Method and device for automatically generating Chinese address element knowledge graph

Abstract

The embodiment of the invention provides a method and device for automatically generating a Chinese address element knowledge graph of a generative large language model, and the method comprises the steps: constructing a segmentation frame of Chinese address elements based on the semantic features of the address elements, and segmenting a complete address into an address element sequence based on the segmentation frame; the method comprises the following steps: defining address elements based on national standards, and constructing an ontology framework of a Chinese address element knowledge graph with a hierarchical structure and semantic constraints based on the definition of the address elements; based on the ontology framework, performing text classification on the address element sequence by adopting a GPT-based light supervision classification method; according to the ontology framework and the text classification, constructing a Chinese address element knowledge graph based on a Neo4j graph database; and adopting a semantic-spatial similarity calculation method to optimize the synonymous and synonymous entities in the Chinese address element knowledge graph.

Description

[0001] Method and apparatus for automatic generation of Chinese address element knowledge graph Technical Field This document relates to the field of computer technology, and in particular to a method and apparatus for automatically generating a Chinese address element knowledge graph based on a generative large language model. Background Technology

[0002] With the accelerating urbanization, booming e-commerce, and rapid popularization of mobile internet technology in my country, Chinese address information data has experienced explosive growth, playing a crucial role in various application scenarios such as real-scene 3D construction, high-precision map navigation, and intelligent logistics services. However, due to the complex structure of Chinese addresses, the lack of unified standards, diverse expression methods, and the presence of missing or incorrect key information, difficulties arise in the integration and matching of Chinese address information from different departments, platforms, or systems, severely impacting the quality and service capabilities of Chinese address information data. The construction of Chinese address element knowledge graphs has become key to the semantic fusion and intelligent services of multi-source heterogeneous Chinese address data. Currently, Chinese address element analysis mainly focuses on the grammatical analysis and processing of the structure, composition, and associations between Chinese words. Address knowledge graph construction primarily represents the hierarchical relationships of structured address information data, lacking the ability to construct semantic relationships and generate graphs for large-scale unstructured and even heterogeneous address information, thus limiting the promotion and application of Chinese address graphs.

[0003] Existing technologies introduce ontological concepts into the field of geographic information. Attempts are made to incorporate geographic ontology into spatial address verification tasks, and a preliminary knowledge representation framework for geospatial entities has been constructed. Although a knowledge graph is not directly built, it provides a theoretical foundation for address knowledge graph construction. In recent years, several explorations have been conducted in address knowledge graph construction in China. Existing technologies integrate the four levels of administrative divisions, place names, aliases, and old names of a city to form a standard address for that city. Using "province," "city," "district," "street," "old name," and "alias" as ontology classes, and extracting hierarchical relationships such as "belonging to," "old name," and "alias" from the standard address as ontology relationships, a standard address knowledge graph is constructed, supporting Chinese address correction and completion. This work mainly relies on the integration of structured address data, with knowledge extraction dependent on manual methods, lacking automated extraction and knowledge fusion capabilities. Existing technologies incorporate POI semantic information, extracting the ontology "province," "city," "district," "street," and "road" from a standard address database based on rules, along with the hierarchical relationship "located in." They also obtain the ontology "POI" and its attributes "first-level industry" and "second-level industry" from electronic maps, constructing a POI address knowledge graph. Address matching is then achieved through a similarity matching algorithm based on entity relationship weights, improving the address matching rate. However, its ontology design is relatively simple, lacking sufficient characterization of semantic and spatial levels, making it difficult to adapt to the current complex and diverse massive amounts of Chinese addresses. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for automatically generating Chinese address element knowledge graphs of generative large language models, aiming to solve the above-mentioned problems in the prior art.

[0005] This invention provides a method for automatically generating a Chinese address element knowledge graph of a generative large language model, comprising: A segmentation framework for Chinese address elements is constructed based on the semantic features of address elements. Based on the segmentation framework, the complete address is segmented into a sequence of address elements. Address elements are defined based on national standards, and an ontology framework for a Chinese address element knowledge graph with hierarchical structure and semantic constraints is constructed based on the definition of address elements. Based on the aforementioned ontology framework, a lightly supervised classification method based on GPT is used to perform text classification on the address element sequence. Based on the ontology framework and the text classification, a Chinese address element knowledge graph is constructed using the Neo4j graph database. The semantic-spatial similarity calculation method is used to optimize the synonymous entities in the Chinese address element knowledge graph.

[0006] This invention provides an automatic generation device for Chinese address element knowledge graphs based on generative large language models, comprising: The segmentation module is used to construct a segmentation framework for Chinese address elements based on the semantic features of address elements, and to segment the complete address into a sequence of address elements based on the segmentation framework. The definition module is used to define address elements based on national standards, and to construct an ontology framework for a Chinese address element knowledge graph with hierarchical structure and semantic constraints based on the definition of address elements. The classification module is used to perform text classification on the address element sequence based on the ontology framework and using a lightly supervised classification method based on GPT. The construction module is used to construct a Chinese address element knowledge graph based on the Neo4j graph database according to the ontology framework and the text classification. The optimization module is used to optimize synonymous entities in the Chinese address element knowledge graph using a semantic-spatial similarity calculation method.

[0007] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the above-described method for automatically generating a Chinese address element knowledge graph of a generative large language model.

[0008] This invention also provides a computer-readable storage medium storing an information transmission implementation program. When the program is executed by a processor, it implements the steps of the above-described method for automatically generating a Chinese address element knowledge graph of a generative large language model.

