Policy recommendation method and system based on heterogeneous knowledge graph

Abstract

The invention discloses a policy recommendation method and system based on a heterogeneous knowledge graph, and relates to the field of policy recommendation, and the method comprises the steps: obtaining an original policy text from a plurality of heterogeneous policy data sources, and extracting policy-related information; mapping the policy-related information into policy entity nodes; mapping the enterprise portrait data into enterprise entity nodes; defining and generating multiple types of edges between the policy entity nodes and the enterprise entity nodes based on a preset semantic relationship so as to construct a heterogeneous knowledge graph, constructing a graph neural network model, and executing embedding calculation on various nodes in the graph so as to obtain node representation vectors; and obtaining a matching degree, and outputting a recommended policy entry set according to the matching degree as policy recommendation information of the target enterprise. The method not only breaks through the problems of format dependence, rule stiffness and semantic shallow hierarchy of a traditional policy recommendation system, but also has relatively high universality, interpretability and actual deployment value.

Description

Technical Field

[0001] This specification relates to the field of policy recommendation, and more specifically, this application relates to a policy recommendation method and system based on heterogeneous knowledge graphs. Background Technology

[0002] Currently, SMEs generally face three core pain points in acquiring, understanding, and applying policies at all levels: "fragmented information," "obscure content," and "inaccurate matching." On the one hand, policy releases are heterogeneous, covering various formats such as HTML web pages, PDF announcements, and scanned documents, making it difficult for traditional information systems to achieve structured parsing and efficient unification. On the other hand, enterprise profile data is often distributed across multiple dimensions, including business registration information, performance records, R&D investment, and credit ratings. Existing policy matching methods typically rely on keyword or rule matching, lacking a deep understanding of the comprehensive semantic characteristics of enterprises and the logical conditions of policies. Furthermore, current policy recommendation methods generally lack structure awareness, failing to uncover contextual dependencies and cross-document reasoning logic between policy content, resulting in recommendation results with insufficient accuracy, poor interpretability, and weak generalization ability.

[0003] Some existing technologies have attempted to introduce machine learning-based text classification or rule-based enhanced matching mechanisms, but they still remain at the level of flat text processing and one-way matching, lacking unified multi-source information modeling and knowledge association mapping capabilities, making it difficult to support policy intelligent push tasks in complex scenarios.

[0004] Therefore, it is necessary to propose a policy recommendation method and system based on heterogeneous knowledge graphs to at least solve some of the above-mentioned problems. Summary of the Invention

[0005] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. This summary section is not intended to limit the key and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0006] Firstly, this application proposes a policy recommendation method based on heterogeneous knowledge graphs, including: Obtain the original policy text from multiple heterogeneous policy data sources; Based on the aforementioned original policy texts, relevant policy information was extracted; Obtain enterprise profile data of the target company; Map the above policy-related information to policy entity nodes; Map the above enterprise profile data to enterprise entity nodes; Based on the preset semantic relationship definition, generate multi-type edges between the above policy entity nodes and the above enterprise entity nodes to construct a heterogeneous knowledge graph; Based on the aforementioned heterogeneous knowledge graph, a graph neural network model is constructed to perform embedding computation on various types of nodes in the graph to obtain node representation vectors. These node representation vectors include the node representations of the target enterprise and the node representations of each policy entity. The similarity between the node representations of the target enterprises and the node representations of the policy entities is calculated to obtain the matching degree. Based on the matching degree, a set of recommended policy items is output as policy recommendation information for the aforementioned target companies.

[0007] In one feasible implementation, the above-mentioned extraction of policy-related information based on multiple original policy texts includes: Obtain the format information of the original policy text; Given that the original policy text is in HTML format, the first policy text is extracted based on a text extraction algorithm that integrates text density analysis, title position patterns, and visible area rules. Given that the original policy text is in PDF format, the second policy text is extracted by dividing logical blocks and reorganizing text content based on the layout structure recognition algorithm. The first policy text and the second policy text mentioned above are input into a pre-trained BERT text classification model for semantic discrimination in order to extract policy-related information.

[0008] In one feasible implementation, the above-mentioned text extraction algorithm based on the fusion of text density analysis, title position patterns, and visible area rules extracts the first policy text, including: The original policy text in HTML format is parsed, a DOM node tree is constructed, and the node attribute information of the DOM node tree is extracted. The node attributes include tag type information, text content length information, structural hierarchy information, and style attribute information. Based on the text content length information of each node, the text density value of each node is calculated, and the first set of candidate nodes for the main text is selected according to the preset density threshold. Based on the above tag type information, detect structural nodes with title features in the above DOM node tree, identify the neighboring node regions of the above title feature structural nodes based on the above structural hierarchy information, and mark the nodes containing high text density as the second body text candidate node set. Based on the above style attribute information, determine the visibility status of each node in the above original policy text page, remove nodes with hidden attributes or in non-visible areas, and obtain the third set of candidate nodes for the main text. Based on the aforementioned first set of candidate nodes, second set of candidate nodes, and third set of candidate nodes, a set fusion and sorting process is performed. The candidate nodes are comprehensively scored according to a preset fusion scoring function, and one or more nodes with the highest scores are selected as the aforementioned first policy text.

[0009] In one feasible implementation, the above-mentioned logical block division and text content reorganization based on the layout structure recognition algorithm to extract the second policy text includes: Obtain the original policy text in PDF format, and perform structured parsing on the original policy text to extract page coordinate information and paragraph distribution information; Based on the page coordinate information and paragraph distribution information mentioned above, a logical block division operation is performed to obtain multiple page structure blocks, wherein each of the above structure blocks includes one or more text fragments with visual continuity and semantic consistency. Based on the positional relationship and text connection features of the above structural blocks, the above structural blocks are reordered to generate a rearranged text sequence. In the above-mentioned text sequence, the policy title, publication time, issuing agency, and main clause content of the original PDF policy text are extracted based on keyword template matching and position rules to form the second policy text.

[0010] In one feasible implementation, the first policy text and the second policy text are input into a pre-trained BERT text classification model for semantic discrimination to extract policy-related information, including: The first policy text and the second policy text mentioned above are subjected to text standardization and word segmentation encoding to construct a token ID sequence that conforms to the BERT input format; Input the above token ID sequence into the pre-trained BERT text classification model to extract the global semantic vector corresponding to each text. Based on the aforementioned global semantic vector, multi-class prediction is performed, and the semantic category label of each text is output. Based on the semantic category tags mentioned above, text content belonging to the policy category is selected as the extracted policy-related information.

