Enterprise portrait label intelligent generation method

Abstract

The invention relates to the technical field of big data processing, in particular to an enterprise portrait label intelligent generation method. The method comprises the following steps: acquiring industrial standard reference data; performing tag corpus construction processing according to the industrial standard reference data to obtain a tag corpus; performing domain knowledge modeling according to the tag corpus to obtain a domain knowledge graph; performing original enterprise data analysis according to the domain knowledge graph to obtain a preprocessed enterprise data set; performing data fusion and association on the preprocessed enterprise data set according to a domain knowledge graph to obtain a unified enterprise data set; performing enterprise feature extraction on the unified enterprise data set according to the domain knowledge graph to obtain an enterprise feature vector; and carrying out feature clustering analysis on the enterprise feature vectors to obtain an enterprise clustering graph. According to the method, the problems that a label system is isolated and static in an existing method are effectively solved, and the application value, timeliness and accuracy of the label system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and in particular to a method for intelligently generating enterprise portrait labels. Background Art

[0002] With the advent of the big data era, enterprises have accumulated massive, multi-source, heterogeneous data. In order to better understand and utilize these data, enterprise portrait technology has emerged. As the core element of enterprise portraits, enterprise portrait labels can intuitively describe the characteristics and behaviors of enterprises and provide data support for precision marketing, risk control, customer relationship management and other businesses. Early enterprise portrait label generation mainly relied on manual experience, which was inefficient, costly, and highly subjective. In recent years, with the rapid development of technologies such as machine learning and natural language processing, the intelligent generation method of enterprise portrait labels has gradually become a research hotspot.

[0003] However, the existing intelligent generation methods of enterprise portrait labels have the disadvantages of isolated label systems and static label systems. The manifestation of isolated label systems: There is a lack of semantic association between labels generated by existing methods. Each label exists independently and cannot fully explore the relationship and connection between labels. This leads to limited application value of the label system and makes it difficult to support more in-depth and comprehensive enterprise portrait analysis and application. The manifestation of static label systems: The label systems generated by existing methods are usually static and cannot adapt to the dynamic changes of enterprises in a timely manner. The operating conditions, business models, user groups, etc. of enterprises are constantly changing. Static label systems are difficult to reflect these changes, resulting in reduced timeliness and accuracy of labels. Summary of the invention

[0004] Based on this, it is necessary to provide a method for intelligently generating enterprise portrait labels to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above purpose, a method for intelligently generating enterprise portrait labels includes the following steps:

[0006] Step S1: Obtain industry standard reference data; construct a label corpus based on the industry standard reference data to obtain a label corpus; perform domain knowledge modeling based on the label corpus to obtain a domain knowledge graph;

[0007] Step S2: Analyze the original enterprise data according to the domain knowledge graph to obtain a preprocessed enterprise data set; perform data fusion and association on the preprocessed enterprise data set according to the domain knowledge graph to obtain a unified enterprise data set;

[0008] Step S3: extract enterprise features from the unified enterprise data set according to the domain knowledge graph to obtain an enterprise feature vector;

[0009] Step S4: Perform feature clustering analysis on the enterprise feature vector to obtain an enterprise clustering graph; define cluster cluster labels on the enterprise clustering graph according to the domain knowledge graph to obtain a cluster cluster label set; perform dynamic label generation according to the cluster cluster label set to obtain a dynamic enterprise label library;

[0010] Step S5: construct a tag semantic association network based on the dynamic enterprise tag library and the domain knowledge graph to obtain an initial tag semantic association network; construct an enterprise tag semantic network based on the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network;

[0011] Step S6: Obtain portrait tag application scenario data; identify the optimization direction of the tag system based on the portrait tag application scenario data and the enterprise tag semantic network to obtain tag system optimization data; optimize the enterprise tag semantic network based on the tag system optimization data to obtain enterprise portrait tag data, so as to realize the intelligent generation of enterprise portrait tags.

[0012] The present invention can provide rich industry knowledge and semantic information for the generation of enterprise portrait labels by constructing a label corpus and a domain knowledge graph, ensuring the professionalism, accuracy and completeness of the label system. Through the steps of data source identification, data collection, data preprocessing, data fusion and association, high-quality and structured enterprise data can be obtained, laying a solid data foundation for subsequent feature extraction and label generation. By constructing a feature dimension system, extracting text features, numerical features, category features and enterprise graph features, the multi-dimensional information of the enterprise can be converted into a quantifiable feature vector, providing data support for subsequent cluster analysis and label prediction. Through the steps of cluster analysis, label definition, label annotation and dynamic update, accurate and dynamic labels can be marked for the enterprise, a preliminary portrait of the enterprise can be achieved, and it can be continuously updated with the development of the enterprise. By constructing a label semantic association network and an enterprise label semantic network, the potential semantic relationship between labels can be excavated, and this relationship can be applied to specific enterprise label data to form a semantically rich and clearly structured network structure for supporting the analysis and application of enterprise portraits. By obtaining application scenario requirements, formulating label application strategies, tracking business results, evaluating label system effects, identifying optimization directions and optimizing, the quality and application effects of enterprise portrait labels can be continuously improved, and the intelligent generation and continuous optimization of enterprise portrait labels can be realized, so that they can better serve the actual business needs of enterprises. Therefore, the present invention provides a method for intelligently generating enterprise portrait labels, which is used to solve the shortcomings of existing methods for intelligently generating enterprise portrait labels, that is, the isolated label system and the static label system. By constructing a label semantic association network and an enterprise label semantic network, and utilizing domain knowledge graphs, external knowledge bases, and label co-occurrence relationships, effective association of label semantics is achieved, and the semantic expression ability of the label system is enriched; at the same time, a dynamic label generation mechanism is introduced, and by clustering the enterprise feature vectors and dynamically updating the labels in combination with the enterprise's real-time behavior data stream, the label system can be dynamically adjusted as the enterprise changes, effectively solving the problems of isolated label systems and static label systems in existing methods, and improving the application value, timeliness, and accuracy of the label system.

[0013] Preferably, step S1 comprises the following steps:

[0014] Step S11: Acquire industry standard reference data; construct a label corpus based on the industry standard reference data to obtain a label corpus;

[0015] Step S12: performing tag concept mining on the tag corpus to obtain a tag concept set;

[0016] Step S13: extracting concept relations from the tag corpus according to the tag concept set to obtain concept relation triple data;

[0017] Step S14: Use the concepts in the label concept set as nodes and the relationships in the concept relationship triple data as edges to construct a domain knowledge graph, and perform manual verification of the graph to obtain a domain knowledge graph.

[0018] The present invention obtains industry standard reference data from authoritative data sources, and uses natural language processing technology and word vector models to extract, clean and merge labels, so as to construct a concise, accurate and wide-coverage label corpus, providing a high-quality data foundation for subsequent label concept mining and domain knowledge map construction. By analyzing the label corpus using the topic model, labels with similar semantics can be classified into the same topic to form a label concept set, thereby realizing the abstraction and generalization of label information, helping to better understand the semantic relationship between labels, and providing a basis for subsequent concept relationship extraction and knowledge map construction. Based on the label concept set, the semantic relationship between labels in the label corpus can be automatically identified using syntactic analysis technology, and represented in the form of structured triples, thereby realizing the accurate characterization of the label semantic relationship, and providing data support for the construction of structured and knowledgeable domain knowledge maps. By importing the label concept set and the concept relationship triple data into the graph database, a domain knowledge map with clear structure and rich semantics can be constructed, which is used to store and manage professional knowledge in the industry field, and supports knowledge reasoning and query, providing strong knowledge support for subsequent enterprise portrait label generation and application. The manual verification process can further improve the accuracy and completeness of the domain knowledge graph.

[0019] Preferably, step S11 includes the following steps:

[0020] Step S111: Acquire industry standard reference data; construct a labeling system according to the industry standard reference data to obtain an initial labeling system;

[0021] Step S112: Perform enterprise competitive product analysis based on the initial label system to obtain excellent enterprise label system data;

[0022] Step S113: optimizing the initial label system using the excellent enterprise label system data to obtain a label system;

[0023] Step S114: construct a label corpus according to the label system to obtain a label corpus.

[0024] The present invention can construct a preliminary labeling system framework by acquiring industry standard reference data, and combining industry classification standards and keyword extraction, covering the basic attributes and characteristics of enterprises in the industry, and providing a basis for subsequent optimization of the labeling system. By analyzing the labeling system of excellent enterprises, we can learn from their experience in label design, label classification, label definition, etc., provide reference and direction for optimizing the initial labeling system, and help build a labeling system that is more in line with industry best practices. The initial labeling system is optimized using the labeling system data of excellent enterprises, such as label replacement, label merging, label deletion and other operations, which can remove redundant labels, merge similar labels, and delete irrelevant labels, thereby improving the simplicity, accuracy and practicality of the labeling system. According to the finalized labeling system, text data containing relevant keywords is collected in a targeted manner, and a label corpus that is highly relevant to the labeling system can be constructed, which provides a high-quality data foundation for subsequent label concept mining and semantic analysis, and improves the semantic expression ability of the labeling system.

[0025] Preferably, step S21 includes the following steps:

[0026] Step S21: Identify data sources related to enterprise portrait labels based on the domain knowledge graph to obtain label data; classify the label data sources to obtain a classified data source list;

[0027] Step S22: Collect and store data according to the classified data source list to obtain the original enterprise data set;

[0028] Step S23: performing data preprocessing on the original enterprise data set to obtain a preprocessed enterprise data set;

[0029] Step S24: performing data conversion on the preprocessed enterprise data set according to the domain knowledge graph to obtain a converted enterprise data set; performing data mapping of a corresponding label system on the converted enterprise data set to obtain a structured enterprise data set;

[0030] Step S25: Perform data fusion and association on the structured enterprise data sets to obtain a unified enterprise data set.