[0009] By employing embodiments of the present invention, address element knowledge graphs are integrated with natural language processing and knowledge representation technologies, which can effectively improve the organization, storage, management, and intelligent understanding capabilities of address data. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the automatic generation method of Chinese address element knowledge graph of generative large language model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the Chinese address segmentation framework according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the address element knowledge graph ontology design process according to an embodiment of the present invention; Figure 4 This is the address element knowledge graph ontology diagram of an embodiment of the present invention; Figure 5 This is a schematic diagram of the address element knowledge graph ontology model according to an embodiment of the present invention; Figure 6 This is a schematic diagram of an example of GPT-TC according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the construction process of the address element knowledge graph according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the automatic generation device for Chinese address element knowledge graph of generative large language model according to an embodiment of the present invention; Figure 9 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0012] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0013] Method Implementation Examples According to embodiments of the present invention, a method for automatically generating a Chinese address element knowledge graph of a generative large language model is provided. Figure 1 This is a flowchart of the automatic generation method for Chinese address element knowledge graph of generative large language model according to an embodiment of the present invention, such as... Figure 1 As shown, the method for automatically generating a Chinese address element knowledge graph of a generative large language model according to an embodiment of the present invention specifically includes: Step S101: Construct a segmentation framework for Chinese address elements based on the semantic features of address elements; based on the segmentation framework, segment the complete address into a sequence of address elements; specifically including: A Chinese address element segmentation framework is constructed based on the semantic features of address elements. The address information is segmented into candidate address elements through Chinese word segmentation. MacBert (Masked Correction BERT) is used to embed the candidate address elements into word vectors with semantic features. A linear layer is used to map the word vector of each candidate address element to a score of BIO tags. The dependency relationship between the tags is modeled with the help of Conditional Random Field (CRF). Finally, the Viterbi algorithm is used to decode the optimal tag sequence, realizing the segmentation of Chinese addresses into address element sequences.

[0014] Step S102 involves defining address elements based on national standards, and constructing an ontology framework for a Chinese address element knowledge graph with hierarchical structure and semantic constraints based on these definitions; specifically including: The address element ontology is defined based on national standards, and the ontology class names and hierarchical structure are determined, specifically including: administrative division name, provincial division name, province, autonomous region, municipality, special administrative region, prefecture-level division name, prefecture-level city, region, autonomous prefecture, league, county-level division name, county, autonomous county, county-level city, banner, autonomous banner, municipal district, forest area, township / street-level division name, town, township, ethnic township, street, street, courtyard number, unit door number, floor, household room number, and point of interest; Define the attributes of the address feature ontology and design the semantic relationships between address feature ontologs. The attributes of the address feature ontology include: unique identifier, name, longitude, latitude, alias, and object attribute "located". Based on the ontology class name and hierarchical structure, the attributes of the address element ontology, and the semantic relationships, a model of the ontology framework of the Chinese address element knowledge graph is constructed, and data types and constraints are set.

[0015] Step S103: Based on the address element sequence and the ontology framework, a lightly supervised classification method based on GPT (Generative Pre-Trained Transformer) is used to perform text classification on the address element sequence; specifically including: We employ a lightly supervised classification method based on GPT, leveraging the language understanding and generation capabilities of a large language model to transform the address element classification task into a natural language generation task with context and examples. Through designed prompts and a small number of examples, we achieve automatic annotation of address element sequences, perform semantic classification on actual address corpora, identify the address element category of each word, and finally map the classification results into structured entities and relationships.

[0016] Step S104, based on the ontology framework and the text classification, the construction of a Chinese address element knowledge graph based on the Neo4j graph database; specifically includes: Import the constructed address feature ontology file into the Neo4j graph database, initialize the class, object attribute, and data attribute structure of the graph, and generate a semantic framework consistent with the address feature ontology structure through the Neo4j graph database to provide constraints and type support for the subsequent insertion of address entities; Using address element ontology as nodes and structural "located" relationships between address elements as edges, entities and relationships in the Neo4j graph database are constructed. Each address data is represented as a top-down path structure in the Neo4j graph database. The path structure corresponds to the hierarchical structure of the address ontology, reflecting the spatial and semantic subordinate relationships between address element entities. Finally, a Chinese address element knowledge graph composed of nodes and relationship edges is obtained.

[0017] Step S105 involves optimizing the synonymous entities in the Chinese address element knowledge graph using a semantic-spatial similarity calculation method. Specifically, this includes: A semantic-spatial similarity calculation method is adopted. A set of downstream adjacent address entities is constructed based on the "located" relationship of address entities. The semantic matching degree is evaluated by the similarity of the set of downstream adjacent address entities. A spatial contour is constructed by combining a set of address points with latitude and longitude information, and the geographical overlap between address entities is calculated. A semantic-spatial collaborative constraint mechanism is introduced between similar address elements. The similarity results are integrated with a minimum value strategy. Based on the fusion judgment rule, the address entities are accurately fused.

[0018] The technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] The following embodiments of the present invention focus on the construction process of an address element knowledge graph, including key methods and implementation paths for Chinese address segmentation, detailed address element ontology design, address element information extraction and classification, address element graph construction, and fusion of entities with the same name.

[0020] First, in the Chinese address segmentation part of this invention, based on the semantic features of address elements, a Chinese address segmentation framework combining Jieba, MacBert, and CRF is proposed to accurately segment Chinese addresses into sequences of address elements. In the address ontology design and modeling part, address elements are defined according to national standards, constructing an ontology framework with hierarchical structure and semantic constraints, clarifying core classes, attributes, and relationships. Second, in the address information extraction part, a lightly supervised classification method based on GPT is proposed to efficiently identify multi-granularity entities in address element sequences. Then, a knowledge graph is constructed based on the Neo4j graph database, ensuring data visualization and queryability. Finally, to address the redundancy problem of synonymous entities, a semantic-spatial similarity mechanism is introduced to achieve the fusion and unified expression of address elements. Detailed explanations follow.

[0021] 1. Chinese URL splitting Because Chinese characters are continuous strings without word boundaries, and Chinese addresses have complex structures and lack standardized definitions, they suffer from issues such as missing and redundant address elements, making Chinese address segmentation more difficult. To fully utilize the contextual semantic features of address elements, a Chinese address segmentation framework combining Jieba word segmentation, a pre-trained language model MacBERT with full-word masking, and a Conditional Random Field (CRF) is proposed. The specific process of this framework is as follows: Figure 2 As shown.