[0011] In one feasible implementation, the above-mentioned generation of multiple types of edges between the policy entity nodes and the enterprise entity nodes based on a preset semantic relationship definition, in order to construct a heterogeneous knowledge graph, includes: Based on the attribute information of the aforementioned policy entity nodes and the profile information of the aforementioned enterprise entity nodes, a set of semantic relationship definitions is constructed. Attribute matching is performed between the aforementioned enterprise entity nodes and the aforementioned policy entity nodes. Based on the preset semantic relationships, one or more edge relationships describing their semantic associations are established to generate multi-type edges in the heterogeneous graph. The aforementioned edge relationships are represented in the form of triples, with additional edge weight information and credibility information, to construct the heterogeneous knowledge graph.

[0012] In one feasible implementation, the steps for determining the aforementioned edge weight information and the aforementioned credibility information include: Calculate the attribute matching score, behavior co-occurrence score, and semantic relevance score between the aforementioned enterprise entity nodes and the aforementioned policy entity nodes; Based on the preset edge type confidence prior, the fusion score is calculated as the edge weight information according to the above attribute matching score, above behavior co-occurrence score and above semantic relevance score of each edge. Based on the edge weight information, data source credibility, and edge type confidence coefficient mentioned above, the credibility information of the edge is calculated.

[0013] In one feasible implementation, the above-mentioned similarity calculation between the node representation of the target enterprise and the node representation of each policy entity to obtain a matching degree includes: Based on the node representation vectors of the target enterprises and the representation vectors of the policy entity nodes, a semantic similarity score is calculated. Based on the structural information in the heterogeneous knowledge graph, semantic path relationship features between the target enterprise node and each policy entity node are extracted, and path relationship scores are calculated. The path relationship features include path length, path type weight, and path existence. Obtain the enterprise clustering results of the target enterprise, count the historical recommendation or matching frequency of other enterprise nodes in the cluster to the policy entity node, and calculate the cluster relevance score; Based on the semantic similarity score, the path relationship score, and the clustering relevance score, a weighted fusion is performed to generate the final matching degree between the target enterprise and each of the policy entity nodes.

[0014] In one feasible implementation, the set of recommended policy items output based on the matching degree, as the policy recommendation information for the target enterprise, includes: Based on the above matching degree, the candidate policies are initially ranked; Based on the preliminary ranking results, a multi-objective ranking strategy is further introduced to comprehensively score and optimize the ranking of candidate policies. The multi-objective ranking strategy includes at least one of the following: setting ranking weights based on the enterprise's growth stage to improve the scores of policy items that are more suitable for the current stage; reducing the ranking priority of complex policies in early-stage enterprise recommendations based on the difficulty of policy application, approval cycle, or acceptance threshold; and dynamically fine-tuning the ranking results by combining the enterprise's historical policy preferences and the application behavior trends of similar enterprises. Based on the comprehensive ranking results above, the top K items with the highest scores are selected to output a set of recommended policy items, which serve as the policy recommendation information for the target companies.

[0015] Secondly, this application proposes a policy recommendation system based on heterogeneous knowledge graphs, including: The first acquisition unit is used to acquire the original policy text from multiple heterogeneous policy data sources; The extraction unit is used to extract policy-related information based on multiple original policy texts mentioned above. The second acquisition unit is used to acquire enterprise profile data of the target enterprise; The first mapping unit is used to map the above policy-related information into policy entity nodes; The second mapping unit is used to map the above-mentioned enterprise profile data into enterprise entity nodes; The first construction unit is used to generate multiple types of edges between the aforementioned policy entity nodes and the aforementioned enterprise entity nodes based on a preset semantic relationship definition, so as to construct a heterogeneous knowledge graph. The second building unit is used to build a graph neural network model based on the above heterogeneous knowledge graph, and to perform embedding calculations on various types of nodes in the graph to obtain node representation vectors, wherein the above node representation vectors include the node representations of the target enterprise and the node representations of each policy entity; The calculation unit is used to calculate the similarity between the node representation of the target enterprise and the node representation of each policy entity to obtain the matching degree. The recommendation unit is used to output a set of recommended policy items based on the matching degree, which serve as policy recommendation information for the target enterprise.

[0016] In summary, this embodiment proposes a policy recommendation method based on heterogeneous knowledge graphs. It supports differentiated structural parsing of HTML policy web pages and PDF policy documents. By integrating text density analysis, DOM structure features, and layout logic reorganization, it automatically extracts the policy text content and core element fields, providing a unified input data format for subsequent semantic modeling and knowledge graph construction. By mapping policy content and enterprise profiles to semantic nodes, and generating multi-type edge relationships based on predefined multi-dimensional semantic relationships (such as regional adaptation, industry adaptation, and funding thresholds), a clear and semantically rich heterogeneous knowledge graph is constructed, significantly enhancing the semantic and structural alignment capabilities between policies and enterprises. Compared to traditional rule-based or flat vector-based recommendation methods, this solution introduces a graph neural network model to learn context-aware vector representations of nodes in the graph. This allows the representation of each enterprise node and policy node to comprehensively consider neighbor structure, edge type, and attribute semantics, improving the accuracy and interpretability of recommendation matching. This solution incorporates structural information such as "path relationship features" and "enterprise clustering results" as auxiliary factors during similarity calculation. This allows recommendations to be based not only on node semantic similarity but also on the graph structure connections and group behavior characteristics between enterprises and policies, thereby improving the personalization and stability of the recommendation results. When outputting recommendation results, the system supports a multi-objective fusion ranking mechanism based on enterprise development stage, policy difficulty, and historical preferences, achieving highly controllable and configurable strategy push paths that adapt to various government and enterprise service scenarios. In summary, this embodiment not only overcomes the format dependency, rule rigidity, and shallow semantic issues of traditional policy recommendation systems but also possesses high versatility, interpretability, and practical deployment value. It is particularly suitable for SME policy support scenarios, regional industrial guidance path analysis, and multi-level government service systems, demonstrating significant social benefits and technological innovation value. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a policy recommendation method based on heterogeneous knowledge graphs is provided for embodiments of this application. Figure 2 This application provides a flowchart illustrating a method for extracting policy-related information in an embodiment of the present application. Figure 3 This application provides a flowchart illustrating a method for extracting a first policy text in an embodiment of the present application. Figure 4This application provides a flowchart illustrating a method for extracting a second policy text. Figure 5 A flowchart illustrating a method for extracting policy-related information based on a pre-trained BERT text classification model, provided in this application embodiment; Figure 6 This application provides a flowchart illustrating a method for constructing a heterogeneous knowledge graph. Figure 7 This is a flowchart illustrating a method for determining edge weight information and the aforementioned credibility information, provided in an embodiment of this application. Figure 8 This is a flowchart illustrating a method for obtaining matching degree according to an embodiment of this application. Figure 9 This application provides a flowchart illustrating a method for determining policy recommendation information in an embodiment. Figure 10 This is a schematic diagram of a policy recommendation device based on heterogeneous knowledge graphs, provided as an embodiment of this application. Detailed Implementation