[0031] The present invention can more accurately identify the data sources related to the enterprise portrait tags by utilizing the label concepts and semantic relationships contained in the domain knowledge graph, and classify and manage them according to the data source type to form a classified data source list, provide clear guidance for subsequent data collection work, and improve the efficiency and pertinence of data collection. According to the classified data source list, different data collection methods are adopted for different types of data sources, and complete and accurate original enterprise data can be efficiently obtained and stored in the data warehouse, providing sufficient data basis for subsequent data preprocessing and analysis. Through preprocessing operations such as data cleaning, format unification, and standardization, noise and redundant information in the original enterprise data can be eliminated, the accuracy, consistency and comparability of the data can be improved, and high-quality data guarantee can be provided for subsequent data analysis and mining. According to the domain knowledge graph, the preprocessed enterprise data is semantically converted, the data expressing the same semantics in different data sources can be unified, and data mapping is performed according to a predefined label system to form a structured enterprise data set, which lays the foundation for subsequent data fusion and association analysis. Through data fusion and correlation analysis, scattered information about the same enterprise in different data sources can be integrated to form a unified and complete enterprise view, and the potential correlation between different data can be explored to provide data support for building a comprehensive enterprise portrait.

[0032] Preferably, step S3 comprises the following steps:

[0033] Step S31: define feature dimensions according to the domain knowledge graph to obtain a feature dimension system; associate label features with the feature dimension system and the domain knowledge graph to obtain a label feature mapping table;

[0034] Step S32: extracting text features from the unified enterprise data set to obtain an enterprise text feature set; performing text vectorization on the enterprise text feature set to obtain an enterprise text feature vector;

[0035] Step S33: extracting numerical features from the unified enterprise data set to obtain an enterprise numerical feature set; and standardizing the enterprise numerical feature set to obtain a standardized enterprise numerical feature set;

[0036] Step S34: performing category feature encoding on the unified enterprise data set to obtain a coded enterprise category feature set;

[0037] Step S35: construct an enterprise relationship graph based on the unified enterprise data set to obtain an enterprise relationship graph; extract enterprise graph features from the enterprise relationship graph to obtain an enterprise graph feature vector;

[0038] Step S36: According to the label feature mapping table, feature vectors of the enterprise text feature vector, the standardized enterprise numerical feature set, the coded enterprise category feature set and the enterprise graph feature vector are fused and feature vector dimension reduction is performed to obtain the enterprise feature vector.

[0039] The present invention can classify label concepts into different feature dimensions by utilizing the concept hierarchy and semantic relationship in the domain knowledge graph, construct a multi-level feature dimension system, and establish a mapping relationship between labels and data features to form a label feature mapping table, thereby integrating domain knowledge into the enterprise feature extraction process and improving the accuracy and interpretability of feature extraction. Through text feature extraction and vectorization representation, unstructured text data can be converted into numerical vectors that can be understood and processed by computers, retaining the semantic information of text data, and providing a basis for subsequent clustering analysis and label prediction. Standardization of numerical features can eliminate the influence of dimensions between different features, avoid certain features from dominating because of large values, improve the fairness and comparability of features, and enhance the accuracy of subsequent analysis results. Using unique hot encoding to process category features can convert discrete category information into continuous numerical vectors, facilitate computer processing and calculation, and avoid introducing order relationships between categories, maintaining the independence of category information. Constructing an enterprise relationship graph and extracting enterprise graph feature vectors can dig out potential associations between enterprises, and represent them as quantifiable features, enrich enterprise feature dimensions, and provide more comprehensive information for enterprise portraits. By fusing different types of feature vectors, we can construct an enterprise feature vector that contains rich information. By reducing the feature dimension through dimensionality reduction processing, we can eliminate redundant information, improve the efficiency and effectiveness of subsequent analysis, and finally obtain a low-dimensional vector representation for characterizing enterprise characteristics.

[0040] Preferably, step S4 comprises the following steps:

[0041] Step S41: Select a clustering algorithm according to the enterprise feature vector to obtain a selected clustering algorithm; optimize clustering parameters of the selected clustering algorithm to obtain clustering parameters;

[0042] Step S42: clustering the enterprise feature vectors using the clustering parameters and the selected clustering algorithm to obtain enterprise clustering result data; and visualizing the clustering result of the enterprise clustering result data to obtain an enterprise clustering graph;

[0043] Step S43: defining cluster labels for the enterprise cluster graph according to the domain knowledge graph to obtain a cluster label set;

[0044] Step S44: annotating enterprise dynamic labels according to enterprise clustering result data and clustering cluster label sets to obtain enterprise initial label sets;

[0045] Step S45: obtaining the enterprise real-time behavior data stream; performing real-time behavior data analysis on the enterprise real-time behavior data stream to obtain the enterprise key event sequence;

[0046] Step S46: updating the dynamic enterprise tag library of the enterprise initial tag set according to the enterprise key event sequence and the preset tag dynamic update rule library to obtain a dynamic enterprise tag library.

[0047] The present invention can ensure the effectiveness and accuracy of cluster analysis by selecting a suitable clustering algorithm according to the characteristics of the enterprise feature vector and optimizing its parameters, and divide enterprises with similar characteristics into the same cluster cluster, providing a basis for subsequent label definition and annotation. Using the optimized clustering algorithm and parameters to cluster the enterprise feature vector, the enterprise can be grouped according to its characteristics, and the clustering results can be displayed in a visual way, which helps to intuitively understand the structure and characteristic distribution of the enterprise group, and provide a reference for the subsequent label definition. In combination with the domain knowledge map, the enterprise clustering diagram is analyzed, and a representative label can be assigned to each cluster cluster to form a cluster cluster label set, and the domain knowledge is integrated into the label definition process to improve the accuracy and interpretability of the label. According to the cluster cluster to which the enterprise belongs, it can be automatically labeled with a corresponding label to form an initial label set of the enterprise, realize the preliminary portrait of the enterprise, and provide a basis for the subsequent dynamic update of the label. Real-time acquisition and analysis of enterprise behavior data can timely capture changes in enterprise behavior, identify key events in the process of enterprise development, and provide a basis for dynamically updating enterprise labels. According to the key events of the enterprise and the preset tag dynamic update rules, the enterprise's tag information can be dynamically adjusted, such as adding new tags, deleting old tags, updating tag weights, etc., so that the enterprise portrait can be continuously updated as the enterprise develops, thereby improving the timeliness and accuracy of the enterprise portrait.

[0048] Preferably, step S43 includes the following steps:

[0049] Step S431: extract typical enterprises from the enterprise clustering result data to obtain a typical enterprise sample set; manually annotate the typical enterprise sample set to obtain a manually annotated sample set;

[0050] Step S432: training a label propagation model using a manually annotated sample set to obtain a label propagation model;

[0051] Step S433: using the label propagation model to perform full enterprise label prediction on the enterprise clustering result data to obtain enterprise label prediction data;

[0052] Step S434: using a preset confidence threshold to filter the enterprise tag prediction data by tag confidence to obtain high-confidence enterprise tags;

[0053] Step S435: Refine the high-confidence enterprise labels according to the domain knowledge graph to obtain a cluster label set.

[0054] The present invention can construct a high-quality labeled sample set through typical enterprise extraction and manual labeling, provide accurate supervision information for the training of the label propagation model, and improve the prediction accuracy of the model. The label propagation model is trained using the manually labeled sample set, so that the model can learn the correlation between enterprise characteristics and labels, thereby realizing label prediction for unlabeled enterprises. Using the trained label propagation model, label prediction can be performed for all enterprises that have not been manually labeled, thereby achieving full coverage of enterprise labels and improving the efficiency of label labeling. Setting a confidence threshold to filter the label prediction results can screen out labels with high reliability of prediction results, improve the accuracy of labels, and avoid applying low-confidence labels to actual business. According to the domain knowledge graph, high-confidence labels are refined, and the granularity of labels can be refined to a more specific concept level, such as refining "artificial intelligence" into "machine learning", "deep learning", etc., thereby improving the precision and expression ability of labels.

[0055] Preferably, step S5 comprises the following steps:

[0056] Step S51: performing tag concept mapping on the dynamic enterprise tag library and the domain knowledge graph to obtain a tag concept mapping table;

[0057] Step S52: Performing external knowledge base text mining on the tags that are not concept-mapped in the tag concept mapping table to obtain relevant tag concept data; performing unmatched tag processing on the tag concept mapping table according to the relevant tag concept data to obtain an extended tag concept mapping table;

[0058] Step S53: Calculate the label semantic similarity according to the extended label concept mapping table and the domain knowledge graph to obtain a label semantic similarity matrix;

[0059] Step S54: constructing a label semantic association network for the label semantic similarity matrix using a preset semantic similarity threshold to obtain an initial label semantic association network;

[0060] Step S55: constructing an enterprise tag semantic network according to the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network.

[0061] The present invention maps the labels in the dynamic enterprise label library with the concept nodes in the domain knowledge graph, so that the labels can be associated with the domain knowledge, providing a basis for the construction of the label semantic association network. For labels that cannot be directly mapped with the domain knowledge graph, text mining can be performed through an external knowledge base to supplement the missing information in the domain knowledge graph, and the unmapped labels can be associated with the domain knowledge graph to improve the degree of association between the labels and the domain knowledge. Based on the extended label concept mapping table, the semantic similarity between the labels is calculated, and the semantic distance between the labels can be quantified, providing a data basis for the construction of the label semantic association network. According to the preset semantic similarity threshold, the semantically similar labels are connected, and the initial label semantic association network can be constructed to mine the potential semantic association relationship between the labels, providing a framework for the construction of the enterprise label semantic network. According to the initial label semantic association network and the dynamic enterprise label library, the enterprise label semantic network can be constructed, and the semantic association relationship between the labels is applied to the specific enterprise label data to form a semantically rich and clearly structured network structure for supporting the analysis and application of enterprise portraits.