[0022] (1) Jieba pre-segmentation (candidate address element generation) In the original address text, word boundaries are not clear, and there are many compound place names (such as "Zhongguancun East Road" and "Xisanqi, Haidian District"). To reduce the complexity of subsequent model processing, the Jieba word segmentation tool is first used to perform preliminary segmentation of the address text, obtaining a set of candidate address elements. Jieba relies on a user dictionary and a prefix graph algorithm, which can effectively identify common geographic entity words and generate reasonable word boundary sequences as the basic units for subsequent semantic modeling.

[0023] (2) MacBERT embedding representation (semantic feature extraction) The candidate address elements segmented by Jieba are input into the pre-trained language model MacBERT to capture their semantic features within the entire address context, obtaining the contextual semantic representation of each element. Considering the inconsistent lengths of different candidate address element sequences, to achieve batch modeling and training, each address sequence needs to be padded to a fixed length of 30 (to meet the maximum length of the existing pre-segmented data). The specific processing is as follows: 1) Sub-word encoding and word-level semantic mapping Candidate address elements are input into MacBERT, where each address element is further divided into several sub-tokens, and a corresponding contextual semantic vector is output. However, since the labels in the address segmentation task are based on BIO label prediction using address elements (words), it is necessary to map the sub-token vectors output by MacBERT to word-level representations. To achieve this alignment, the mapping strategy in this invention is as follows: average pooling is performed on the output vectors of all sub-words to obtain the word-level representation of the candidate address element, providing word-level input for the subsequent CRF module.

[0024] 2) Semantic representation sequence construction and sequence alignment The word vectors of all address elements are concatenated into a word-level semantic representation sequence and uniformly padded to a preset length of 30 to accommodate batch training requirements. To ensure that the model only models real address elements and to avoid padded vectors interfering with subsequent decoding, a CRF mask is constructed using an attention mask to indicate whether each position is a valid token. Based on this, the resulting semantic vector sequence of address elements will be used as input to the CRF layer for subsequent sequence labeling and optimal path decoding.

[0025] (3) Conditional Random Field (CRF) annotation and decoding (address element identification) The semantic vectors output in step (2) are mapped to the BIO label space, and the label scores (logits) for each word are obtained through a fully connected linear layer. Based on this, a Conditional Random Field (CRF) is used to model the transition relationships between labels to optimize the global consistency of the label sequence. The CRF effectively avoids unreasonable label combinations (such as omitting "B-county" before "I-county"), improving label validity and overall accuracy. Finally, the Viterbi algorithm is used to perform optimal path decoding at the CRF layer, outputting the segmentation labels for each address element, thus achieving structured segmentation of Chinese addresses.

[0026] 2. Address Element Ontology Design and Modeling The ontology design of address elements is an iterative process, continuously modifying and improving the ontology based on the scope of the knowledge being described and the problems to be solved. Simultaneously, the ontology is a model reflecting the real world, requiring its design to be intuitive, scalable, and maintainable. This invention adheres to the following principles in the ontology design of address elements: (1) Determine the domain and scope of the ontology based on knowledge requirements or problem tasks: that is, the starting point of the ontology design and improvement process is to fully model the semantic knowledge contained in the address. For example, if the research goal is to realize address element parsing, the ontology class needs to cover administrative divisions, roads, courtyard numbers, floors, etc.; if the task is to realize address matching through POI attributes, the ontology attributes need to include "first-level industry classification" and "second-level industry classification".

[0027] (2) Naming consistency and readability: The naming of ontology classes, attributes, and relationships should be semantically clear and consistent, conform to natural language habits, and avoid ambiguity and polysemy. In the naming process, the names in national standard documents should be used first, and the missing English names should be used from authoritative knowledge bases or open data platforms.

[0028] (3) Ontology reuse and knowledge reuse: that is, the ontology design should reuse existing mature ontology or standards as much as possible, while referring to existing research on ontology work.

[0029] (4) Alignment and integration of existing ontology resources: that is, directly aligning with existing domain authoritative ontology during the ontology design process, and improving the universality and scalability of address element ontology by linking existing ontology concepts.

[0030] Based on the above design principles, such as Figure 3 As shown, the specific design process of the technical solution of this invention is as follows: (1) Ontology class definition and hierarchical structure construction The Address Element Knowledge Graph aims to fully model Chinese address elements, meaning that each address element has a clear category. This includes administrative division names, such as provincial, prefecture, county, township / street level administrative divisions, and community / village (residential) committee names, as well as their specific types, such as province, autonomous region, municipality, and special administrative region, and street, courtyard number, unit door number, floor, household room number, and point of interest.

[0031] Referring to ontology design principles and following ontology design procedures, this embodiment of the invention defines ontology classes and class hierarchies for the address element knowledge graph, as shown in Table 1.

[0032] Table 1. Ontology classes and class hierarchy of the address element knowledge graph

[0033] (2) Ontology attribute design and entity relationship modeling Having clarified the categories and hierarchical structure of address element ontology, the next step is to define the attributes of address element ontology and design semantic relationships between entities to meet the requirements of downstream tasks such as address resolution, information fusion, and entity alignment. For each address element ontology, its data attributes include a unique identifier, name, longitude, latitude, and alias; object attributes are located at [location missing]. Based on the GeoNames ontology, international geographic information standards, and national standards, the design of the address element knowledge graph data attributes and object attributes is shown in Table 2.

[0034] Table 2. Attribute Design of Address Element Knowledge Graph

[0035] It should be noted that the default type of the property is data property, and * indicates that the property is an object property; constraints are set through Protégé's restriction type.

[0036] Based on the ontology class definitions in Table 1 and the attribute designs in Table 2, this embodiment of the invention uses Protégé to construct the ontology model and sets the data types and constraints. Figure 4 This is a schematic diagram of the ontology class for the address element knowledge graph and the ontology class for GeoNames. Figure 5 This refers to the ontology classes, data attributes, and object attributes of the address element knowledge graph ontology model.