[0018] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0019] Figure 1 This application provides a flowchart illustrating a policy recommendation method based on heterogeneous knowledge graphs, which may specifically include: S110. Obtain the original policy text from multiple heterogeneous policy data sources; S120. Based on the above-mentioned original policy texts, extract relevant policy information; S130. Obtain the target company's corporate profile data; S140. Map the above policy-related information to policy entity nodes; S150. Map the above enterprise profile data to enterprise entity nodes; S160. Generate multi-type edges between the above-mentioned policy entity nodes and the above-mentioned enterprise entity nodes based on the preset semantic relationship definition, so as to construct a heterogeneous knowledge graph. S170. Based on the above heterogeneous knowledge graph, construct a graph neural network model, perform embedding calculations on various types of nodes in the graph to obtain node representation vectors, wherein the above node representation vectors include the node representations of the target enterprise and the node representations of each policy entity; S180. Calculate the similarity between the node representations of the target enterprises and the node representations of the policy entities to obtain the matching degree. S190. Output a set of recommended policy items based on the matching degree, as the policy recommendation information for the target enterprise mentioned above.

[0020] For example, this embodiment discloses a policy recommendation method based on heterogeneous knowledge graphs. Its core idea is to achieve intelligent policy push to target enterprises through heterogeneous graph modeling, graph neural network embedding, and multi-dimensional similarity calculation.

[0021] First, in step S110, the system connects to multiple heterogeneous policy data sources and automatically obtains the original policy texts. These data sources may include government websites, local industrial policy announcement systems, industry platforms, or policy white papers published in PDF format. The obtained policy texts are in various formats, including HTML web pages, PDF files, and Word documents.

[0022] Subsequently, in S120, the system preprocesses the original policy text to extract relevant policy information. Specifically, for HTML text, a text extraction algorithm that integrates text density analysis, title position patterns, and visible area rules is used to remove noise such as navigation bars and advertisements, extracting the main text content. For PDF files, a layout structure recognition algorithm is used to divide logical blocks and reorganize the text to extract structured policy text. All extracted policy text is then uniformly input into a pre-trained BERT text classification model for semantic discrimination to filter out non-policy text and extract key content such as policy clauses, applicable objects, issuing agencies, and support methods.

[0023] In step S130, the system collects enterprise profile data of the target enterprise, including attributes such as registered capital, industry category, years of establishment, main business, number of patents, credit rating, and geographical information.

[0024] Next, in steps S140 and S150, the system maps the aforementioned policy-related information to policy entity nodes and the target enterprise's profile information to enterprise entity nodes. The mapped nodes contain a structured representation of attributes and maintain unique identification within the graph.

[0025] In S160, the system generates various types of edge relationships between policy entity nodes and enterprise entity nodes based on a pre-defined set of semantic relationship definitions. For example, edge relationships with different semantic categories such as "industry matching," "financial support," "regional coverage," and "qualification requirements" are used to connect nodes. Each edge relationship is represented in the form of a triple, with additional edge weights and credibility information to enhance the semantic integrity and reliability of the knowledge graph, ultimately constructing a complete heterogeneous knowledge graph.

[0026] Next, in S170, the system takes the completed heterogeneous knowledge graph as input and constructs a graph neural network model (such as R-GCN, HAN, etc.) to perform embedding calculations on various types of nodes in the graph, obtaining a node representation vector with a unified dimension. The node representation not only includes its own attribute information, but also incorporates the semantic features of its contextual neighbors in the graph.

[0027] In S180, the system calculates the similarity between the embedding vector of the target enterprise node and the embedding vectors of all policy entity nodes to obtain the matching degree between each policy node and the enterprise. The similarity calculation not only considers semantic similarity, but also introduces structural enhancement features (such as semantic path relationships and enterprise cluster relevance) as auxiliary factors to improve the accuracy and diversity of recommendations.

[0028] Finally, in S190, the system sorts the data according to the matching degree and introduces a multi-objective sorting mechanism for comprehensive optimization, taking into account factors such as the enterprise's growth stage, the difficulty of policy implementation, and historical behavioral preferences, and outputs the optimal sorting result to form a personalized policy recommendation information set for the target enterprise.

[0029] In summary, this embodiment proposes a policy recommendation method based on heterogeneous knowledge graphs. It supports differentiated structural parsing of HTML policy web pages and PDF policy documents. By integrating text density analysis, DOM structure features, and layout logic reorganization, it automatically extracts the policy text content and core element fields, providing a unified input data format for subsequent semantic modeling and knowledge graph construction. By mapping policy content and enterprise profiles to semantic nodes, and generating multi-type edge relationships based on predefined multi-dimensional semantic relationships (such as regional adaptation, industry adaptation, and funding thresholds), a clear and semantically rich heterogeneous knowledge graph is constructed, significantly enhancing the semantic and structural alignment capabilities between policies and enterprises. Compared to traditional rule-based or flat vector-based recommendation methods, this solution introduces a graph neural network model to learn context-aware vector representations of nodes in the graph. This allows the representation of each enterprise node and policy node to comprehensively consider neighbor structure, edge type, and attribute semantics, improving the accuracy and interpretability of recommendation matching. This solution incorporates structural information such as "path relationship features" and "enterprise clustering results" as auxiliary factors during similarity calculation. This allows recommendations to be based not only on node semantic similarity but also on the graph structure connections and group behavior characteristics between enterprises and policies, thereby improving the personalization and stability of the recommendation results. When outputting recommendation results, the system supports a multi-objective fusion ranking mechanism based on enterprise development stage, policy difficulty, and historical preferences, achieving highly controllable and configurable strategy push paths that adapt to various government and enterprise service scenarios. In summary, this invention not only overcomes the format dependency, rule rigidity, and shallow semantic issues of traditional policy recommendation systems but also possesses high versatility, interpretability, and practical deployment value. It is particularly suitable for SME policy support scenarios, regional industrial guidance path analysis, and multi-level government service systems, demonstrating significant social benefits and technological innovation value.

