Precise policy service pushing method and system for small and medium-sized enterprises

Abstract

The invention belongs to the technical field of data processing, and discloses a policy service accurate pushing method and system for medium and small enterprises. Comprising the following steps of: constructing a matching rule base covering four types of conditions of regions, scales, industries and products by performing clause-level atomization disassembly and multi-dimensional knowledge expansion on enterprise industry and commerce information on policy texts, and realizing cross-dimensional consistency matching of policy conditions and enterprise characteristics; according to the method, construction of a logical dependency relationship graph is introduced, atomic judgment conditions are used as graph nodes, a core admission chain is identified through dependency strength weights and topological sorting, matching confidence is quantified based on path weights and a logical propagation algorithm, and enterprise policy applicability and information gaps are accurately identified; the method has intelligence and accuracy, deep semantic analysis and applicability evaluation can be carried out on policy terms, and personalized difference pushing is effectively achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for precisely delivering policy services to small and medium-sized enterprises. Background Technology

[0002] With the deepening of digital government construction and the intelligentization of public services, policy services have entered a stage of precise and personalized supply. Currently, government departments at all levels release thousands of support policies annually, covering areas such as tax incentives, financing support, innovation subsidies, and talent introduction. Existing policy service platforms cover tens of millions of market entities, and policy information push services reach hundreds of millions of times. These measures have demonstrated significant value in reducing the cost for enterprises to access policies and improving the efficiency of government services, but they also face the serious challenge of accurately matching policies with enterprises.

[0003] Existing technologies generally lack a deep semantic understanding of policy provisions and a comprehensive, accurate profiling mechanism for enterprise characteristics, resulting in low accuracy in policy applicability assessments and unsatisfactory push notification effects. Specifically, traditional push methods mainly rely on superficial rules such as keyword matching and industry tag classification, focusing only on the surface attribute information of enterprise registration, while ignoring the complex logical constraints within policy provisions and the potential suitability characteristics of enterprises. Furthermore, existing systems struggle to accurately analyze the logical dependencies between conditions, failing to identify which are core admission conditions and which are auxiliary requirements, let alone assess the comprehensive suitability of enterprises across multiple dimensions. Existing technologies are particularly vulnerable in cross-dimensional comprehensive assessment scenarios—when policy provisions involve composite judgments based on multiple types of conditions such as region, scale, industry, and product, the lack of a unified confidence measurement framework and differentiated push strategies results in generic and untargeted push content. This lack of accurate matching capability not only reduces enterprises' sense of policy access and application success rate but also leads to inefficient allocation of policy resources. Many preferential policies for enterprises fail to fully play their role due to poor information transmission, constituting a key bottleneck in the intelligent upgrading of public services.

[0004] In view of this, the present invention proposes a method and system for precise delivery of policy services to small and medium-sized enterprises to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: A method for precisely delivering policy services to SMEs, characterized by including: Obtain basic business registration information and policy documents of the target company; The policy clauses within the policy text data are decomposed into clause-level atomic decompositions to obtain several atomic judgment conditions. The judgment basis type and judgment threshold in the atomic judgment conditions are extracted, and a logical dependency graph between the atomic judgment conditions is constructed simultaneously. The judgment basis type includes regional conditions, scale conditions, industry conditions, and product conditions. The basic business information is expanded into multiple dimensions to obtain enterprise extended information; and a conditional information matching rule base is constructed based on the enterprise extended information and the judgment criteria type. The matching confidence of each policy clause with the corresponding enterprise is obtained according to the conditional information matching rule base. Based on the matching confidence level and logical dependency graph, suitable policy clauses are selected, and personalized policy push content is generated for the corresponding enterprises. The personalized policy push content is then pushed in a differentiated manner based on the matching confidence level.

[0006] Furthermore, the process of performing clause-level atomization decomposition includes: Syntactic dependency analysis is performed on the policy clauses in the policy text data to identify the subject-verb-object semantic units in each policy clause. By extracting the modification and dominance relationships between the subject-verb-object semantic units, a clause semantic dependency tree is obtained. The policy clauses are separated into conditional and conclusive clauses to obtain the corresponding conditional clauses. The limiting and constraining components in the conditional clauses are then extracted based on the semantic dependency tree of the clauses. Traverse the logical connector nodes in the semantic dependency tree of the clause, determine the semantic boundary of the atomic decision condition based on the position of the logical connector node, and divide each policy clause into several semantically independent atomic decision conditions based on it. For each atomic decision condition, the components are extracted to obtain the decision object, decision dimension, decision operator and decision value in the atomic decision condition; The judgment criteria type for the corresponding atomic judgment conditions is determined based on the judgment dimension, judgment object, and constraint components, and constraint strength is assigned to each atomic judgment condition. The decision operator and the decision value are combined to form the decision threshold. A structured representation is constructed for each atomic decision condition, including the decision basis type, the decision threshold, and the position index of the atomic decision condition in the original policy clause.