[0037] 3. Address information extraction based on generative large language models This invention proposes a text classification method, GPT-TC (Text Classification Based on GPT), based on the generative pre-trained language model GPT, focusing on the core task of address element classification in the address element graph construction process. This method leverages the powerful language understanding and generation capabilities of large language models to transform the traditional address element classification task into a natural language generation task. Through carefully designed prompts and a small number of examples, it automatically labels address element sequences. Compared to traditional discriminative methods, GPT-TC can more flexibly adapt to multi-level labeling systems and introduces a semantic backoff mechanism to handle ambiguous or difficult-to-distinguish words, thereby improving the robustness and generalization ability of the overall classification. Based on this, the GPT-TC method is used to perform semantic classification on actual address corpora, identifying the address element category of each word. Finally, the classification results are mapped to structured entities and relationships, and an address knowledge graph conforming to the ontology structure is constructed using the Neo4j graph database. This achieves automatic conversion and organization from unstructured text to a semantic graph, laying the foundation for subsequent address element fusion and downstream tasks.

[0038] (1) Text classification method based on GPT (GPT-TC) In the field of natural language processing, generative pre-trained transformers (GPTs) have demonstrated strong adaptability to various downstream tasks. Compared to traditional discriminative models, GPTs can not only understand complex contextual semantics, but also complete tasks through natural language prompts in scenarios lacking explicit supervision.

[0039] This capability provides a new approach for the automatic classification of address element information. To fully utilize the generation capabilities of large language models, this invention proposes a text classification method based on GPT, namely GPT-TC (Text Classification Based on GPT). Figure 6 GPT-TC transforms the classification task into a natural language generation task with context and examples. By constructing appropriate prompts and combining them with a few examples (Few-shot demonstrations), it achieves parallel classification of multiple input sequence of address elements. This method not only adapts to traditional flat classification structures but also supports multi-level label systems, providing flexible expressive capabilities for multi-level label classification with inconsistent granularity. The prompts mainly consist of the following parts: 1) Task Definition: First, define the model's role and establish background knowledge of GPT. Then, concisely describe to the model that the current task is text classification and the contextual relationships of the input. The specific format is as follows: You are a text classification expert. This task involves semantically classifying each word in the input text word sequence based on a given multi-level labeling system, <contextual relationships of a single text word sequence>. Here, <contextual relationships of a single text word sequence> describes the semantic or logical relationships within a single text word sequence. For example, for an address element sequence, the entire address element sequence constitutes a complete address.

[0040] 2) Label System Description: Provide the label system used for the classification task and give the classification principles. The specific format is as follows: The classification principle is: if the finest-grained category can be identified, output the corresponding category; if the finest-grained category cannot be identified, output its parent category; if it still cannot be determined, output <default>The tagging system is as follows: <Multi-level tagging system>. Among them, <default>The default label class for unrecognizable time is set according to different text classification tasks. For the address element classification of the embodiment of the application, the default label is "address element", and the multi-level label system is the hierarchical structure of the address element knowledge graph ontology class.

[0041] 3) Few-shot example: construct a sample containing an input text token sequence and its corresponding label to show the expected input-output pattern of the model and enhance the generation consistency of the model through example learning. In order to improve the classification effect and task robustness, GPT-TC adopts the Few-shot Prompting strategy, which adds several example samples in the prompt words, and at the same time guides the LLM to imitate the output style of the example samples, so as to facilitate the model to analyze the output content of the LLM. The specific format is as follows: output in JSON array form, return a JSON object, and each key is an input text token sequence, and the value is an array of corresponding token classification. All text tokens need to be covered, and the order is consistent with the input. Among them, text is the original token, and label is its semantic category. The following is an example sample: input: <text token sequence>1, <text token sequence>2, …, <text token sequence>k output: { <text token sequence>1: { ["text": <token>11, "label": <label>11], …}, … <text token sequence>k: { ["text": <token>1k, "label": <label>1k], …}} Among them, <text token sequence> is a plurality of to-be-analyzed tokens with context relationship, and if there is no context relationship, it should be placed in different text token sequences.

[0042] (2) Address element category classification based on GPT-TC The embodiment of the application proposes an address element text classification method GPT-TC based on a generative pre-training language model. By constructing a prompt (Prompt) and introducing a Few-shot example, the traditional address element classification task is converted into a generative task. On this basis, the GPT-TC of the application example further labels the semantic categories of the segmented address element sequence. This process aims to provide structured information support at the semantic level for knowledge graph construction.

[0043] For the address elements contained in the address element sequence, the semantic understanding ability of the large language model (LLM) is used to identify the address element category to which each token belongs. Unlike traditional classification methods, GPT-TC has the following two significant advantages, making it more adaptable and robust in complex structure address element labeling tasks: 1) Support for multi-level label system and automatic fallback mechanism The address element ontology has a strict hierarchical structure, for example, under "administrative division name", it can be further divided into "provincial administrative division", "prefecture-level administrative division", "county-level administrative division", and "township / street-level division name"; "provincial administrative division" further includes "provinces", "autonomous regions", "municipalities directly under the central government", etc. In actual tasks, some address tokens may be difficult to determine their most specific category due to ambiguous expression, insufficient context, or not conforming to a specific format. To address this issue, GPT-TC proposes a semantic fallback mechanism: if the most specific category cannot be determined, the model can fallback and predict its superior category label, ensuring that the classification task is always completed within the ontology label system, avoiding the generation of "unknown" categories. This mechanism can improve the robustness of the model on ambiguous samples.

[0044] 2) Utilize the knowledge generalization ability of large language models Traditional classification models rely on prior knowledge in the training corpus and perform poorly on rare or uncommon address elements. GPT-type language models, through large-scale pre-training, have extensive world knowledge and can understand that "Zhujiang Road" belongs to the "street" type and "Surveying and Mapping Mansion" is an "interest point". After being limited by the prompt, GPT-TC can activate its internal knowledge and improve its semantic discrimination ability for address entities, especially for complex composite entities.

