Enterprise service policy matching method and system based on knowledge graph

Abstract

This invention relates to the technical field of knowledge graphs, and provides a method and system for matching enterprise service policies based on knowledge graphs. The method includes: acquiring enterprise profile data of the enterprise to be matched, and converting the enterprise profile data into enterprise profile fact nodes; wherein, the enterprise profile fact nodes are structured data nodes representing the registered region, industry, operating status, qualification certificates, personnel size, operating income, R&D investment, or submitted materials information of the enterprise to be matched; performing conditional matching on the enterprise profile fact nodes according to the relational expression in the policy knowledge graph to obtain the enterprise service policy matching result; and outputting matched policy information, information on unmet conditions, or supplementary material prompts based on the enterprise service policy matching result. This invention improves the accuracy and reliability of enterprise service policy matching.

Description

Technical Field

[0001] This invention relates to the technical field of knowledge graphs, and specifically to a method and system for matching enterprise service policies based on knowledge graphs. Background Technology

[0002] In the field of enterprise service policy matching, existing technologies typically use knowledge graph technology to solve the problem of association between multi-source heterogeneous data. By extracting entities and splitting conditions, relationships are established between policy entities and conditions such as qualifications, scale, and industry. Then, enterprise profile data is called to perform graph reasoning to generate candidate enterprises and matching criteria.

[0003] However, in actual local government policy releases, the main text and appendices are often released separately. The main text typically only outlines the supported targets and application directions, while specific threshold conditions, exclusion clauses, and time limits are placed in appendice tables or application guidelines. In the process of policy analysis and knowledge graph modeling, particularly in the section where main text clauses and appendice conditions are combined, complex processing logic often arises. When the policy text references conditions stipulated in the appendices, and the appendices contain parallel conditions, exceptions, and negative restrictions, existing condition splitting rules usually only segment based on sentence fragments or keyword granularity. This approach struggles to coordinate the analysis of citations in the main text and joint constraints in the appendices, resulting in the separation of general conditions in the main text from detailed restrictions in the appendices. Because cross-carrier joint constraint relationships cannot be accurately identified, policy entities may be directly connected to several isolated condition nodes in the knowledge graph, lacking subordinate, parallel, or preconditional relationships between conditions, and even mistaking exception clauses in the appendices as universally applicable conditions. Therefore, in the application scenario where the policy text cites the eligibility conditions of the appendix, the existing policy condition entity splitting rules cannot jointly model the citation relationship in the text and the parallel, exception and time restrictions in the appendix. This results in the policy eligibility relationship in the knowledge graph being split into disconnected condition edges, which in turn causes the enterprise adaptation and matching to be mismatched based on incomplete condition chains.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] This application discloses a knowledge graph-based enterprise service policy matching method and system, which aims to solve the problem of enterprise adaptation and matching mismatch caused by the inability to collaboratively parse the reference relationship and joint constraints between the policy text and attachments in the prior art.

[0006] The technical solution of this application is as follows: Firstly, this application discloses a knowledge graph-based enterprise service policy matching method, including: Obtain the full text of the company's policies, as well as the policy attachments associated with the full text; Identify the citation relationships between the policy text and the policy appendices, and integrate the contextual constraints of the policy text with the specific content of the policy appendices based on the citation relationships to obtain a unified parsed text block; Semantic analysis is performed on the unified parsed text block to identify the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and a condition logic tree representing the hierarchical relationship and dependency relationship of the fine-grained conditions is constructed. The conditional logic tree is converted into a relational representation in a policy knowledge graph. This relational representation includes the associations between policy entities and fine-grained conditions, the logical operational relationships between fine-grained conditions, and the time constraints corresponding to the fine-grained conditions. The policy knowledge graph is a structured graph data used to store the associations between policy entities, fine-grained conditions, logical operational relationships, time constraints, and enterprise profile fields. Nodes in the policy knowledge graph include policy entity nodes, condition nodes, time constraint nodes, and enterprise profile field nodes. Edges represent the applicability relationships between policy entities and fine-grained conditions, the logical operational relationships between fine-grained conditions, and the time constraints corresponding to the fine-grained conditions. Obtain enterprise profile data of the companies to be matched and convert the enterprise profile data into enterprise profile fact nodes; among them, the enterprise profile fact nodes are structured data nodes used to represent the registered region, industry, operating status, qualification certificates, personnel size, operating income, R&D investment or submitted materials information of the companies to be matched. Based on the relational expressions in the policy knowledge graph, conditional matching is performed on the factual nodes of the enterprise profile to obtain the enterprise service policy matching results; Based on the matching results of enterprise service policies, output information on matching policies, information on conditions not met, or information on supplementary materials.

[0007] This technical solution enables collaborative parsing of cross-carrier joint constraints of policy text and appendices, constructing a conditional logic tree containing complete conditional hierarchy and dependencies. This solves the matching mismatch problem caused by the disconnection of conditional relationships in knowledge graphs, and significantly improves the accuracy of enterprise policy matching.

[0008] Secondly, this application also discloses a knowledge graph-based enterprise service policy matching system for performing knowledge graph-based enterprise service policy matching, including: The document attachment retrieval module is used to retrieve the enterprise's policy document and the policy attachments associated with the policy document; The text block management module is used to identify the citation relationship between the policy text and the policy appendix, and to integrate the contextual constraint information of the policy text and the specific content of the policy appendix according to the citation relationship to obtain a unified parsed text block; The Condition Logic Tree module is used to perform semantic analysis on the unified parsed text block, identify the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and construct a condition logic tree that represents the hierarchical relationship and dependency relationship of the fine-grained conditions. The relation representation conversion module is used to convert the conditional logic tree into relation representations in the policy knowledge graph. The relation representations include the association between policy entities and fine-grained conditions, the logical operation relationships between fine-grained conditions, and the time constraint relationships corresponding to the fine-grained conditions. The Enterprise Profile Fact Node module is used to obtain enterprise profile data of the enterprise to be matched and convert the enterprise profile data into enterprise profile fact nodes; The service policy matching module is used to perform conditional matching on the fact nodes of the enterprise profile based on the relationship expression in the policy knowledge graph, and obtain the enterprise service policy matching results. The results information output module is used to output matching policy information, information on unmet conditions, or prompts for supplementary materials based on the enterprise service policy matching results.

[0009] This technical solution enables collaborative parsing of policy text and appendices, automatic construction of conditional logic trees, and precise matching based on knowledge graphs through a modular architecture, thereby providing a complete system-level solution and ensuring the operability and engineering practicality of the method.

[0010] Beneficial Effects: This application identifies the citation relationships between the policy text and its appendices and integrates them into a unified parsing text block. Deep semantic analysis is then performed on this unified parsing text block to identify fine-grained conditions and their logical connections. A conditional logic tree containing hierarchical and dependency relationships is constructed and converted into a relational expression within a policy knowledge graph. This enables collaborative parsing of the joint constraints of the policy text and its appendices. Compared to existing technologies that only segment based on sentence fragments or keywords, leading to disjointed conditional relationships, this application effectively identifies cross-carrier joint constraint relationships, avoids isolated conditional nodes in the knowledge graph, ensures accurate modeling of subordinate, parallel, or preconditional relationships between conditions, prevents exceptions in the appendices from being mistakenly treated as universally applicable conditions, and thus performs enterprise adaptation matching based on a complete conditional chain, significantly improving the accuracy and reliability of enterprise service policy matching. Attached Figure Description

[0011] Figure 1 This is a flowchart of a knowledge graph-based enterprise service policy matching method according to one embodiment of the present invention; Figure 2 This is a flowchart of a knowledge graph-based enterprise service policy matching method according to another embodiment of the present invention; Figure 3 This is a system block diagram of a knowledge graph-based enterprise service policy matching system according to another embodiment of the present invention; Explanation of reference numerals in the attached figures: 1. Knowledge graph-based enterprise service policy matching system; 11. Text attachment acquisition module; 12. Text block management module; 13. Conditional logic tree module; 14. Relationship expression conversion module; 15. Enterprise profile fact node module; 16. Service policy matching module; 17. Result information output module. Detailed Implementation

[0012] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0013] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0014] Combination Figure 1 As shown, this application discloses a knowledge graph-based enterprise service policy matching method, including the following steps: S1, obtain the company's policy text and related policy attachments; S2 identifies the citation relationship between the policy text and the policy appendix, and integrates the contextual constraint information of the policy text with the specific content of the policy appendix based on the citation relationship to obtain a unified parsed text block; S3 performs semantic analysis on the unified parsed text block, identifies the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and constructs a condition logic tree representing the hierarchical relationship and dependency relationship of the fine-grained conditions. S4. The conditional logic tree is converted into a relational representation in the policy knowledge graph. The relational representation includes the association between policy entities and fine-grained conditions, the logical operation relationship between fine-grained conditions, and the time constraint relationship corresponding to the fine-grained conditions. The policy knowledge graph is a structured graph data used to store the association between policy entities, fine-grained conditions, logical operation relationships, time constraint relationships, and enterprise profile fields. The nodes in the policy knowledge graph include policy entity nodes, condition nodes, time constraint nodes, and enterprise profile field nodes. The edges are used to represent the applicability relationship between policy entities and fine-grained conditions, the logical operation relationship between fine-grained conditions, and the time constraint relationship corresponding to the fine-grained conditions. S5: Obtain the enterprise profile data of the enterprise to be matched and convert the enterprise profile data into enterprise profile fact nodes; wherein, the enterprise profile fact nodes are structured data nodes used to represent the registered region, industry, operating status, qualification certificates, personnel size, operating income, R&D investment or submitted materials information of the enterprise to be matched. S6. Based on the relationship expression in the policy knowledge graph, perform condition matching on the fact nodes of the enterprise profile to obtain the enterprise service policy matching result; S7 outputs matching policy information, information on unmet conditions, or prompts for supplementary materials based on the enterprise service policy matching results.

