integrating dynamic features, it addresses the blind spots of traditional methods in industrial POI classification.

[0128] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A POI address type recognition method based on knowledge graph and multi-neural network, characterized by: The following steps are involved: S1: Build a POI address knowledge graph: S1a: Obtain POI address data from multiple source data; S1b: Preprocess the data, including cleaning duplicate and erroneous data, and aligning entities through name similarity, address, and latitude and longitude matching; S1c: Use the BiLSTM+CRF sequence labeling model for address parsing, identify entity boundaries and types in addresses, and deduce spatial inclusion relationships; S1d: Construct a knowledge graph containing three types of entities and three types of relationships, where proximity is determined by the distance between POIs being less than 500 meters; S2: POI address type identification: S2a: Perform address parsing on the classified address text and extract POI entities; S2b: If the POI node exists in the knowledge graph, directly return its category attribute; S2c: If it does not exist, the type is identified through the rule model: the POI name is matched based on the keyword library and the three-level classification result is output; S2d: If the rule model cannot be classified, the fuzzy POI recognition model integrated with the knowledge graph is used for processing. S3: Handling implicit POI addresses: When the address parsing fails to extract a POI, the address text is used as the query condition to find the associated POI node in the knowledge graph and return the category; If there is no associated node, classification is performed according to the S2 trigger rule model or the fuzzy POI recognition model; S4: Trigger the graph update mechanism: add category attributes to the newly added POI nodes and store them in the knowledge graph.

2. The POI address type identification method based on knowledge graph and multi-neural network according to claim 1 is characterized in that: The entity alignment in S1 specifically includes merging entities by calculating name similarity, comparing addresses and longitude and latitude based on name differences of the same POI in different data sources.

3. The POI address type identification method based on knowledge graph and multi-neural network according to claim 1 is characterized in that: In the knowledge graph of S1: The three types of entities include AOI nodes, POI nodes and address nodes, and the three types of relationships include positional relationships, proximity relationships and inclusion relationships; Among them, AOI nodes and POI nodes contain six attributes: alias, abbreviation, category, latitude and longitude, creation time, and abandonment time.

4. The POI address type identification method based on knowledge graph and multi-neural network according to claim 1 is characterized in that: In S2, when the rule model cannot classify, the fuzzy POI recognition model is used for processing. First, BERT is used to encode the address text to generate text features, then the related information is queried from the knowledge graph and converted into text descriptions, and the geographic location features are generated by BERT. Then, the text features and geographic location features are fused, and features are jointly extracted through CNN and BiLSTM-Attention. The dynamic weight fusion mechanism is used to generate the final feature vector, and finally the classification result is output through the fully connected layer and Softmax.

5. The POI address type identification method based on knowledge graph and multi-neural network according to claim 1 is characterized in that: The rule model of S2 is based on a three-level classification system, which includes 3 major categories and 371 minor categories. The keyword library covers POI aliases, abbreviations and dialect expressions.

6. The POI address type identification method based on knowledge graph and multi-neural network according to claim 4 is characterized in that: The dynamic weight fusion mechanism in the fuzzy POI recognition model of S2 includes: Output feature I of CNN cnn And BiLSTM-Attention output feature I lstm Perform linear transformation and splice respectively; Through the learnable weight matrix W β And the Sigmoid function generates the fusion coefficient β, where the calculation formula is β=Sigmoid(W β [W1I cnn ;W2I lstm ]+b β ); According to the fusion coefficient β, I cnn and I lstm Perform weighted summation to obtain the final feature I, which is calculated as I = β·I cnn +(1-β)·I lstm ; Finally, the final feature I is fed into a multi-classifier consisting of a fully connected layer and a Softmax function to transform it into the final classification decision: y = Softmax (W f I+b f ).

7. The POI address type identification method based on knowledge graph and multi-neural network according to claim 6 is characterized in that: The S2 CNN uses three one-dimensional convolution kernels with different window sizes to extract multi-granularity features and then fuse them through maximum pooling.

Images