[0007] Furthermore, the process of constructing a logical dependency graph includes: Atomic decision conditions are mapped to condition nodes, and each condition node is labeled with the decision basis type, decision threshold, and constraint strength; Iterate through all atomic decision conditions within the same policy clause, and identify logical relationships based on the logical connectors between the atomic decision conditions. The logical relationships include logical AND, logical OR, and logical NOT. A logical relationship dependency graph is constructed using the identified logical relationships as edges and the mapped condition nodes as graph nodes; Calculate the dependency strength of each logical relation edge in the logical relation dependency graph, perform topological sorting on the logical relation dependency graph based on the dependency strength of the logical relation edges, identify the key condition paths in the logical relation dependency graph, use them as core admission chains, and assign path weights to each condition node in the core admission chains.

[0008] Furthermore, the process of expanding knowledge in multiple dimensions includes: The industry categories in the basic business information are semantically expanded to obtain the complete classification path corresponding to the industry category, and the industry knowledge features of each level in the complete classification path are extracted. The industry knowledge features include the national economic industry classification code, the position in the industrial chain, and the direction of industrial policy corresponding to the industry category. Based on the registered capital in the basic business information, the scale level is inferred to obtain the scale level of the target enterprise, and the scale knowledge characteristics of the corresponding enterprise are obtained. The scale knowledge characteristics include asset size range, financing ability rating, risk resistance ability label and market competitiveness label. Based on the enterprise address in the basic business information, regional economic attributes are analyzed to obtain regional knowledge features. These regional knowledge features include the administrative region to which the target enterprise belongs and the regional economic development level, regional industrial structure characteristics, and regional policy incentives within the corresponding region. Based on the main products in the basic business information, a technical field mapping is performed to obtain product knowledge features. The product knowledge features include technical field tags, innovation attribute tags, and market application scenario tags corresponding to the main products. The industry knowledge characteristics, scale knowledge characteristics, regional knowledge characteristics, and product knowledge characteristics are verified for knowledge consistency. Conflict information between different knowledge sources is detected, and the detected conflict information is resolved to obtain enterprise extended information. Analyze the confidence level of each knowledge feature in the enterprise's extended information, and mark knowledge features with confidence levels below a preset confidence threshold as features to be verified.

[0009] Furthermore, the process of constructing the conditional information matching rule base includes: Iterate through the judgment criteria types corresponding to the atomic judgment conditions. If the judgment criteria type is a regional condition, obtain the regional range identifier within the corresponding judgment threshold and perform text matching with the regional knowledge features in the enterprise extended information to obtain the regional matching rules. If the judgment criteria are scale-based conditions, extract the numerical range within the corresponding judgment threshold and compare it with the scale knowledge features in the enterprise's extended information to obtain the scale matching rules. If the judgment criteria are industry-specific conditions, extract the industry description text within the corresponding judgment threshold, and perform semantic similarity calculation between it and the industry knowledge features in the enterprise's extended information to obtain the industry matching rules. If the judgment criteria are product-type conditions, then extract the product feature requirements from the corresponding judgment threshold, perform multi-label matching with the product knowledge features in the enterprise's extended information, calculate the label overlap, and generate product matching rules. The region matching rules, scale matching rules, industry matching rules and product matching rules are subjected to conflict detection to identify conflicting rule pairs with logical contradictions. The conflicting rule pairs are prioritized in combination with the core access chain, and high-priority conflicting rules are selected and organized with non-conflicting rules into a condition information matching rule library. The rule type of each matching rule in the condition information matching rule base is verified, and the rule type includes deterministic matching rules and fuzzy matching rules; and rule weights are assigned to each matching rule according to the dependency strength of each logical relation edge in the logical dependency graph.

[0010] Furthermore, the process of obtaining the matching confidence level includes: Obtain the completeness of the collected basic business information and assign initial weights based on it; Based on the matching rules in the conditional information matching rule base, the atomic decision conditions of each policy clause are verified to obtain the initial verification score of each atomic decision condition. The initial verification score is corrected by logical propagation based on the logical relationship edges in the logical relationship dependency graph to obtain the corrected verification score; The modified verification score is multiplied by the rule weight of the corresponding matching rule to obtain the weighted verification score; and the weighted verification scores of all atomic judgment conditions within the same policy clause are summed to obtain the original matching score. Obtain the number of knowledge features marked as features to be verified within the enterprise's extended information, calculate the information discount coefficient based on it, and calculate the discounted matching score by combining the information discount coefficient with the original matching score. The satisfaction of the core admission chain is calculated. The satisfaction of the core admission chain, the discount matching score and the initial weight are multiplied together and the result is normalized to obtain the matching confidence.