[0015] Specifically, obtaining the policy text for enterprises, as well as the associated policy attachments, serves as the data entry point for the policy information analysis process. The policy text can be documents such as policy notices, application announcements, implementation rules, or management methods issued by government departments. Policy attachments can be documents accompanying the policy text, such as explanations of application conditions, lists of application materials, scoring criteria, form templates, or operation guides. In practical applications, the system can obtain the policy text and attachments through web crawlers, data interfaces, or manual uploads. For example, the system can periodically scan the notice and announcement sections of official government websites such as the Science and Technology Bureau and the Industry and Information Technology Bureau to download the latest policy text documents and obtain the associated policy attachments based on the attachment links, attachment numbers, or attachment names within the policy text. Taking the document "Notice on Conducting Applications for the 2023 High-tech Enterprise Cultivation Pool.pdf" as an example, after obtaining the policy text, the system can further obtain the associated "Attachment 1: Explanation of Application Conditions.docx" and "Attachment 2: List of Application Materials.xlsx," enabling the subsequent analysis process to simultaneously cover the general requirements in the policy text and the specific implementation content in the policy attachments.

[0016] After obtaining the policy text and its appendices, the system identifies the citation relationships between them and integrates the contextual constraints of the policy text with the specific content of the appendices to obtain a unified parsed text block. Citation relationships refer to the relationships in the policy text that point to, explain, or constrain the policy appendices, such as phrases like "see appendix for details," "see attached table," "implement according to the appendix," and "specific requirements are in Appendix 1." By identifying these citation relationships, the system can determine the logical connection between a paragraph, clause, or condition in the policy text and the corresponding policy appendix. Instead of simply appending the full text of the policy appendix to the end of the policy text, the system organizes the specific content of the policy appendix as a content unit associated with the citation location in the text, ensuring a correspondence between the contextual constraints in the text and the specific content in the appendix within the same parsed structure. Through this process, the application requirements, material lists, condition thresholds, and implementation rules, originally scattered throughout the policy text and appendices, can be integrated into a unified parsed text block, thereby reducing information fragmentation caused by the separation of the text and appendices.

[0017] To more clearly illustrate the construction process of the unified parsing text block, the following specific implementation method can be adopted. The system first performs text extraction or optical character recognition on the policy text to obtain the original text sequence. Then, the system scans the policy text using a preset citation pattern library. The citation pattern library stores common citation expression templates, such as regular expression patterns like "See Attachment [Number]", "See Attachment [Number]", and "Specific [Requirements / Standards / Conditions] See Attachment [Name]". When the system detects a statement in the policy text such as "Specific application conditions are detailed in Attachment 1, 'Application Guidelines'", it extracts the attachment identifier "Attachment 1" and the attachment name "Application Guidelines". Next, the system searches the acquired file list for policy attachments whose filenames or titles contain "Attachment 1" or "Application Guidelines", locating the corresponding attachment file. Then, the system extracts the specific content of the policy attachment and inserts a citation marker anchor at the corresponding citation position in the policy text, storing the specific content of the policy attachment as a child node of the citation marker anchor. The resulting unified parsed text block is represented in the data structure as a structured document object containing contextual references. The policy body and the policy appendix are logically linked through reference identifiers, so that specific conditions such as "registered capital not less than X million yuan" in the appendix can establish a correct semantic association with general descriptions such as "meets the conditions for financial support" in the body.

[0018] After obtaining the unified parsed text block, semantic analysis is performed on it to identify the fine-grained conditions and logical relationships between them. A conditional logic tree representing the hierarchical and dependency relationships of these fine-grained conditions is then constructed. Fine-grained conditions are the smallest conditional units extracted from the unified parsed text block that can directly participate in enterprise qualification or policy matching judgments, such as "registered for at least 3 years," "the enterprise has at least 5 authorized invention patents," and "no major safety accidents recorded in the past 3 years." When identifying fine-grained conditions, the system not only extracts the conditional phrases themselves but also analyzes the connections and constraints between conditions, such as logical connectives like "simultaneously," "and," "or," "any," "but," "except," and "must not." By parsing these logical connectives, it can determine whether different fine-grained conditions are related by "AND," "OR," or "NOT," or whether there are exceptions, exclusions, or prerequisite dependencies. Finally, the system organizes the identified fine-grained conditions and logical relationships into a conditional logic tree, enabling policy qualification requirements to be expressed in a structured manner.

[0019] For example, the unified parsing text block contains the following content: "Applicant companies must simultaneously meet the following conditions: (i) have been registered and established in this city for more than 2 years; (ii) belong to the software and information technology service industry; (iii) have annual operating revenue of not less than RMB 5 million or annual R&D investment of not less than RMB 1 million. In addition, companies listed in the list of companies with abnormal operations are not allowed to apply." When the system performs semantic analysis, it can identify the following fine-grained conditions: Condition A is "has been registered and established in this city for more than 2 years"; Condition B is "belongs to the software and information technology service industry"; Condition C1 is "annual operating revenue of not less than RMB 5 million"; Condition C2 is "annual R&D investment of not less than RMB 1 million"; Condition D is "listed in the list of companies with abnormal operations". At the same time, the system identifies that "simultaneously meet" means that there is an "AND" relationship between Condition A, Condition B and Condition (iii), "OR" means that there is an "OR" relationship between Condition C1 and Condition C2, and "cannot apply" means that Condition D is not applied. Based on the above analysis, the conditional logic tree can be constructed as follows: the root node corresponds to the policy entity, and the first-level logical operator "AND" is connected under the root node. This "AND" relationship includes condition A, condition B, a sub-logical operator "OR", and a sub-logical operator "NOT". Among them, the "OR" relationship includes condition C1 and condition C2, and the "NOT" relationship connects to condition D. Through this conditional logic tree, the qualification constraint structure of the original policy, which requires "several conditions to be met simultaneously, some of which can be met selectively, and there are exclusionary restrictions", can be fully expressed.

[0020] After constructing the conditional logic tree, it is converted into a relational representation in the policy knowledge graph. The policy knowledge graph can be implemented using Neo4j or other graph databases to store policy entities, fine-grained conditions, logical operators, and the relationships between them. Specifically, policies themselves can be created as policy entity nodes; for example, "entry into the high-tech enterprise cultivation pool" can be created as a node of type "POLICY," and this node can be configured with attributes such as policy name, issuing unit, validity period, and applicable region. Each fine-grained condition can be created as a condition node; for example, "registered for at least 2 years" can be created as a node of type "CONDITION," and this node can be configured with attributes such as condition description, condition type, threshold parameter, and unit. Logical relationships such as "AND," "OR," and "NOT" can be created as logical operator nodes; for example, a node of type "OPERATOR" can be created, and the operation type can be configured as AND, OR, or NOT. Subsequently, the system establishes the "HAS_ROOT" relationship from policy entity nodes to root logical operator nodes, and the "HAS_CHILD" relationship from logical operator nodes to their child condition nodes or child logical operator nodes. Through the above transformation, the hierarchical relationships and dependencies in the conditional logic tree can be stored as relational expressions in the policy knowledge graph, enabling the subsequent matching process to directly traverse and determine based on the graph structure.

[0021] After the policy knowledge graph is constructed, the system acquires the enterprise profile data of the companies to be matched and converts this data into enterprise profile fact nodes. Enterprise profile data is structured data describing a company's basic information, operating status, qualifications and capabilities, financial indicators, intellectual property, personnel size, R&D investment, credit records, etc. It can originate from enterprise databases, enterprise information management platforms, business registration data interfaces, tax data interfaces, or data self-reported by the company. For example, an enterprise's profile data might include: "Registered Region" as "This City", "Industry" as "Software and Information Technology Services", "Operating Status" as "Normal", "Years Established" as "3 Years", "Previous Year's Revenue" as "6 Million Yuan", and "Previous Year's R&D Investment" as "800,000 Yuan". The system can create this enterprise as an "ENTERPRISE" node and create separate "FACT" fact nodes for each fact in the enterprise profile data, such as "Years Established Fact Node", "Industry Fact Node", and "Revenue Fact Node". Each enterprise profile fact node can be configured with attributes such as actual value, unit of measurement, data source, and update time, and is connected to enterprise nodes through the "HAS_FACT" relationship. In this way, enterprise profile data is transformed into a graph fact structure that can perform relationship matching and condition determination with the policy knowledge graph.