[0062] Preferably, step S55 comprises the following steps:

[0063] Step S551: mining tag co-occurrence relationships on the dynamic enterprise tag library to obtain a tag co-occurrence relationship matrix;

[0064] Step S552: constructing a tag co-occurrence network for the tag co-occurrence relationship matrix using a preset co-occurrence strength threshold to obtain a tag co-occurrence network;

[0065] Step S553: ​​fusing the initial tag semantic association network and the tag co-occurrence network into a semantic association network and a co-occurrence network to obtain a fused tag semantic network;

[0066] Step S554: performing semantic network optimization on the fused label semantic network to obtain an optimized label semantic network;

[0067] Step S555: Generate an enterprise tag semantic network based on the optimized tag semantic network to obtain the enterprise tag semantic network.

[0068] The present invention can discover the implicit association between tags by mining the tag co-occurrence relationship of the dynamic enterprise tag library. Even if two tags do not have a direct semantic association, if they often appear together in the tag list of the enterprise, it also indicates that there may be some potential connection between them. By constructing a tag co-occurrence network using a preset co-occurrence strength threshold, tag pairs with higher co-occurrence frequency can be connected to form a network structure that reflects the co-occurrence relationship between tags, providing supplementary information for the construction of a tag semantic association network. By fusing the initial tag semantic association network and the tag co-occurrence network, the semantic association and co-occurrence relationship between tags can be comprehensively considered to construct a more comprehensive and accurate tag semantic association network. By optimizing the fused tag semantic network, such as deleting isolated nodes, merging semantically duplicated nodes, adjusting the weights of edges, etc., the structural rationality and information expression ability of the network can be further improved, providing a better network foundation for the generation of enterprise tag semantic networks. Based on the optimized label semantic network, an enterprise label semantic network can be generated, and the semantic association relationship between labels can be applied to specific enterprise label data to form a semantically rich and clearly structured network structure to support the analysis and application of enterprise portraits, such as identifying the potential characteristics of enterprises, predicting enterprise behavior, and discovering the association relationship between enterprises.

[0069] Preferably, step S6 comprises the following steps:

[0070] Step S61: Acquire portrait tag application scenario data; perform application scenario demand analysis on the portrait tag application scenario data to obtain application scenario demand data;

[0071] Step S62: Formulate a label application strategy based on the application scenario requirement data and the enterprise label semantic network to obtain a label application strategy;

[0072] Step S63: tracking and monitoring the business effect according to the tag application strategy to obtain business effect monitoring data; constructing an evaluation index system according to the application scenario demand data to obtain an evaluation index system;

[0073] Step S64: perform label system effect evaluation based on the business effect monitoring data and the evaluation index system to obtain a label system evaluation report;

[0074] Step S65: identifying the label system optimization direction based on the label system evaluation report to obtain label system optimization data;

[0075] Step S66: Optimize the enterprise portrait label of the enterprise label semantic network according to the label system optimization data to obtain the enterprise portrait label data to realize the intelligent generation of enterprise portrait labels.

[0076] The present invention can clarify the specific requirements of the label system in different application scenarios, such as label coverage, label accuracy, label real-time, etc., by obtaining the portrait label application scenario data and performing demand analysis, so as to provide a basis for the formulation of label application strategies and the construction of evaluation index systems. According to the application scenario requirements and the enterprise label semantic network, reasonable label application strategies can be formulated, such as label selection strategies, label combination strategies, label weight strategies, etc., so as to improve the effectiveness and pertinence of labels in specific application scenarios. Tracking and monitoring the label application effect, and constructing an evaluation index system according to the application scenario requirements, can provide data support and evaluation criteria for the label system effect evaluation, such as tracking the conversion rate of labels in marketing activities, or evaluating the contribution of labels to risk control models, etc. According to the business effect monitoring data and the evaluation index system, the effect of the label system can be comprehensively and objectively evaluated, and a label system evaluation report can be formed to provide direction and basis for the optimization of the label system. By analyzing the label system evaluation report, it is possible to identify in which aspects of the label system needs to be improved, such as insufficient label coverage, low label accuracy, untimely label updates, etc., to form label system optimization data and guide subsequent optimization work. By optimizing data according to the label system and performing targeted optimization on the enterprise label semantic network, such as adding new labels, adjusting label relationships, updating label weights, etc., the quality and application effect of enterprise portrait labels can be continuously improved, and the intelligent generation and continuous optimization of enterprise portrait labels can be realized.

[0077] The present invention can provide rich industry knowledge and semantic information for the generation of enterprise portrait labels by constructing a label corpus and a domain knowledge graph, ensuring the professionalism, accuracy and completeness of the label system. Through the steps of data source identification, data collection, data preprocessing, data fusion and association, high-quality and structured enterprise data can be obtained, laying a solid data foundation for subsequent feature extraction and label generation. By constructing a feature dimension system, extracting text features, numerical features, category features and enterprise graph features, the multi-dimensional information of the enterprise can be converted into a quantifiable feature vector, providing data support for subsequent cluster analysis and label prediction. Through the steps of cluster analysis, label definition, label annotation and dynamic update, accurate and dynamic labels can be marked for the enterprise, a preliminary portrait of the enterprise can be achieved, and it can be continuously updated with the development of the enterprise. By constructing a label semantic association network and an enterprise label semantic network, the potential semantic relationship between labels can be excavated, and this relationship can be applied to specific enterprise label data to form a semantically rich and clearly structured network structure for supporting the analysis and application of enterprise portraits. By obtaining application scenario requirements, formulating label application strategies, tracking business results, evaluating label system effects, identifying optimization directions and optimizing, the quality and application effects of enterprise portrait labels can be continuously improved, and the intelligent generation and continuous optimization of enterprise portrait labels can be realized, so that they can better serve the actual business needs of enterprises. Therefore, the present invention provides a method for intelligently generating enterprise portrait labels, which is used to solve the shortcomings of existing methods for intelligently generating enterprise portrait labels, that is, the isolated label system and the static label system. By constructing a label semantic association network and an enterprise label semantic network, and utilizing domain knowledge graphs, external knowledge bases, and label co-occurrence relationships, effective association of label semantics is achieved, and the semantic expression ability of the label system is enriched; at the same time, a dynamic label generation mechanism is introduced, and by clustering the enterprise feature vectors and dynamically updating the labels in combination with the enterprise's real-time behavior data stream, the label system can be dynamically adjusted as the enterprise changes, effectively solving the problems of isolated label systems and static label systems in existing methods, and improving the application value, timeliness, and accuracy of the label system. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 A schematic diagram of the steps of a method for intelligently generating enterprise portrait labels;

[0079] Figure 2 for Figure 1 Detailed implementation steps of step S4 in FIG.

[0080] Figure 3 for Figure 1 Detailed implementation steps of step S5 are shown in the flowchart.

[0081] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0082] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0083] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0084] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0085] To achieve this, please refer to Figures 1 to 3 , a method for intelligently generating enterprise portrait labels, comprising the following steps:

[0086] Step S1: Obtain industry standard reference data; construct a label corpus based on the industry standard reference data to obtain a label corpus; perform domain knowledge modeling based on the label corpus to obtain a domain knowledge graph;

[0087] Step S2: Analyze the original enterprise data according to the domain knowledge graph to obtain a preprocessed enterprise data set; perform data fusion and association on the preprocessed enterprise data set according to the domain knowledge graph to obtain a unified enterprise data set;

[0088] Step S3: extract enterprise features from the unified enterprise data set according to the domain knowledge graph to obtain an enterprise feature vector;

[0089] Step S4: Perform feature clustering analysis on the enterprise feature vector to obtain an enterprise clustering graph; define cluster cluster labels on the enterprise clustering graph according to the domain knowledge graph to obtain a cluster cluster label set; perform dynamic label generation according to the cluster cluster label set to obtain a dynamic enterprise label library;

[0090] Step S5: construct a tag semantic association network based on the dynamic enterprise tag library and the domain knowledge graph to obtain an initial tag semantic association network; construct an enterprise tag semantic network based on the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network;

[0091] Step S6: Obtain portrait tag application scenario data; identify the optimization direction of the tag system based on the portrait tag application scenario data and the enterprise tag semantic network to obtain tag system optimization data; optimize the enterprise tag semantic network based on the tag system optimization data to obtain enterprise portrait tag data, so as to realize the intelligent generation of enterprise portrait tags.

[0092] In the embodiment of the present invention, reference Figure 1 The above is a schematic diagram of the steps of the method for intelligently generating enterprise portrait labels of the present invention. In this example, the method for intelligently generating enterprise portrait labels includes the following steps:

[0093] Step S1: Obtain industry standard reference data; construct a label corpus based on the industry standard reference data to obtain a label corpus; perform domain knowledge modeling based on the label corpus to obtain a domain knowledge graph;

[0094] In an embodiment of the present invention, first, industry classification standards, policies and regulations, research reports and other document materials are obtained from authoritative data sources such as the National Bureau of Statistics and industry association websites as industry standard reference data. Keywords are extracted using natural language processing tools, and semantic similarity is calculated using a word vector model to construct a streamlined label corpus. Then, the LDA topic model is used to perform topic analysis on the label corpus to obtain a label concept set. Next, a syntactic analysis tool is used to extract the semantic relationship between labels to obtain concept relationship triple data. Finally, a graph database is used to construct the label concepts and semantic relationships into a domain knowledge graph, and manual verification is performed.