[0011] Furthermore, the process of providing differentiated push notifications includes: Based on the matching confidence level, relevant policy clauses are filtered, and policy clauses with a matching confidence level greater than a preset confidence threshold are marked as candidate policies to be pushed. Based on the matching confidence and core access chain satisfaction, the candidate push policy is divided into high-confidence policy, medium-confidence policy and low-confidence policy; For high-confidence policies, extract the atomic judgment conditions corresponding to the condition nodes that have passed the clause verification in the logical dependency relationship graph, convert the atomic judgment conditions into a description of the access elements that the enterprise has met, and extract the policy title, core policy support content, list of conditions that the enterprise has met, brief description of the application process, and application deadline from the application requirements of the policy text data to obtain the application guidance document; For policies with medium confidence, supplementary information guidance text is generated based on the dependency graph. The supplementary information guidance text includes the policy title, the core support content of the policy, the conditions that enterprises may meet, the list of information items that need to be supplemented, and the method of supplementing information. For policies with low confidence, generate policy understanding expansion text, which includes the policy title, the core support content of the policy, an explanation of the main conditions that the enterprise does not currently meet, and suggestions for the enterprise to improve. The push notifications for all candidate policies are personalized and integrated, sorted by priority according to high-confidence, medium-confidence, and low-confidence policies, and a comprehensive push notification is generated. The format of the comprehensive push notification is optimized and then pushed to the target enterprise.

[0012] Furthermore, the process of extracting the limiting and constraining components in the conditional clause based on the aforementioned clause semantic dependency tree includes: Each semantic unit in the semantic dependency tree of the clause is tagged with part of speech to obtain candidate limiting components, which include adjective modifiers, quantifiers and time words. Traverse and extract the central words used for modification from the candidate limiting components, and determine whether the corresponding central words are nouns related to enterprise attributes; By using nouns related to enterprise attributes, numerical and range expressions are extracted from candidate bounded components. Candidate bounded components containing numerical or range expressions are marked as bounded components. The numerical content and range boundaries of the bounded components are then extracted and denoted as the quantitative attributes of the bounded components. Pattern matching is performed on the sentence structure in the semantic dependency tree of the clauses to identify and mark the condition markers in the semantic units, and the semantic units containing the condition markers are marked as constraint components; and constraint strength is assigned to the atomic decision conditions modified by the constraint components; wherein, the condition markers include necessary condition markers and sufficient condition markers.

[0013] Furthermore, the process of performing knowledge consistency verification includes: Using the enterprises in the basic business information as the central entity nodes, and the industry knowledge features, scale knowledge features, regional knowledge features, and product knowledge features as attribute nodes, and establishing the association edges between the central entity nodes and attribute nodes, we obtain an enterprise feature knowledge graph. Semantic disambiguation is performed on the attribute nodes in the enterprise feature knowledge graph to identify conflicting attribute values ​​from different knowledge sources under the same attribute dimension; For each conflicting attribute value, trace its corresponding knowledge source, obtain the publishing institution, update time and data coverage of the corresponding knowledge source, and calculate the authority score of the corresponding knowledge source based on it; For conflicting attribute dimensions, compare the authority scores of the knowledge sources corresponding to each conflicting attribute value, select the attribute value corresponding to the knowledge source with the highest authority score as the main attribute value, and retain the other conflicting attribute values ​​as candidate attribute values. Calculate the consistency check score between the main attribute value and the basic business information; when the consistency check score is less than the preset consistency threshold, perform weighted fusion on the main attribute value and the candidate attribute value to obtain the fused attribute value as the enterprise feature information after resolving the conflict.

[0014] A targeted policy service delivery system for SMEs, including: The data acquisition module is used to obtain basic business registration information and policy documents of the target company. The data parsing module performs clause-level atomic decomposition on the policy clauses within the policy text data to obtain several atomic judgment conditions. It extracts the judgment basis type and judgment threshold from the atomic judgment conditions and simultaneously constructs a logical dependency graph between the atomic judgment conditions. The judgment basis type includes regional conditions, scale conditions, industry conditions, and product conditions. The knowledge expansion module expands the basic business information into multi-dimensional knowledge to obtain enterprise extended information; and constructs a conditional information matching rule base based on the enterprise extended information and the judgment basis type, and obtains the matching confidence of each policy clause with the corresponding enterprise according to the conditional information matching rule base. The data push module filters out suitable policy clauses based on the matching confidence level and logical dependency relationship graph, and generates personalized policy push content for the corresponding enterprise based on them. The personalized policy push content is then pushed in a differentiated manner based on the matching confidence level.