[0022] Based on the relational representation in the policy knowledge graph, conditional matching is performed on the fact nodes of the enterprise profile to obtain the policy matching results for enterprise services. The matching process can be understood as substituting the factual data in the fact nodes of the enterprise profile into the corresponding conditional logic tree of the policy knowledge graph for logical operations. The system can use a recursive traversal algorithm, starting from the root logical operator node corresponding to the policy entity node, visiting its child nodes layer by layer, and performing corresponding judgments according to the node type. Taking the above enterprise as an example, the system first visits condition A, queries the "years of establishment" in the fact node of the enterprise profile, finds it to be 3 years, and compares it with the requirement of "2 years" in condition A, and determines that the condition is met; then it visits condition B, queries the "industry" in the fact node of the enterprise profile, finds it to be "software and information technology services", which is consistent with the requirement of condition B, and determines that the condition is met; then it visits the sub-logical operator "OR", and judges conditions C1 and C2 respectively, where condition C1 requires "the previous year's operating revenue is not less than 5 million yuan", the "previous year's operating revenue" in the fact node of the enterprise profile is 6 million yuan, so condition C1 is determined to be met. Since any sub-condition in the "OR" relation is satisfied for the branch to be true, the system can determine that the branch is satisfied based on the short-circuit evaluation principle. Finally, the system accesses condition D under the "NOT" operator and queries the "Operating Status" in the enterprise profile fact node, which is "Normal," and does not meet the condition of "being listed in the list of abnormal operations." Therefore, the "NOT" operation result is satisfied. Since all "AND" relation branches under the root node are satisfied, the system finally obtains the enterprise service policy matching result that the enterprise meets the corresponding policy conditions.

[0023] Based on the enterprise service policy matching results, the system outputs matching policy information, information on unmet conditions, or supplementary material prompts. If the enterprise service policy matching result indicates a successful match, the system can output the corresponding matching policy information, such as indicating that the enterprise meets the application conditions for "entry into the High-tech Enterprise Cultivation Pool," and providing the policy name, application portal, application deadline, link to the original policy text, and required materials. If the enterprise service policy matching result indicates a failed match, the system can output information on unmet conditions based on the failed condition nodes in the condition logic tree. For example, if an enterprise's operating revenue in the previous year was 4 million yuan and its R&D investment in the previous year was 800,000 yuan, the system can indicate that the unmet condition is "the previous year's operating revenue is not less than 5 million yuan, or the previous year's R&D investment is not less than 1 million yuan." If a necessary factual data is missing from the enterprise profile fact nodes, such as the number of authorized invention patents or the record of safety accidents in the past three years, the system can output supplementary material prompts, suggesting that the enterprise supplement the corresponding supporting materials, data fields, or application attachments so that the policy matching can be performed again.

[0024] Through the above processing, the technical solution of this application first constructs a unified parsing text block to integrate the contextual constraints in the policy text with the specific content in the policy appendices, thus fundamentally solving the information fragmentation problem caused by the separation of the policy text and policy appendices. Subsequently, by identifying fine-grained conditions and the logical relationships between them, a conditional logic tree is constructed, enabling the accurate expression of complex conditional relationships such as "AND," "OR," "NOT," and exclusionary restrictions in the policy. Furthermore, the conditional logic tree is converted into relational expressions in a policy knowledge graph, and enterprise profile data is converted into enterprise profile fact nodes, allowing policy conditions and enterprise fact data to be matched within a unified graph structure. Compared to keyword matching that treats policy conditions as isolated text, this application preserves the hierarchical and dependency relationships between policy conditions, thereby reducing policy matching deviations caused by missing appendices, missing logical relationships, or incorrect condition parsing, and improving the accuracy and interpretability of enterprise service policy matching results.

[0025] Optionally, the steps of performing semantic analysis on the unified parsed text block, identifying the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and constructing a condition logic tree representing the hierarchical relationships and dependencies of the fine-grained conditions include: Lexical analysis is performed on the unified parsed text blocks to obtain word sequences and general part-of-speech tags within the word sequences. The system first performs word segmentation, employing a segmentation algorithm combining dictionary and statistical models to divide continuous text into independent word units. For example, for the text "Application enterprises must simultaneously meet the following conditions," the segmentation result is "Application / Enterprise / Need / Simultaneously / Meet / The following / Conditions." Then, the system performs part-of-speech tagging on each word, annotating its general grammatical attributes such as noun (n), verb (v), numeral (m), and quantifier (q), resulting in a word sequence with part-of-speech tags.

[0026] The process involves labeling words in a word sequence that have already been tagged with general part-of-speech tags with policy domain semantic types, resulting in a word sequence with policy domain semantic types. This step relies on a pre-defined policy domain semantic labeling model. This model is obtained as follows: First, a large amount of historical policy texts are collected as training corpus, and policy research experts are invited to manually label the entity types, such as labeling "national high-tech enterprise" as "qualification level type," "operating income" as "financial indicator type," and "the past 3 years" as "time range type." Then, a Conditional Random Field (CRF), Bidirectional Long Short-Term Memory (BiLSTM) network, or Transformer-based sequence labeling algorithm is used to train the policy domain semantic labeling model using the labeled corpus. In practical applications, the system inputs the segmented sequence into the model, and the model outputs the domain semantic type label for each word or word fragment. For example, for "registered capital of 1 million yuan or more," the system will label "registered capital" as "registered capital indicator," "1 million yuan" as "amount value," and "above" as "comparison operator."

[0027] Based on word sequences with policy-domain semantic types, the system performs semantic analysis on conditional statements in unified parsing text blocks to identify conditional phrases representing the applicant, business status, qualification requirements, time range, geographical scope, numerical threshold, supporting documents, and exclusion scenarios, resulting in fine-grained conditional recognition results. The system identifies conditional phrases through a combination of pattern matching and rule-based reasoning. Specifically, the system maintains a conditional recognition rule base, storing conditional pattern templates obtained through inductive analysis, such as "[Qualification Level Type] + Enterprise", "[Financial Indicator Type] + [Comparison Operator] + [Amount Value]", and "[Time Range Type] + Within + [Event Description]". When a word sequence combination matching these patterns is scanned, the system recognizes it as a complete conditional phrase. For example, for "belongs to the software and information technology services industry," the system identifies "software and information technology services industry" as an "industry classification type," and combined with the preceding verb "belongs to," determines that this is a conditional phrase about industry attributes.

[0028] The system standardizes each fine-grained condition in the fine-grained condition identification results to obtain standardized fine-grained conditions. Since the same meaning may have multiple expressions, the system needs to unify them into a standard format. The system has a pre-set standardized mapping table, which is established by manually compiling synonyms, near-synonyms, and different expressions. For example, expressions such as "not less than," "not lower than," "greater than or equal to," and "at least" are mapped to the comparison symbol "≥"; "not more than," "not higher than," "less than or equal to," and "at most" are mapped to "≤"; "1 million yuan" and "1,000,000 yuan" are mapped to the standard numerical value "1,000,000" and marked with the unit "yuan." For time duration conditions such as "full 3 years," the system standardizes it to "duration ≥ 3 years" based on the condition type. According to this mapping table, the system converts the identified condition phrases into a standard format, such as standardizing "registered capital not less than 1 million yuan" to "registered capital ≥ 1,000,000 yuan."

[0029] The system performs syntactic analysis on word sequences with policy domain semantic types to construct a syntactic dependency tree. A dependency parser is used to analyze the grammatical dependencies between various components of a sentence. For example, in the sentence "Applicant companies must simultaneously meet the following conditions," the dependency parser identifies "meet" as the core predicate verb, "company" as the subject dependent on "meet," "simultaneously" as an adverbial modifying "meet," and "conditions" as the object dependent on "meet." These dependencies are represented in a tree structure, with the root node being the predicate headword. Other nodes are connected to the headword or other nodes via dependency arcs, and the dependency relationship type is labeled, such as subject-predicate, verb-object, adverbial-head, and coordinate relationships.

[0030] The system traverses the syntactic dependency tree, identifies logical connectives representing coordination, selection, negation, progression, premise, exception, or supplementary explanation, and determines the logical connections between normalized fine-grained conditions. The system searches for logical markers with specific dependency relationships by traversing the dependency tree. For example, when words like "simultaneously," "and," or "moreover" modify multiple parallel conditional phrases, or multiple conditional phrases are in an enumeration structure at the same level, the system determines that these conditions have an "AND" relationship; when words like "or" or "or" connect conditions, the system determines that they have an "OR" relationship; when structures like "except for" or "but except for" are found, the system determines that they have an exception negation relationship. The system records the start and end positions and the type of these logical connections.