[0095] Step S2: Analyze the original enterprise data according to the domain knowledge graph to obtain a preprocessed enterprise data set; perform data fusion and association on the preprocessed enterprise data set according to the domain knowledge graph to obtain a unified enterprise data set;

[0096] In an embodiment of the present invention, first, the data sources related to the enterprise portrait labels are identified based on the label concepts and semantic relationships in the domain knowledge graph, and classified according to the data source type to form a classified data source list. Then, according to the data source type, different data collection methods are used to obtain the original enterprise data and store it in the data warehouse. Next, the original enterprise data is preprocessed by data cleaning, format unification, standardization, etc. Finally, according to the domain knowledge graph, the preprocessed enterprise data is semantically converted and labeled, and data fusion and association analysis are performed to finally obtain a unified enterprise data set.

[0097] Step S3: extract enterprise features from the unified enterprise data set according to the domain knowledge graph to obtain an enterprise feature vector;

[0098] In an embodiment of the present invention, first, based on the domain knowledge graph, label concepts with similar semantics or association relationships are classified into the same feature dimension, a multi-level feature dimension system is constructed, and a mapping relationship between labels and data features is established to form a label feature mapping table. Then, text features, numerical features, and category features are extracted from the unified enterprise data set, and text vectorization, numerical feature standardization, and category feature encoding are performed respectively. At the same time, an enterprise relationship graph is constructed based on the enterprise relationship data, and a graph embedding algorithm is used to extract enterprise graph feature vectors. Finally, all feature vectors are spliced ​​and dimensionally reduced to obtain the final enterprise feature vector.

[0099] Step S4: Perform feature clustering analysis on the enterprise feature vector to obtain an enterprise clustering graph; define cluster cluster labels on the enterprise clustering graph according to the domain knowledge graph to obtain a cluster cluster label set; perform dynamic label generation according to the cluster cluster label set to obtain a dynamic enterprise label library;

[0100] In an embodiment of the present invention, first, a suitable clustering algorithm is selected according to the characteristics of the enterprise feature vector, and the parameters of the clustering algorithm are optimized using the grid search method. Then, the enterprise feature vector is clustered using the optimized clustering algorithm and parameters, and the clustering results are visualized using a data visualization tool to form an enterprise clustering diagram. Next, based on the center distance of the enterprise in the cluster cluster, a typical enterprise is selected for manual labeling to form a manually labeled sample set. The label propagation model is trained using the manually labeled sample set, and label prediction is performed for all enterprises. Low-confidence labels are filtered out, and high-confidence labels are refined according to the domain knowledge graph, and finally a cluster cluster label set is obtained. Finally, an initial label is assigned to the enterprise according to the cluster cluster to which it belongs, and the initial label is dynamically updated according to the enterprise's real-time behavior data stream and a preset label dynamic update rule library to form a dynamic enterprise label library.

[0101] Step S5: construct a tag semantic association network based on the dynamic enterprise tag library and the domain knowledge graph to obtain an initial tag semantic association network; construct an enterprise tag semantic network based on the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network;

[0102] In an embodiment of the present invention, first, the tags in the dynamic enterprise tag library are matched with the concept nodes in the domain knowledge graph to establish a tag concept mapping table. For tags that are not successfully matched, text mining is performed using an external knowledge base to try to associate them with existing concept nodes in the domain knowledge graph to form an extended tag concept mapping table. Then, based on the extended tag concept mapping table, the semantic similarity between tags is calculated using a word vector model to construct a tag semantic similarity matrix. According to a preset semantic similarity threshold, semantically similar tags are connected to construct an initial tag semantic association network. Finally, based on the tag co-occurrence relationship in the dynamic enterprise tag library, a tag co-occurrence network is constructed, and the initial tag semantic association network is fused and optimized with the tag co-occurrence network, and finally an enterprise tag semantic network is constructed according to the optimized network structure.

[0103] Step S6: Acquire portrait tag application scenario data; identify the optimization direction of the tag system according to the portrait tag application scenario data and the enterprise tag semantic network to obtain tag system optimization data; optimize the enterprise portrait tag of the enterprise tag semantic network according to the tag system optimization data to obtain enterprise portrait tag data, so as to realize the intelligent generation of enterprise portrait tags;

[0104] In an embodiment of the present invention, first, the application requirements of enterprise portrait tags in different business scenarios are collected to form a structured application scenario description document, and the specific requirements of the tag system in different application scenarios are extracted to form application scenario requirement data. Then, based on the application scenario requirement data and the enterprise tag semantic network, a tag application strategy is formulated, including a tag selection strategy, a tag combination strategy, a tag weight strategy, etc. Next, according to the tag application strategy, the enterprise portrait tag is applied to a specific business scenario, and the business effect is tracked and monitored, and an evaluation index system is constructed according to the application scenario requirement data. According to the business effect monitoring data and the evaluation index system, the application effect of the tag system in different application scenarios is evaluated to form a tag system evaluation report. Finally, the tag system evaluation report is analyzed to identify the direction in which the tag system needs to be improved, and the enterprise tag semantic network is optimized according to the optimization direction, including adding new tags, adjusting tag relationships, updating tag weights, etc., to realize the intelligent generation of enterprise portrait tags.

[0105] Preferably, step S1 comprises the following steps:

[0106] Step S11: Acquire industry standard reference data; construct a label corpus based on the industry standard reference data to obtain a label corpus;

[0107] Step S12: performing tag concept mining on the tag corpus to obtain a tag concept set;

[0108] Step S13: extracting concept relations from the tag corpus according to the tag concept set to obtain concept relation triple data;

[0109] Step S14: Use the concepts in the label concept set as nodes and the relationships in the concept relationship triple data as edges to construct a domain knowledge graph, and perform manual verification of the graph to obtain a domain knowledge graph.

[0110] In an embodiment of the present invention, first, industry classification standards and codes (such as "National Economic Industry Classification"), as well as relevant policies, regulations, research reports and other document materials of various industries are obtained from authoritative data sources such as the National Bureau of Statistics and industry association websites as industry standard reference data. Then, natural language processing tools (such as HanLP) are used to perform operations such as word segmentation, part-of-speech tagging, and named entity recognition on these documents to extract keywords related to enterprise attributes and business characteristics. These keywords constitute an initial tag set. Next, the Word2Vec model is used to train word vectors for the tags in the initial tag set based on the industry corpus, and the semantic similarity between the tags is calculated. Tags with similar semantics are merged to remove redundant and duplicate tags, and finally a streamlined tag corpus is formed. The LDA topic model is used to perform topic analysis on the tags in the tag corpus, and tags with similar semantics are classified under the same topic. Each topic represents a tag concept, such as "enterprise scale", "business risk", etc. By setting parameters such as the number of topics and the number of iterations, the granularity and accuracy of tag concept mining can be controlled. Finally, all tag concepts constitute a tag concept set. Based on the label concept set, the text in the label corpus is subjected to dependency syntactic analysis using a syntactic analysis tool (such as Stanford CoreNLP) to extract the semantic relationship between the labels. For example, analyzing the sentence "Artificial intelligence enterprises are facing huge market competition pressure" can extract the concept relationship triple: (artificial intelligence enterprises, facing, market competition pressure). After analyzing all the texts, concept relationship triple data containing rich semantic relationships can be obtained. The domain knowledge graph is constructed using the Neo4j graph database. Each label concept in the label concept set obtained in step S12 is added as a node to the graph database. Then, the concept relationship triple data obtained in step S13 is traversed, and the relationship in each triple is used as an edge to connect the corresponding two label concept nodes, and the relationship type label is added to the edge. For example, for the triple (artificial intelligence enterprise, facing, market competition pressure), an edge connecting the two nodes "artificial intelligence enterprise" and "market competition pressure" is added to the graph database, and the edge type is set to "facing". Finally, the constructed domain knowledge graph is manually verified, the wrong nodes and relationships are corrected, and the missing information is supplemented, and finally a complete domain knowledge graph is obtained.

[0111] Preferably, step S11 includes the following steps:

[0112] Step S111: Acquire industry standard reference data; construct a labeling system according to the industry standard reference data to obtain an initial labeling system;

[0113] Step S112: Perform enterprise competitive product analysis based on the initial label system to obtain excellent enterprise label system data;

[0114] Step S113: optimizing the initial label system using the excellent enterprise label system data to obtain a label system;

[0115] Step S114: construct a label corpus according to the label system to obtain a label corpus.

[0116] In an embodiment of the present invention, the "National Economic Industry Classification" standard is obtained from the website of the National Bureau of Statistics, and the code and description documents of the corresponding industry are downloaded. The Python crawler program is used to crawl the industry research reports, enterprise rankings and other web page information publicly released on the website of the industry association (such as the China Software Industry Association). Regular expressions are used to clean and extract the web page content to obtain keywords related to enterprise attributes, such as "registered capital", "business scope", "products and services", etc. These keywords are classified and sorted according to the industry to which they belong, and the initial label system is constructed in combination with the "National Economic Industry Classification" standard. For example, keywords such as "software development" and "system integration" are classified under the "software and information technology service industry". First, determine the target industry and competing companies. For example, select the "artificial intelligence" industry, and determine leading companies such as SenseTime and Megvii Technology as competitive analysis objects. Then, by visiting the official website of the competing company, checking the publicly released corporate reports, analyzing the competitive analysis reports released by industry research institutions, etc., collect the label information used by the competing company. For example, labels such as "smart city", "smart business", and "smart finance" can be collected from the "solution" page of the SenseTime official website. All collected tag information is sorted and deduplicated to form excellent enterprise tag system data. The Jaccard similarity algorithm is used to calculate the similarity between each tag in the initial tag system and all tags in the excellent enterprise tag system data. For tags with similarity lower than the preset threshold (e.g., 0.5), they are marked as tags to be optimized. For tags to be optimized, the following three optimization strategies can be adopted: (1) Tag replacement: Use the Word2Vec model to calculate the word vector similarity between the tag to be optimized and all tags in the excellent enterprise tag system data, and select the tag with the highest similarity for replacement. (2) Tag merging: Merge multiple tags to be optimized with similar semantics into a new tag. (3) Tag deletion: For tags to be optimized that are less relevant to business goals, directly delete them from the initial tag system. After the above optimization process, the final tag system is obtained. Based on the final tag system, the Scrapy framework is used to build a web crawler program to collect text data containing keywords in the tag system from enterprise information query platforms (e.g., Qichacha, Tianyancha), news websites, social media and other channels. For example, for the tag "artificial intelligence", a keyword list can be set, including "artificial intelligence", "AI", "machine learning", "deep learning", etc. The crawler automatically crawls relevant web page content from the target website based on the keyword list, and performs data cleaning and deduplication to finally form a tag corpus.