[0015] The technical effects and advantages of this invention's method and system for precisely delivering policy services to SMEs are as follows: This invention, by establishing a mechanism for atomizing policy clauses and constructing logical dependency graphs, can effectively analyze the deep semantic structure and complex logical constraints of policy conditions. This semantic understanding capability enhances the accuracy of policy applicability judgments, accurately identifies core access conditions and auxiliary requirements, and clarifies the logical dependencies between conditions, thereby providing refined condition decomposition support for policy matching. This invention helps to build a comprehensive enterprise profile system, compensating for the incompleteness of basic business information and improving the information symmetry and matching success rate between enterprises and policies. By reconstructing the intelligent matching link between policies and enterprises, this invention significantly reduces the policy search costs for enterprises and the promotion costs for government departments, improves the efficiency of policy resource allocation, and provides a technical foundation for building an efficient, transparent, and inclusive policy service system. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method for precisely delivering policy services to small and medium-sized enterprises according to the present invention; Figure 2 This is a schematic diagram of the policy service precision delivery system for small and medium-sized enterprises according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0018] Please see Figure 1 As shown in this embodiment, the method for precisely delivering policy services to SMEs includes: The system acquires basic business registration information and policy documents of target enterprises through a pre-defined government data collection network. This network employs a multi-source heterogeneous data synchronous acquisition mechanism, continuously collecting key data sources including enterprise registration information, tax registration information, and the latest policy documents during the policy release cycle. Enterprise registration information includes basic business attributes such as enterprise name, registered address, registered capital, business scope, and legal representative. Policy documents cover support policies, preferential measures, and application guidelines issued by governments at all levels, collectively forming a data set with both an enterprise foundation layer and a policy environment layer. It should be noted that all data acquisition in this invention is completed through legal and compliant means, adhering to national data security, government information disclosure regulations, and enterprise information disclosure regulations. Data transmission and storage utilize encryption technology and comply with relevant national cryptographic management regulations. Strict hierarchical protection is implemented for collected sensitive enterprise information to ensure that enterprise trade secrets are not leaked during policy matching. Furthermore, the generation of policy interpretation content strictly follows the original semantics and does not distort the original intent of the policy.

[0019] The policy clauses within the policy text data are decomposed into clause-level atomic decompositions to obtain several atomic judgment conditions. The judgment basis types and judgment thresholds within these atomic judgment conditions are extracted, and a logical dependency graph between the atomic judgment conditions is constructed simultaneously. Clause-level atomic decomposition is the core technology of intelligent policy text understanding, breaking down complex policy clauses into semantically independent and logically clear minimum judgment units. Through natural language processing technology, the limiting conditions, constraints, and logical relationships within the policy clauses are identified, forming a structured set of judgment conditions. The judgment basis types include four main categories: regional conditions, scale conditions, industry conditions, and product conditions, each corresponding to different judgment thresholds and verification methods.

[0020] Multi-dimensional knowledge expansion of basic business registration information yields extended enterprise information. Based on this extended information and the types of judgment criteria, a conditional information matching rule base is constructed. The matching confidence level between various policy clauses and corresponding enterprises is obtained from this rule base. Multi-dimensional knowledge expansion is a key step in building a comprehensive enterprise profile. Through knowledge graph technology and external data sources, simple business registration information is expanded into a rich set of information covering industry characteristics, scale characteristics, geographical characteristics, and product characteristics. The conditional information matching rule base establishes a mapping relationship between atomic judgment conditions and enterprise characteristics, providing technical support for accurate matching.

[0021] The system filters out suitable policy clauses based on matching confidence and logical dependency graphs, and generates personalized policy push content for corresponding enterprises. The personalized policy push content is pushed in a differentiated manner based on the matching confidence. Differentiated push is the core mechanism to improve user experience. Based on the matching confidence, different types of policy information services are provided to enterprises, including application guidance, information supplement guidance, and policy understanding expansion.

[0022] In embodiments of the present invention, the process of performing clause-level atomization decomposition includes: Syntactic dependency analysis is performed on policy clauses in policy text data. A dependency parser is used to construct a syntactic dependency tree for each policy clause. Syntactic dependency analysis is a fundamental step in understanding the semantic structure of policy clauses and can accurately identify the grammatical relations and component organization within sentences. The analysis process begins with word segmentation and part-of-speech tagging of the policy clauses. Using Chinese word segmentation tools, the policy clauses are divided into a sequence of words, and each word is tagged with its part of speech (noun, verb, adjective, numeral, etc.). A dependency parser is then used to identify the grammatical relationships within the word sequence, identifying the subject-verb-object semantic units within each policy clause. The subject is typically the policy's applicable entity (e.g., micro and small enterprises, high-tech enterprises), the verb is the policy condition or action (e.g., should, may apply), and the object is the policy support content or requirement (e.g., tax reduction, subsidies). By analyzing the modification and dominance relationships between semantic units, a semantic dependency tree is constructed. The dependency tree uses words as nodes and dependency relationships as edges, with edge types including subject-verb, verb-object, attributive, and coordinate relationships. The dependency tree comprehensively describes the syntactic structure of the policy clauses, providing an accurate grammatical framework for subsequent semantic analysis and condition extraction.