[0030] In one feasible implementation, such as Figure 2 As shown, step S120 above extracts policy-related information based on multiple original policy texts, including: S1201. Obtain the format information of the original policy text; S1202. If the original policy text is in HTML format, the first policy text is extracted based on a text extraction algorithm that integrates text density analysis, title position pattern, and visible area rules. S1203. If the original policy text is in PDF format, the second policy text is extracted by dividing the text into logical blocks and reorganizing the text content based on the layout structure recognition algorithm. S1204. Input the first policy text and the second policy text mentioned above into the pre-trained BERT text classification model for semantic discrimination in order to extract policy-related information.

[0031] For example, in one feasible implementation, to address the multi-source heterogeneity of multiple original policy text data, this embodiment achieves efficient extraction and unified representation of policy-related information through a joint modeling method combining structured extraction and semantic recognition. The specific steps are as follows: The system first obtains the format information of the original policy text to determine whether it originates from an HTML webpage or a PDF document. For HTML-formatted policy text, the system employs a fusion-based text extraction algorithm. This algorithm combines text density analysis, title position pattern recognition, and visible area rules. Specifically, the system first parses the HTML structure and constructs a DOM node tree, extracting information such as tag type, text length, structural level, and style attributes for each node. Then, it calculates the text density value based on the node text length information and filters the candidate node set using a set density threshold. Simultaneously, it identifies contextually adjacent nodes based on the structural features of the title position and further filters out hidden elements or non-display area content based on the node's visibility. Finally, through candidate set fusion and scoring mechanisms, the set of node content with the highest score is selected as the first policy text.

[0032] For the original policy text in PDF format, the system employs a processing workflow based on page layout recognition. First, page coordinates and paragraph distribution information are extracted. Logical blocks are then divided based on visual continuity and semantic relevance, dividing the text into multiple structural blocks. Next, the system reorganizes these structural blocks according to their spatial order and text connection features, generating a sequence of main text with the correct reading order. Finally, through keyword template matching and positional rule recognition, the policy title, publication date, issuing agency, and main clause content are extracted from the rearranged main text, serving as the second policy text.

[0033] After obtaining the first and second policy texts, the system inputs them uniformly into a pre-trained BERT text classification model for semantic discrimination. The specific process includes: first, standardizing the text preprocessing (e.g., removing special characters, standardizing line breaks, etc.) and performing word segmentation encoding to generate a Token ID sequence conforming to the BERT input format; then, inputting this sequence into the BERT model to obtain a global semantic vector representation; the system uses this vector representation to perform multi-class classification operations to determine whether each text belongs to a valid policy document; finally, selecting text entries that meet the policy semantic tags, extracting the core policy fields and text content, and using this as the structured policy-related information output for this step.

[0034] In summary, this implementation method significantly improves the accuracy and adaptability of policy information parsing through the synergistic application of multi-format compatible structure extraction technology and deep semantic discrimination algorithm, providing a high-quality semantic input foundation for subsequent knowledge graph construction and policy recommendation.

[0035] In one feasible implementation, such as Figure 3 The above-mentioned step S1202 extracts the first policy text based on a text extraction algorithm that integrates text density analysis, title position patterns, and visible area rules, including: S12021. Parse the original policy text in HTML format, construct a DOM node tree, and extract the node attribute information of the DOM node tree. The node attributes include tag type information, text content length information, structural hierarchy information, and style attribute information. S12022. Based on the above text content length information of each node, calculate the text density value of each node, and select the first set of candidate nodes for the main text according to the preset density threshold. S12023. Detect structural nodes with title features in the DOM node tree based on the above tag type information, identify the neighboring node regions of the above title feature structural nodes based on the above structural hierarchy information, and mark the nodes containing high text density as the second body text candidate node set. S12024. Based on the above style attribute information, determine the visibility status of each node in the above original policy text page, remove nodes with hidden attributes or in non-visible areas, and obtain the third set of candidate nodes for the main text. S12025. Based on the above-mentioned first set of candidate nodes for the main text, the above-mentioned second set of candidate nodes for the main text, and the above-mentioned third set of candidate nodes for the main text, perform set fusion and sorting processing, and comprehensively score the candidate nodes according to the preset fusion scoring function, and select one or more nodes with the highest scores as the above-mentioned first policy text.

[0036] For example, in one feasible implementation, to achieve high-quality extraction of the main text from original policy text in HTML format, the system introduces a joint text extraction algorithm that integrates text density analysis, title position patterns, and visible area rules. This method fully utilizes the structured characteristics, multi-dimensional node attributes, and page presentation state of HTML documents to achieve interference removal and high-confidence extraction of the main text content.

[0037] The system first performs structural parsing on the original HTML policy text, constructs a complete DOM node tree, and extracts the following attributes from each node: Tag type information: such as , , , etc.; Text content length information (Text Len): The number of plain text characters contained in the node; Depth: The depth of a node's hierarchy within the DOM tree; Style attribute information (StyleAttr): such as display:none, visibility:hidden, position:absolute, etc. in CSS.

[0038] This information is used for subsequent density analysis, structural location, and visual status assessment.

[0039] The typical formula for calculating the text density value for each node in the DOM tree is as follows: in: For nodes The number of text characters in the TextLen; For nodes The number of child nodes.

[0040] Set an experience threshold through the system. Nodes with a density value greater than the threshold are selected as the first candidate node set for the main text. This strategy can effectively filter out densely structured areas such as menu bars and navigation bars that lack text.

[0041] The system further identifies a set of nodes with title features. Nodes such as , , or those containing keywords such as "notice," "announcement," or "policy." For each title node...<h1> It identifies structurally adjacent regions (such as sibling nodes or nodes below them) and detects nodes whose text density exceeds a certain threshold, marking them as the second set of candidate nodes for the main text. This operation is defined using the following nearest neighbor search function: The system determines whether a node is visible on the page based on style attributes. A node is considered invisible and removed if it has the following styles: display: none, visibility: hidden, position: absolute, and is not within the visible area. From all visible nodes, those with higher density are further selected to form the third set of candidate nodes for the main text. .

[0042] The system will The three sets are joined together to obtain the complete set of candidate nodes. and for each node Calculate the fusion scoring function The confidence level for identifying it as part of the main text is determined using the following scoring function: in: RelPos is the normalized value for the node text length. The score is based on the relative position of the node on the page (e.g., a higher score is given for nodes closer to the center). The visible weight of a node (can be considered as 1, not 0); These are preset weighting coefficients used to balance the influence of the three dimensions.

[0043] The final system sorts candidate nodes by Sort the values ​​in descending order, select the text content of one or more nodes with the highest scores, and concatenate them to generate the final first policy text.

[0044] The method proposed in this embodiment integrates multiple features such as structural density, semantic adjacency, and visual presentation. It can effectively filter non-text content such as navigation bars, advertisements, and recommendation slots. It can still effectively locate the main text area for web pages with complex structures or non-standard tags. It can significantly improve the accuracy of main text recognition by combining context and visual features.

