Industrial map-oriented enterprise data identification method and corresponding product

Abstract

This application relates to the field of commercial data processing, providing a method and corresponding products for enterprise data identification in industry graphs. The method includes: asynchronously collecting multi-source, heterogeneous raw enterprise data from the internet and big data resource services, generating a standardized enterprise data set, and storing it in a cloud database; generating a basic entity set based on a pre-built industry knowledge ontology model, and labeling each entity with structured attribute information; analyzing the basic entity set using an enterprise relationship mining model to mine direct and implicit relationships between entities; fusing, deduplicating, and weighting the mined direct and implicit relationships, and dynamically constructing and updating an industry knowledge graph with entities as nodes and weighted relationships as edges; and responding to query requests for a target industry by intelligently retrieving and performing multi-dimensional analysis of the industry knowledge graph, outputting enterprise identification results that meet the criteria.

Description

Technical Field

[0001] This application relates to the field of business data processing, and in particular to a method and corresponding product for enterprise data identification oriented towards industry mapping. Background Technology

[0002] Industry maps, as a structured form of knowledge representation, can clearly depict entities such as enterprises, products, technologies, and capital within a specific industry and their complex relationships. They are of significant value for industry analysis, investment decisions, supply chain management, and risk warning. The core and prerequisite for constructing accurate, dynamic, and comprehensive industry maps lies in accurately identifying relevant enterprise entities from massive and heterogeneous data sources and clarifying the diverse relationships between them.

[0003] Currently, several technical solutions exist in the industry for enterprise data identification and correlation analysis. One type of solution relies primarily on predefined rule templates or keyword matching to extract enterprise information and directly stated relationships from structured or semi-structured data. This approach is highly efficient when processing well-formatted data, but struggles to uncover implicit, indirect, deep relationships from unstructured text (such as news articles and reports). Furthermore, the rule system suffers from high maintenance costs and insufficient flexibility and scalability. Another type of solution attempts to apply machine learning or simple deep learning models, such as training classification models to identify entity types or calculating text similarity to infer potential connections between enterprises. This approach improves the ability to extract information from complex text to some extent, but still faces challenges in handling multi-source heterogeneous data fusion, distinguishing relationship types, and achieving accuracy and interpretability in relational reasoning. Especially when facing rapidly changing industry dynamics and massive incremental data, existing solutions often have limitations in dynamic knowledge updates, the accuracy of cross-source information fusion, and intelligent analysis capabilities.

[0004] In summary, existing technologies in the fields of enterprise data identification and industry graph construction mainly suffer from the following shortcomings: 1) Insufficient depth in identification and relationship mining, making it difficult to effectively integrate rule clarity and semantic understanding capabilities to simultaneously cover explicit direct relationships and deep implicit relationships; 2) The constructed knowledge graphs are highly static, lacking an efficient and accurate dynamic update mechanism, making it difficult to respond to data changes and industry dynamics in real time; 3) The overall process's intelligence and automation level needs to be improved, especially in the fusion processing of multi-source heterogeneous data and complex intelligent analysis based on graphs, making it difficult to support efficient industry insights. Summary of the Invention

[0005] This application provides a method and corresponding product for enterprise data identification based on industry graphs. By integrating multi-source data, ontology recognition, intelligent mining and dynamic graph construction, it achieves comprehensive, accurate and dynamic identification and analysis of enterprises and their industry relationships.

[0006] On the one hand, this application provides a method for enterprise data identification based on industry graphs. The method is executed by software deployed on a cloud-based integrated application runtime support platform, and includes:

[0007] Step S1: Asynchronously collect multi-source heterogeneous enterprise raw data from the Internet and big data resource services, clean, integrate and standardize the enterprise raw data, generate a standardized enterprise dataset and store it in the cloud database;

[0008] Step S2: Based on the pre-built industry knowledge ontology model, perform entity recognition and attribute extraction on the enterprise data set to obtain a basic entity set including enterprise entities, product entities and technology entities, and label each entity with structured attribute information.

[0009] Step S3: Analyze the basic entity set using an enterprise relationship mining model to mine the direct and implicit relationships between entities; wherein, the enterprise relationship mining model is implemented by integrating a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks. The direct relationship identification module is used to identify the direct relationships between entities based on keyword templates related to supply chain, investment and cooperation. The implicit relationship prediction module is used to calculate the semantic association degree of entities in a high-dimensional feature space based on entity attributes and the identified direct relationships, so as to predict the implicit relationships between entities.

[0010] Step S4: The direct and implicit relationships mined in Step S3 are fused, redundancy is removed, and weights are calculated to dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges.

[0011] Step S5: In response to a query request targeting a specific industry, perform intelligent retrieval and multi-dimensional analysis on the industry knowledge graph, and output the enterprise identification results that meet the conditions.

[0012] On the other hand, this application provides an enterprise data identification device for industry maps, the device comprising:

[0013] The data acquisition module is used to asynchronously collect multi-source heterogeneous raw enterprise data from the Internet and big data resource services, clean, integrate and standardize the raw enterprise data, and generate a standardized enterprise dataset which is then stored in the cloud database.

[0014] The extraction module is used to perform entity recognition and attribute extraction on the enterprise data set based on a pre-built industry knowledge ontology model, to obtain a basic entity set including enterprise entities, product entities and technology entities, and to label each entity with structured attribute information.

[0015] The mining module is used to analyze the basic entity set using an enterprise relationship mining model to mine the direct and implicit relationships between entities. The enterprise relationship mining model is implemented by integrating a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks. The direct relationship identification module is used to identify the direct relationships between entities based on keyword templates related to supply chain, investment, and cooperation. The implicit relationship prediction module is used to calculate the semantic association degree of entities in a high-dimensional feature space based on entity attributes and the identified direct relationships, so as to predict the implicit relationships between entities.

[0016] The construction module is used to fuse, remove redundancy and calculate weights for the direct and implicit relationships, and dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges.

[0017] The analysis module is used to respond to query requests targeting a specific industry by performing intelligent retrieval and multi-dimensional analysis on the industry knowledge graph and output enterprise identification results that meet the conditions.

[0018] Thirdly, this application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the technical solution of the enterprise data identification method for industry mapping described above.

[0019] Fourthly, this application provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described enterprise data identification method for industry mapping.

[0020] As can be seen from the technical solution provided in this application, on the one hand, the enterprise relationship mining model, which integrates a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks, can not only efficiently and accurately identify explicitly stated direct relationships in text (e.g., investment, cooperation), but also effectively mine potential, non-directly stated implicit relationships (e.g., competition, potential upstream and downstream) through deep reasoning of entity associations at the semantic level using graph neural networks, thereby constructing an industry knowledge graph with more comprehensive relationship types and richer layers. On the other hand, when constructing the graph, the graph is generated by fusing direct and implicit relationships, removing redundancy, and calculating weights. More importantly, this method can dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges based on the aforementioned process. This means that when new data flows in, the system can trigger corresponding identification, mining, and calculation processes to incrementally update the graph, ensuring that the graph content keeps pace with changes in the data source and reflects the latest industry dynamics in a timely manner. This makes the analysis and decision-making based on the graph more timely and valuable. Thirdly, by responding to query requests targeting specific industries, the system performs intelligent retrieval and multi-dimensional analysis of the industry knowledge graph, outputting enterprise identification results that meet the criteria. This transforms the complex underlying knowledge graph construction results into upper-level applications for user queries. Intelligent retrieval allows users to naturally locate their points of interest, while multi-dimensional analysis deeply analyzes the industry landscape from multiple perspectives such as community structure, centrality, and relational paths. The final structured identification results are intuitive and clear, greatly enhancing the usability of the industry graph and providing users with powerful data-driven decision support. In summary, the technical solution of this application, by integrating multi-source data, ontology recognition, intelligent mining, and dynamic graph construction, achieves comprehensive, accurate, and dynamic identification and analysis of enterprises and their industry relationships. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or 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 of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the enterprise data identification method for industry maps provided in the embodiments of this application;

[0023] Figure 2 This is a schematic diagram of the structure of the enterprise data identification device for industry mapping provided in the embodiments of this application;

[0024] Figure 3This is a schematic diagram of the device provided in the embodiments of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] In this specification, adjectives such as "first" and "second" are used only to distinguish one element or action from another, without necessarily requiring or implying any actual such relationship or order. Where circumstances permit, reference to an element or component or step (etc.) should not be construed as being limited to only one of the elements, components, or steps, but may be one or more of the elements, components, or steps, etc.