[0023] Separating conditions from conclusions in policy clauses involves identifying conditional semantic components within the clauses. This separation is a semantic parsing process that divides the applicable conditions section of a policy clause from the policy outcome section. The separation process identifies conditional semantic markers in policy clauses through conditional identifier matching. Conditional identifiers (including words such as if, when, for, meeting the conditions, satisfying the requirements, etc.) are used to express conditional constraints. Typical expression patterns of policy clauses are captured through sentence pattern recognition, such as: providing support to…enterprises, etc. The semantic scope covered by conditional identifiers and sentence patterns is extracted into conditional clauses, which contain all the constraint information required for policy applicability judgment. Based on the clause semantic dependency tree, limiting and constraining components are extracted from the conditional clauses. Limiting components are semantic units in the conditional clauses with quantitative constraint functions, including numerical expressions (e.g., not exceeding a specified number of people), time expressions (e.g., the past three years), and regional expressions (e.g., western region). Constraining components are semantic units in the conditional clauses with logical constraint functions, including necessary condition identifiers (e.g., must, should, require), sufficient condition identifiers (e.g., can, can), and exclusionary condition identifiers (e.g., must not, prohibited, except). Accurate extraction of limiting and constraining components is the key foundation for constructing precise atomic judgment conditions.

[0024] The process traverses the logical connector nodes in the clause semantic dependency tree, determining the semantic boundaries of atomic decision conditions based on their position and scope. Logical connectors are the semantic links connecting different decision conditions, and their identification and analysis directly affect the segmentation accuracy of atomic decision conditions. The traversal process starts from the root node of the clause semantic dependency tree, using a depth-first search strategy to visit all nodes and identify logical connector nodes. Logical connectors include coordinate relation words (such as "and," "and," and "as well as"), selection relation words (such as "or," "or," and "any"), progression relation words (such as "and," and "simultaneously"), and contrast relation words (such as "but," and "unless"). The syntactic scope of each logical connector word is analyzed to determine the range of semantic components it connects. Scope analysis is based on the transitivity of dependency relations and the completeness of semantic components. Based on the position of the logical connector nodes, the conditional clause is segmented into several semantically independent atomic decision conditions. Each atomic decision condition represents a smallest granularity decision unit and does not contain internal logical connections. The segmentation process ensures that each atomic decision condition is semantically complete and logically independent, providing clear semantic boundaries for subsequent structured representation and matching rule construction.

[0025] Each atomic decision condition undergoes component extraction and structured representation to form a standardized condition description format. Component extraction is the process of classifying, labeling, and structuring the semantic elements in the atomic decision conditions. The extraction process identifies four core components in the atomic decision conditions: the decision object is the target entity of the condition constraint, typically an attribute or characteristic of an enterprise, such as its registered address; the decision dimension is the classification dimension to which the decision object belongs, categorized according to policy management and enterprise characteristics, including geographical, scale, industry, and product dimensions; the decision operator is the comparison operator connecting the decision object and the decision value, including greater than, less than, equal to, belong to, contain, and not exceed; the decision value is the specific threshold or standard of the condition constraint, which can be a numerical value, range, enumeration (e.g., Beijing, Shanghai, Shenzhen), or description (e.g., high-tech fields); based on the decision dimension, decision object, and constraint components, the decision basis type of the corresponding atomic decision condition is determined. This decision basis type includes geographical, scale, industry, and product conditions; geographical conditions primarily constrain the geographical location and administrative affiliation of the enterprise. The criteria are categorized into three types: scale-based conditions primarily constrain a company's operating scale and financial indicators; industry-based conditions primarily constrain a company's industry classification and business scope; and product-based conditions primarily constrain a company's product characteristics and technical attributes. Each atomic decision condition is assigned a constraint strength, quantified based on the semantic strength of the constraint component. Necessary conditions correspond to high constraint strength, sufficient conditions to medium constraint strength, and general descriptions to low constraint strength. Decision operators are combined with decision values ​​to form decision thresholds; for example, the decision threshold for "annual operating revenue greater than 5 million yuan" is greater than 5 million yuan. A structured representation is constructed for each atomic decision condition. This structured representation uses a standardized data format and includes key attributes such as the type of decision basis, decision threshold, constraint strength, and the position index of the atomic decision condition in the original policy clause. This provides a standardized data foundation for subsequent logical dependency construction and matching rule generation.