[0117] Preferably, step S21 includes the following steps:

[0118] Step S21: Identify data sources related to enterprise portrait labels based on the domain knowledge graph to obtain label data; classify the label data sources to obtain a classified data source list;

[0119] Step S22: Collect and store data according to the classified data source list to obtain the original enterprise data set;

[0120] Step S23: performing data preprocessing on the original enterprise data set to obtain a preprocessed enterprise data set;

[0121] Step S24: performing data conversion on the preprocessed enterprise data set according to the domain knowledge graph to obtain a converted enterprise data set; performing data mapping of a corresponding label system on the converted enterprise data set to obtain a structured enterprise data set;

[0122] Step S25: Perform data fusion and association on the structured enterprise data sets to obtain a unified enterprise data set.

[0123] In an embodiment of the present invention, first, all nodes (label concepts) and edges (conceptual relationships) in the domain knowledge graph are traversed. For each node, its semantic information is analyzed to identify the data source related to the label concept. For example, for the label concept "enterprise scale", the relevant data source may include fields such as "registered capital" and "number of employees" in the "enterprise basic information database". For each edge, its relationship type and connected nodes are analyzed to identify the data source that can reflect the relationship. For example, for an edge with a relationship type of "provide" and connecting nodes of "enterprise" and "product", the relevant data source may include the "product introduction" page of the "enterprise official website". All identified data sources are classified according to type, such as "enterprise internal database", "public data platform", "Internet public data", etc., and finally a classified data source list is formed. Based on the classified data source list obtained in step S21, different data collection methods are used for different types of data sources. For the internal database of the enterprise, the database is directly accessed through a database connection tool (such as Navicat) and SQL statements are executed to obtain data. For the public data platform, the API interface provided by it is called to obtain data. For Internet public data, a web crawler program is built using the Scrapy framework to crawl data from the target website. The collected data is cleaned and converted according to a predefined format and stored in a data warehouse to form an original enterprise data set. First, the original enterprise data set is cleaned, including removing duplicate data, missing value processing, and outlier processing. For example, a density-based clustering algorithm (DBSCAN) is used to identify and remove abnormal data points in the data set. Second, the data format is unified, such as converting date data to the "YYYY-MM-DD" format. Finally, the data is standardized, such as using the Min-Max standardization method to scale numerical data of different dimensions to the [0,1] interval. After the above preprocessing operations, a preprocessed enterprise data set is obtained. First, according to the conceptual relationship in the domain knowledge graph, the data in the preprocessed enterprise data set is semantically converted. For example, there is a relationship "subsidiary" in the domain knowledge graph. By matching the enterprise name and the "subsidiary" relationship, the parent company and subsidiary information stored in different data sources can be associated. Second, according to the predefined label system, the converted enterprise data is mapped to the corresponding label. For example, the "registered capital" field in the enterprise basic information database is mapped to the label "enterprise scale". After data conversion and label mapping, a structured enterprise data set is obtained. The data fusion algorithm is used to fuse the data with the same label from different data sources in the structured enterprise data set. For example, for the label "enterprise scale", the weighted average method can be used to perform weighted fusion of data from different data sources based on the reliability and data quality of the data source. At the same time, according to the conceptual relationship in the domain knowledge graph, the data between different labels are analyzed for association.For example, we can analyze the relationship between the label "enterprise size" and the label "profitability" to mine data patterns. After data fusion and association analysis, we can finally obtain a unified enterprise data set.

[0124] Preferably, step S3 comprises the following steps:

[0125] Step S31: define feature dimensions according to the domain knowledge graph to obtain a feature dimension system; associate label features with the feature dimension system and the domain knowledge graph to obtain a label feature mapping table;

[0126] Step S32: extracting text features from the unified enterprise data set to obtain an enterprise text feature set; performing text vectorization on the enterprise text feature set to obtain an enterprise text feature vector;

[0127] Step S33: extracting numerical features from the unified enterprise data set to obtain an enterprise numerical feature set; and standardizing the enterprise numerical feature set to obtain a standardized enterprise numerical feature set;

[0128] Step S34: performing category feature encoding on the unified enterprise data set to obtain a coded enterprise category feature set;

[0129] Step S35: construct an enterprise relationship graph based on the unified enterprise data set to obtain an enterprise relationship graph; extract enterprise graph features from the enterprise relationship graph to obtain an enterprise graph feature vector;

[0130] Step S36: According to the label feature mapping table, feature vectors of the enterprise text feature vector, the standardized enterprise numerical feature set, the coded enterprise category feature set and the enterprise graph feature vector are fused and feature vector dimension reduction is performed to obtain the enterprise feature vector.

[0131] In an embodiment of the present invention, first, the label concepts and the relationships between the concepts in the domain knowledge graph are analyzed, and the label concepts with similar semantics or association relationships are classified under the same feature dimension. For example, label concepts such as "registered capital", "number of employees", and "operating income" are classified under the feature dimension of "enterprise scale". In this way, all label concepts are divided into different feature dimensions to build a multi-level feature dimension system. Then, for each label in the feature dimension system, its data type and data feature are analyzed, and associated with the corresponding concepts and relationships in the domain knowledge graph, and a mapping relationship between the label and the data feature is established to form a label feature mapping table. For example, the label "enterprise scale" is mapped to data features such as "registered capital" (numeric type), "number of employees" (numeric type), and "enterprise profile" (text type). First, text type fields such as "enterprise profile", "business scope", and "product introduction" are extracted from the unified enterprise data set to form an enterprise text feature set. Then, the TF-IDF algorithm is used to calculate the weight of each word in the enterprise text feature set, and stop words and low-frequency words are filtered out. Finally, the Word2Vec model is used to convert each word into a fixed-dimensional word vector, and all word vectors are weighted averaged to obtain the enterprise text feature vector representing the enterprise text information. Numerical fields such as "registered capital", "number of employees", "operating income", "total profit", etc. are extracted from the unified enterprise data set to form the enterprise numerical feature set. Then, the Min-Max standardization method is used to standardize each feature in the enterprise numerical feature set, scale the feature value to the [0,1] interval, eliminate the influence of the dimensions between different features, and obtain the standardized enterprise numerical feature set. Category fields such as "industry", "enterprise type", "listing status", etc. are extracted from the unified enterprise data set. Each category feature is encoded using the one-hot encoding method and converted into a binary vector. For example, assuming that the "industry" feature has three categories: "manufacturing", "service industry", and "agriculture", then one-hot encoding is used to encode "manufacturing" as [1,0,0], "service industry" as [0,1,0], and "agriculture" as [0,0,1]. Finally, the encoded enterprise category feature set is obtained. Based on the enterprise relationship data in the unified enterprise data set, such as "investment relationship", "cooperative relationship", "competitive relationship", etc., an enterprise relationship graph is constructed. The nodes in the graph represent enterprises, and the edges represent the relationship between enterprises. Then, the enterprise relationship graph is analyzed using a graph embedding algorithm (such as DeepWalk), and each enterprise node is mapped to a low-dimensional vector space to obtain the enterprise graph feature vector. The enterprise graph feature vector can reflect the structural characteristics and importance of the enterprise in the relationship network.According to the label feature mapping table obtained in step S31, the enterprise text feature vector, standardized enterprise numerical feature set, coded enterprise category feature set and enterprise graph feature vector obtained in steps S32, S33, S34 and S35 are concatenated to form a high-dimensional feature vector. In order to reduce the feature dimension, the principal component analysis (PCA) is used to reduce the dimension of the high-dimensional feature vector, extract the main feature information, and obtain the final enterprise feature vector.

[0132] Preferably, step S4 comprises the following steps:

[0133] Step S41: Select a clustering algorithm according to the enterprise feature vector to obtain a selected clustering algorithm; optimize clustering parameters of the selected clustering algorithm to obtain clustering parameters;

[0134] Step S42: clustering the enterprise feature vectors using the clustering parameters and the selected clustering algorithm to obtain enterprise clustering result data; and visualizing the clustering result of the enterprise clustering result data to obtain an enterprise clustering graph;

[0135] Step S43: defining cluster labels for the enterprise cluster graph according to the domain knowledge graph to obtain a cluster label set;

[0136] Step S44: annotating enterprise dynamic labels according to enterprise clustering result data and clustering cluster label sets to obtain enterprise initial label sets;

[0137] Step S45: obtaining the enterprise real-time behavior data stream; performing real-time behavior data analysis on the enterprise real-time behavior data stream to obtain the enterprise key event sequence;

[0138] Step S46: updating the dynamic enterprise tag library of the enterprise initial tag set according to the enterprise key event sequence and the preset tag dynamic update rule library to obtain a dynamic enterprise tag library.