[0045] According to the hierarchical structure of the address element ontology and the prompt construction method of GPT-TC, the following prompts are constructed for the address element category classification task: you are an address element classification expert, this task classifies each address element token in the input address element sequence according to the given multi-level label system, the input single address sequence is a complete address segmentation. The classification principle is: if the most specific category can be identified, output that category; if the most specific category cannot be identified, output its superior category; if it still cannot be determined, output "address element". The label system is as follows: - Address element - Administrative division name - Provincial division name - Province - Autonomous region - Municipality directly under the central government - Special administrative region - Prefecture-level division name - Prefecture-level city - Region - Autonomous prefecture - League - County-level division name - County - Autonomous county - County-level city - Flag - Autonomous Banner - Municipal districts - Forest area - Special Administrative Region - Township / street-level administrative division names - Town - Township - Ethnic Township - street - Streets and alleys - Courtyard Number - Unit number - Floor - Room number Points of interest Output as a JSON array, returning a single JSON object where each key is a sequence of input address elements and the values ​​are arrays of the corresponding address element categories. All text terms must be overwritten, in the same order as the input. Here, `text` represents the original address element terms, and `label` represents their semantic categories. The following is a sample example: Input: Beijing | Haidian District | Lianhuachi West Road | No. 28 Courtyard | Surveying and Mapping Building, Beijing | Haidian District | Beitaiping Road | No. 16 Courtyard | Surveying and Mapping Science Research Institute { "Beijing | Haidian District | Lianhuachi West Road | No. 28 Courtyard | Surveying and Mapping Building":{ {"text": "Beijing Municipality", "label": "Municipality"}, {"text": "Haidian District", "label": "Municipal District"}, {"text": "Lianhuachi West Road", "label": "street and alley"}, {"text": "Courtyard No. 28", "label": "Courtyard Number"}, {"text": "Surveying Building", "label": "Point of Interest"} }, "Beijing | Haidian District | No. 16, Beitaiping Road | Surveying and Mapping Science Research Institute":{ {"text": "Beijing Municipality", "label": "Municipality"}, {"text": "Haidian District", "label": "Municipal District"}, {"text": "North Taiping Road", "label": "Jieluxiang"}, {"text": "Courtyard No. 16", "label": "Courtyard Number"}, {"text": "Academy of Surveying and Mapping Science", "label": "Point of Interest"} } } 4. Construction of Address Element Knowledge Graph Based on Neo4j To achieve the structured semantic representation and graph storage management of address data, this paper selects the graph database Neo4j as the carrier platform for the address knowledge graph, and uses its n10s plugin and APOC plugin to complete the import and semantic modeling of the ontology. The specific process is as Figure 7 shown.

[0046] (1)Chinese Address Segmentation and Address Element Classification The graph entities come from the address text data in the address corpus. The original input is a continuous Chinese address string. In the embodiment of the present invention, first, based on the address element segmentation method FISCA-MS, the complete address is segmented into an address element sequence (such as "Beijing City|Haidian District|Lianhuachi West Road|Courtyard No. 28|Surveying and Mapping Building"), which is convenient for structuring address elements. Then, the GPT-TC classification method is used to perform text classification on the segmented address element sequence. This model can automatically judge the address element category to which each token belongs based on the prompt words and a small number of examples, and output the label result in JSON structure, so as to achieve the full automation classification of address text into semantic structures.

[0047] (2)Ontology Import and Graph Database Initialization Neo4j supports the import and semantic interpretation of RDF / OWL files through the n10s plugin (Neosemantics). In the embodiment of the present invention, the n10s.onto.import.ontology method is used to import the previously constructed address element ontology file (in RDF format) into the graph database, and initialize the class (Class), object property (ObjectProperty), and data property (DatatypeProperty) structures of the graph. After this process, the Neo4j graph database has a semantic framework consistent with the ontology structure, providing constraints and type support for the subsequent insertion of address entities.

[0048] (3)Entity Creation and Relationship Construction After semantic classification, this embodiment of the invention uses address elements as nodes and the structural "located" relationship between address elements as edges to construct entities and relationships in the graph database. Each address data is represented in the graph database as a top-down path structure: (:Municipality{name: "Beijing", id: 1})-[:located]->(:District{name: "Haidian District", id: 2})-[:located]->(:Street / Road{name: "Lianhuachi West Road", id: 3})-[:located]->(:Courtyard Number{name: "No. 28 Courtyard", id: 4})-[:located]->(:Point of Interest{name: "Surveying Building", id: 5,lat: 39.89657888973683, lon: 116.29652008804757}).

[0049] This path corresponds to the hierarchical structure of the address ontology, reflecting both the spatial and semantic hierarchical relationships between address element entities. Entity creation and relationship building are automated using the Cypher query language, with the core process including: ① Node creation (MERGE nodes and their rdf:type types); ② Edge connection (MERGE "located" relationship); ③ Attribute binding (e.g., name, source address, unique identifier, etc.); Example Cypher statements are as follows: MERGE (p:Municipality{name: "Beijing"}) MERGE (c:District{name: "Haidian District"}) MERGE (p)-[:Located in]->(c)…… Ultimately, this embodiment of the invention yielded an address element knowledge graph ACKG consisting of 13,376 nodes and 13,375 relational edges.

[0050] 4. Optimization of Address Element Knowledge Graph Based on Semantic-Spatial Similarity Computation With the continuous growth of the number of entities in address element knowledge graphs, the diverse forms of expression for a single address element often result in multiple nodes for the same address element, leading to redundancy in the graph structure and scattered entity relationships, which restricts the overall quality and application value of the graph. To address this issue, this invention proposes an address element fusion method based on semantic-spatial similarity calculation. This method quantifies the similarity between different address elements from two dimensions: semantic adjacency relationships and spatial contour similarity, identifies potential duplicate entities, and fuses them accordingly. First, a set of downstream adjacent entities is constructed based on the "located" relationship of address entities, and the semantic matching degree is evaluated by the similarity of the set. Second, a spatial contour is constructed by combining a set of address points with latitude and longitude information, and the geographical overlap between entities is calculated. Finally, a semantic-spatial collaborative constraint mechanism is introduced between similar address elements, using a minimum value strategy to synthesize the similarity results and formulate robust fusion judgment rules, thereby achieving accurate fusion of fine-grained address entities and improving the consistency and expressive power of the graph.