[0031] Based on the results of conditional logical associations, the hierarchical relationships and dependencies between normalized fine-grained conditions are determined. The system determines the condition hierarchy based on the hierarchical structure of the syntactic dependency tree and the nesting relationships of logical associations. For example, if a composite condition representing an "AND" relationship contains two sub-conditions connected by "OR", then the "OR" relationship becomes a child node of the "AND" relationship, forming a nested structure. The system constructs the complete hierarchical structure through recursive analysis.

[0032] Based on normalized fine-grained conditions, logical associations, hierarchical relationships, and dependencies, a conditional logic tree is constructed. The system adopts a tree-structured data structure, creating a root node to represent the overall policy, and then creating logical operator nodes and conditional leaf nodes level by level according to the hierarchical relationships. For each logical operator node, its operation type (AND / OR / NOT) and list of child nodes are recorded; for each conditional leaf node, its normalized conditional expression, original text, threshold parameters, and other attributes are recorded.

[0033] The system identifies and parses time adverbs in unified text blocks, attaching them as time constraints to corresponding condition nodes or as status verification nodes in the conditional logic tree. It detects time adverbs in the text using time expression recognition rules, such as "between January 1, 2023 and December 31, 2023" or "within 3 years from the date of establishment." For relative time expressions, the system calculates the absolute time range by combining the benchmark time information in the policy text; when the policy text does not explicitly provide a benchmark time sufficient to determine the absolute time range, it retains the expression as a relative time constraint. Then, the system attaches these time constraints as special attributes to relevant condition nodes or inserts them as independent time verification nodes into appropriate positions in the conditional logic tree to ensure the time validity of the enterprise's status can be verified during matching.

[0034] Optional, combined Figure 2 As shown, the steps for traversing the syntactic dependency tree, identifying logical connectives representing coordination, selection, negation, progression, premise, exception, or supplementary explanation, and determining the logical connections between normalized fine-grained conditions by combining the hierarchical position and scope of the logical connectives in the syntactic dependency tree, to obtain the results of conditional logical connections include: A1 traverses the syntactic dependency tree to identify explicit logical association words that are directly connected to the normalized fine-grained conditions. A2. Determine the scope of explicit logical connectives in the syntactic dependency tree based on the connection relationships between the parent nodes, child nodes, and adjacent nodes at the same level of the explicit logical connectives. A3, based on the type and scope of the explicit logical connectives, determine the explicit logical connections between normalized fine-grained conditions connected by the explicit logical connectives. A4, based on a pre-defined set of policy domain logical conventions, performs semantic similarity assessment on adjacent normalized fine-grained conditions that are not connected by explicit logical association terms; wherein, the set of policy domain logical conventions is used to store implicit logical relationship judgment rules commonly found in policy texts. Implicit logical relationship judgment rules include one or more of the following: adjacent qualification conditions are assumed to be parallel, the preceding subject conditions have a premise constraint relationship on the subsequent material conditions, the exception statement has an exclusion relationship with the aforementioned conditions, and the supplementary explanatory statement has a supplementary limiting relationship with adjacent conditions; A5. Based on the semantic similarity between adjacent normalized fine-grained conditions, their relative positions in the syntactic dependency tree, and their syntactic roles, determine whether there is an implicit logical relationship between adjacent normalized fine-grained conditions. A6. When it is determined that there is an implicit logical association, insert a virtual logical operator node in the syntactic dependency tree, and determine the hierarchical position and scope of the virtual logical operator node according to the type of implicit logical association. A7, based on the hierarchical position and jurisdiction of the virtual logic operator node, integrates explicit and implicit logical associations to generate a standardized fine-grained conditional logical association result.

[0035] The process involves traversing the syntactic dependency tree to identify explicit logical connectives directly linked to the normalized fine-grained conditions, and determining their scope and logical relationships. Then, a core mechanism is introduced: based on a pre-defined set of policy domain logical conventions, adjacent conditions not connected by explicit logical connectives are analyzed. This set of logical conventions is a rule base built through manual analysis and summarization of a large number of historical policy texts. Specifically, various enterprise support policy texts published over the past few years are selected as samples. Policy analysis experts analyze the correspondence between the arrangement of conditions and logical relationships in each policy, abstracting common implicit logical patterns into rule entries and storing them in the logical convention set.

[0036] Below are some examples of rules from the logical convention set: Rule 1: Adjacent qualification conditions are assumed to be parallel. That is, consecutive conditions under bulleted or numbered lists are assumed to be in an "AND" relationship. Specifically, when the system detects multiple condition phrases preceded by sequential numbers (such as "(I)", "1.", "Item 1", etc.) or bullets (such as "•", "·", etc.), and there are no explicit logical connectors (such as "or" or "or") between these condition phrases, it automatically infers that these conditions are in a logical "AND" relationship. For example, when the text contains "(I) Registered and established in this city for more than 2 years; (II) Belongs to the software and information technology service industry;", the system automatically inserts a virtual "AND" logical node between condition (I) and condition (II).

[0037] Rule 2: Prerequisite conditions for the applicant have a prerequisite constraint relationship with subsequent material conditions. For example, if the preceding text states "the applicant is a micro or small enterprise," and the following text states "a micro or small enterprise certification is required," then the former is a prerequisite for the latter. The specific determination method is as follows: when the system identifies a conditional phrase involving the applicant's qualifications (such as enterprise size, enterprise nature, industry sector), and subsequently identifies a conditional phrase involving the submission of supporting materials, and these supporting materials are directly related to the applicant's qualifications (e.g., both contain the keyword "micro or small enterprise"), then the former is determined to be a prerequisite for the latter. In the conditional logic tree, this is represented as the material condition node depending on the applicant condition node, forming a dependency relationship.

[0038] The system uses these rules to determine whether implicit logical relationships exist between adjacent conditions that are not explicitly connected. When a relationship is found, the system adds a virtual logical operator node, such as a virtual "AND" node, during the construction of the conditional logic tree to explicitly represent the inferred logical relationship. Finally, the system integrates the explicit logical relationships with the implicit logical relationships inferred from the set of conventions to generate the final, complete conditional logical relationship result. This method of combining explicit and implicit logic overcomes the limitation of traditional methods that can only handle explicit logical conjunctions. It enables the system to understand the "subtext" of policy texts, extracting hidden logical relationships from the structure of the article and the conventions of the domain, thereby constructing a conditional logic tree that truly reflects the policymaker's intentions. This ensures the depth and breadth of policy analysis, making the final matching result more reliable and intelligent.

[0039] Optionally, the step of performing lexical analysis on the unified parsed text block to obtain the word sequence and the general part-of-speech tagging in the word sequence specifically includes: Identify word fragments from the unified parsed text block that are not effectively included in the general dictionary or are incorrectly segmented in the current context, and use these word fragments as candidate new words.

[0040] Based on the frequency of occurrence of candidate new words in the unified parsed text block, the context, and the semantic relevance between candidate new words and known policy domain terms, candidate new words are screened and weighted to obtain the screened candidate new words.

[0041] The selected candidate words and their weights are added to the domain dictionary of the lexical analyzer to obtain the updated domain dictionary; the word segmentation rules of the lexical analyzer are adjusted according to the context of the candidate words to obtain the adjusted word segmentation rules.

[0042] Based on the updated domain dictionary and the adjusted word segmentation rules, lexical analysis is performed on the unified parsed text block to obtain word sequences and general part-of-speech tags.

[0043] This specific implementation plan introduces a dynamic new word discovery and dictionary update mechanism. First, in the candidate new word identification stage, the system can employ a statistical method. Specifically, the system first performs N-gram segmentation on the unified parsed text block, extracting all continuous strings of length 2 to A characters as a candidate string set, where A is set based on the corpus's linguistic features and the target term length distribution. For each candidate string, the system calculates its internal cohesion and boundary degrees of freedom. Taking mutual information calculation as an example, the system statistically analyzes the co-occurrence frequency and independent occurrence frequency of adjacent substrings in the candidate string within the overall corpus, and calculates point mutual information or normalized cohesion scores based on the candidate string's segmentation method. When this score exceeds a preset threshold, the candidate string is considered to have high internal cohesion. Simultaneously, the system calculates the information entropy of the sets of adjacent characters to the left and right of the candidate string. When the information entropy of both left and right adjacent characters is greater than a preset threshold, it indicates that the string has high boundary degrees of freedom and possesses the characteristics of an independent word. For example, when processing a batch of policy texts, the system will extract continuous strings related to "specialized, refined, and innovative" and calculate the cohesion between substrings corresponding to different segmentation positions. In numerous policy texts, these four characters typically appear in fixed combinations, exhibiting a high degree of internal cohesion. Meanwhile, the left side of "specialized, refined, distinctive, and innovative" may contain words like "cultivate" or "apply," while the right side may contain words like "enterprise" or "project," resulting in a richer set of adjacent characters and a higher degree of boundary freedom. Therefore, when both cohesion and freedom meet the threshold conditions, the system will identify it as a candidate new word.