[0139] As an example of the present invention, refer to Figure 2 As shown, in this example, step S4 includes:

[0140] Step S41: Select a clustering algorithm according to the enterprise feature vector to obtain a selected clustering algorithm; optimize clustering parameters of the selected clustering algorithm to obtain clustering parameters;

[0141] In an embodiment of the present invention, first, a suitable clustering algorithm is selected based on factors such as the dimension, sparsity, and clustering target of the enterprise feature vector. For example, if the dimension of the enterprise feature vector is high and the data distribution is uneven, the density-based clustering algorithm DBSCAN can be selected; if the clustering results need to be obtained quickly and there are no special requirements for the cluster shape, the K-Means algorithm can be selected. After determining the clustering algorithm, the parameters of the selected clustering algorithm are optimized using the grid search method. For example, for the K-Means algorithm, it is necessary to determine the value of the number of clusters K. The clustering effect under different K values ​​can be evaluated by the elbow rule or the silhouette coefficient method, and the optimal K value can be selected as the clustering parameter.

[0142] Step S42: clustering the enterprise feature vectors using the clustering parameters and the selected clustering algorithm to obtain enterprise clustering result data; and visualizing the clustering result of the enterprise clustering result data to obtain an enterprise clustering graph;

[0143] In an embodiment of the present invention, the clustering algorithm and parameters determined in step S41 are used to perform cluster analysis on the enterprise feature vectors. For example, the K-Means algorithm is used to divide all enterprise feature vectors into different cluster clusters, and each cluster cluster represents a class of enterprises with similar characteristics. After clustering is completed, the cluster number to which each enterprise belongs and the coordinates of the center point of each cluster cluster can be obtained, and this information constitutes the enterprise clustering result data. Then, a data visualization tool (such as the Python matplotlib library) is used to visualize the enterprise clustering result data, such as drawing a scatter plot, distinguishing different clusters with different colors or shapes, and intuitively displaying the enterprise clustering results to form an enterprise clustering diagram.

[0144] Step S43: defining cluster labels for the enterprise cluster graph according to the domain knowledge graph to obtain a cluster label set;

[0145] In an embodiment of the present invention, first, on the enterprise clustering diagram, the data distribution characteristics of each cluster are observed, and the typical characteristics of the enterprises in the cluster are analyzed. Then, one or more representative labels are defined for each cluster in combination with the semantic information of the label concept in the domain knowledge graph. For example, if the enterprises in a cluster have the label of "high-tech enterprise", the label of "artificial intelligence" and the feature of high "R&D investment", the cluster can be defined as "high-tech innovative enterprise". The labels of all clusters are aggregated to form a cluster label set.

[0146] Step S44: annotating enterprise dynamic labels according to enterprise clustering result data and clustering cluster label sets to obtain enterprise initial label sets;

[0147] In the embodiment of the present invention, each enterprise is divided into a corresponding cluster according to the enterprise clustering result data obtained in step S42. Then, the labels in the cluster label set obtained in step S43 are assigned to all enterprises in the cluster as the initial labels of the enterprises. For example, all enterprises in the cluster "high-tech innovative enterprises" are labeled with "high-tech" and "innovative". In this way, the initial labeling of all enterprises is completed to form the initial label set of enterprises.

[0148] Step S45: obtaining the enterprise real-time behavior data stream; performing real-time behavior data analysis on the enterprise real-time behavior data stream to obtain the enterprise key event sequence;

[0149] In the embodiment of the present invention, the latest behavior data of the enterprise is obtained in real time from channels such as the enterprise's business system, user behavior logs, and external data sources, such as new product information released by the enterprise, user browsing behavior on the enterprise website, and interactive data of the enterprise on social media, etc. These data constitute the enterprise's real-time behavior data stream. Then, the enterprise's real-time behavior data stream is analyzed in real time, for example, using a complex event processing engine (CEP) to identify key events such as financing, mergers and acquisitions, and product releases that occur in the enterprise, and record the time, participants, impact, and other information of the event to form a key event sequence for the enterprise.

[0150] Step S46: updating the dynamic enterprise tag library of the enterprise initial tag set according to the enterprise key event sequence and the preset tag dynamic update rule library to obtain a dynamic enterprise tag library;

[0151] In an embodiment of the present invention, a tag dynamic update rule library is predefined to dynamically adjust enterprise tags according to key events of the enterprise. For example, a rule is set that "if a financing event occurs in the enterprise, the weight of its 'development potential' tag is increased." When step S45 identifies that a key event has occurred in the enterprise, the initial tag set of the enterprise is updated according to the preset tag dynamic update rule library. For example, if a large financing event occurs in enterprise A, according to the preset rules, the weight of the "development potential" tag of enterprise A can be increased, or new tags such as "capital favor" can be added to it. In this way, the dynamic update of enterprise tags is achieved, and finally a dynamic enterprise tag library is formed.

[0152] Preferably, step S43 includes the following steps:

[0153] Step S431: extract typical enterprises from the enterprise clustering result data to obtain a typical enterprise sample set; manually annotate the typical enterprise sample set to obtain a manually annotated sample set;

[0154] Step S432: training a label propagation model using a manually annotated sample set to obtain a label propagation model;

[0155] Step S433: using the label propagation model to perform full enterprise label prediction on the enterprise clustering result data to obtain enterprise label prediction data;

[0156] Step S434: using a preset confidence threshold to filter the enterprise tag prediction data by tag confidence to obtain high-confidence enterprise tags;

[0157] Step S435: Refine the high-confidence enterprise labels according to the domain knowledge graph to obtain a cluster label set.

[0158] In an embodiment of the present invention, first, for each cluster, the enterprises are sorted according to the center distance in the cluster, and several enterprises closest to the center point are selected as typical enterprises. For example, each cluster selects the top 5% of enterprises to form a typical enterprise sample set. Then, each enterprise in the typical enterprise sample set is manually labeled and labeled with one or more labels that best describe its characteristics. For example, experts can label it with "high-tech enterprise", "Internet finance", "artificial intelligence" and other labels based on the business scope, product services, market positioning and other information of the enterprise. The label information of all typical enterprises constitutes a manually labeled sample set. The manually labeled sample set is used as training data to train the label propagation model. The label propagation model is a graph-based semi-supervised learning algorithm that can propagate labels to unlabeled samples based on the label information of a small number of labeled samples. During the model training process, each enterprise is regarded as a node in the graph, and the similarity between enterprises is used as the weight of the edge. For example, the cosine similarity between enterprise feature vectors can be used as the weight of the edge. Through iterative calculation, the label information will be propagated along the high-weight edge, and finally the labels are predicted for all enterprises. Use the trained label propagation model to predict labels for all enterprises that have not been manually labeled in the enterprise clustering result data. The model predicts one or more labels for the enterprise based on its position in the graph and its similarity to the labeled enterprises, and gives the predicted probability of each label. This information constitutes the enterprise label prediction data. A label confidence threshold is pre-set, for example, 0.8. For the enterprise label prediction data obtained in step S433, filters out labels with a predicted probability lower than the confidence threshold, and only retains labels with a predicted probability higher than the confidence threshold as high-confidence enterprise labels. For example, if the label prediction results of enterprise A are: "High-tech Enterprise" (probability 0.9), "Internet Finance" (probability 0.7), "Artificial Intelligence" (probability 0.6), after setting the confidence threshold to 0.8, only the "High-tech Enterprise" label is retained. According to the concept hierarchy in the domain knowledge graph, the high-confidence enterprise labels obtained in step S434 are refined. For example, if the high-confidence enterprise label contains the label "artificial intelligence", the label can be refined into more specific labels, such as "machine learning", "deep learning", "computer vision", etc., based on the subordinate concepts of "artificial intelligence" in the domain knowledge graph. After the label refinement process, the label set of each cluster is finally obtained to form the cluster label set.

[0159] Preferably, step S5 comprises the following steps:

[0160] Step S51: performing tag concept mapping on the dynamic enterprise tag library and the domain knowledge graph to obtain a tag concept mapping table;

[0161] Step S52: Performing external knowledge base text mining on the tags that are not concept-mapped in the tag concept mapping table to obtain relevant tag concept data; performing unmatched tag processing on the tag concept mapping table according to the relevant tag concept data to obtain an extended tag concept mapping table;

[0162] Step S53: Calculate the label semantic similarity according to the extended label concept mapping table and the domain knowledge graph to obtain a label semantic similarity matrix;

[0163] Step S54: constructing a label semantic association network for the label semantic similarity matrix using a preset semantic similarity threshold to obtain an initial label semantic association network;

[0164] Step S55: constructing an enterprise tag semantic network according to the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network.

[0165] As an example of the present invention, refer to Figure 3 As shown, in this example, step S5 includes:

[0166] Step S51: performing tag concept mapping on the dynamic enterprise tag library and the domain knowledge graph to obtain a tag concept mapping table;

[0167] In an embodiment of the present invention, all tags in the dynamic enterprise tag library are traversed, and each tag is matched with a tag concept in the domain knowledge graph. For example, the tag "artificial intelligence" is mapped to the "artificial intelligence" concept node in the domain knowledge graph. If the tag can find the corresponding concept node in the domain knowledge graph, a mapping relationship between the tag and the concept node is established, and the mapping result is stored in the tag concept mapping table. For example, a record in the tag concept mapping table is: ("artificial intelligence", artificial intelligence concept node).