[0051] (1) Calculation of semantic similarity of address elements with similarity Semantic similarity between address elements is a key factor in entity fusion, especially in Chinese address corpora. The same address entity may have multiple referents due to name aliases, abbreviations, and different granularities of expression. Traditional string matching methods are difficult to handle such semantic variations.

[0052] In the Chinese address knowledge graph, the "location" relationship between address elements not only reflects the spatial hierarchy but also contains rich semantic information. Taking "Haidian District" and "Haidian" as examples, although the names are slightly different, they both point to the same address in reality, exhibiting a high degree of semantic consistency. Observing the names of address elements associated with these two entities, especially all the lower-level address elements "located" in them, reveals a high degree of redundancy in their adjacent entity sets. As new addresses are imported into the graph and the data volume increases, all address elements located in the entity "Haidian" are almost completely contained within all address elements located in the entity "Haidian District." This approximate inclusion relationship reflects the semantic similarity between the two entities.

[0053] Based on this, embodiments of the present invention introduce a set containing similarity to represent the semantic similarity between two address elements. The semantic representation of address elements is constructed based on the adjacency set formed by the "located" relationship between address element entities, and the semantic similarity between entities is calculated accordingly. Specifically, assuming there are two address element entities, Entity A and Entity B, to be compared in the address knowledge graph, their respective semantic adjacency sets are constructed using the following method: ① For A, collect all entities C that satisfy "C is located in A"; ② For each C, continue recursively collecting all cases that satisfy " Entity located in C” ; ③ Repeat this process continuously until no new entities are added; ④ Form set A from all collected entities; Similarly, construct the semantic adjacency set B of entity B. This process can be understood as starting from the address entity and recursively expanding downstream along the "located" relationship to capture all its more granular address element information. Then, the semantic similarity between entity element A and entity element B is: (Formula 4-1) (2) Calculation of spatial similarity of address elements based on spatial contour overlap In actual address information, although some address elements differ significantly in semantic representation, they should still be considered the same or similar entities because they point to the same or highly overlapping geographic spatial areas. Therefore, combining the spatial boundary information of address elements to assess the degree of overlap in their geographical locations is an important supplementary means to achieve entity fusion. To this end, this embodiment of the invention introduces spatial contour overlap based on set inclusion similarity to measure the similarity of address element entities in geographic space. For address element entities A and B, the specific calculation method is as follows: ① Calculate the semantic adjacency set A of A; ②Preserve the point set with latitude and longitude coordinates ; ③Construction minimum convex hull ; Similarly, we obtain the corresponding minimum convex hull of B. Then the spatial similarity between entity element A and entity element B is: (Formula 4-2) in, This represents the area of ​​a polygon, with the numerator being the area of ​​the overlapping region of the two convex hulls and the denominator being the smaller of the two areas. This definition follows the core idea of ​​semantic similarity of address elements, which measures the spatial consistency between two contours by the degree to which the smaller contour is approximately "contained" by the larger contour.

[0054] (3) Address element fusion under semantic-spatial co-constraints Since the address knowledge graph construction method in this embodiment relies on the classification results of address element sequences by GPT-TC, the final graph structure consists of high-granularity entity nodes. Therefore, the core challenge faced by the graph is the fusion of fine-grained address element entities. That is, multiple address elements with different expressions but actually referring to the same entity (such as "Sanlitun Subdistrict, Chaoyang District" or "Sanlitun Office") may be represented as multiple nodes in the graph, resulting in structural redundancy and information dispersion, which seriously affects the consistency of the graph and the effectiveness of downstream applications.

[0055] To address this issue, a semantic-spatial co-constraint address element fusion method (SSCC-Fusion) is proposed based on semantic and spatial similarity of address elements. This method employs a conservative fusion strategy to determine whether to fuse address element entities: ① In terms of fusion conditions, fusion is only performed when two address elements belong to the same address category (e.g., streets, buildings, etc.) to effectively reduce the risk of cross-type misfusion; ② In terms of fusion decision, fusion is only performed when address element entities have significant matching in both semantic and spatial similarity. The specific fusion determination process is as follows: Let the semantic similarity between address element entities A and B be Simsesem(A, B), and their spatial similarity be Simspatial(A, B). The final similarity between them is defined as follows: (Formula 4-3) when When address elements A and B are merged, the threshold τ can be set to different values ​​depending on the granularity of the address elements. The purpose of adopting the minimum similarity strategy is to ① emphasize high-confidence matching: as long as there is a high degree of matching in either semantic or spatial dimensions, they can be determined to be the same entity, avoiding false fusion caused by weak matching; ② enhance consistency constraints and tolerate some missing information: in actual address knowledge graphs, some entities may lack complete spatial coordinates or semantic context, leading to inaccurate or missing similarity calculations in a certain dimension. To ensure the reliability and conservatism of the fusion results, the SSCC-Fusion method uses a minimum similarity strategy, that is, taking the smaller value between semantic similarity and spatial similarity as the fusion basis. This strategy requires entities to have high similarity in both semantic and spatial dimensions to avoid false fusion caused by accidental similarity in only one dimension, thereby more effectively ensuring the consistency and accuracy of the merged entities. Although this strategy is relatively conservative and may sacrifice some fusion recall, it has significant advantages in application scenarios with high graph quality requirements. As the amount of map data gradually accumulates, the similarity of multiple address elements pointing to the same real entity in two dimensions will gradually converge. Through periodic map cleaning and fusion updates, a gradually improving fusion effect can still be achieved, enhancing the stability and consistency of the map representation.