[0044] Next, the system performs multi-dimensional evaluation when screening and weighting candidate new words. Specifically, the system first counts the frequency of candidate new words in the policy text corpus. If the frequency is lower than the minimum support count, it is directly filtered. For candidate new words that pass the frequency screening, the system uses a pre-trained policy domain word vector model to calculate semantic relevance. This word vector model is obtained by collecting policy texts from the past B years as training data. After cleaning and word segmentation preprocessing, a Skip-gram model architecture is adopted, setting the word vector dimension to , the context window size to , and the minimum word frequency to . Iterative training rounds are conducted, and the model's effectiveness is verified by combining downstream validation task performance or manual sampling evaluation results. When the model's effectiveness meets the preset standards, the domain word vector model is obtained. For example, for the candidate new word "chain leader enterprise," the system counts its frequency of occurrence in the policy text corpus. Simultaneously, the system analyzes the co-occurring words within its context window and finds that it frequently appears alongside known policy domain terms such as "industrial chain," "supply chain," and "leading enterprise." Furthermore, the system utilizes the trained word vector model to calculate the cosine similarity between the word vector of "chain leader enterprise" and the word vectors of known terms such as "leading enterprise" and "core enterprise." When the similarity exceeds a threshold, it indicates that they are highly semantically related. Combining this information, the system calculates a comprehensive score according to a preset weight allocation strategy. When the score exceeds a threshold, a higher weight value is given to the candidate new word, confirming it as a valuable new word in the domain.

[0045] Finally, the system adds the term "chain leader enterprise" and its weight to a custom domain dictionary. In practice, the domain dictionary is stored in a structured database, with each record containing fields such as word text, part-of-speech tag, weight value, source tag, and update timestamp. During lexical analysis, the word segmenter employs a segmentation mechanism combining dictionary matching and statistical models, prioritizing queries to the domain dictionary. When "chain leader enterprise" is matched and its dictionary weight and contextual judgment result meet the segmentation conditions, it is directly segmented as a whole, rather than incorrectly split into two words. Simultaneously, the segmentation rule library dynamically adds a segmentation rule, with an example format: prohibiting segmentation between "chain leader" and "enterprise," and setting the rule priority to high. The system loads this rule through the rule engine and enforces this segmentation constraint during subsequent word segmentation.

[0046] Optionally, the step of performing policy domain semantic type annotation on words in the word sequence that have already been labeled with general part-of-speech tags to obtain a word sequence with policy domain semantic types specifically includes: Based on a pre-defined set of policy domain semantic type rules, the semantic type of words in the word sequence that have been marked with general part-of-speech tags is identified.

[0047] When a word or phrase matches a matching rule in the policy domain semantic type rule set, the word or phrase is labeled with the corresponding policy domain semantic type according to the matched rule.

[0048] The output is a word sequence with both general part-of-speech tagging and policy domain semantic type, which is a word sequence with policy domain semantic type.

[0049] The core of this specific implementation plan lies in constructing a rich and multi-dimensional set of semantic type rules for the policy domain. In practice, this rule set is stored using a hierarchical data structure, with each rule containing attributes such as rule identifier, rule type, matching pattern, semantic type label, priority score, and activation status. The rule set construction process is as follows: first, domain experts define basic rule templates based on common expressions in policy texts; then, high-frequency patterns are extracted from an annotated policy corpus using a semi-automatic method and added to the rule set; finally, after manual review and testing, an initial preset rule set is formed.

[0050] The following examples illustrate different types of rules: Keyword matching rules: These are the most basic rules used to identify terms with a clear target meaning. For example, the rule base defines that if a word is an entry in the dictionary of "business license," "audit report," or other material names, then its semantic type is "material condition type." In practice, the system maintains a keyword dictionary, stored using a hash table structure, to achieve fast lookup.

[0051] Syntactic structure matching rules: These rules combine words and their syntactic relationships to identify more complex expressions. For example, a rule might be defined as: finding a verb-object phrase with "own," "possess," or "obtain" as the core verb; if its object contains words like "patent," "copyright," or "trademark," then the entire verb-object phrase (e.g., "owns 5 invention patents") is labeled as either an "intellectual property condition type" or a "qualification condition type," with the specific type determined by a predefined semantic system. In practice, the system uses a dependency parser to analyze the sentence structure, extract the subject-verb-object relationship, and then match it with the rule template.

[0052] Part-of-speech tagging rules: These rules utilize general part-of-speech sequences to identify specific types of phrases. For example, a rule could be defined as: a sequence indicating duration, such as "numeral + time quantifier" or an equivalent part-of-speech tagging pattern, should be labeled as "time span type". The system identifies part-of-speech tagging sequence patterns using regular expressions or finite state automata.

[0053] Numerical expression matching rules: Such rules usually use regular expressions to precisely match various numerical formats. For example, a rule can be defined as: match the pattern in the form of "([0-9]+(.[0-9]+)?)\s*(ten thousand? yuan)", extract the numerical part therein as a numerical value, and label the entire phrase as "amount numerical type". In specific writing, this regular expression will consider different writing forms such as full-width and half-width numbers, Chinese numerals, and thousands separators.

[0054] Time expression matching rules: Regular expressions are also adopted for matching dates and time ranges. For example, the rules can match absolute dates in the format of "YYYY year MM month DD day", or relative time ranges such as "nearly X years", and uniformly label them as "time condition type".

[0055] Material name matching rules: This can be regarded as an extension of keyword matching, which maintains a dedicated and dynamically updatable material name dictionary. When a phrase in the text completely or fuzzily matches an entry in the dictionary (such as "Annual Corporate Income Tax Return"), it is labeled as "material condition type".

[0056] Optionally, the step of performing policy domain semantic type recognition on words that have been added with general part-of-speech tags in the word sequence according to a preset policy domain semantic type rule set specifically comprises: Sorting the matching rules in the preset policy domain semantic type rule set by matching priority to obtain a rule matching priority sorting result.

[0057] Performing conflict detection on the matching rules in the preset policy domain semantic type rule set to obtain a rule conflict detection result.

[0058] When the rule conflict detection result indicates that a rule conflict exists, adjust or merge the related conflicting rules according to a preset conflict resolution strategy, so as to obtain an optimized policy domain semantic type rule set.

[0059] When the rule conflict detection result indicates that no rule conflict exists, update the priority of the matching rules in the policy domain semantic type rule set according to the rule matching priority sorting result, and take the policy domain semantic type rule set after priority update as the optimized policy domain semantic type rule set. Policy domain semantic type recognition is performed on the words that have been added with general part-of-speech tags in the word sequence according to the optimized policy domain semantic type rule set.

[0060] This specific implementation plan introduces a rule governance mechanism. First, it prioritizes matching rules. In practice, the system quantifies and scores rules based on their specificity. Scoring factors include the length of the rule pattern, the number of qualifiers, whether regular expressions are used, and the degree of constraint. For example, a rule pattern matching "high-tech enterprise" has a length of A characters and contains stronger semantic constraints, scoring 1 / 2; while a rule pattern matching "enterprise" has a length of B characters and weaker constraints, scoring 1 / 2. The system sorts rules from highest to lowest score, with rules of higher specificity receiving higher priority.

[0061] Secondly, conflict detection is performed. In practice, the system traverses rule pairs in the rule base using a pairwise comparison method. For each pair of rules, the system first checks whether their matching patterns have overlap, i.e., whether there are text fragments that can be matched by both rules simultaneously. Then, it checks whether their output semantic types are the same. If the patterns overlap but the output types are different, it is marked as a mutual exclusion conflict. If the patterns overlap and the matching set of one rule is covered by the other rule, or the two rules overlap in their main applicable scenarios and have the same output type, it is marked as a redundant conflict. For example, if the system finds that rule A (matching pattern "last three years") and rule B (matching pattern "within three years") can both function in scenarios expressing time ranges and both output "time range type", it can further determine whether to merge or retain their respective applicable ranges. As another example, if the system finds that rule C (matching "high-tech enterprise" and outputting "qualification condition type") and rule D (matching "high-tech enterprise" and outputting "subject type") have the same matching pattern but different output types, this is a mutual exclusion conflict.

[0062] Then, conflict resolution strategies are executed based on the conflict detection results. For redundant conflicts, the system employs rule merging or rule orchestration strategies to merge two rules into a more general rule, or retain them as different sub-rules and attach them to a unified template. For example, the matching mode can be adjusted to a rule template that can cover multiple time expressions, while disabling the original duplicate rules or redirecting them to the unified template. For mutually exclusive conflicts (such as rules C and D), the system retains the higher-priority rule (rule C) based on the preset priority sorting results and modifies the lower-priority rule (rule D). For example, an exclusion condition "but does not contain the complete phrase related to 'high-tech enterprise'" can be added to rule D, so that it no longer matches "high-tech enterprise", or the enabled status of rule D can be directly set to disabled.

[0063] Optionally, the step of sorting the matching rules in the preset policy domain semantic type rule set by matching priority to obtain the rule matching priority ranking result specifically includes: Obtain metadata information for each rule from a pre-defined policy domain semantic type rule set. Determine the rule specificity and coverage of each rule based on its matching pattern.