[0027] For ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale.

[0028] Currently, several technical solutions exist in the industry for enterprise data identification and correlation analysis. One type of solution relies primarily on predefined rule templates or keyword matching to extract enterprise information and directly stated relationships from structured or semi-structured data. This approach is highly efficient when processing well-formatted data, but struggles to uncover implicit, indirect, deep relationships from unstructured text (such as news articles and reports). Furthermore, the rule system suffers from high maintenance costs and insufficient flexibility and scalability. Another type of solution attempts to apply machine learning or simple deep learning models, such as training classification models to identify entity types or calculating text similarity to infer potential connections between enterprises. This approach improves the ability to extract information from complex text to some extent, but still faces challenges in handling multi-source heterogeneous data fusion, distinguishing relationship types, and achieving accuracy and interpretability in relational reasoning. Especially when facing rapidly changing industry dynamics and massive incremental data, existing solutions often have limitations in dynamic knowledge updates, the accuracy of cross-source information fusion, and intelligent analysis capabilities. In summary, existing technologies in the fields of enterprise data identification and industry graph construction mainly suffer from the following shortcomings: 1) Insufficient depth in identification and relationship mining, making it difficult to effectively integrate rule clarity and semantic understanding capabilities to simultaneously cover explicit direct relationships and deep implicit relationships; 2) The constructed knowledge graphs are highly static, lacking an efficient and accurate dynamic update mechanism, making it difficult to respond to data changes and industry dynamics in real time; 3) The overall process's intelligence and automation level needs to be improved, especially in the fusion processing of multi-source heterogeneous data and complex intelligent analysis based on graphs, making it difficult to support efficient industry insights.

[0029] To address the aforementioned problems in existing technologies, this application proposes a method for enterprise data identification based on industry maps, which can be executed by software deployed on a cloud-based integrated application runtime support platform. Its flowchart is attached. Figure 1 As shown, it mainly includes steps S1 to S5, which are detailed below:

[0030] Step S1: Asynchronously collect multi-source heterogeneous enterprise raw data from the Internet and big data resource services, clean, integrate and standardize the enterprise raw data, generate standardized enterprise datasets and store them in the cloud database.

[0031] Traditional data collection often employs sequential, synchronous methods, which are inefficient and struggle to guarantee the timeliness of data when faced with massive and dispersed internet data sources. Furthermore, data obtained from different sources (e.g., government public platforms, news websites, corporate websites, API interfaces) exhibits significant differences in format, structure, and encoding; directly merging and using these data can lead to errors in subsequent analysis.

[0032] To this end, this application designs an efficient and intelligent data acquisition and preprocessing pipeline. Specifically, the asynchronous acquisition of multi-source heterogeneous enterprise raw data from the Internet and big data resource services in the above embodiments can be achieved through steps S1.1 to S1.3, as detailed below:

[0033] Step S1.1: Configure a distributed asynchronous crawler framework to crawl initial data in parallel from designated public bidding websites, enterprise credit information disclosure platforms, and industry news websites.

[0034] Traditional web crawlers are mostly single-threaded synchronous operations, and their crawling efficiency is limited by network latency and website anti-crawling policies. This application adopts a distributed asynchronous crawler framework, such as an architecture based on Scrapy-Redis or Celery. "Distributed" means that multiple crawler nodes can simultaneously crawl data from different target websites or different pages of the same website, significantly improving throughput. "Asynchronous" means that after initiating a network request, the crawler can continue processing other tasks or initiating new requests without waiting for a response, thus making full use of network I / O latency and avoiding CPU idle time. By configuring targeted parsing rules (XPath or CSS selectors) for different types of websites (e.g., government information disclosure platforms, news portals), this framework can extract raw data such as company names, registered capital, shareholder information, bidding records, and news events in a structured manner from key data sources such as public bidding websites, enterprise credit information disclosure platforms, and industry news websites.

[0035] Step S1.2: Listen for and capture the data stream from the enterprise's public API interface, and align the data stream with the initial crawled data using timestamps and source tags.

[0036] In addition to web page data, many platforms (such as the open platforms of Tianyancha and Qichacha) or the companies themselves provide API interfaces that return data in more standardized JSON or XML formats. Listening to and capturing data streams from publicly available enterprise API interfaces means the system will establish a continuous data stream monitoring service, calling these APIs at a preset frequency. Since API data and web crawler data may describe the same event but arrive at different times, "timestamp alignment" is crucial. The system will attach two timestamps to each data record: the data generation time (e.g., news release time, business registration change date) and the data acquisition time. For multiple records describing the same entity and the same event, the system will align them based on the generation time of the event in the most reliable data source. Simultaneously, each data record will be "source-tagged," recording its original origin (e.g., "Source A - Bidding website," "Source B - Enterprise credit API"), providing a basis for subsequent data credibility assessment and weight calculation.

[0037] Step S1.3: Based on the preset data source priority and update frequency strategy, dynamically schedule the crawler framework and API listening process to achieve incremental collection and real-time updates of multi-source data.

[0038] To balance data freshness and system load, this application introduces an intelligent scheduling mechanism. Preset data source priorities and update frequency strategies define the importance and update characteristics of different data sources. For example, business registration change information sources might be set to "high priority, daily updates," while industry news sources might be set to "medium priority, hourly updates." The scheduler dynamically schedules the crawler framework and API listening processes according to this strategy. For example, during off-peak periods, high-priority but low-frequency deep crawling tasks are centrally scheduled; during peak periods, high-frequency API listening and incremental crawling are primarily scheduled. Incremental collection of multi-source data refers to the crawler or listener being able to identify data that has been added or changed since the last collection, and only crawling this part of the data, greatly saving bandwidth and computing resources. Real-time updates of multi-source data are reflected in the listening of high-priority, high-frequency data sources (e.g., breaking news, stock price streams). The system can capture and send these data into subsequent processes almost in real time, ensuring the timeliness of the graph.

[0039] Through steps S1.1 to S1.3, a highly efficient, real-time, and schedulable multi-source asynchronous data acquisition system is essentially constructed. In other words, it doesn't simply describe "acquiring data from multiple sources," but rather specifically forms a parallel, incremental, and responsive priority-based intelligent acquisition closed loop through three steps: "configuring a distributed asynchronous crawler framework," "monitoring API data streams," and "dynamic scheduling." This solves the balance problem between the timeliness and completeness of data acquisition and system resource consumption in the context of massive and ever-changing internet data, providing a high-quality data stream guarantee for subsequent processing.

[0040] As one embodiment of this application, the cleaning, integration, and standardization of raw enterprise data to generate a standardized enterprise dataset and its merging and storage in a cloud database can be achieved through the following steps S1.4 to S1.6:

[0041] Step S1.4: Using a joint deduplication module based on regular expressions and named entity recognition models, duplicate records pointing to the same entity are merged from the original enterprise data from different sources.

[0042] Since the same company may be mentioned multiple times in different data sources, or the same event may be reported by different media, deduplication is crucial. This application employs a joint deduplication module based on regular expressions and a named entity recognition model. Regular expressions are used to process data with fixed patterns, such as unified social credit codes and phone numbers, enabling precise matching. For text such as company names and product names, a named entity recognition model fine-tuned on a business corpus is used for identification and standardization. For example, "TX Technology" and "Shenzhen TX Computer Systems Co., Ltd." are identified as the entity "TX" using the NER model. The deduplication strategy is based on composite keys: for example, when the semantic similarity between the "standardized company name" and "core event description" of two records exceeds a threshold, they are considered duplicates. During merging, the most complete record fields with the highest source authority are retained, and supplementary information from other records is integrated.

[0043] Step S1.5: Use the business rule base and anomaly detection algorithm to correct and complete the deduplicated data records, in order to correct obviously erroneous fields and fill in the missing values ​​of key fields.