[0056] Device Example 1 According to embodiments of the present invention, an automatic generation device for Chinese address element knowledge graphs based on generative large language models is provided. Figure 8 This is a schematic diagram of an automatic generation device for a Chinese address element knowledge graph of a generative large language model according to an embodiment of the present invention, such as... Figure 8 As shown, the automatic generation device for Chinese address element knowledge graph of generative large language model according to an embodiment of the present invention specifically includes: The segmentation module 80 is used to construct a segmentation framework for Chinese address elements based on the semantic features of the address elements, and to segment the complete address into a sequence of address elements based on the segmentation framework; specifically, it is used for: A Chinese address element segmentation framework is constructed based on the semantic features of address elements. The address information is segmented into candidate address elements through Chinese word segmentation. MacBert is used to embed the candidate address elements into word vectors with semantic features. A linear layer is used to map the word vector of each candidate address element to a score of BIO tags. Conditional random field (CRF) is used to model the dependency relationship between the BIO tags. Finally, the Viterbi algorithm is used to decode the optimal tag sequence, realizing the segmentation of Chinese addresses into address element sequences. Module 82 defines address elements based on national standards, and constructs an ontology framework for a hierarchical knowledge graph of Chinese address elements with semantic constraints based on these definitions; specifically, it is used for: The address element ontology is defined based on national standards, and the ontology class names and hierarchical structure are determined, specifically including: administrative division name, provincial division name, province, autonomous region, municipality, special administrative region, prefecture-level division name, prefecture-level city, region, autonomous prefecture, league, county-level division name, county, autonomous county, county-level city, banner, autonomous banner, municipal district, forest area, township / street-level division name, town, township, ethnic township, street, street, courtyard number, unit door number, floor, household room number, and point of interest; Define the attributes of the address element ontology and design the semantic relationships between address element ontologs. The attributes of the address element ontology include: unique identifier, name, longitude, latitude, alias, and object attribute "located". Based on the ontology class name and hierarchical structure, the attributes of the address element ontology, and the semantic relationships, a model of the ontology framework of the Chinese address element knowledge graph is constructed, and data types and constraints are set. Classification module 84 is used to perform text classification on the address element sequence based on the ontology framework and using a lightly supervised classification method based on GPT; specifically, it is used for: We adopt a lightly supervised classification method based on GPT, which utilizes the language understanding and generation capabilities of large language models to transform the address element classification task into a natural language generation task with context and examples. Through designed prompts and examples, we achieve automatic annotation of address element sequences, perform semantic classification on actual address corpora, identify the address element category of each word, and finally map the classification results into structured entities and relationships.

[0057] Module 86 is used to construct a Chinese address element knowledge graph based on the Neo4j graph database according to the ontology framework and the text classification; specifically, it is used for: Import the constructed address feature ontology file into the Neo4j graph database, initialize the class, object attribute, and data attribute structure of the graph, and generate a semantic framework consistent with the address feature ontology structure through the Neo4j graph database to provide constraints and type support for the subsequent insertion of address entities; Using address element ontology as nodes and structural "located" relationships between address elements as edges, entities and relationships in the Neo4j graph database are constructed. Each address data is represented as a top-down path structure in the Neo4j graph database. The path structure corresponds to the hierarchical structure of the address ontology, reflecting the spatial and semantic subordinate relationships between address element entities. Finally, a Chinese address element knowledge graph composed of nodes and relationship edges is obtained. Optimization module 88 is used to optimize synonymous entities in the Chinese address element knowledge graph using a semantic-spatial similarity calculation method. Specifically, it is used for: A semantic-spatial similarity calculation method is adopted. A set of downstream adjacent address entities is constructed based on the "located" relationship of address entities. The semantic matching degree is evaluated by the similarity of the set of downstream adjacent address entities. A spatial contour is constructed by combining a set of address points with latitude and longitude information, and the geographical overlap between address entities is calculated. A semantic-spatial collaborative constraint mechanism is introduced between similar address elements. The similarity results are integrated with a minimum value strategy. Based on the fusion judgment rule, the address entities are accurately fused.

[0058] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operation of each module can be understood with reference to the description of the method embodiments, and will not be repeated here.

[0059] Device Example 2 This invention provides an electronic device, such as... Figure 9 As shown, it includes: a memory 90, a processor 92, and a computer program stored in the memory 90 and executable on the processor 92, wherein the computer program, when executed by the processor 92, performs the steps as described in the method embodiment.

[0060] Device Example 3 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 92, performs the steps described in the method embodiment.

[0061] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.< / default> < / default>

Claims

1. A method for automatically generating a Chinese address element knowledge graph of a generative large language model, characterized in that, The method comprises the following steps: Based on the semantic characteristics of address elements, a segmentation framework of Chinese address elements is constructed, and based on the segmentation framework, a complete address is segmented into a sequence of address elements; Based on the definition of address elements according to national standards, an ontology framework of Chinese address element knowledge graph with hierarchical structure and semantic constraints is constructed based on the definition of address elements; Based on the ontology framework, a light supervision classification method based on GPT is used to classify the address element sequence; According to the ontology framework and the text classification, the Neo4j graph database is used to construct the Chinese address element knowledge graph; The semantic-spatial similarity calculation method is used to optimize the synonymous entities in the Chinese address element knowledge graph.

2. The method for automatically generating a Chinese address element knowledge graph of a generative large language model according to claim 1, characterized in that, Based on the semantic characteristics of address elements, a segmentation framework of Chinese address elements is constructed, and based on the segmentation framework, a complete address is segmented into a sequence of address elements; Based on the semantic characteristics of address elements, a segmentation framework of Chinese address elements is constructed, and based on the segmentation framework, a complete address is segmented into a sequence of address elements; 3. The method for automatically generating a Chinese address element knowledge graph of a generative large language model according to claim 1, characterized in that, Based on the definition of address elements according to national standards, an ontology framework of Chinese address element knowledge graph with hierarchical structure and semantic constraints is constructed based on the definition of address elements; Based on the definition of address elements according to national standards, an ontology framework of Chinese address element knowledge graph with hierarchical structure and semantic constraints is constructed based on the definition of address elements; Based on the ontology class name and hierarchical structure, the attributes of the address element ontology and the semantic relationship between the address element ontologies, the model of the ontology framework of the Chinese address element knowledge graph is constructed, and the data types and constraint conditions are set. Based on the ontology framework, a light supervision classification method based on GPT is used to classify the address element sequence, and the address information extraction is realized. The text classification of the address element sequence is realized, which comprises the following steps:

4. The method for automatically generating a Chinese address element knowledge graph of a generative large language model according to claim 1, characterized in that, Based on the ontology framework, a light supervision classification method based on GPT is used to classify the address element sequence, and the address information extraction is realized. The text classification of the address element sequence is realized, which comprises the following steps: ​ 5. The method for automatically generating a Chinese address element knowledge graph of a generative large language model according to claim 3, characterized in that, According to the ontology framework and the text classification, the Chinese address element knowledge graph construction based on the Neo4j graph database specifically comprises: The file of the constructed address element ontology is imported into the Neo4j graph database, the class, object attribute and data attribute structure of the graph are initialized, a semantic framework consistent with the address element ontology structure is generated through the Neo4j graph database, and constraint and type support are provided for subsequent insertion of address entities; The address element ontology is taken as a node, and a structural "location" relationship between address elements is taken as an edge, an entity and a relationship in the Neo4j graph database are constructed, each piece of address data is represented as a path structure from top to bottom in the Neo4j graph database, the path structure corresponds to the hierarchical structure of the address ontology, and the spatial and semantic subordinate relationships between address element entities are reflected, and finally a Chinese address element knowledge graph composed of nodes and relationship edges is obtained.

6. The method for automatically generating a Chinese address element knowledge graph of a generative large language model according to claim 3, characterized in that, The synonym and alias entities in the Chinese address element knowledge graph are optimized by using a semantic-spatial similarity calculation method, specifically comprising: The semantic-spatial similarity calculation method is used to construct a downstream adjacent address entity set based on the "location" relationship of the address entity, the matching degree in the semantic level is evaluated through the similarity of the downstream adjacent address entity set, a spatial contour is constructed in combination with the address point set with latitude and longitude information, the geographical overlap degree between address entities is calculated, a semantic-spatial collaborative constraint mechanism is introduced between address elements of the same type, the similarity results are integrated by using a minimum value strategy, and accurate fusion of address entities is performed based on a fusion judgment rule.

7. A Chinese address element knowledge graph automatic generation device of a generative large language model, characterized in that, It comprises: The segmentation module is configured to construct a segmentation framework of Chinese address elements based on semantic features of the address elements, and segment a complete address into an address element sequence based on the segmentation framework; The definition module is configured to define the address elements based on national standards, and construct an ontology framework of the Chinese address element knowledge graph with hierarchical structure and semantic constraints based on the definition of the address elements; The classification module is configured to perform text classification on the address element sequence based on the ontology framework and using a GPT-based light supervision classification method; The construction module is configured to construct the Chinese address element knowledge graph based on the Neo4j graph database according to the ontology framework and the text classification; The optimization module is configured to optimize synonym and alias entities in the Chinese address element knowledge graph by using a semantic-spatial similarity calculation method.

8. The Chinese address element knowledge graph automatic generation device of the generative large language model according to claim 7, wherein The segmentation module is specifically configured to: construct a segmentation framework of Chinese address elements based on semantic features of the address elements, segment address information into candidate address elements through Chinese word segmentation, embed the candidate address elements into word vectors with semantic features by using MacBert, map the word vectors of each candidate address element into scores of BIO labels through a linear layer, model the dependency relationship between the BIO labels by using a conditional random field (CRF), and finally decode an optimal label sequence by using a Viterbi algorithm to realize segmentation of Chinese addresses into address element sequences. The definition module is specifically used for: Defining the address element ontology based on national standards, determining the ontology class name and hierarchical structure, and specifically including: administrative division name, provincial division name, province, autonomous region, municipality directly under the central government, special administrative region, prefectural division name, prefecture-level city, district, autonomous prefecture, league, county-level division name, county, autonomous county, county-level city, banner, autonomous banner, municipal district, forest area, township / street-level division name, town, township, ethnic township, street, street, lane, courtyard number, unit door number, floor, room number, and point of interest; Defining the attributes of the address element ontology and designing the semantic relationship between the address element ontologies, wherein the attributes of the address element ontology include: unique identifier, name, longitude, latitude, alias, and object attribute "located in"; According to the ontology class name and hierarchical structure, the attributes of the address element ontology, and the semantic relationship, a model of the ontology framework of the Chinese address element knowledge graph is constructed, and data types and constraint conditions are set; The classification module is specifically used for: Using a GPT-based light supervision classification method, utilizing the language understanding and generation capabilities of a large language model, converting an address element classification task into a natural language generation task with context and examples, realizing automatic labeling of address element sequences through designed prompt words and examples, performing semantic classification on actual address corpus, identifying the address element categories of each word element, and finally mapping the classification results to structured entities and relationships; The construction module is specifically used for: Importing the constructed address element ontology file into a Neo4j graph database, initializing the class, object attribute, and data attribute structure of the graph, and generating a semantic framework consistent with the address element ontology structure through the Neo4j graph database to provide constraint and type support for subsequent address entity insertion; Taking the address element ontology as a node and the structural "located in" relationship between address elements as an edge, constructing entities and relationships in the Neo4j graph database, wherein each address data in the Neo4j graph database represents a path structure from top to bottom, the path structure corresponds to the hierarchical structure of the address ontology, and reflects the spatial and semantic subordinate relationship between address element entities, and finally a Chinese address element knowledge graph composed of nodes and relationship edges is obtained; The optimization module is specifically used for: Using a semantic-spatial similarity calculation method, constructing a downstream adjacent address entity set based on the "located in" relationship of address entities, evaluating the matching degree in the semantic level through the similarity of the downstream adjacent address entity set; constructing a spatial profile combined with the address point set with longitude and latitude information, calculating the geographical overlap degree between address entities; introducing a semantic-spatial collaborative constraint mechanism between address elements of the same type, comprehensively integrating the similarity results with a minimum value strategy, and performing accurate fusion of address entities based on a fusion judgment rule.

9. An electronic device, comprising: It includes: A memory, a processor, and a computer program stored on the memory and executable on the processor, which, when executed by the processor, implements the steps of the generative large language model Chinese address element knowledge graph automatic generation method of any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores an information transmission implementation program, and the program, when executed by the processor, implements the steps of the method for automatically generating a Chinese address element knowledge graph of a generative large language model according to any one of claims 1 to 6.

Images