[0064] Calculate the rule activity score based on metadata information. Calculate the rule stability score based on metadata information. Calculate the rule's overall priority score based on the rule activity score, rule stability score, rule specificity, rule coverage, and domain expert ratings.

[0065] Based on the rule comprehensive priority score, the matching rules in the policy domain semantic type rule set are prioritized and sorted to obtain the rule matching priority ranking result.

[0066] This specific implementation plan establishes a comprehensive "evaluation profile" for each rule and calculates its overall priority based on this profile. First, the system records rich metadata information for each rule, including the rule's creation time, modification history, creator ID, cumulative trigger count, most recent trigger time, number of error annotation feedbacks, and an initial benchmark score given by domain experts. The initial benchmark score can be represented using a range-based scoring system.

[0067] Secondly, the system quantifies and evaluates the characteristics of the rules themselves: Rule Specificity: This measures the accuracy of the rule. Specifically, the system calculates features such as the number of required terms, constraints, restrictive structures in the regular expression, and the level of logical operations within the rule's matching pattern. For example, a rule matching the pattern "The previous year's R&D investment as a percentage of operating revenue is not less than X%" contains specific business terms and numerical constraints, resulting in a high specificity score; while a rule that only matches "not less than X%" has a lower specificity score.

[0068] Rule Coverage: This measures the generality of a rule. In the evaluation, the system applies the rule to a large-scale historical corpus (such as policy texts from the most recent C years) and counts the number of different text segment categories that the rule can successfully match, or the normalized coverage rate. Higher coverage indicates a broader rule scope. For example, a rule template that can match "the last three years," "the last five years," and "the last year" has a higher coverage score than a rule that can only match a single time period.

[0069] The system then evaluates the dynamic performance of the rules: the rule activity score, which reflects the "timeliness" of the rules. Specifically, the system counts the number of times each rule was successfully triggered in the most recent time period and normalizes the trigger counts for different corpus sizes. A higher trigger count generally results in a higher activity score. For example, if a rule is frequently triggered in the most recent observation period, its activity score is high, indicating that it remains effective for newly released policies.

[0070] Rule stability score: Reflects the "maturity" of a rule. In the specific evaluation, the system analyzes the rule's modification history. If a rule has been modified less than a certain number of times since its creation, and its accuracy or consistency metrics on the test set reach a threshold, then the stability score is high; if it has been modified frequently, the stability score decreases.

[0071] Finally, the system calculates the comprehensive priority score for each rule by weighted summation of all these dimensions of data. In specific implementation, the system uses the following calculation logic: the comprehensive score equals the expert score multiplied by a weight coefficient, plus the rule specificity multiplied by a weight coefficient, plus the rule coverage multiplied by a weight coefficient, plus the rule activity multiplied by a weight coefficient, and plus the rule stability multiplied by a weight coefficient. The contribution direction of rule coverage can be set according to the target optimization direction: when emphasizing exact matching, a suppression term can be introduced for overly broad coverage; when emphasizing recall, it can be used as a positive gain term. The method for determining the weight coefficients is as follows: a set of values ​​is initially set using expert experience, then tested on the validation set, and the weights are adjusted based on feedback from annotation accuracy, recall, or comprehensive evaluation indicators until the target effect is achieved. The weight ratio can be adjusted according to actual needs; for example, when pursuing high precision, the weights can be appropriately increased, and other weights can be adjusted accordingly.

[0072] Optionally, the step of performing conflict detection on the matching rules in the preset policy domain semantic type rule set to obtain the rule conflict detection results specifically includes: Obtain the matching pattern, output semantic type, and rule source information for each rule in the preset policy domain semantic type rule set.

[0073] For each rule's matching pattern, the system analyzes the keywords, syntactic structures, and part-of-speech combinations contained within the pattern to construct a rule feature description representing the matching pattern. In practice, the system first parses the rule matching pattern, extracting its core elements. For keywords, the system converts them into vector representations using TF-IDF or a bag-of-words model. For syntactic structures, the system records the part-of-speech sequence patterns and dependency constraints defined in the rule, such as structured templates like "verb + object + number". For part-of-speech combinations, the system extracts the specified part-of-speech tag sequences from the rule. Subsequently, the system merges these heterogeneous features into a multi-dimensional rule feature vector or structured descriptor. For example, for the rule "match 'high-tech enterprise'", its feature description includes: keyword vectors (high-tech, enterprise), syntactic structure (noun phrase), and part-of-speech combination (adjective + noun).

[0074] The system pairs any two rules from the policy domain semantic type rule set and compares the similarity of the corresponding rule feature descriptions between each pair. When the similarity exceeds a preset similarity threshold, it determines that there is pattern overlap between the corresponding rule pairs. In practice, the system uses a double loop to traverse the rule base and calculates the feature similarity for each pair of rules. For keyword vectors, cosine similarity is used; for syntactic structures and part-of-speech combinations, edit distance or structural similarity algorithms are used. When the overall similarity score exceeds the threshold (usually set between 0.7 and 0.9), the system marks the rule pair as having pattern overlap.

[0075] The system analyzes whether there is a mutually exclusive relationship between the output semantic types of the rules. In practice, the system maintains a hierarchical relationship graph of semantic types, defining the mutual exclusion, inclusion, or intersection relationships between various semantic types. For example, "eligibility condition type" and "subject type" may be mutually exclusive in a specific context, while "time condition type" and "regional condition type" are usually independent dimensions. When the output semantic types of two rules are marked as mutually exclusive nodes in the hierarchical relationship graph, or belong to mutually exclusive subclasses under the same parent class, the system determines that a mutual exclusion relationship exists.

[0076] For rule pairs with overlapping or mutually exclusive matching patterns, the overlap or contradiction between the rule pairs is quantitatively marked by combining the rule source information, resulting in rule conflict marking information. The rule conflict marking information includes the identifier of the conflicting rule pair, the type of rule conflict (pattern overlap, semantic mutual exclusion, or composite), the degree of conflict (quantified by similarity value), the potential scope of impact (statistically counting the number of texts that have triggered both rules simultaneously in history), and information on the differences in rule sources (e.g., one from expert definition, and the other from automatic extraction).

[0077] Based on the rule conflict marker information, rule conflict detection results are generated. The detection results are output in structured data format, including a list of conflicting rule pairs, detailed attributes of each conflict, and suggested priority processing order (usually sorted in descending order of conflict severity).

[0078] Optionally, when the rule conflict detection result indicates the existence of a rule conflict, the relevant conflicting rules are adjusted or merged according to a preset conflict resolution strategy to obtain an optimized policy domain semantic type rule set. This step specifically includes: Read the rule conflict detection results and the rule conflict marker information corresponding to the rule conflict detection results.

[0079] Based on the rule conflict marker information, read the corresponding rule conflict type. Rule conflict types include overlapping matching patterns, mutually exclusive output semantic types, or differences in rule sources.

[0080] When rule conflicts occur due to overlapping matching patterns and consistent output semantic types, the matching conditions of the conflicting rules are merged to obtain a merged rule. In practice, the system employs pattern generalization algorithms, such as extracting the longest common subsequence of the matching patterns of two rules, or constructing a more general regular expression that covers the semantics of both. The merged rule inherits the common semantic type of the original rules and updates its metadata information.

[0081] When the rule conflict type is overlapping matching patterns and mutually exclusive output semantic types, the rule with higher matching priority is retained according to the rule matching priority sorting result, and the rule with lower matching priority is given additional restrictions to obtain the rule with added restrictions.

[0082] When the rule conflict type is a difference in rule source, the system determines whether to retain, revise, or merge the rule based on the difference in rule source information in the rule conflict marker, the credibility of the rule source, and the domain expert rating. Specifically, the system assigns basic credibility weights to different sources (such as expert definitions, automatic learning, and external imports) and calculates the overall credibility by combining expert ratings (0-100 points). When one rule comes from an authoritative expert and has a high rating, while another comes from automatic extraction and has a low rating, the system tends to retain the former and revise the latter. When the credibility of both sources is comparable but their expressions are complementary, the system may choose to merge them to form a more complete rule.

[0083] Based on the merged rules, rules with added restrictions, retained rules, revised rules, or merged rules resulting from differences in rule sources, the policy domain semantic type rule set is updated to obtain an optimized policy domain semantic type rule set. In specific implementation, the system first writes the processed rules to a temporary area of ​​the rule base, performs consistency checks to ensure no new conflicts arise, then atomically replaces the original rules, and records version change logs to support rule backtracking.

[0084] This application also proposes a knowledge graph-based enterprise service policy matching system for performing knowledge graph-based enterprise service policy matching, combined with... Figure 3 As shown, the knowledge graph-based enterprise service policy matching system 1 includes: The document and attachment acquisition module 11 is used to acquire the main policy text of an enterprise, as well as the policy attachments associated with the main policy text. This module is specifically implemented as a data acquisition engine, which periodically crawls policy documents published on government websites via web crawlers or API interfaces, parses the HTML or PDF document structure, and extracts the main text content and attachment download links. The attachment acquisition submodule automatically downloads the relevant documents based on the links and stores them in a distributed file system.