[0044] Data often contains errors (e.g., incorrect registered capital unit) and omissions (e.g., an empty business scope). This application utilizes a "business rule base" for error correction. For example, the rule base might define "registered capital should be less than RMB 1 trillion," and records exceeding this value will be flagged and verified. Simultaneously, "anomaly detection algorithms," such as cluster-based outlier detection, are used to identify outliers in numerical fields (e.g., revenue) or categorical fields. For missing key fields, multiple methods are employed to complete them: for categorical fields, the mode of other companies within the same industry is used; for numerical fields, the industry average is used, or a regression model is built to predict based on other fields.

[0045] Step S1.6: Convert the format and map the fields of the corrected and completed data records according to the predefined global data pattern, and store them uniformly as a standardized enterprise data set with the same field structure.

[0046] In this embodiment, the predefined global data schema is a centralized schema definition that specifies all fields and their data types and constraints (e.g., fields: company_name (string), registered_capital (floating-point number, unit: RMB 10,000), business_scope (list of strings)) that the final standardized enterprise data set should contain. Step S1.6 performs "format conversion and field mapping" on data records from different sources with different formats according to this global schema. For example, "registered capital: 50 million" from source A and "RegCapital: 50 million" from source B are both converted to registered_capital: 5000.0. Finally, all data is "uniformly stored" in a "cloud database" (e.g., HBase, MongoDB, or a cloud-native distributed database), forming a clean, consistent, high-quality data foundation that can be used for subsequent in-depth analysis.

[0047] Step S2: Based on the pre-built industry knowledge ontology model, perform entity recognition and attribute extraction on the enterprise data set to obtain a basic entity set including enterprise entities, product entities, and technology entities, and label each entity with structured attribute information.

[0048] After obtaining standardized data, it is necessary to extract knowledge units—entities and their attributes—from it. Traditional entity recognition often uses general domain models, which have low recognition rates for industry-specific terms (e.g., specific technical terms, product models), and the extracted attributes are scattered and lack correlation. To address this issue, this application introduces a "pre-built industry knowledge ontology model." An industry knowledge ontology is a formalized and clearly defined conceptual model that defines important concepts within a specific industry domain (e.g., Company, Product, Technology), the attributes of these concepts (e.g., Company has revenue, employee count), and the relationships between concepts (e.g., Company produces Product). By encoding domain knowledge into the ontology, explicit semantic guidance is provided for information extraction.

[0049] Specifically, step S2 can be implemented through steps S2.1 to S2.3, as detailed below:

[0050] Step S2.1: Input the unstructured text description field from the standardized enterprise data set into the multi-model voting recognition module, which consists of the BERT model, the BiLSTM-CRF model, and the dictionary matching model based on industry knowledge ontology.

[0051] For unstructured text (such as company profiles and news content), this application employs a multi-model voting recognition module to ensure robustness and high recall in entity recognition. The BERT model, BiLSTM-CRF model, and dictionary matching model based on industry knowledge ontology each have their own strengths, which are detailed below:

[0052] 1) BERT model: Utilizing its powerful deep semantic understanding capabilities, it can accurately determine the entity type of words based on the context, and is particularly good at handling polysemy and long-distance dependencies.

[0053] 2) BiLSTM-CRF model: It is very effective in sequence labeling tasks and can capture the sequence features of text and the constraints between labels very well.

[0054] 3) Dictionary matching model based on industry knowledge ontology: Construct a dictionary of all entity concepts defined in the ontology (such as company directory, product database, technical terminology table) for fast and accurate matching, ensuring that no known entities in the domain are missed.

[0055] These three models process the same text in parallel and output lists of the entities and types they have identified.

[0056] Step S2.2: The multi-model voting identification module outputs the entities and types identified by each model, and a weighted voting strategy is used to fuse and resolve conflicts in the identification results to obtain a preliminary entity list.

[0057] The weighted voting strategy used in the above embodiments for fusing and resolving conflicts in the recognition results can be as follows: Dynamic weights are assigned to the BERT model, BiLSTM-CRF model, and dictionary matching model, respectively. These dynamic weights are periodically adjusted based on the recognition accuracy of each model on historical source data. When at least two models identify the same entity type for the same text segment, the entity and its type are added to the basic entity set. When there is a conflict in the entity types identified by the models, the entity type output by the model with the highest current dynamic weight is selected as the final result, and this conflict case is recorded in the training sample library for subsequent reinforcement training of the low-weight model. Specifically, each of the three models is assigned a dynamic weight, which is periodically adjusted based on the recognition accuracy of each model on historical source data. For example, in a quarter's financial news data, the BERT model performs better, and its weight is increased. The fusion rule is: for a text segment, when at least two models identify the same entity type for the same text segment, the entity and its type are added to the basic entity set. This fusion rule ensures the reliability of the results. When conflicting entity types are identified by different models, the entity type output by the model with the highest current dynamic weight is selected as the final result. This mechanism respects the judgment of the most authoritative model under a specific data distribution. All conflict cases are recorded and used for subsequent reinforcement training of the low-weight model, forming a closed loop for continuous model optimization.

[0058] By introducing an optimization mechanism of dynamic weight adjustment and conflict resolution, multi-model voting is no longer a one-time static rule, but an adaptive learning system that can periodically adjust itself based on the "recognition accuracy on historical source data". This mechanism can not only optimize the current recognition effect, but also transform conflict cases into training samples for subsequent model enhancement, forming a positive feedback loop from recognition to model iteration, giving entity recognition capabilities the potential for continuous evolution.

[0059] Step S2.3: Based on the attribute constraint rules defined in the industry knowledge ontology model, parse the context information associated with each entity in the preliminary entity list, extract and populate it to form structured attribute information.

[0060] After identifying entities, their attributes need to be extracted. Attribute constraint rules are defined in the ontology. For example, for the Company entity, the rule might specify that the foundingDate attribute should be extracted from the context containing patterns such as "founded in" and "registration date," and converted to YYYY-MM-DD format. The system parses the context (sentences or paragraphs) surrounding the entity and extracts values ​​according to these rules using dependency parsing, pattern matching, or pre-trained information extraction models. For example, from the text "Company A was founded on April 1, 1976 by Shi B et al.", the attribute foundingDate:1976-04-01 can be extracted for the entity "Company A". Ultimately, a set of key-value pairs of "structured attribute information" is formed for each entity, and together with the entity, constitutes a high-quality "base entity set," preparing it for relation mining.

[0061] As can be seen from step S2 and the specific implementation steps S2.1 to S2.3 in the example above, by integrating entity recognition and attribute extraction with multi-model and knowledge ontology voting, the cutting-edge deep learning model (BERT, BiLSTM-CRF) is combined with the traditional method based on domain knowledge (ontology dictionary), and the decision is made through a weighted voting strategy. This method is not a simple stacking of models, but achieves the complementary advantages of "data-driven" and "knowledge-driven" through the voting mechanism, which significantly improves the accuracy and robustness of entity recognition and attribute extraction in unstructured text, especially in professional domain contexts, and overcomes the deficiency of single model's insufficient generalization ability in specific scenarios.

[0062] Step S3: Analyze the basic entity set using the enterprise relationship mining model to uncover the direct and implicit relationships between entities. The enterprise relationship mining model is implemented by integrating a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks. The direct relationship identification module is used to identify the direct relationships between entities based on keyword templates related to supply chain, investment, and cooperation. The implicit relationship prediction module is used to calculate the semantic association degree of entities in the high-dimensional feature space based on entity attributes and the identified direct relationships, so as to predict the implicit relationships between entities.