[0045] In one feasible implementation, such as Figure 4 As shown, step S1203 above uses a layout structure recognition algorithm to perform logical block division and text content reorganization to extract the second policy text, including: S12031. Obtain the original policy text in PDF format, and perform structured parsing on the original policy text to extract page coordinate information and paragraph distribution information; S12032. Based on the above page coordinate information and the above paragraph distribution information, perform a logical block division operation to obtain multiple page structure blocks, wherein each of the above structure blocks includes one or more text fragments with visual continuity and semantic consistency. S12033. Based on the positional relationship and text connection features of the above structural blocks, the above structural blocks are reordered to generate a rearranged text sequence. S12034. In the above-mentioned text sequence, the policy title, publication time, issuing agency and text content of the original policy text in the above-mentioned PDF format are extracted based on keyword template matching and position rules to form the second policy text.

[0046] For example, in order to efficiently and accurately extract structured policy information from original policy documents in PDF format, the system introduces a text logic block segmentation and reorganization method based on page layout recognition to ensure accurate recovery of the main text area in complex formatted documents.

[0047] First, the system uses layout analysis tools (such as PDFBox, DocParser, pdfplumber, etc.) to perform structured parsing of the original policy text in PDF format, obtaining the text segments (TextSpan) and their page attributes for each page. The system extracts the following element information: Page coordinate information : indicates the first The top-left corner coordinates and width and height of the text fragment; Paragraph distribution information: Paragraph boundaries are defined by analyzing the vertical spacing and font consistency between consecutive text. You can also use font information (font size, bold / italic attributes) to help determine the boundary between the title and the body text. The purpose of this step is to restore the original PDF page content into structured "content units" and "spatial location information".

[0048] The system divides text segments into layout blocks based on the principles of "visual continuity" and "semantic consistency." The division can be based on the following criteria: 1. Vertical distance constraint: If the vertical distance between two adjacent text units is less than a certain threshold... If the initial horizontal coordinates of two elements differ, they are considered to belong to the same structural block; 2. Horizontal consistency: If the initial horizontal coordinates of two elements differ by a certain value... If they are not identical, they can be aggregated into the same block; 3. Semantic consistency assessment: Use text embedding models to determine the semantic continuity of the context.

[0049] The partitioning logic can be formally represented by the following clustering judgment rules: in, For text units; For the first A structural block; Dist The Euclidean distance of the text unit in spatial coordinates; For horizontal alignment measurement; This is the threshold for merging structural blocks.

[0050] The final set of structure blocks It is a logical segment that has positional continuity and content consistency.

[0051] PDF files may have disordered structural blocks due to multi-column layout or horizontal misalignment. The system can reorder the structural blocks according to the following rules: first, sort by page number in ascending order; within the same page, prioritize sorting by y-axis value from top to bottom; if there are side-by-side structures (such as two-column layouts), then sort by horizontal coordinate. Auxiliary sorting or grid partitioning methods can be used. The rearrangement process yields an ordered sequence of text segments, TextSeq. This is used for subsequent information extraction.

[0052] Within the text segment sequence, the system extracts standard policy fields using a combination of keyword template matching and location information rules. Priority is given to matching structural blocks located in the first few lines of the first page, with larger font sizes, and frequently containing keywords such as "notice" or "regulations." Regular expressions are used to match common time formats (e.g., "October 5, 2023") and common organization names (e.g., "Provincial Development and Reform Commission"). Paragraphs appearing in clause formats such as "Article 1," "Article 2," etc., are then extracted from the remaining text segments.

[0053] For example, the rules for recognizing "publication time" in the main text can be expressed in the following form: The main body of the clauses can be combined with structural rules and clause keywords: Finally, the system organizes the extracted policy title, publication date, issuing agency, and main clause content into structured fields to form a complete second policy text.

[0054] The method proposed in this embodiment can automatically identify the logical structure even if the policy PDF adopts a multi-column, staggered, or template format. Combined with keywords and position constraints, it can significantly improve the accuracy of field extraction. The extracted structured fields can be directly used to construct policy entity attributes, and have a good foundation for data standardization.

[0055] In one feasible implementation, such as Figure 5 As shown, step S1204 above inputs the first policy text and the second policy text into a pre-trained BERT text classification model for semantic discrimination to extract policy-related information, including: S12041. Perform text standardization and word segmentation encoding on the above-mentioned first policy text and the above-mentioned second policy text to construct a token ID sequence that conforms to the BERT input format; S12042. Input the above token ID sequence into the pre-trained BERT text classification model to extract the global semantic vector corresponding to each text. S12043. Based on the above global semantic vector, perform multi-class prediction and output the semantic category label of each text. S12044. Based on the above semantic category tags, select the text content belonging to the policy category as the extracted policy-related information.

[0056] For example, in order to accurately identify paragraphs or clauses that are truly relevant to the policy content from the policy text extracted from HTML or PDF, the system introduces a pre-trained BERT text classification model for semantic discrimination and structured recognition.

[0057] For the extracted first and second policy texts, standardization and encoding preprocessing are first performed to meet the input requirements of the BERT model. Standardization and encoding preprocessing operations include, but are not limited to: 1. Text cleaning: Remove line breaks, special characters, and residual HTML tags; 2. Word segmentation: The text is segmented into sub-words using the WordPiece word segmenter corresponding to BERT; 3. Construct the input vector: including: input_ids: The sequence of token IDs corresponding to the tokens after word segmentation; attention_mask: A mask used to indicate the location of a valid token; token_type_ids (optional): A token used to distinguish sentence pairs.

[0058] The encoded input vector is fed into a pre-trained and fine-tuned BERT classification model to extract the global semantic representation vector for each text segment. Generally, the output vector at the corresponding position of the [CLS] token represents the semantic features of the entire text segment. This process can be represented as: in: The output vector representing [CLS] (i.e., the semantic representation of the paragraph); This represents the size of the hidden dimension of BERT (e.g., 768).

[0059] Next, the semantic vectors are fed into the output layer (usually a fully connected layer + Softmax) to predict the category of the text: in: For the predicted category distribution, This represents the total number of categories; These are the weight matrix and bias vector of the classifier, respectively, which are obtained by minimizing the cross-entropy loss during the training phase.

[0060] The output prediction labels include: "Policy Title", "Main Text Clauses", "Supplementary Provisions / Explanations", "Non-Policy Content", and "Table / Graphic Information".