[0026] In embodiments of the present invention, the process of extracting the limiting and constraining components in a conditional clause based on a clause semantic dependency tree includes: Each semantic unit in the semantic dependency tree of the clauses is tagged with part-of-speech tags to obtain candidate limiting components. These candidate limiting components include adjective modifiers, quantifiers, and time-related components. Part-of-speech tagging is a fundamental step in identifying limiting components. By analyzing the part-of-speech features of semantic units, components that may contain limiting information are initially screened. The tagging process uses a deep learning-based part-of-speech tagger, which can accurately identify complex part-of-speech structures in Chinese policy texts. Adjective modifiers typically contain attribute limiting information, such as "large" in "large enterprise"; quantifiers contain numerical limiting information; and time-related components contain time-related limiting information, such as "two years" in "established for two years". These candidate components provide a candidate set for subsequent precise extraction.

[0027] The process involves traversing and extracting the headwords used for modification from candidate limiting components, and determining whether each headword is a noun related to enterprise attributes. Headword extraction is a crucial step in identifying the limiting object. Through dependency relation analysis, the core noun modified by each modifying component is identified. The extraction process follows the modification edges of the semantic dependency tree, tracing from the modifying component to the modified component to find the endpoint noun of the modification relationship. The determination of enterprise attribute-related nouns is based on a predefined enterprise attribute dictionary, including core concepts such as enterprise, company, operating revenue, registered capital, employees, and business scope. Only limiting components that modify enterprise attribute-related nouns are considered valid limiting information. This filtering mechanism ensures the relevance and accuracy of the extraction results.

[0028] By extracting numerical and range expressions from candidate limiting components using nouns related to enterprise attributes, candidates containing numerical or range expressions are marked as limiting components. The numerical content and range boundaries of these limiting components are then extracted and recorded as their quantitative attributes. Numerical and range expressions are the core content of the limiting components, containing the specific quantitative requirements of policy clauses. The extraction process employs a combination of regular expressions and named entity recognition to accurately identify various numerical formats, including integers, decimals, fractions, and percentages, as well as various range expressions such as "not exceeding," "below," "greater than," and "between." The extracted quantitative attributes include complete information such as numerical value, comparison operators, and units of measurement, providing accurate criteria for subsequent condition matching. The extraction results are stored in a structured format for easy programmatic processing and calculation.

[0029] Pattern matching is performed on the sentence structure in the semantic dependency tree of the clauses to identify and label conditional markers in semantic units, and semantic units containing conditional markers are marked as constraint components. Constraint strengths are assigned to the atomic decision conditions modified by the constraint components. Conditional markers include necessary condition markers and sufficient condition markers. Conditional markers are key clues for identifying constraint components; specific marker words are searched for in semantic units using pattern matching technology. Necessary condition markers include words such as "must," "should," "need," and "require," indicating mandatory requirements. Sufficient condition markers include words such as "may," "allow," "encourage," and "priority," indicating non-mandatory requirements. Pattern matching employs a rule-based method, establishing a complete conditional marker lexicon covering various common expressions in policy texts.

[0030] The allocation of constraint strength is based on the semantic strength of condition markers, employing a three-level classification: strong constraints (weight 1.0) correspond to necessary condition markers, medium constraints (weight 0.7) correspond to advisory condition markers, and weak constraints (weight 0.5) correspond to priority condition markers. Constraint strength directly impacts subsequent matching calculations and confidence assessments, ensuring a reasonable trade-off between different levels of stringency. Through precise extraction of limiting and constraining components, the semantic information of policy clauses is fully transformed into a structured representation, laying a solid foundation for intelligent policy matching.

[0031] In embodiments of the present invention, the process of constructing a logical dependency graph includes: Atomic decision conditions are mapped to condition nodes, and each condition node is labeled with the decision basis type, decision threshold, and constraint strength. Condition node mapping is the process of converting atomic decision conditions into graph data structure nodes, providing a standardized node representation for subsequent graph theory analysis and dependency calculation. The mapping process first assigns a unique node identifier to each atomic decision condition, and the node identifier adopts a hierarchical naming rule, including the policy document number, policy clause number, and the sequence number of the atomic decision condition within the clause. A data structure corresponding to the condition node is created, which includes core attribute fields such as node identifier, decision basis type, decision threshold, constraint strength, and source clause location index. An attribute vector is constructed for each condition node, which contains the decision basis type code, the vectorized representation of the decision threshold, the constraint strength value, and auxiliary identification information, forming a complete digital representation of the condition node. A condition node index table is established, which adopts a hash table data structure to support fast lookup based on node identifier and classification retrieval based on decision basis type, providing an efficient data access foundation for subsequent logical relationship identification and graph construction operations.