[0063] After obtaining a high-quality basic entity set, the core challenge lies in mining the rich relationship network between entities. Traditional methods either rely solely on manual rules, which are insufficient to cover complex and ever-changing expressions and implicit relationships; or they rely solely on statistical models, which lack interpretability and are not precise enough in capturing explicit relationships. To address this, this application introduces a fusion architecture, utilizing an enterprise relationship mining model to analyze the basic entity set and mine direct and implicit relationships between entities. Specifically, a direct relationship identification module identifies direct relationships between entities based on keyword templates related to supply chain, investment, and cooperation, while an implicit relationship prediction module calculates the semantic correlation degree of entities in a high-dimensional feature space based on entity attributes and the identified direct relationships, thereby predicting implicit relationships between entities. By fusing the predefined rule-based direct relationship identification module with the graph neural network-based implicit relationship prediction module, this approach creatively combines the determinism of the "white box" with the inferential capabilities of the "black box": the rule module efficiently, accurately, and interpretably captures explicit relationships, while the graph neural network module deeply mines implicit relationships by calculating semantic correlation degree in a high-dimensional feature space. This integrated architecture achieves a complementarity between breadth and depth, ensuring high-precision capture of explicit relationships while expanding the ability to discover potential complex associations, significantly improving the comprehensiveness and intelligence of relationship mining.

[0064] Specifically, the implementation of step S3 includes two closely connected modules (i.e., the direct relationship identification module and the implicit relationship prediction module) and an optimization process, which are described in detail below.

[0065] The direct relationship identification module aims to efficiently and accurately identify explicitly stated relationships from text. Its core is a scalable rule knowledge base. The specific process includes steps S3.1a to S3.1c:

[0066] Step S3.1a: Construct a rules knowledge base that includes supply chain keywords, investment relationship terms, and cooperation agreement templates.

[0067] The rule knowledge base in this application embodiment is not only a list of keywords, but each template is a structured object. For example, an "investment" relationship template may contain: a list of trigger words ("investment", "capital injection", "equity participation"), a core syntactic pattern ( <subj>right <obj>Perform [trigger words]), and relationship role mapping ( <subj>Mapped to investors, <obj>(Mapped to the investee). The template library is organized by relationship type (supply chain, investment, cooperation, etc.).

[0068] Step S3.1b: Perform dependency parsing on text fragments containing entities from at least two basic entity sets to extract relational phrases.

[0069] The specific implementation of step S3.1b is as follows: For a sentence, a dependency parser (such as Stanford CoreNLP, HanLP, or LTP) is first used to parse it, obtaining a syntax tree rooted at the verb. The system locates all entity mentions in the sentence that belong to the "basic entity set". For any sentence containing at least two entities, the algorithm finds the core predicate (usually a verb or preposition) on the shortest dependency path connecting these entities. Then, the extracted relation phrase is a substructure starting from this predicate, containing key modifiers (such as negation words, modal verbs) and the grammatical relations connecting the entities. For example, in the sentence "Company A announced a strategic investment of hundreds of millions of yuan in Company B", dependency parsing will identify the core verb "investment", its subject is "Company A", its object is "Company B", and it carries the adverbial "strategy". The extracted relation phrase can be represented as a triple (Company A, strategic investment, Company B) and its dependency path.

[0070] Step S3.1c: Match the extracted relational phrases with the templates in the rule knowledge base.

[0071] This is a multi-layered matching process: 1) Lexical trigger matching: Checking whether the core predicate in the relational phrase appears in the trigger word list of any template; 2) Structural matching: Comparing the dependency path of the relational phrase with the syntactic patterns defined in the template, calculating graph edit distance or rule-based transformation, and determining structural consistency; 3) Semantic matching: Using word vectors or a sentence encoder to calculate the semantic similarity between the trigger words and template examples. Successful matching is typically determined based on a comprehensive scoring function.

[0072]

[0073] in, It is an indicator function, if the trigger word in it... In the template, then The result is 1 if the value is 1, otherwise it is 0. It is the structural distance. It is semantic similarity. , and This is a weighting coefficient. If a match is successful, meaning the overall score exceeds a preset threshold, a direct relationship edge with a clear relationship type is created between the two entities based on the matched template type. The system records the relationship type, direction (determined by the role mapping in the template), evidence text, and match confidence level.

[0074] The implicit relationship prediction module based on graph neural networks for deep reasoning is used to discover potential associations in text that are not explicitly stated but can be inferred from entity attributes and existing relationship networks. Its theoretical basis is mapping entities and relationships to a continuous high-dimensional feature space, where semantic associations are manifested as geometric proximity between vectors. Based on entity attributes and identified direct relationships, the semantic association degree of entities in the high-dimensional feature space is calculated to predict implicit relationships between entities, including steps S3.2a to S3.2c:

[0075] Step S3.2a: Combine the entities, entity attributes, and identified direct relationships in the basic entity set to construct an initial heterogeneous graph G=(V, E, R).

[0076] Here, the node set V represents an entity. Each node... It has a feature vector This vector is obtained by encoding its entity attributes (e.g., industry classification, regional code, logarithmic scale of registered capital, etc.) using a learnable multilayer perceptron (MLP). The edge set E represents the direct relationships identified by S3.1, and each edge... Associate a relation type ( It is a predefined set of relation types.

[0077] Step S3.2b: Use a graph convolutional network (GCN) or graph attention network (GAT) model to perform multiple rounds of message passing and feature aggregation on the initial heterogeneous graph to learn the deep vector representation of each entity.

[0078] Taking GAT as an example, at layer l, the representation update of node i is achieved by performing attention-weighted aggregation on the representations of its neighbor node j:

[0079]

[0080] in, It is a neighbor of node i. It is a learnable weight matrix. It's an activation function. Attention weights. The attention mechanism learns the importance of neighbor node j to node i. After L layers of message passing, each entity obtains its final depth vector representation. It integrates rich structural and attribute information of the L-hop neighbors.

[0081] Step S3.2c: Calculate the cosine similarity or neural network-based association score between any two entity vector representations. Entity pairs with scores exceeding a preset threshold are predicted to have a latent relationship, and the predicted relationship type and confidence level are labeled for them.

[0082] Specifically, for entity pair (i,j), its association score can be cosine similarity. Simultaneously, to predict relation types, a separate relation classifier can be trained, which uses vector pairs... Or their splicing [ ; [Input] is [value], output is the probability distribution of each type of relation. The final determination of implicit relations is: when the association score... > ( (where the threshold is used), and the prediction probability of the relation classifier for a certain relation type t. > If, then, there is a hidden relationship of type t between entities i and j, with a confidence level of t. .

[0083] The implicit relationship prediction module in the above embodiment constructs an initial heterogeneous graph by combining entities, attributes, and known relationships. It then uses advanced models such as GCN / GAT to perform deep feature learning and semantic correlation calculation. From a technical perspective, it realizes the calculable and quantifiable definition and mining of the abstract concept of implicit relationship. This enables the system to discover deep industry connections in the text that are not explicitly stated but are implied through complex network structures, such as potential competition and technology-derived relationships, greatly expanding the depth and breadth of relationship mining.

[0084] To ensure the model can adapt to changes in data distribution and continuously improve, the enterprise relationship mining model in the above embodiment also includes a relationship mining optimization process, detailed in steps S3.3a to S3.3c below:

[0085] Step S3.3a: Periodically use the industry knowledge graph constructed in step S4 as new training data to incrementally train the graph neural network model in the implicit relationship prediction module.

[0086] Specifically, high-weight, multi-source verified relation triples are selected from the graph as positive samples, and negative samples are generated by using "triple destruction" techniques (such as randomly replacing the head and tail entities). These negative samples are then added to the model's training set for incremental training, enabling the model's embedding representation and relation prediction capabilities to keep pace with the times.

[0087] Step S3.3b: The relationship identification cases generated by the direct relationship identification module on the new data and whose confidence level is higher than the set threshold are converted into new rule templates after manual review and added to the rule knowledge base.

[0088] For example, when the system frequently identifies a cooperative relationship with high confidence from the sentence "A and B reach a joint research and development agreement", but the existing "cooperation" template does not cover the triggering pattern of "joint research and development", after review, "joint research and development" and its common syntax patterns will be abstracted into a new sub-template and incorporated into the "cooperation" relationship template library to achieve self-expansion of the rules.

[0089] Step S3.3c: Based on the incrementally trained graph neural network model and the updated rule knowledge base, re-perform relation mining on the full set of basic entities to iteratively optimize the accuracy and completeness of relations in the industry knowledge graph.