[0061] Based on the classification prediction results, the system filters out representative paragraphs from the original text that belong to the "policy text category". For example, only paragraphs with semantic tags such as "text clause", "title", and "summary of key points" are retained, while irrelevant content such as the preface, appendix, and chart descriptions are removed.

[0062] This can be represented by the following rules: in: Indicates the first A text paragraph; For the first The predicted labels for the policy text; ValidPolicyLabels is the set of labels that are considered to be part of the policy text (such as clauses, policy summaries, titles, etc.).

[0063] According to the method proposed in this embodiment, the pre-trained BERT text classification model can understand the complex semantic relationships between long texts, is suitable for judging the meaning changes in policy language, effectively removes invalid paragraphs (such as appendices, footer instructions, and signature areas) in template policy documents, can process text paragraphs extracted from HTML, and also supports the judgment of logical paragraph content after PDF reconstruction. The model can be further fine-tuned to adapt to specific policy fields (such as science and technology, taxation, and environmental protection) to form targeted recognition capabilities.

[0064] In one feasible implementation, such as Figure 6 As shown, step S160 generates multiple types of edges between the policy entity nodes and the enterprise entity nodes based on a preset semantic relationship definition to construct a heterogeneous knowledge graph, including: S1601. Based on the attribute information of the above-mentioned policy entity nodes and the profile information of the above-mentioned enterprise entity nodes, construct a set of semantic relationship definitions; S1602. Perform attribute matching between the above-mentioned enterprise entity nodes and the above-mentioned policy entity nodes, establish one or more edge relationships describing their semantic associations based on preset semantic relationships, and generate multi-type edges in the heterogeneous graph. S1603. Represent the above edge relationships in the form of triples, and add edge weight information and credibility information to construct the above heterogeneous knowledge graph.

[0065] In one feasible implementation, such as Figure 7 As shown, the steps for determining the aforementioned edge weight information and the aforementioned confidence information include: S210. Calculate the attribute matching score, behavior co-occurrence score, and semantic relevance score between the above-mentioned enterprise entity nodes and the above-mentioned policy entity nodes; S220. Based on the preset edge type confidence prior, calculate the fusion score as the edge weight information according to the above attribute matching score, above behavior co-occurrence score and above semantic relevance score of each edge. S230. Based on the above edge weight information, data source credibility and edge type confidence coefficient, calculate the credibility information of the edge.

[0066] For example, the system first predefines a semantic relation set R based on the attribute information of policy entity nodes (such as the applicable industry, applicable enterprise size, applicable region, release time, policy type, etc.) and the profile information of enterprise entity nodes (such as industry classification, registered capital, personnel size, location, revenue scale, development stage, etc.). Each relation type It has corresponding structural matching conditions and logical rules, for example: Adaptation relationship (IsApplicableTo): If the enterprise belongs to the industry Policy support for the industry; Coverage relationship (IsCoveredBy): If the enterprise is located in a certain region Policy coverage area; Indirect Relevance (IsLatentlyRelevantTo): If there is an indirect graph path between the firm and the policy; HasPolicyHistory: This refers to a company's past successful applications for this type of policy.

[0067] For each pair of enterprise nodes With policy nodes Based on the semantic rule matching results mentioned above, the system generates several triplet edge relationships: Each edge is recorded as follows: starting entity, relation type, and ending entity, and all are stored in the edge set. This is used for subsequent graph model training.

[0068] To improve the learning quality of graph neural networks, the system processes each semantic edge. The following three scoring indicators are introduced: 1. Attribute Similarity Score in, and These respectively represent the enterprises and policies in the [number]th [month]. The value of each attribute dimension This is an indicator function; 1 indicates an exact match of the attribute, otherwise 0. It represents the number of attributes being compared.

[0069] 2. Behavior Co-Frequency Score It indicates the relative proportion of the number of times a company has applied for or consulted about this type of policy in its overall application behavior, and has the ability to normalize. Indicates the enterprise I have actually applied for or consulted about policies in the past. The number of occurrences reflects the co-occurrence frequency of its behavior. Indicates enterprise The total number of all historical applications or policy actions (i.e., the overall activity level of the enterprise).

[0070] 3. Semantic Similarity Score in, These are the embedding vectors of enterprise nodes and policy nodes, encoded using a large language model or graph neural network, respectively. This represents the cosine similarity function.

[0071] The final edge weights are calculated using the following nonlinear combination function: in And satisfy This is obtained through cross-validation or empirical optimization.

[0072] Using the hyperbolic tangent function Guarantee the result falls within the interval This enhances the ability to differentiate and express non-linear expressions.

[0073] Each edge also has an added credibility level. This represents the confidence level at which the semantic edge is adopted in the reasoning. It is calculated as follows: in: : Adjustment parameters for prior confidence of edge type; The credibility priority of the edge type source can be set to 0.9 for government information disclosure data and 0.6 for web crawlers. : Edge weight confidence mapping in sigmoid function form Control the slope, It is the credibility threshold.

[0074] Through the above steps, each semantic edge not only possesses structural expressive capabilities but also learnable dynamic attributes (weights, credibility), thus providing a solid foundation for subsequent edge attention mechanisms, path backtracking reasoning, and multi-objective recommendation optimization in graph neural networks.

[0075] In one feasible implementation, such as Figure 8 As shown, step S180 above calculates the similarity between the node representation of the target enterprise and the node representation of each policy entity to obtain the matching degree, including: S1801. Based on the node representation vectors of the target enterprises and the representation vectors of the policy entity nodes, calculate the semantic similarity score. S1802. Based on the structural information in the heterogeneous knowledge graph, extract the semantic path relationship features between the target enterprise node and each of the policy entity nodes, and calculate the path relationship score. The path relationship features include path length, path type weight, and path existence. S1803. Obtain the enterprise clustering results of the target enterprise, count the historical recommendation or matching frequency of other enterprise nodes in the cluster to the policy entity node, and calculate the clustering relevance score. S1804. Based on the above semantic similarity score, the above path relationship score and the above clustering relevance score, a weighted fusion is performed to generate the final matching degree between the above target enterprise and each of the above policy entity nodes.

[0076] For example, the process of calculating the similarity between the node representation of the target enterprise and the node representation of each policy entity to obtain the matching degree integrates semantic representation, graph structure path features and enterprise clustering behavior priors.

[0077] After embedding nodes into a heterogeneous knowledge graph using a graph neural network model, the system obtains the representation vector of the target enterprise node. ;No. The representation vector of each policy entity node .

[0078] Calculate semantic scores using cosine similarity: Where • denotes the vector dot product. This represents the L2 norm. This score measures the semantic closeness between businesses and policies, and its value ranges from [value missing]. .