[0085] The text block management module 12 is used to identify the citation relationships between the policy text and policy attachments, and to integrate the contextual constraint information of the policy text with the specific content of the policy attachments based on the citation relationships to obtain a unified parsed text block. This module includes a citation relationship identification engine, which identifies attachment citations in the text based on regular expressions or natural language processing models; and a text integration engine, which embeds the attachment content into the logical structure of the text according to the citation context, constructing a unified parsed text block data object containing a complete semantic context.

[0086] The Conditional Logic Tree module 13 is used to perform semantic analysis on the unified parsed text block, identify the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and construct a conditional logic tree representing the hierarchical relationship and dependency relationship of the fine-grained conditions. This module integrates a lexical analyzer, a syntactic analyzer, and a semantic annotation engine. As mentioned above, through new word discovery, domain semantic annotation, dependency parsing, and other technologies, it parses the text into a structured conditional logic tree, which is stored in a graph database or tree data structure.

[0087] The relation representation transformation module 14 is used to convert the conditional logic tree into relational representations in the policy knowledge graph. These representations include the associations between policy entities and fine-grained conditions, the logical operation relationships between fine-grained conditions, and the time constraints corresponding to the fine-grained conditions. This module implements graph data transformation logic, mapping the nodes of the conditional logic tree to entity nodes (policy entities, condition entities, time entities, etc.) in the knowledge graph, and mapping edges to relational edges (application relationships, logical operation relationships, and time constraint relationships). The data is stored using Cypher statements or RDF triples in a graph database (such as Neo4j).

[0088] The Enterprise Profile Fact Node Module 15 is used to acquire enterprise profile data of the enterprises to be matched and convert the enterprise profile data into enterprise profile fact nodes. This module connects to an enterprise information database or business registration data interface, extracts structured data such as enterprise registration information, qualification certificates, and financial data, and maps this data into fact nodes in a knowledge graph, establishing matching links with policy condition nodes.

[0089] The service policy matching module 16 is used to perform conditional matching on the fact nodes of the enterprise profile based on the relational expressions in the policy knowledge graph, and obtain the enterprise service policy matching result. This module implements a graph traversal algorithm or a logical reasoning engine, traverses the conditional logic tree along the relational edges in the knowledge graph, compares the enterprise profile data with the conditional thresholds, calculates the logical operation result, and outputs the judgment of successful or unsuccessful matching.

[0090] The results output module 17 is used to output matched policy information, information on unmet conditions, or supplementary material prompts based on the enterprise service policy matching results. This module includes a results formatting engine that generates a structured report based on the matching results, including a list of successfully matched policies, detailed explanations of unmet conditions, and supplementary material prompts, which is presented to the end user through an API interface or user interface.

[0091] 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 knowledge graph-based method for matching enterprise service policies, characterized in that, include: Obtain the full text of the company's policy, as well as the policy attachments associated with the full text of the policy; Identify the reference relationship between the policy text and the policy appendix, and integrate the contextual constraint information of the policy text and the specific content of the policy appendix according to the reference relationship to obtain a unified parsed text block; Semantic analysis is performed on the unified parsed text block to identify the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and a condition logic tree representing the hierarchical relationship and dependency relationship of the fine-grained conditions is constructed. The conditional logic tree is converted into a relational representation in a policy knowledge graph. This relational representation includes the association between policy entities and fine-grained conditions, the logical operation relationships between fine-grained conditions, and the time constraint relationships corresponding to the fine-grained conditions. The policy knowledge graph is structured graph data used to store the associations between policy entities, fine-grained conditions, logical operation relationships, time constraint relationships, and enterprise profile fields. The nodes in the policy knowledge graph include policy entity nodes, condition nodes, time constraint nodes, and enterprise profile field nodes. Edges are used to represent the applicability relationships between policy entities and fine-grained conditions, the logical operation relationships between fine-grained conditions, and the time constraint relationships corresponding to the fine-grained conditions. Obtain enterprise profile data of the enterprises to be matched, and convert the enterprise profile data into enterprise profile fact nodes; wherein, the enterprise profile fact nodes are structured data nodes used to represent the registered region, industry, operating status, qualification certificates, personnel size, operating income, R&D investment or submitted materials information of the enterprises to be matched. Based on the relational representation in the policy knowledge graph, conditional matching is performed on the fact nodes of the enterprise profile to obtain the enterprise service policy matching result; Based on the enterprise service policy matching results, output matching policy information, information on unmet conditions, or information prompting for supplementary materials.

2. The knowledge graph-based enterprise service policy matching method according to claim 1, characterized in that, The steps of performing semantic analysis on the unified parsed text block, identifying the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and constructing a condition logic tree representing the hierarchical relationship and dependency relationship of the fine-grained conditions include: Lexical analysis is performed on the unified parsed text block to obtain a word sequence and general part-of-speech tags in the word sequence; The words in the word sequence that have been marked with the general part-of-speech tag are labeled with policy domain semantic types to obtain a word sequence with policy domain semantic types; the policy domain semantic types include one or more of the following: subject type, qualification condition type, time condition type, regional condition type, numerical condition type, material condition type, or exclusion condition type. Based on the word sequence with policy domain semantic type, the conditional statements in the unified parsing text block are semantically analyzed to identify conditional phrases representing the applicant, business status, qualification requirements, time range, geographical range, numerical threshold, supporting materials and exclusion scenarios, and to obtain fine-grained condition recognition results. The fine-grained conditions in the fine-grained condition identification results are normalized to obtain normalized fine-grained conditions; the normalization process includes unifying the condition name, condition object, comparison method, numerical unit, and time range; Syntactic analysis is performed on word sequences with policy domain semantic types to construct syntactic dependency trees; Traverse the syntactic dependency tree, identify logical connectives representing coordination, selection, negation, progression, premise, exception, or supplementary explanation, and determine the logical connections between the normalized fine-grained conditions by combining the hierarchical position and scope of the logical connectives in the syntactic dependency tree, thus obtaining the conditional logical connection results. Based on the conditional logic association results, the hierarchical relationship and dependency relationship between the normalized fine-grained conditions are determined; wherein, the hierarchical relationship is used to represent the superior-subordinate relationship between the main condition, sub-condition and supplementary condition, and the dependency relationship is used to represent whether one fine-grained condition is conditional on the fulfillment of another fine-grained condition; Based on the normalized fine-grained conditions, the conditional logical association results, the hierarchical relationship, and the dependency relationship, a conditional logic tree is constructed; wherein, the normalized fine-grained conditions are used as condition nodes, the logical operation relationships corresponding to the logical association terms are used as logical operator nodes, and the condition nodes and the logical operator nodes are connected according to the hierarchical relationship and the dependency relationship. Identify time adverbs in the unified parsed text block and determine the fine-grained conditions corresponding to the time adverbs; The time adverbial is attached as a time constraint attribute to the corresponding condition node, or the time adverbial is connected to the condition logic tree as a status verification node; wherein, the status verification node is used to verify the condition of the corresponding fine-grained condition within the time range defined by the time adverbial. The condition fulfillment includes condition satisfaction, condition validity, or continuous fulfillment. The condition satisfaction result indicates that the corresponding field in the enterprise profile fact node matches the condition object, comparison method, and condition value in the fine-grained condition. The condition validity result indicates that the qualifications, materials, operating status, or application qualifications involved in the fine-grained condition are in an unexpired state within the time range. The continuous fulfillment result indicates that when the fine-grained condition requirement is met continuously for a continuous period, it is continuously met within the time range according to a preset verification cycle.

3. The enterprise service policy matching method based on knowledge graph according to claim 2, characterized in that, The steps of traversing the syntactic dependency tree, identifying logical connectives representing coordination, selection, negation, progression, premise, exception, or supplementary explanation, and determining the logical connections between the normalized fine-grained conditions by combining the hierarchical position and scope of the logical connectives in the syntactic dependency tree to obtain the conditional logical connection results include: The syntactic dependency tree is traversed to identify explicit logical association words that are directly connected to the normalized fine-grained conditions; Based on the connection relationships between the explicit logical association words and their parent nodes, child nodes, and adjacent nodes at the same level in the syntactic dependency tree, the scope of jurisdiction of the explicit logical association words in the syntactic dependency tree is determined. Based on the type of the explicit logical association vocabulary and the scope of jurisdiction, determine the explicit logical associations between the normalized fine-grained conditions connected by the explicit logical association vocabulary; Based on a pre-defined set of policy domain logic conventions, semantic similarity is evaluated for adjacent normalized fine-grained conditions that are not connected by explicit logical association terms. The set of policy domain logic conventions is used to store implicit logical relationship judgment rules commonly found in policy texts. These implicit logical relationship judgment rules include one or more of the following: adjacent qualification conditions are assumed to be parallel; the preceding subject condition has a prerequisite constraint relationship on the subsequent material condition; the exception statement has an exclusion relationship with the aforementioned condition; and the supplementary explanatory statement has a supplementary limiting relationship with adjacent conditions. Based on the semantic similarity between adjacent normalized fine-grained conditions, their relative positions in the syntactic dependency tree, and their syntactic roles, determine whether there is an implicit logical relationship between adjacent normalized fine-grained conditions. When an implicit logical association is determined to exist, a virtual logical operator node is inserted into the syntactic dependency tree, and the hierarchical position and scope of the virtual logical operator node are determined according to the type of the implicit logical association. Based on the hierarchical position and jurisdiction of the virtual logic operator nodes, the explicit logical associations and the implicit logical associations are combined to generate the conditional logical association results between the normalized fine-grained conditions.