[0090] This process can be triggered at fixed intervals (e.g., monthly) or when rule / model updates accumulate to a certain amount, ensuring that the knowledge quality of the graph continues to improve.

[0091] Step S4: The direct and implicit relationships mined in Step S3 are merged, redundancy is removed, and weights are calculated. Based on this, an industry knowledge graph with entities as nodes and weighted relationships as edges is dynamically constructed and updated.

[0092] Through steps S3.3a to S3.3c in the example above, the industry knowledge graph constructed in step S4 is used as new training data to feed back into the model in step S3, and the rule knowledge base is updated simultaneously, thus forming a complete iterative optimization closed loop of "mining -> building -> training -> re-mining". This design makes the entire system no longer an open-loop pipeline, but an intelligent system that can continuously improve and evolve with data accumulation and use, effectively solving the problem that the relationship mining model and rule base may "age" over time, and ensuring long-term effectiveness.

[0093] Step S4: The direct and implicit relationships mined in Step S3 are merged, redundancy is removed, and weights are calculated. Based on this, an industry knowledge graph with entities as nodes and weighted relationships as edges is dynamically constructed and updated.

[0094] Regarding the industry's concern about how to integrate the massive amounts of potentially conflicting or redundant relationships mined into a unified, clean, measurably important, and time-evolving knowledge graph, rather than a static, flat data snapshot, to truly reflect industry dynamics and support accurate analysis, existing technologies either employ static graph construction schemes (i.e., directly constructing the graph from the results of a single mining operation, requiring a full reconstruction for subsequent updates) or unweighted or simply weighted graph schemes (i.e. ignoring differences in relationship strength and reliability, or using only a single indicator (such as co-occurrence frequency) for simple weighting). However, the former cannot adapt to continuous data changes, quickly becomes outdated, fails to reflect real-time industry dynamics, and its practical value diminishes over time. The latter cannot distinguish the importance and reliability of different relationships, and when performing retrieval, recommendation, or analysis based on the graph, it cannot effectively sort and filter the results, reducing the accuracy of analysis and the effectiveness of decision support. To address these issues in the existing technology, the technical solution adopted in this application is to fuse, remove redundancy from, and calculate weights of the direct and implicit relationships mined in step S3, and thereby dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges.

[0095] Specifically, fusing direct and implicit relationships between entities can involve merging the "direct relationships" from the rule module and the "implicit relationships" from the graph neural network module into a unified set of relationships. In the middle, calculating the weights of direct and implicit relationships between entities is key to giving the graph "intelligence," which specifically includes the following steps S4.1a to S4.1c:

[0096] Step S4.1a: Construct a weight evaluation vector v for each relation. The dimensions of the weight evaluation vector include: the authority score of the evidence source. Data freshness score The frequency of the relationship, n, and the association confidence, c, calculated by the implicit relationship prediction module.

[0097] in, Assign values ​​based on the authority level table of the data source (e.g., government websites = 1.0, mainstream media = 0.8, general websites = 0.5); Decays over time, for example ,in, It is the number of days. is the attenuation coefficient, n is the normalized frequency; the association confidence level c comes from step S3.2.

[0098] Step S4.1b: Input the weight evaluation vector v into a pre-trained multilayer perceptron (MLP) weight calculation model, and the MLP weight calculation model outputs a comprehensive weight value w.

[0099] The structure of this MLP model can be: ,in, , , and The parameters are used to train the system by learning manually labeled relational importance tags (e.g., important, average, or unimportant) on historical data.

[0100] Step S4.1c: Use the comprehensive weight value w as the final weight of the relation edge, which will be used for sorting and filtering in subsequent graph analysis and retrieval.

[0101] To address the industry challenges of high costs and poor timeliness in updating large-scale knowledge graphs, as an embodiment of this application, the dynamic construction and updating of an industry knowledge graph with entities as nodes and weighted relationships as edges can be achieved through steps S4.2a to S4.2c, as detailed below.

[0102] Step S4.2a: Calculate the semantic similarity between the newly mined relations and the existing relations in the graph, perform redundancy removal and merging on relations with similarity exceeding the merging threshold, and update the weights and evidence sources for the merged relations.

[0103] The specific algorithm is as follows: for relation sets... Each new relationship (Entities involved) , (and type t), find all connections in the existing graph. and And a set of relations of type t Then, calculate. and The semantic similarity between the evidence text descriptions of each relationship is calculated (using models such as Sentence-BERT to obtain text vectors and calculate cosine similarity). If the maximum similarity exceeds a threshold... (For example, =0.9), then consider It is redundant, so it is merged with the most similar existing relationship. The merging operation includes: merging the evidence text lists, adding the frequencies, and recalculating the weights.

[0104] Step S4.2b: Establish an event-driven mechanism. When data updates are detected in a specific high-priority data source in step S1, the local re-identification of affected entities, relationship mining, and graph subgraph updates are automatically triggered.

[0105] Specifically, taking the enterprise entities directly involved in the data update as the central nodes, and based on the existing topology of the industry knowledge graph, all related entities within their N-hop neighborhoods are identified, forming a set of subgraphs to be updated. Then, only entities within the set of subgraphs to be updated are re-executed using the entity identification, attribute extraction, and relationship mining processes described in the previous example. This is an efficient local computation. Finally, the newly mined subgraph is compared with the original graph, and the differences are updated atomically using a transaction mechanism to ensure data consistency. The difference comparison includes newly added, disappeared, and weighted relationships and attributes.

[0106] Step S4.2c: Create a version snapshot for each update of the map and provide map backtracking and trend analysis functions based on the time dimension.

[0107] Each time an update transaction is committed, the system creates a version number (e.g., a timestamp) for the entire graph or the affected subgraph and stores the differences. This allows users to query the graph status at any historical point in time or analyze the changing trends of specific entities or relationships over time.

[0108] As can be seen from steps S4.2a to S4.2c in the above example, the event-driven mechanism responds to changes in the data source and achieves partial re-identification and updating, rather than full reconstruction. At the same time, the introduction of version snapshot and change trend analysis functions enables the map to not only reflect the current state, but also trace historical evolution, providing time-dimensional insights for industry dynamic analysis, and upgrading from static snapshots to dynamically evolving digital twins.

[0109] Step S5: In response to a query request targeting a specific industry, perform intelligent retrieval and multi-dimensional analysis of the industry knowledge graph, and output the enterprise identification results that meet the criteria.

[0110] The completed dynamic industry knowledge graph is a valuable data treasure trove, and step S5 is key to transforming it into actionable insights. Its goal is to translate users' high-level business intentions (e.g., "identify potential technology partners in the new energy vehicle battery field") into computations that can be performed on the graph, outputting intuitive and in-depth analytical results. This overcomes the limitations of traditional enterprise query tools, which often result in flat outcomes and lack of correlation analysis.

[0111] The goal of intelligent retrieval of industry knowledge graphs is to accurately map users' natural and fuzzy query intentions to structured queries on the knowledge graph. The process mainly includes steps S5.1a to S5.1c, which are detailed below:

[0112] Step S5.1a: Provide a natural language query interface to receive the user's query request text.

[0113] For example, a user might enter, "Help me find startups that are potential competitors of Company A in the field of artificial intelligence chip design."

[0114] Step S5.1b: Perform intent recognition and entity linking on the query request text, and parse it into one or more structured query statements that can be executed on the graph.

[0115] Step S5.1b is a core technical step, and the intent recognition, entity linking, and query parsing involved are explained below:

[0116] 1) Intent identification: This involves using a text classification model finely tuned to an industry consulting corpus (such as one based on BERT) to identify the core intent category of a query, such as "finding competitors," "finding suppliers," "finding investment targets," or "analyzing technology paths." For example, the intent identified from the above query could be "discovering competitive relationships."

[0117] 2) Entity linking primarily involves entity recognition of the query text (reusing the model from step S2) and linking the identified entity mentions (e.g., "Company A", "AI chip design") with entity nodes in the knowledge graph. This requires calculating the semantic similarity between the mention text and candidate entity names, aliases, and descriptions, and considering context disambiguation. After successful linking, "Company A" corresponds to a specific company node in the graph, while "AI chip design" may correspond to a technology node or a concept node.