[0168] Step S52: Performing external knowledge base text mining on the tags that are not concept-mapped in the tag concept mapping table to obtain relevant tag concept data; performing unmatched tag processing on the tag concept mapping table according to the relevant tag concept data to obtain an extended tag concept mapping table;

[0169] In an embodiment of the present invention, for labels in the label concept mapping table that cannot be successfully mapped with concept nodes in the domain knowledge graph, text mining is performed using an external knowledge base. For example, the Baidu Encyclopedia API interface is used to query the encyclopedia entry of the label, and the summary information, related concepts, classification information, etc. in the entry are extracted as relevant label concept data. Then, based on the relevant label concept data, an attempt is made to associate the unmapped label with the existing concept nodes in the domain knowledge graph. For example, if the unmapped label is "metaverse", through Baidu Encyclopedia query, its related concepts can be obtained as "virtual reality", "augmented reality", "blockchain", etc. Based on these related concepts, the "metaverse" label can be associated with the corresponding concept nodes in the domain knowledge graph, and the association results can be updated to the label concept mapping table to form an extended label concept mapping table.

[0170] Step S53: Calculate the label semantic similarity according to the extended label concept mapping table and the domain knowledge graph to obtain a label semantic similarity matrix;

[0171] In an embodiment of the present invention, the semantic similarity between tags is calculated based on the extended tag concept mapping table using the Word2Vec model. First, all concept nodes in the domain knowledge graph and all tags in the extended tag concept mapping table are regarded as words to form a dictionary. Then, the domain corpus is used to train word vectors for all words in the dictionary to obtain a word vector representation for each word. Finally, the cosine similarity between any two word vectors is calculated as the semantic similarity of the two words, and the semantic similarity of all words is stored in the tag semantic similarity matrix.

[0172] Step S54: constructing a label semantic association network for the label semantic similarity matrix using a preset semantic similarity threshold to obtain an initial label semantic association network;

[0173] In an embodiment of the present invention, a semantic similarity threshold is preset, for example, 0.7. The label semantic similarity matrix is ​​traversed, and if the semantic similarity between two labels is higher than the preset threshold, an edge connecting the two labels is added to the label semantic association network. For example, if the semantic similarity between the label "artificial intelligence" and the label "machine learning" is 0.8, which is higher than the preset threshold of 0.7, an edge connecting the two labels is added to the network. In this way, the initial label semantic association network is finally constructed.

[0174] Step S55: constructing an enterprise tag semantic network according to the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network;

[0175] In an embodiment of the present invention, all enterprises in the dynamic enterprise tag library are traversed to obtain a tag list corresponding to each enterprise. According to the initial tag semantic association network, the tags in the enterprise tag list are associated to construct an enterprise tag semantic network. For example, if the tag list of enterprise A is: "artificial intelligence", "machine learning", and "deep learning", then according to the initial tag semantic association network, these three tags can be connected to form a sub-network representing the semantic relationship of enterprise A tags. By integrating the tag semantic sub-networks of all enterprises, a complete enterprise tag semantic network is formed.

[0176] Preferably, step S55 comprises the following steps:

[0177] Step S551: mining tag co-occurrence relationships on the dynamic enterprise tag library to obtain a tag co-occurrence relationship matrix;

[0178] Step S552: constructing a tag co-occurrence network for the tag co-occurrence relationship matrix using a preset co-occurrence strength threshold to obtain a tag co-occurrence network;

[0179] Step S553: ​​fusing the initial tag semantic association network and the tag co-occurrence network into a semantic association network and a co-occurrence network to obtain a fused tag semantic network;

[0180] Step S554: performing semantic network optimization on the fused label semantic network to obtain an optimized label semantic network;

[0181] Step S555: Generate an enterprise tag semantic network based on the optimized tag semantic network to obtain the enterprise tag semantic network.

[0182] In an embodiment of the present invention, first, a tag co-occurrence matrix is ​​constructed to store the frequency of co-occurrence of any two tags in a dynamic enterprise tag library. All enterprises in the dynamic enterprise tag library are traversed, and the number of co-occurrences of two tags in each enterprise tag list is counted, and the statistical results are updated to the tag co-occurrence matrix. For example, if the tag list of enterprise A is: "artificial intelligence", "machine learning", and the tag list of enterprise B is: "artificial intelligence", "deep learning", then the number of co-occurrences of the tag "artificial intelligence" and the tag "machine learning" is 1, and the number of co-occurrences of the tag "artificial intelligence" and the tag "deep learning" is 1. After the statistics are completed, the tag co-occurrence matrix is ​​normalized, and the number of co-occurrences is converted into co-occurrence probabilities to compare the co-occurrence strengths between different tags. A co-occurrence strength threshold is pre-set, for example, 0.1. The tag co-occurrence relationship matrix is ​​traversed, and if the co-occurrence probability between two tags is higher than the preset threshold, an edge connecting the two tags is added to the tag co-occurrence network, and the weight of the edge is the co-occurrence probability. For example, if the co-occurrence probability between the label "artificial intelligence" and the label "machine learning" is 0.2, which is higher than the preset threshold value of 0.1, an edge connecting the two labels is added to the label co-occurrence network, and the weight of the edge is 0.2. In this way, a label co-occurrence network is constructed, and the edges in the network indicate the co-occurrence relationship between the labels, and the weight of the edge indicates the co-occurrence strength. The initial label semantic association network obtained in step S54 is fused with the label co-occurrence network obtained in step S552 to generate a fused label semantic network. The fusion method can adopt a weighted average method, that is, different weights are assigned to the initial label semantic association network and the label co-occurrence network according to different network types, and then the weights of the corresponding edges in the two networks are weighted averaged to obtain the weights of the corresponding edges in the fused label semantic network. For example, a weight of 0.7 can be assigned to the initial label semantic association network, and a weight of 0.3 can be assigned to the label co-occurrence network. Optimize the fused label semantic network obtained in step S553, for example: remove isolated nodes, for nodes with highly similar semantics, such as "artificial intelligence" and "AI", they can be merged into one node, and the edges connected to them can be merged; merge semantically repeated nodes, for nodes with highly similar semantics, such as "artificial intelligence" and "AI", they can be merged into one node, and the edges connected to them can be merged; adjust the weight of the edge, according to the weight distribution of the edge, adjust the weight of the edge, for example, delete the edge with too low weight, reduce the weight of the edge with too high weight, etc. Traverse all enterprises in the dynamic enterprise label library and obtain the label list corresponding to each enterprise. According to the optimized label semantic network, associate the labels in the enterprise label list to build an enterprise label semantic network.For example, if the tag list of enterprise A is: "artificial intelligence", "machine learning", and "deep learning", the path connecting these three tags can be found according to the optimized tag semantic network, and the semantic distance between the three tags can be calculated according to the edge weights on the path, and finally a sub-network representing the semantic relationship of the tags of enterprise A can be constructed. Integrating the tag semantic sub-networks of all enterprises will form a complete enterprise tag semantic network.

[0183] Preferably, step S6 comprises the following steps:

[0184] Step S61: Acquire portrait tag application scenario data; perform application scenario demand analysis on the portrait tag application scenario data to obtain application scenario demand data;

[0185] Step S62: Formulate a label application strategy based on the application scenario requirement data and the enterprise label semantic network to obtain a label application strategy;

[0186] Step S63: tracking and monitoring the business effect according to the tag application strategy to obtain business effect monitoring data; constructing an evaluation index system according to the application scenario demand data to obtain an evaluation index system;

[0187] Step S64: perform label system effect evaluation based on the business effect monitoring data and the evaluation index system to obtain a label system evaluation report;

[0188] Step S65: identifying the label system optimization direction based on the label system evaluation report to obtain label system optimization data;

[0189] Step S66: Optimize the enterprise portrait label of the enterprise label semantic network according to the label system optimization data to obtain the enterprise portrait label data to realize the intelligent generation of enterprise portrait labels.

[0190] In an embodiment of the present invention, the application requirements of enterprise portrait tags in different business scenarios are collected, such as precision marketing, risk control, customer relationship management and other scenarios. For each application scenario, its business objectives, target users, data requirements and other information are clarified to form a structured application scenario description document. For example, for precision marketing scenarios, it is necessary to clarify the target customer group, marketing products or services, available marketing channels and other information. All application scenario description documents are analyzed to extract the specific requirements of the tag system in different application scenarios, such as tag coverage, tag accuracy, tag real-time, etc., to form application scenario demand data. According to the application scenario demand data obtained in step S61 and the enterprise tag semantic network obtained in step S55, a tag application strategy is formulated, including a tag selection strategy, a tag combination strategy, a tag weight strategy, etc. For example, for precision marketing scenarios, tags that are highly relevant to the target customer group can be selected from the enterprise tag semantic network according to the characteristics of the target customer group, and multiple tags can be combined and used according to the semantic similarity and co-occurrence relationship between the tags to improve the recognition accuracy of the target customer group. At the same time, different weights can be set for the tags according to their importance in different application scenarios, so that the tags can be prioritized in subsequent applications. According to the label application strategy formulated in step S62, the enterprise portrait label is applied to a specific business scenario, and the business effect is tracked and monitored. For example, after the enterprise portrait label is used in the precision marketing scenario, it is necessary to record the number of times each label is used in the marketing activity, the conversion rate of the target customer, the return on investment of the marketing activity, and other data as business effect monitoring data. At the same time, according to the application scenario demand data obtained in step S61, an evaluation index system is constructed to evaluate the application effect of the label system in different application scenarios. For example, for the precision marketing scenario, indicators such as the conversion rate of the target customer and the return on investment of the marketing activity can be included in the evaluation index system. According to the business effect monitoring data collected in step S63 and the evaluation index system constructed in step S63, the application effect of the label system in different application scenarios is evaluated to form a label system evaluation report. For example, the average conversion rate of each label in different marketing activities and the impact of different label combinations on the return on investment of the marketing activity can be counted, and the advantages and disadvantages of the label system in meeting the precision marketing needs can be analyzed. The label system evaluation report needs to comprehensively and objectively reflect the application effect of the label system and provide a basis for the subsequent optimization of the label system. The label system evaluation report obtained in step S64 is analyzed to identify the areas where the label system needs to be improved in terms of label coverage, label accuracy, label real-time performance, etc., to form label system optimization data.For example, if the evaluation report shows that the enterprise tag coverage rate of certain industries is low, it is necessary to supplement the relevant tag data for these industries; if the prediction accuracy of some tags is low, it is necessary to analyze the reasons and optimize the tag generation model; if the tag update frequency cannot meet the needs of real-time marketing, it is necessary to optimize the tag dynamic update mechanism. According to the tag system optimization data obtained in step S65, the enterprise tag semantic network is optimized, including: adding new tags, supplementing the missing tag data according to the optimization direction of tag coverage, and adding the new tags to the enterprise tag semantic network; adjusting the tag relationship, according to the optimization direction of tag accuracy, adjusting the semantic similarity and co-occurrence relationship between tags, such as deleting erroneous tag associations, adding new tag associations, etc.; updating tag weights, according to the optimization direction of tag real-time, adjusting the weights of tags in different application scenarios, such as increasing the weights of tags with higher real-time requirements and reducing the weights of tags with lower real-time requirements, etc.