[0079] The system mines the reachability path characteristics between target enterprise nodes and each policy node in a heterogeneous knowledge graph, including: shortest path length. The set of semantic type weights for each edge in the path: Path reachability identifier: If it is reachable, then it is 1.

[0080] The formula for the comprehensive path relationship score is as follows: The shorter the path and the higher the semantic edge weight on the path, the stronger the matching strength; if it is unreachable, the score is 0.

[0081] After clustering enterprise nodes according to their profile features, we set the target enterprise to belong to the cluster. For each enterprise node in the cluster... To determine whether it has successfully matched or recommended the first One policy: in, It is an indicator function, if Previously with policy A score of 1 indicates a successful match, while a score of 0 indicates a mismatch. This score reflects the "recommendation tendency of similar companies".

[0082] The system performs a weighted fusion of the scores from the three sub-modules to obtain the scores of the target company and the second sub-module. The final match between the policies: in: For the weighting coefficients, satisfying It can be optimized according to the recommendation scenario, such as semantic-driven recommendation, which can be set... Structural support recommendations can be set 0.4, etc.

[0083] In one feasible implementation, such as Figure 9 As shown, step S190 above outputs a set of recommended policy items based on the matching degree, which serves as the policy recommendation information for the target enterprise, including: S1901. Based on the above matching degree, the candidate policies are initially ranked; S1902. Based on the preliminary ranking results mentioned above, a multi-objective ranking strategy is further introduced to comprehensively score and optimize the ranking of candidate policies. The multi-objective ranking strategy includes at least one of the following: setting ranking weights based on the enterprise's growth stage to improve the scores of policy items that are more suitable for the current stage; reducing the ranking priority of complex policies in early-stage enterprise recommendations based on the difficulty of policy application, approval cycle, or acceptance threshold; and dynamically fine-tuning the ranking results by combining the enterprise's historical policy preferences and the application behavior trends of similar enterprises. S1903. Based on the above comprehensive ranking results, select the top K items in the score and output the recommended set of policy items as the policy recommendation information for the target enterprise.

[0084] For example, firstly, based on the matching score calculated in the previous stage... For the set of candidate policies Perform a preliminary descending sort to obtain an initial sorted list: Based on the initial ranking described above, a multi-objective weighting factor is introduced to adjust the overall score. The overall score function is defined as follows: The meanings of each item are as follows: For the target company and the first The matching degree of each policy (from graph similarity fusion calculation); The growth stage suitability score reflects the degree to which the policy is suitable for the target company's current growth stage (such as startup, expansion, or maturity). The policy application complexity factor has a value range of [value range missing]. A higher value indicates that the policy is more difficult to apply for; The range of values ​​represents the support level for a company's historical preferences or the behavior of similar companies. The higher the value, the more popular the policy is; These are weighting coefficients, adjusted according to business needs to meet [the requirements]. .

[0085] For example, for a startup, the system can be configured... To enhance the control over the adaptability to different stages and the complexity of operations; if a certain policy With a matching degree of 0.75, a stage adaptation score of 0.9, a complexity of 0.8, and a historical preference of 0.6, the overall score is: The system is based on a comprehensive score. The candidate policies are then ranked, and the top-scoring policies are selected. The policy recommendations comprise the following set of results: This embodiment achieves dynamic prioritization by incorporating growth stage and historical behavioral characteristics. High-threshold policies are postponed for recommendation, aligning with the company's capacity to bear the burden.

[0086] The second aspect, such as Figure 10 The diagram shown illustrates the structure of a policy recommendation system based on heterogeneous knowledge graphs proposed in this application, including: The first acquisition unit 21 is used to acquire the original policy text from multiple heterogeneous policy data sources; Extraction unit 22 is used to extract policy-related information based on multiple original policy texts mentioned above; The second acquisition unit 23 is used to acquire enterprise profile data of the target enterprise; The first mapping unit 24 is used to map the above-mentioned policy-related information into policy entity nodes; The second mapping unit 25 is used to map the above-mentioned enterprise profile data into enterprise entity nodes; The first construction unit 26 is used to generate multiple types of edges between the above-mentioned policy entity nodes and the above-mentioned enterprise entity nodes based on the preset semantic relationship definition, so as to construct a heterogeneous knowledge graph. The second construction unit 27 is used to construct a graph neural network model based on the above heterogeneous knowledge graph, and to perform embedding calculations on various types of nodes in the graph to obtain node representation vectors, wherein the above node representation vectors include the node representations of the target enterprise and the node representations of each policy entity; The calculation unit 28 is used to calculate the similarity between the node representation of the target enterprise and the node representation of each policy entity to obtain the matching degree. Recommendation unit 29 is used to output a set of recommended policy items based on the matching degree, which serves as the policy recommendation information for the aforementioned target enterprise.

[0087] It is understood that the policy recommendation system proposed in this application based on heterogeneous knowledge graphs can also perform the methods described in any of the first aspects.

[0088] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A policy recommendation method based on heterogeneous knowledge graphs, characterized in that, include: Obtain the original policy text from multiple heterogeneous policy data sources; Based on multiple original policy texts, extract policy-related information; Obtain enterprise profile data of the target company; Map the policy-related information to policy entity nodes; Map the enterprise profile data to enterprise entity nodes; Based on the preset semantic relationship definition, multi-type edges are generated between the policy entity nodes and the enterprise entity nodes to construct a heterogeneous knowledge graph; Based on the heterogeneous knowledge graph, a graph neural network model is constructed, and embedding calculations are performed on various types of nodes in the graph to obtain node representation vectors, wherein the node representation vectors include the node representations of the target enterprise and the node representations of each policy entity; The similarity between the node representation of the target enterprise and the node representation of each policy entity is calculated to obtain the matching degree. Based on the matching degree, a set of recommended policy items is output as the policy recommendation information for the target enterprise.

2. The policy recommendation method based on heterogeneous knowledge graphs according to claim 1, characterized in that, The extraction of policy-related information based on multiple original policy texts includes: Obtain the format information of the original policy text; When the original policy text is in HTML format, the first policy text is extracted based on a text extraction algorithm that integrates text density analysis, title position patterns, and visible area rules. When the original policy text is in PDF format, the second policy text is extracted by logical block division and text content reorganization based on the layout structure recognition algorithm. The first policy text and the second policy text are input into a pre-trained BERT text classification model for semantic discrimination in order to extract policy-related information.