[0118] 3) Query parsing, which mainly refers to the system calling a predefined "query template" to generate a graph query statement based on the identified intent and linked entities. For example, for the intent of "competitive relationship discovery" and the linked entity (company A, technology field T), the template may generate a Cypher (graph query language) statement: find entities that are in the same technology field T as "company A", and that do not have a direct cooperative relationship between the two, but have similar semantic descriptions of products or technologies, and sort them by centrality.

[0119] Step S5.1c: Execute a structured query statement on the industry knowledge graph, use the graph index and the weight of the relation edges to accelerate the retrieval and sort the results, and return a preliminary retrieval result set.

[0120] Graph databases utilize indexes on nodes and relation types to quickly locate the starting node and traverse paths that meet the conditions. The sorting of results considers not only the matching degree of the query conditions but also relation weights (calculated in step S4). For example, results connected to the queried entity through high-weight relations are returned first, thereby improving the relevance and importance of the results.

[0121] For the initial search result set, or a user-specified entity / set of entities, in-depth analysis from a network science perspective is required to reveal its structural characteristics, i.e., multi-dimensional analysis of the industry knowledge graph, which goes beyond simple neighbor queries. As an embodiment of this application, multi-dimensional analysis of the industry knowledge graph can be achieved through the following steps S5.2a to S5.2c:

[0122] Step S5.2a: For the graph subgraphs corresponding to the preliminary retrieval result set, execute the community detection algorithm to identify enterprise clusters with tight internal connections and sparse external connections, and mark each cluster as a potential strategic alliance or competitive group.

[0123] Commonly used algorithms, such as the Louvain or Leiden algorithms, divide communities through modularity optimization. For example, in the "semiconductor industry" subgraph, the algorithm might automatically identify a cluster centered on "manufacturing" and a cluster centered on "chip design." The system calculates characteristics for each community (such as dominant industry and average company size) and labels them as "potential ecosystem alliances" or "niche competitive markets," providing users with a macro-level perspective on the industry landscape.

[0124] Step S5.2b: On the graph subgraph, calculate the centrality index of key entity nodes and rank entities according to the centrality index to identify core enterprises or key technology holders in the supply chain.

[0125] The aforementioned centrality metrics are key to quantifying the importance of nodes in a network. This application calculates various centrality metrics, including degree centrality, betweenness centrality, and eigenvector centrality:

[0126] 1) Degree centrality: This refers to the number of direct connections, reflecting the direct influence of a firm.

[0127] 2) Betweenness centrality: This metric measures how frequently a node lies on the shortest path between other node pairs, reflecting its control as a "hub" or "bridge". ,in, It is the total number of shortest paths from node s to t. It is the number of shortest paths through v. High betweenness numbers indicate that a firm may be a critical link in the supply chain.

[0128] 3) Eigenvector centrality: The eigenvector centrality index considers not only the number of neighbors, but also the importance of the neighbors, reflecting the degree of embedding of the enterprise in the "core circle".

[0129] By combining these indicators, we can accurately identify "chain leader" enterprises or hidden champions.

[0130] Step S5.2c: On the graph subgraph, perform a shortest path or weighted shortest path query between specified entity pairs to analyze potential association paths and influence chains between enterprises.

[0131] This feature reveals the intermediary path when users want to understand how two seemingly unrelated companies (A and D) become related. The shortest path query returns the path with the fewest edges. The weighted shortest path query uses the reciprocal of the edge weight w (1 / w) or (1+ε-w) (where ε is a minimum) as the distance to find the path with the strongest overall relationship. For example, the path "A–(supply)–>B–(cooperation)–>C–(investment)–>D" reveals a possible capital or influence transmission path from A to D. This is crucial for risk transmission analysis and business opportunity discovery.

[0132] The "raw results" of the aforementioned intelligent retrieval and multi-dimensional analysis are transformed into deliverables that users can directly use. This application provides multi-level and multi-form outputs, that is, the enterprise identification results that meet the conditions in the above embodiments may include the following steps S5.3a to S5.3c:

[0133] Step S5.3a: Input the preliminary search results set and / or multi-dimensional analysis results into a pre-configured report generation template to automatically generate a structured analysis report containing a list of key entities, a relationship network diagram, and statistical analysis charts.

[0134] The report generation engine selects a template based on the query intent (e.g., competitive analysis report, supply chain assessment report). The template includes pre-defined analysis dimensions, such as "Top 10 Core Enterprises" and "Cooperative Network Density Trends." The system then populates the template with specific data (entity names, centrality values, community division results) and automatically generates bar charts, radar charts, etc., from the chart library, ultimately assembling a complete PDF or Word document.

[0135] Step S5.3b: Visualize the graph subgraphs, enterprise clusters, core enterprise rankings, and related paths using force-directed graphs, Sankey diagrams, or tree diagrams, and provide an interactive exploration interface.

[0136] Visualization is central to understanding complex networks. Force-directed graphs vividly illustrate the push-pull relationships between entities, making communities and core nodes readily apparent. Sankey diagrams excel at showing the flow of traffic or relationships between entities (e.g., the flow of investment funds). Interactive interfaces allow users to click on nodes / edges to view details, drag layouts, and expand / collapse communities, enabling "explorable" knowledge delivery.

[0137] Step S5.3c: Encapsulate the structured analysis report and visualization results, and output them via API interface or file format.

[0138] The system offers multiple delivery methods: 1) returning structured data in JSON format via a RESTful API for integration with other business systems; 2) directly displaying interactive visualizations and downloadable report documents on the web interface; and 3) automatically pushing reports and charts as email attachments. This ensures that analytical results can be seamlessly integrated into the user's workflow.

[0139] To ensure Figure 1 In addition to ensuring the long-term stable operation and performance optimization of the example method, this application also includes advanced features such as graph quality monitoring and alerting, and cloud platform performance optimization strategies in its specific implementation:

[0140] 1) Knowledge Graph Quality Monitoring and Alarming: The system periodically calculates several quality indicators for the industry knowledge graph, including the growth rate of the total number of entities, relation density, average data freshness, and relation prediction confidence distribution. Dynamic thresholds are set for each indicator (e.g., the weekly decline in relation density should not exceed 10%). When any indicator remains abnormal for a set period, an alarm event is triggered. After an alarm is triggered, the system can automatically or semi-automatically trigger corresponding data source checks, model retraining, or knowledge graph reconstruction processes according to predefined strategies, forming an operational closed loop.

[0141] 2) Cloud platform performance optimization strategies, specifically including: When the method is executed on the cloud-based converged application runtime support platform, a fine-grained microservice architecture is adopted. The data acquisition task in step S1, the relationship mining task in step S3, and the retrieval and analysis task in step S5 are decoupled and processed in parallel by different microservice instances. Frequently accessed basic entity sets and hotspot subgraphs of the industry knowledge graph are cached using a distributed in-memory database, reducing millisecond-level query latency to microsecond-level. GPU-accelerated computing resource scheduling is provided for graph neural network model computation, multi-hop relationship reasoning, and community detection algorithms. When computationally intensive tasks are triggered, GPU resources are automatically requested and released after the task is completed, thereby optimizing costs while ensuring performance.

[0142] From the above appendix Figure 1 As illustrated by the example of the enterprise data identification method for industry graphs, on the one hand, by utilizing an enterprise relationship mining model that integrates a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks, it can not only efficiently and accurately identify explicitly stated direct relationships in text (e.g., investment, cooperation), but also effectively mine potential, indirect implicit relationships (e.g., competition, potential upstream and downstream relationships) through deep semantic reasoning of entity associations via graph neural networks. This results in the construction of an industry knowledge graph with more comprehensive relationship types and richer layers. On the other hand, the graph is generated by fusing direct and implicit relationships, removing redundancy, and calculating weights. More importantly, this method can dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges based on the aforementioned process. This means that when new data flows in, the system can trigger corresponding identification, mining, and calculation processes to incrementally update the graph, ensuring that the graph content keeps pace with changes in the data source and reflects the latest industry dynamics in a timely manner. This makes the analysis and decision-making based on the graph more timely and valuable. Thirdly, by responding to query requests targeting specific industries, the system performs intelligent retrieval and multi-dimensional analysis of the industry knowledge graph, outputting enterprise identification results that meet the criteria. This transforms the complex underlying knowledge graph construction results into upper-level applications for user queries. Intelligent retrieval allows users to naturally locate their points of interest, while multi-dimensional analysis deeply analyzes the industry landscape from multiple perspectives such as community structure, centrality, and relational paths. The final structured identification results are intuitive and clear, greatly enhancing the usability of the industry graph and providing users with powerful data-driven decision support. In summary, the technical solution of this application, by integrating multi-source data, ontology recognition, intelligent mining, and dynamic graph construction, achieves comprehensive, accurate, and dynamic identification and analysis of enterprises and their industry relationships.