[0032] The process iterates through all atomic decision conditions within the same policy clause, identifying logical relationships based on the logical connectors between these conditions. Logical relationship identification is a crucial process for analyzing the semantic connection patterns between atomic decision conditions, accurately capturing the complex logical structure of policy conditions. The identification process first obtains a complete list of atomic decision conditions for the policy clause, arranged in the order of their appearance in the original clause. Based on the clause's semantic dependency tree, the specific location and scope of logical connectors are located. Logical connectors include those representing logical AND ("and", "and", "simultaneously", and "as well as"), logical OR ("or", "or", "any", and "optional"), and logical NOT ("not", "must not", and "except for"). The grammatical scope of each logical connector is analyzed to determine the atomic decision condition it connects to. The grammatical scope is defined by the dominance and modification relationships identified through dependency parsing. The identified logical relationships are then standardized and classified. Logical AND relationships represent… Strong constraints requiring multiple conditions to be met simultaneously; weak constraints where satisfying any one condition is sufficient; and exclusive constraints where a certain condition must not be satisfied. A logical relation descriptor is constructed, containing relation type identifiers, semantic weights of connectives, scope boundaries, and a list of condition node identifiers. Complex logical structures are handled; for complex policy clauses containing multi-level nested logical relations, a recursive decomposition method is used to break down the complex relations into multiple simple binary or multi-ary logical relations, ensuring that the semantic boundaries of each logical relation are clear and unambiguous. A logical relation mapping table is established, recording all logical connections between condition nodes, providing a complete relation data source for subsequent graph edge construction.

[0033] A logical dependency graph is constructed using the identified logical relations as edges and the mapped condition nodes as graph nodes. The logical dependency graph is a graph-theoretic representation of the logical structure of policy conditions, intuitively showing the dependency constraints between conditions. The construction process first initializes a directed graph data structure, employing a hybrid representation of adjacency matrices and adjacency lists. Adjacency matrices support fast relation lookups, while adjacency lists support efficient traversal operations. All condition nodes are added to the graph structure as graph nodes, each carrying complete attribute information for its corresponding condition node. Each logical relation record in the logical relation mapping table is traversed, converting the logical relation into directed edges in the graph. The direction of the edges is determined based on the directionality of the logical dependency. For logical AND relations, edges connecting all relevant condition nodes are created, representing strong dependencies between these nodes; for logical OR relations, edges connecting candidate condition nodes are created, representing selective dependencies between these nodes; for logical NOT relations, special edges representing exclusion constraints are created. Edges identify mutually exclusive condition nodes; each graph edge is labeled with edge attributes, including logical relationship type, semantic weight of connector, and location information of source logical relationships; the graph's topology is processed, detecting and handling possible loop structures, preserving reasonable logical loops (such as combinations of mutually dependent conditions) and marking conflicting logical loops; a multi-level index structure for the graph is constructed, including a classification index based on node type, a relationship index based on edge type, and a hierarchical index based on dependency hierarchy, supporting graph analysis and query needs of different dimensions; finally, a complete logical dependency graph is formed, where nodes represent atomic decision conditions, edges represent logical relationships between conditions, and the entire graph structure accurately reflects the inherent logical organization and constraint hierarchy of policy provisions.

[0034] The computational logic relation depends on the dependency strength of each logical relation edge in the graph, quantifying the tightness of the logical relationships between conditions. Dependency strength calculation is a quantitative process for evaluating the importance and constraint strength of logical relations, providing a numerical basis for subsequent path analysis and weight allocation. The calculation process employs a multi-factor weighted fusion model, based on the mathematical formula for strength calculation: In the formula, Represents condition nodes and The strength of the dependency between edges, The coefficients are the relation type coefficients (where logical AND, logical OR, and logical NOT correspond to 1.0, 0.5, and -1.0, respectively). and Representing condition nodes respectively and The corresponding constraint strength; , , and This represents the pre-defined weighting coefficient; Represents condition nodes and The semantic similarity between the two is calculated by the cosine similarity of the word vectors of the conditional text; the calculated dependency strength is normalized to avoid extreme values ​​from interfering with subsequent analysis; a dependency strength classification system is established, which divides the dependency strength into four levels: strong dependency, medium dependency, weak dependency, and slight dependency, to provide a classification standard for graph analysis and path recognition.