3. The policy recommendation method based on heterogeneous knowledge graphs according to claim 2, characterized in that, The text extraction algorithm based on fusion text density analysis, title position patterns, and visible area rules extracts the first policy text, including: The original policy text in HTML format is parsed, a DOM node tree is constructed, and the node attribute information of the DOM node tree is extracted. The node attributes include tag type information, text content length information, structural hierarchy information, and style attribute information. Based on the text content length information of each node, the text density value of each node is calculated, and the first set of candidate nodes for the main text is selected according to the preset density threshold. Based on the tag type information, detect structural nodes with title features in the DOM node tree, identify the neighboring node regions of the structural nodes with title features based on the structural hierarchy information, and mark the nodes containing high text density as the second set of candidate nodes for the main text. Based on the style attribute information, determine the visibility status of each node in the original policy text page, remove nodes with hidden attributes or in non-visible areas, and obtain a third set of candidate nodes for the main text. Based on the first set of candidate nodes for the main text, the second set of candidate nodes for the main text, and the third set of candidate nodes for the main text, a set fusion and sorting process is performed. The candidate nodes are comprehensively scored according to a preset fusion scoring function, and one or more nodes with the highest scores are selected as the first policy text.

4. The policy recommendation method based on heterogeneous knowledge graphs according to claim 2, characterized in that, The algorithm based on page layout recognition performs logical block division and text content reorganization to extract the second policy text, including: Obtain the original policy text in PDF format, and perform structured parsing on the original policy text to extract page coordinate information and paragraph distribution information; Based on the page coordinate information and the paragraph distribution information, a logical block division operation is performed to obtain multiple page structure blocks, wherein each structure block includes one or more text fragments with visual continuity and semantic consistency. Based on the positional relationship and text connection features of the structural blocks, the structural blocks are sequentially reorganized to generate a rearranged text sequence. In the main text sequence, the policy title, publication time, issuing agency, and main clause content are extracted from the original policy text in PDF format based on keyword template matching and position rules to form the second policy text.

5. The policy recommendation method based on heterogeneous knowledge graphs according to claim 2, characterized in that, The step of inputting the first policy text and the second policy text into a pre-trained BERT text classification model for semantic discrimination to extract policy-related information includes: The first policy text and the second policy text are subjected to text standardization and word segmentation encoding to construct a token ID sequence that conforms to the BERT input format; The token ID sequence is input into the pre-trained BERT text classification model to extract the global semantic vector corresponding to each text. Multi-class prediction is performed based on the global semantic vector, and the semantic category label of each text is output. Text content belonging to the policy category is filtered out based on the semantic category tags and used as the extracted policy-related information.

6. The policy recommendation method based on heterogeneous knowledge graphs according to claim 1, characterized in that, The generation of multi-type edges between the policy entity nodes and the enterprise entity nodes based on a preset semantic relationship definition to construct a heterogeneous knowledge graph includes: Based on the attribute information of the policy entity nodes and the profile information of the enterprise entity nodes, a set of semantic relationship definitions is constructed; Attribute matching is performed between the enterprise entity nodes and the policy entity nodes, and one or more edge relationships describing their semantic associations are established based on preset semantic relationships to generate multi-type edges in the heterogeneous graph. The edge relationships are represented in the form of triples, and edge weight information and credibility information are added to construct the heterogeneous knowledge graph.

7. The policy recommendation method based on heterogeneous knowledge graphs according to claim 6, characterized in that, The steps for determining the edge weight information and the credibility information include: Calculate the attribute matching score, behavior co-occurrence score, and semantic relevance score between the enterprise entity node and the policy entity node; Based on a preset edge type confidence prior, a fusion score is calculated as the edge weight information according to the attribute matching score, behavior co-occurrence score and semantic relevance score of each edge. Based on the edge weight information, data source confidence, and edge type confidence coefficient, the confidence information of the edge is calculated.

8. The policy recommendation method based on heterogeneous knowledge graphs according to claim 1, characterized in that, The step of calculating the similarity between the node representation of the target enterprise and the node representation of each policy entity to obtain the matching degree includes: Based on the node representation vector of the target enterprise and the representation vector of each policy entity node, a semantic similarity score is calculated; Based on the structural information in the heterogeneous knowledge graph, semantic path relationship features between the target enterprise node and each policy entity node are extracted, and path relationship scores are calculated. The path relationship features include path length, path type weight, and path existence. Obtain the enterprise clustering results to which the target enterprise belongs, count the historical recommendation or matching frequency of other enterprise nodes in the cluster to the policy entity node, and calculate the clustering relevance score; Based on the semantic similarity score, the path relationship score, and the clustering relevance score, a weighted fusion is performed to generate the final matching degree between the target enterprise and each of the policy entity nodes.

9. The policy recommendation method based on heterogeneous knowledge graphs according to claim 1, characterized in that, The set of recommended policy items output based on the matching degree, as the policy recommendation information for the target enterprise, includes: Based on the matching degree, the candidate policies are initially ranked; Based on the preliminary ranking results, a multi-objective ranking strategy is further introduced to comprehensively score and optimize the ranking of candidate policies. The multi-objective ranking strategy includes at least one of the following: setting ranking weights based on the enterprise's growth stage to improve the scores of policy items that are more suitable for the current stage; reducing the ranking priority of complex policies in early-stage enterprise recommendations based on the implementation difficulty, approval cycle, or acceptance threshold of policy applications; and dynamically fine-tuning the ranking results by combining the enterprise's historical policy preferences and the application behavior trends of similar enterprises. Based on the comprehensive ranking results, the top K items with the highest scores are selected to output a set of recommended policy items, which serve as the policy recommendation information for the target enterprise.

10. A policy recommendation method based on heterogeneous knowledge graphs, characterized in that, include: The first acquisition unit is used to acquire the original policy text from multiple heterogeneous policy data sources; The extraction unit is used to extract policy-related information based on multiple original policy texts; The second acquisition unit is used to acquire enterprise profile data of the target enterprise; The first mapping unit is used to map the policy-related information into policy entity nodes; The second mapping unit is used to map the enterprise profile data into enterprise entity nodes; The first construction unit is used to generate multiple types of edges between the policy entity nodes and the enterprise entity nodes based on a preset semantic relationship definition, so as to construct a heterogeneous knowledge graph. The second construction unit is used to construct a graph neural network model based on the heterogeneous knowledge graph, and perform embedding calculations on various types of nodes in the graph to obtain node representation vectors, wherein the node representation vectors include the node representations of the target enterprise and the node representations of each policy entity; A calculation unit is used to calculate the similarity between the node representation of the target enterprise and the node representation of each policy entity to obtain the matching degree. The recommendation unit is used to output a set of recommended policy items based on the matching degree, which serve as policy recommendation information for the target enterprise.

Images