[0191] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0192] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for intelligently generating enterprise portrait labels, characterized in that: The following steps are involved: Step S1: Obtain industry standard reference data; The labeled corpus is constructed and processed according to the industry standard reference data to obtain the labeled corpus; the domain knowledge modeling is performed according to the labeled corpus to obtain the domain knowledge graph; Step S2: Analyze the original enterprise data according to the domain knowledge graph to obtain a preprocessed enterprise data set; According to the domain knowledge graph, the pre-processed enterprise data set is fused and associated to obtain a unified enterprise data set; Step S3: extract enterprise features from the unified enterprise data set according to the domain knowledge graph to obtain an enterprise feature vector; Step S4: Perform feature clustering analysis on the enterprise feature vector to obtain an enterprise clustering graph; define cluster cluster labels on the enterprise clustering graph according to the domain knowledge graph to obtain a cluster cluster label set; perform dynamic label generation according to the cluster cluster label set to obtain a dynamic enterprise label library; Step S5: construct a tag semantic association network based on the dynamic enterprise tag library and the domain knowledge graph to obtain an initial tag semantic association network; construct an enterprise tag semantic network based on the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network; Step S6: Obtaining portrait tag application scenario data; Based on the portrait label application scenario data and the enterprise label semantic network, the label system optimization direction is identified to obtain the label system optimization data; based on the label system optimization data, the enterprise label semantic network is optimized to obtain the enterprise portrait label data to realize the intelligent generation of enterprise portrait labels.

2. The method for intelligently generating enterprise portrait labels according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Acquire industry standard reference data; construct a label corpus based on the industry standard reference data to obtain a label corpus; Step S12: performing tag concept mining on the tag corpus to obtain a tag concept set; Step S13: extracting concept relations from the tag corpus according to the tag concept set to obtain concept relation triple data; Step S14: Use the concepts in the label concept set as nodes and the relationships in the concept relationship triple data as edges to construct a domain knowledge graph, and perform manual verification of the graph to obtain a domain knowledge graph.

3. The method for intelligently generating enterprise portrait labels according to claim 2, characterized in that: Step S11 includes the following steps: Step S111: Acquire industry standard reference data; construct a labeling system according to the industry standard reference data to obtain an initial labeling system; Step S112: Perform enterprise competitive product analysis based on the initial label system to obtain excellent enterprise label system data; Step S113: optimizing the initial label system using the excellent enterprise label system data to obtain a label system; Step S114: construct a label corpus according to the label system to obtain a label corpus.

4. The method for intelligently generating enterprise portrait labels according to claim 1, characterized in that: Step S21 includes the following steps: Step S21: Identify data sources related to enterprise portrait labels based on the domain knowledge graph to obtain label data; classify the label data sources to obtain a classified data source list; Step S22: Collect and store data according to the classified data source list to obtain the original enterprise data set; Step S23: performing data preprocessing on the original enterprise data set to obtain a preprocessed enterprise data set; Step S24: performing data conversion on the preprocessed enterprise data set according to the domain knowledge graph to obtain a converted enterprise data set; performing data mapping of a corresponding label system on the converted enterprise data set to obtain a structured enterprise data set; Step S25: Perform data fusion and association on the structured enterprise data sets to obtain a unified enterprise data set.

5. The method for intelligently generating enterprise portrait labels according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: define feature dimensions according to the domain knowledge graph to obtain a feature dimension system; associate label features with the feature dimension system and the domain knowledge graph to obtain a label feature mapping table; Step S32: extracting text features from the unified enterprise data set to obtain an enterprise text feature set; performing text vectorization on the enterprise text feature set to obtain an enterprise text feature vector; Step S33: extracting numerical features from the unified enterprise data set to obtain an enterprise numerical feature set; and standardizing the enterprise numerical feature set to obtain a standardized enterprise numerical feature set; Step S34: performing category feature encoding on the unified enterprise data set to obtain a coded enterprise category feature set; Step S35: construct an enterprise relationship graph based on the unified enterprise data set to obtain an enterprise relationship graph; extract enterprise graph features from the enterprise relationship graph to obtain an enterprise graph feature vector; Step S36: According to the label feature mapping table, feature vectors of the enterprise text feature vector, the standardized enterprise numerical feature set, the coded enterprise category feature set and the enterprise graph feature vector are fused and feature vector dimension reduction is performed to obtain the enterprise feature vector.

6. The method for intelligently generating enterprise portrait labels according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: Select a clustering algorithm according to the enterprise feature vector to obtain a selected clustering algorithm; optimize clustering parameters of the selected clustering algorithm to obtain clustering parameters; Step S42: clustering the enterprise feature vectors using the clustering parameters and the selected clustering algorithm to obtain enterprise clustering result data; and visualizing the clustering result of the enterprise clustering result data to obtain an enterprise clustering graph; Step S43: defining cluster labels for the enterprise cluster graph according to the domain knowledge graph to obtain a cluster label set; Step S44: annotating enterprise dynamic labels according to enterprise clustering result data and clustering cluster label sets to obtain enterprise initial label sets; Step S45: obtaining the enterprise real-time behavior data stream; performing real-time behavior data analysis on the enterprise real-time behavior data stream to obtain the enterprise key event sequence; Step S46: updating the dynamic enterprise tag library of the enterprise initial tag set according to the enterprise key event sequence and the preset tag dynamic update rule library to obtain a dynamic enterprise tag library.

7. The method for intelligently generating enterprise portrait labels according to claim 6, characterized in that: Step S43 includes the following steps: Step S431: extract typical enterprises from the enterprise clustering result data to obtain a typical enterprise sample set; manually annotate the typical enterprise sample set to obtain a manually annotated sample set; Step S432: training a label propagation model using a manually annotated sample set to obtain a label propagation model; Step S433: using the label propagation model to perform full enterprise label prediction on the enterprise clustering result data to obtain enterprise label prediction data; Step S434: using a preset confidence threshold to filter the enterprise tag prediction data by tag confidence to obtain high-confidence enterprise tags; Step S435: Refine the high-confidence enterprise labels according to the domain knowledge graph to obtain a cluster label set.

8. The method for intelligently generating enterprise portrait labels according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: performing tag concept mapping on the dynamic enterprise tag library and the domain knowledge graph to obtain a tag concept mapping table; Step S52: Performing external knowledge base text mining on the tags that are not concept-mapped in the tag concept mapping table to obtain relevant tag concept data; performing unmatched tag processing on the tag concept mapping table according to the relevant tag concept data to obtain an extended tag concept mapping table; Step S53: Calculate the label semantic similarity according to the extended label concept mapping table and the domain knowledge graph to obtain a label semantic similarity matrix; Step S54: constructing a label semantic association network for the label semantic similarity matrix using a preset semantic similarity threshold to obtain an initial label semantic association network; Step S55: constructing an enterprise tag semantic network according to the initial tag semantic association network and the dynamic enterprise tag library to obtain an enterprise tag semantic network.

9. The method for intelligently generating enterprise portrait labels according to claim 8, characterized in that: Step S55 includes the following steps: Step S551: mining tag co-occurrence relationships on the dynamic enterprise tag library to obtain a tag co-occurrence relationship matrix; Step S552: constructing a tag co-occurrence network for the tag co-occurrence relationship matrix using a preset co-occurrence strength threshold to obtain a tag co-occurrence network; Step S553: ​​fusing the initial tag semantic association network and the tag co-occurrence network into a semantic association network and a co-occurrence network to obtain a fused tag semantic network; Step S554: performing semantic network optimization on the fused label semantic network to obtain an optimized label semantic network; Step S555: Generate an enterprise tag semantic network based on the optimized tag semantic network to obtain the enterprise tag semantic network.

10. The method for intelligently generating enterprise portrait labels according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: Acquire portrait tag application scenario data; perform application scenario demand analysis on the portrait tag application scenario data to obtain application scenario demand data; Step S62: Formulate a label application strategy based on the application scenario requirement data and the enterprise label semantic network to obtain a label application strategy; Step S63: tracking and monitoring the business effect according to the tag application strategy to obtain business effect monitoring data; constructing an evaluation index system according to the application scenario demand data to obtain an evaluation index system; Step S64: perform label system effect evaluation based on the business effect monitoring data and the evaluation index system to obtain a label system evaluation report; Step S65: identifying the label system optimization direction based on the label system evaluation report to obtain label system optimization data; Step S66: Optimize the enterprise portrait label of the enterprise label semantic network according to the label system optimization data to obtain the enterprise portrait label data to realize the intelligent generation of enterprise portrait labels.

Images