[0143] Please see the appendix Figure 2 This application provides an enterprise data identification device for industry maps. The device may include a data acquisition module 201, an extraction module 202, a mining module 203, a construction module 204, and an analysis module 205, as detailed below:

[0144] The acquisition module 201 is used to asynchronously acquire multi-source heterogeneous enterprise raw data from the Internet and big data resource services, clean, integrate and standardize the enterprise raw data, and generate a standardized enterprise dataset which is then stored in the cloud database.

[0145] The extraction module 202 is used to perform entity recognition and attribute extraction on the enterprise data set based on the pre-built industry knowledge ontology model, to obtain a basic entity set including enterprise entities, product entities and technology entities, and to label the structured attribute information of each entity.

[0146] The mining module 203 is used to analyze the basic entity set using the enterprise relationship mining model to mine the direct and implicit relationships between entities. The enterprise relationship mining model is implemented by integrating a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks. The direct relationship identification module is used to identify the direct relationships between entities based on keyword templates related to supply chain, investment and cooperation. The implicit relationship prediction module is used to calculate the semantic association degree of entities in the high-dimensional feature space based on entity attributes and the identified direct relationships in order to predict the implicit relationships between entities.

[0147] Module 204 is used to fuse direct and implicit relationships, remove redundancy and calculate weights, and dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges.

[0148] The analysis module 205 is used to respond to query requests for the target industry, perform intelligent retrieval and multi-dimensional analysis of the industry knowledge graph, and output the enterprise identification results that meet the conditions.

[0149] From the above appendix Figure 2 As illustrated by the example of the enterprise data recognition device for industry graphs, on the one hand, by utilizing an enterprise relationship mining model that integrates a direct relationship recognition module based on predefined rules and an implicit relationship prediction module based on graph neural networks, it can not only efficiently and accurately identify explicitly stated direct relationships in text (e.g., investment, cooperation), but also effectively mine potential, indirect implicit relationships (e.g., competition, potential upstream and downstream) through deep semantic reasoning of entity associations using graph neural networks, thereby constructing an industry knowledge graph with more comprehensive relationship types and richer layers. On the other hand, during graph construction, the graph is generated by fusing direct and implicit relationships, removing redundancy, and calculating weights. More importantly, this method can dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges based on the aforementioned process. This means that when new data flows in, the system can trigger corresponding identification, mining, and calculation processes to incrementally update the graph, ensuring that the graph content keeps pace with changes in the data source and reflects the latest industry dynamics in a timely manner. This makes the analysis and decision-making based on the graph more timely and valuable. Thirdly, by responding to query requests targeting specific industries, the system performs intelligent retrieval and multi-dimensional analysis of the industry knowledge graph, outputting enterprise identification results that meet the criteria. This transforms the complex underlying knowledge graph construction results into upper-level applications for user queries. Intelligent retrieval allows users to naturally locate their points of interest, while multi-dimensional analysis deeply analyzes the industry landscape from multiple perspectives such as community structure, centrality, and relational paths. The final structured identification results are intuitive and clear, greatly enhancing the usability of the industry graph and providing users with powerful data-driven decision support. In summary, the technical solution of this application, by integrating multi-source data, ontology recognition, intelligent mining, and dynamic graph construction, achieves comprehensive, accurate, and dynamic identification and analysis of enterprises and their industry relationships.

[0150] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device 3 in this embodiment mainly includes: a processor 30, a memory 31, and a computer program 32 stored in the memory 31 and executable on the processor 30, such as a program for an enterprise data identification method for an industry map. When the processor 30 executes the computer program 32, it implements the steps described in the above embodiment of the enterprise data identification method for an industry map, for example... Figure 1 The steps S101 to S105 are shown. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 2 The functions of the acquisition module 201, extraction module 202, mining module 203, construction module 204, and analysis module 205 are shown.

[0151] For example, the computer program 32 for enterprise data identification methods for industry graphs mainly includes: asynchronously collecting multi-source heterogeneous raw enterprise data from the Internet and big data resource services; cleaning, integrating, and standardizing the raw enterprise data to generate a standardized enterprise dataset and storing it in a cloud database; based on a pre-built industry knowledge ontology model, performing entity identification and attribute extraction on the enterprise data set to obtain a basic entity set including enterprise entities, product entities, and technology entities, and labeling each entity with structured attribute information; and using an enterprise relationship mining model to analyze the basic entity set and mine the direct and implicit relationships between entities, wherein the enterprise relationship mining model integrates predefined... The system is implemented using a direct relationship identification module and a latent relationship prediction module based on a graph neural network. The direct relationship identification module identifies direct relationships between entities based on keyword templates related to supply chain, investment, and cooperation. The latent relationship prediction module calculates the semantic correlation degree of entities in a high-dimensional feature space based on entity attributes and identified direct relationships to predict latent relationships between entities. The system then fuses, removes redundancy, and calculates weights for the mined direct and latent relationships, dynamically constructing and updating an industry knowledge graph with entities as nodes and weighted relationships as edges. In response to query requests targeting specific industries, the system performs intelligent retrieval and multi-dimensional analysis of the industry knowledge graph, outputting enterprise identification results that meet the conditions. The computer program 32 can be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete this application. One or more modules / units can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 32 in the electronic device 3.For example, computer program 32 can be divided into the functions of acquisition module 201, extraction module 202, mining module 203, construction module 204, and analysis module 205 (a module in the virtual device). The specific functions of each module are as follows: Acquisition module 201 is used to asynchronously acquire multi-source heterogeneous enterprise raw data from the Internet and big data resource services, clean, integrate, and standardize the enterprise raw data, and generate a standardized enterprise dataset which is then stored in a cloud database; Extraction module 202 is used to perform entity recognition and attribute extraction on the enterprise data set based on a pre-built industry knowledge ontology model, obtaining a basic entity set including enterprise entities, product entities, and technology entities, and labeling each entity with structured attribute information; Mining module 203 is used to analyze the basic entity set using an enterprise relationship mining model to mine the relationships between entities. The model integrates direct and implicit relationships. The enterprise relationship mining model achieves this by fusing a direct relationship identification module based on predefined rules with an implicit relationship prediction module based on graph neural networks. The direct relationship identification module identifies direct relationships between entities based on keyword templates related to supply chains, investment, and cooperation. The implicit relationship prediction module calculates the semantic association degree of entities in a high-dimensional feature space based on entity attributes and identified direct relationships to predict implicit relationships between entities. A construction module 204 fuses, removes redundancy, and calculates weights for direct and implicit relationships, dynamically constructing and updating an industry knowledge graph with entities as nodes and weighted relationships as edges. An analysis module 205 responds to query requests for target industries by intelligently retrieving and performing multi-dimensional analysis of the industry knowledge graph, outputting enterprise identification results that meet the criteria.