4. The enterprise service policy matching method based on knowledge graph according to claim 2, characterized in that, The step of performing lexical analysis on the unified parsed text block to obtain a word sequence and general part-of-speech tags in the word sequence includes: Identify words not included in the general dictionary from the unified parsed text block, and use these words as candidate new words. Based on the frequency of occurrence of the candidate new words in the unified parsed text block, the context, and the semantic relevance of the candidate new words to known policy domain terms, the candidate new words are screened and weighted to obtain the screened candidate new words; The selected candidate new words and their weights are added to the domain dictionary of the lexical analyzer to obtain an updated domain dictionary; the word segmentation rules of the lexical analyzer are adjusted according to the context of the candidate new words to obtain adjusted word segmentation rules; wherein, the word segmentation rules are used to determine how continuous character sequences in the unified parsed text block are divided into single words, combined words, or policy domain terms during lexical analysis; Based on the updated domain dictionary and the adjusted word segmentation rules, lexical analysis is performed on the unified parsed text block to obtain word sequences and general part-of-speech tags.

5. The enterprise service policy matching method based on knowledge graph according to claim 2, characterized in that, The step of performing policy domain semantic type annotation on the words in the word sequence that have already been annotated with the general part-of-speech tag to obtain a word sequence with policy domain semantic type includes: Based on a preset policy domain semantic type rule set, the policy domain semantic type is identified for words in the word sequence that have been marked with general parts of speech. The policy domain semantic type rule set includes matching rules for identifying the policy domain semantic type of words or phrases in policy texts. The matching rules include at least one or more of the following: keyword matching rules, syntactic structure matching rules, part of speech combination matching rules, numerical expression matching rules, time expression matching rules, and material name matching rules. When a word or phrase matches a matching rule in the policy domain semantic type rule set, the word or phrase is labeled with the corresponding policy domain semantic type according to the matched rule; The output is a word sequence that simultaneously carries the general part-of-speech tagging and the policy domain semantic type, as a word sequence with policy domain semantic type.

6. The enterprise service policy matching method based on knowledge graph according to claim 5, characterized in that, The step of identifying the policy domain semantic type of words in the word sequence that have been marked with general part-of-speech tags according to a preset policy domain semantic type rule set includes: The matching rules in the preset policy domain semantic type rule set are sorted by matching priority to obtain the rule matching priority sorting result; Conflict detection is performed on the matching rules in the preset policy domain semantic type rule set to obtain the rule conflict detection results; When the rule conflict detection result indicates that a rule conflict exists, the relevant conflict rules are adjusted or merged according to the preset conflict resolution strategy to obtain an optimized policy domain semantic type rule set. When the rule conflict detection result indicates that there is no rule conflict, the matching rules in the policy domain semantic type rule set are prioritized according to the rule matching priority ranking result, and the policy domain semantic type rule set with the updated priority is used as the optimized policy domain semantic type rule set. Based on the optimized policy domain semantic type rule set, policy domain semantic type identification is performed on the words in the word sequence that have been marked with general part-of-speech tags.

7. The enterprise service policy matching method based on knowledge graph according to claim 6, characterized in that, The step of sorting the matching rules in the preset policy domain semantic type rule set by matching priority to obtain the rule matching priority ranking result includes: Obtain metadata information for each rule in a pre-defined policy domain semantic type rule set; the metadata information includes rule creation time, rule modification history, and domain expert ratings; Based on the matching pattern of each rule, the rule specificity and rule coverage of each rule are determined. The rule specificity is used to characterize the degree to which the rule matching conditions limit the semantic type of a specific policy domain. The more specific the rule matching conditions and the clearer the keywords, syntactic structures, or parts of speech combinations involved, the higher the rule specificity. The rule coverage is used to characterize the range of policy text expressions that the rule can match. The more word, phrase, or sentence type types the rule applies to, the larger the rule coverage. Based on the metadata information, a rule activity score is calculated; the rule activity score is used to reflect the frequency with which a rule is matched in recent policy texts. Based on the metadata information, a rule stability score is calculated; the rule stability score is used to reflect the consistency of the matching results corresponding to the rule after multiple modifications. The rule's overall priority score is calculated based on the rule activity score, the rule stability score, the rule specificity, the rule coverage, and the domain expert rating. Based on the comprehensive priority score of the rules, the matching rules in the policy domain semantic type rule set are prioritized to obtain the rule matching priority ranking result.

8. The enterprise service policy matching method based on knowledge graph according to claim 6, characterized in that, The step of performing conflict detection on the matching rules in the preset policy domain semantic type rule set to obtain the rule conflict detection result includes: Obtain the matching pattern, output semantic type, and rule source information for each rule in the preset policy domain semantic type rule set; For each rule's matching pattern, the keywords, syntactic structure, and part-of-speech combinations contained in the matching pattern are analyzed to construct a rule feature description representing the matching pattern; Any two rules in the policy domain semantic type rule set are combined into a rule pair, and the similarity of the corresponding rule feature descriptions between each rule pair is compared. When the similarity exceeds a preset similarity threshold, it is determined that there is an overlap in the matching patterns between the corresponding rule pairs. Analyze whether there is a mutual exclusion relationship between the output semantic types of the rule pairs; For rule pairs with overlapping or mutually exclusive matching patterns, the overlap or contradiction between the rule pairs is quantitatively marked based on the rule source information to obtain rule conflict marking information; the rule conflict marking information includes conflicting rule pairs, rule conflict type, conflict degree, potential impact range, and rule source difference information. Based on the rule conflict marker information, a rule conflict detection result is generated.

9. The enterprise service policy matching method based on knowledge graph according to claim 8, characterized in that, When the rule conflict detection result indicates the existence of a rule conflict, the step of adjusting or merging the relevant conflicting rules according to a preset conflict resolution strategy to obtain an optimized policy domain semantic type rule set includes: Read the rule conflict detection result and the rule conflict marker information corresponding to the rule conflict detection result; Based on the rule conflict marker information, the corresponding rule conflict type is read; the rule conflict type includes overlapping matching patterns, mutually exclusive output semantic types, or differences in rule sources. When the rule conflict type is that the matching patterns overlap and the output semantic types are consistent, the matching conditions of the conflicting rules are merged to obtain the merged rules; When the rule conflict type is that the matching patterns overlap and the output semantic types are mutually exclusive, the rules with higher matching priority are retained according to the rule matching priority sorting result, and the rules with lower matching priority are subject to additional restrictions, resulting in rules with added restrictions. When the rule conflict type is rule source difference, the rule source difference information, the credibility of the rule source, and the domain expert score in the rule conflict marking information are used to determine whether to retain the rule, revise the rule, or merge the rule, and thus obtain the retained rule, the revised rule, or the merged rule formed due to the rule source difference. The policy domain semantic type rule set is updated based on the merged rules, the rules with added restrictions, the retained rules, the revised rules, or the merged rules formed due to differences in the source of the rules, to obtain the optimized policy domain semantic type rule set.

10. A knowledge graph-based enterprise service policy matching system, used to perform knowledge graph-based enterprise service policy matching, characterized in that, include: The document attachment acquisition module is used to acquire the enterprise's policy document and the policy attachments associated with the policy document; The text block management module is used to identify the reference relationship between the policy text and the policy appendix, and to integrate the contextual constraint information of the policy text and the specific content of the policy appendix according to the reference relationship to obtain a unified parsed text block; The conditional logic tree module is used to perform semantic analysis on the unified parsed text block, identify the fine-grained conditions contained in the unified parsed text block and the logical relationships between the fine-grained conditions, and construct a conditional logic tree representing the hierarchical relationship and dependency relationship of the fine-grained conditions. The relation expression conversion module is used to convert the conditional logic tree into relation expressions in the policy knowledge graph. The relation expressions include the association between policy entities and fine-grained conditions, the logical operation relationships between fine-grained conditions, and the time constraint relationships corresponding to the fine-grained conditions. The Enterprise Profile Fact Node Module is used to acquire enterprise profile data of the enterprise to be matched and convert the enterprise profile data into enterprise profile fact nodes; The service policy matching module is used to perform conditional matching on the fact nodes of the enterprise profile based on the relational expression in the policy knowledge graph, so as to obtain the enterprise service policy matching result; The result information output module is used to output matching policy information, unmet conditions information, or supplementary material prompts based on the enterprise service policy matching results.

Images