[0152] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0153] The processor 30 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0154] The memory 31 can be an internal storage unit of the electronic device 3, such as a hard disk or RAM. The memory 31 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 31 can include both internal and external storage units of the electronic device 3. The memory 31 is used to store computer programs and other programs and data required by the electronic device. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0155] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed. That is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above-described device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0156] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0157] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0158] In the embodiments provided in this application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0161] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program for the enterprise data identification method for industry graphs can be stored in a storage medium. When executed by a processor, this computer program can implement the steps of the various method embodiments described above, namely: asynchronously collecting multi-source heterogeneous enterprise raw data from the Internet and big data resource services; cleaning, integrating, and standardizing the enterprise raw data; generating a standardized enterprise dataset and merging it in a cloud database; based on a pre-built industry knowledge ontology model, performing entity identification and attribute extraction on the enterprise data set to obtain a basic entity set including enterprise entities, product entities, and technology entities, and labeling each entity with structured attribute information; and using an enterprise relationship mining model to analyze the basic entities. The system analyzes and mines direct and implicit relationships between entities. The enterprise relationship mining model integrates a direct relationship identification module based on predefined rules with an implicit relationship prediction module based on graph neural networks. The direct relationship identification module identifies direct relationships between entities based on keyword templates related to supply chains, investment, and cooperation. The implicit relationship prediction module calculates the semantic association degree of entities in a high-dimensional feature space based on entity attributes and identified direct relationships to predict implicit relationships. The mined direct and implicit relationships are then fused, redundancy removed, and weighted to dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges. Responding to queries targeting specific industries, the system intelligently retrieves and performs multi-dimensional analysis of the industry knowledge graph, outputting enterprise identification results that meet the criteria. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form. Storage media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of storage media can be appropriately added to or removed according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, storage media may not include electrical carrier signals and telecommunication signals.

[0162] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application. The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the protection scope of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.< / obj> < / subj> < / obj> < / subj>

Claims

1. A method for identifying enterprise data based on an industrial map, characterized in that, The method is executed by software deployed on a cloud-based converged application runtime support platform, and the method includes: Step S1: Asynchronously collect multi-source heterogeneous enterprise raw data from the Internet and big data resource services, clean, integrate and standardize the enterprise raw data, generate a standardized enterprise dataset and store it in the cloud database; Step S2: Based on the pre-built industry knowledge ontology model, perform entity recognition and attribute extraction on the enterprise data set to obtain a basic entity set including enterprise entities, product entities and technology entities, and label each entity with structured attribute information. Step S3: Analyze the basic entity set using an enterprise relationship mining model to mine the direct and implicit relationships between entities; wherein, the enterprise relationship mining model is implemented by integrating a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks. The direct relationship identification module is used to identify the direct relationships between entities based on keyword templates related to supply chain, investment and cooperation. The implicit relationship prediction module is used to calculate the semantic association degree of entities in a high-dimensional feature space based on entity attributes and the identified direct relationships, so as to predict the implicit relationships between entities. Step S4: The direct and implicit relationships mined in Step S3 are fused, redundancy is removed, and weights are calculated to dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges. Step S5: In response to a query request targeting a specific industry, perform intelligent retrieval and multi-dimensional analysis on the industry knowledge graph, and output the enterprise identification results that meet the conditions. 2.The industry graph-oriented enterprise data identification method according to claim 1, characterized in that, The asynchronous collection of multi-source heterogeneous enterprise raw data from Internet and big data resource services includes: By configuring a distributed asynchronous crawler framework, initial data is crawled in parallel from designated public bidding websites, enterprise credit information disclosure platforms, and industry news websites; Listen for and capture data streams from publicly available enterprise API interfaces, and align the data streams with the initial crawled data using timestamps and source tags; Based on the preset data source priority and update frequency strategy, the crawler framework and API listening process are dynamically scheduled to achieve incremental collection and real-time updates of multi-source data.

3. The enterprise data identification method for industry mapping according to claim 1, characterized in that, The pre-built industry knowledge ontology model performs entity recognition and attribute extraction on the enterprise data set to obtain a basic entity set including enterprise entities, product entities, and technology entities, and labels each entity with structured attribute information, including: The unstructured text description field in the standardized enterprise data set is input into a multi-model voting recognition module consisting of a BERT model, a BiLSTM-CRF model, and a dictionary matching model based on the industry knowledge ontology; The multi-model voting identification module outputs the entities and types identified by each model, and uses a weighted voting strategy to fuse and resolve conflicts in the identification results to obtain a preliminary entity list. Based on the attribute constraint rules defined in the industry knowledge ontology model, the context information associated with each entity in the preliminary entity list is parsed, and the structured attribute information is extracted and filled in.

4. The enterprise data identification method for industry mapping according to claim 3, characterized in that, The weighted voting strategy for fusing and resolving conflicts in the identification results specifically includes: Dynamic weights are assigned to the BERT model, the BiLSTM-CRF model, and the dictionary matching model, respectively, wherein the dynamic weights are periodically adjusted based on the recognition accuracy of each model on historical source data; When at least two models identify the same entity type for the same text fragment, the entity and its type are added to the basic entity set. When there is a conflict in the entity types identified by the various models, the entity type output by the model with the highest dynamic weight is selected as the final result, and the conflict case is recorded in the training sample library for subsequent reinforcement training of the low-weight model.

5. The enterprise data identification method for industry mapping according to claim 1, characterized in that, The implicit relationship prediction module predicts implicit relationships between entities based on graph neural networks, specifically including: The entities, entity attributes, and identified direct relationships in the basic entity set are collectively constructed into an initial heterogeneous graph; The initial heterogeneous graph is subjected to multiple rounds of message passing and feature aggregation using a graph convolutional network (GCN) or a graph attention network (GAT) model to learn a deep vector representation of each entity. Calculate the cosine similarity or neural network-based association score between any two entity vector representations. Entity pairs with scores exceeding a preset threshold are predicted to have the implicit relationship, and the predicted relationship type and confidence level are labeled for them.

6. The enterprise data identification method for industry mapping according to claim 5, characterized in that, The enterprise relationship mining model also includes a relationship mining optimization process: The industry knowledge graph constructed in step S4 is periodically used as new training data to incrementally train the graph neural network model in the implicit relationship prediction module. The relationship identification cases generated by the direct relationship identification module on new data, which have a confidence level higher than a set threshold, are converted into new rule templates after manual review and added to the rule knowledge base. Based on the incrementally trained graph neural network model and the updated rule knowledge base, relation mining is re-executed on the full set of basic entities to iteratively optimize the accuracy and completeness of relations in the industry knowledge graph.

7. The enterprise data identification method for industry mapping according to claim 1, characterized in that, The dynamic construction and updating of the industry knowledge graph, which uses entities as nodes and weighted relationships as edges, includes: The newly discovered relationships are semantically similar to existing relationships in the graph. Relationships with similarity exceeding the merging threshold are dereduplicated and merged. The weights and sources of evidence for the merged relationships are then updated. Establish an event-driven mechanism that automatically triggers local re-identification of affected entities, relationship mining, and graph subgraph updates when data updates are detected in a specific high-priority data source in step S1. Create a version snapshot for each update of the map and provide map backtracking and trend analysis functions based on the time dimension.

8. A device for identifying enterprise data in an industry map, characterized in that, The device includes: The data acquisition module is used to asynchronously collect multi-source heterogeneous raw enterprise data from the Internet and big data resource services, clean, integrate and standardize the raw enterprise data, and generate a standardized enterprise dataset which is then stored in the cloud database. The extraction module is used to perform entity recognition and attribute extraction on the enterprise data set based on a pre-built industry knowledge ontology model, to obtain a basic entity set including enterprise entities, product entities and technology entities, and to label each entity with structured attribute information. The mining module is used to analyze the basic entity set using an enterprise relationship mining model to mine the direct and implicit relationships between entities. The enterprise relationship mining model is implemented by integrating a direct relationship identification module based on predefined rules and an implicit relationship prediction module based on graph neural networks. The direct relationship identification module is used to identify the direct relationships between entities based on keyword templates related to supply chain, investment, and cooperation. The implicit relationship prediction module is used to calculate the semantic association degree of entities in a high-dimensional feature space based on entity attributes and the identified direct relationships, so as to predict the implicit relationships between entities. The construction module is used to fuse, remove redundancy and calculate weights for the direct and implicit relationships, and dynamically construct and update an industry knowledge graph with entities as nodes and weighted relationships as edges. The analysis module is used to respond to query requests targeting a specific industry by performing intelligent retrieval and multi-dimensional analysis on the industry knowledge graph and output enterprise identification results that meet the conditions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images