[0035] The logical dependency graph is topologically sorted based on the dependency strength of the logical relationship edges to identify the logical priority sequence of condition nodes. Topological sorting is a classic graph theory algorithm for determining the partial order of nodes in a directed graph, revealing the logical hierarchy and execution order of policy conditions. The sorting process first calculates the weighted in-degree of each condition node in the logical dependency graph. The weighted in-degree considers not only the number of connecting edges but also the dependency strength of the connecting edges. The formula for calculating the weighted in-degree is: In the formula, QJ represents the weighted in-degree; QJ represents the set of all condition nodes in the logical association graph that are directly pointed to by directed edges. The edge type weight is set according to the logical relation type. Typically, the logical AND relation has a weight of 1.0, the logical OR relation has a weight of 0.7, and the logical NOT relation has a weight of 0.9. The priority queue is initialized, and the condition node with a weighted in-degree of 0 is added to the queue as the starting node. The node with the smallest weighted in-degree is taken from the priority queue, added to the topological sort result sequence, and the weighted in-degree of all its successor nodes is updated. The above process is repeated until all condition nodes have been processed, and a complete topological sort sequence is obtained.

[0036] Identify the key condition paths in the logical dependency graph and use them as the core access chain. The key condition path is the path sequence with the highest cumulative dependency strength connecting the starting node to the ending node in the graph, representing the most important logical constraint chain in the policy conditions. The identification process first determines the starting and ending nodes of the logical dependency graph. The starting node is typically the condition node with the smallest weighted in-degree, and the ending node is typically the condition node with the smallest weighted out-degree. Dijkstra's algorithm is used to calculate the path with the maximum dependency strength from the starting node to all other nodes. Unlike traditional shortest path algorithms, this method seeks the path with the largest cumulative dependency strength. All paths with cumulative dependency strength exceeding a preset threshold are collected to form a candidate set of critical paths. The importance of the candidate path set is evaluated, with metrics including path length, dependency strength distribution, diversity of covered condition types, and semantic completeness. The path with the highest cumulative dependency strength and best semantic completeness is selected as the core access chain. The core access chain represents the most critical sequence of conditions in policy applicability judgment; enterprises must meet the main conditions in the core access chain to be eligible for policy applicability. For cases involving multiple parallel critical paths, a path fusion strategy is adopted to merge similar or complementary paths into a composite core access chain, ensuring that the core access chain comprehensively covers the core requirements of the policy.

[0037] Each condition node in the core access chain is assigned a path weight to quantify its importance in policy applicability assessment. Path weight allocation is a process of quantifying the importance of condition nodes based on their position and role in the core access chain, providing a precise weighting basis for subsequent matching confidence calculations. The allocation process first analyzes the topology of the core access chain to determine the positional characteristics of each condition node in the path. These characteristics include the node's sequence number, distance from the starting node, distance from the ending node, and local importance. Based on these positional characteristics, the basic path weight of the node is calculated using the following formula: Basic path weight = (1 / total path length) × (1 + position importance factor × node constraint strength); where the position importance factor is determined based on the criticality of the node in the path, and the front-end nodes and key decision nodes of the path have higher position importance factors; the path weight is adjusted considering the connectivity characteristics of the nodes, and nodes with high connectivity are usually key hubs in the logical structure and should obtain higher path weights. Weight balancing is performed based on the type of node determination criteria to ensure a reasonable balance in path weight allocation for different types of condition nodes, avoiding excessive dominance of one type of condition in the weight distribution; the path weights of all condition nodes are normalized to ensure that the sum of the path weights of all nodes in the core access chain is equal to one, meeting the standard requirements of probability distribution. Establish a path weight allocation record, recording the final path weight of each condition node, the basis for weight calculation, and a description of its role in the core access chain, providing detailed weight parameters for subsequent policy matching and confidence calculation; the resulting core access chain and its path weight system accurately reflect the most important logical constraints and the distribution of condition importance in the policy provisions, laying a solid logical foundation for the accurate assessment of enterprise policy matching.

[0038] In embodiments of the present invention, the process of multi-dimensional knowledge expansion includes: Semantic expansion of industry categories in basic business registration information is performed to construct a complete industry knowledge feature system. This semantic expansion is an information enhancement process that supplements the enterprise's industry profile by leveraging external industry knowledge bases, building upon the enterprise's basic industry information. The expansion process first extracts the enterprise's industry category information from basic business registration information. Industry category information typically exists in the form of economic industry classification codes or industry description text. Based on economic industry classification standards, the enterprise's industry category is standardized and mapped to obtain the complete classification path corresponding to the enterprise's industry category. This complete classification path includes a three-tiered classification system: major category (represented by two digits), intermediate category (represented by three digits), and minor category (represented by four digits). Industry knowledge features at each level of the complete classification path are extracted. These industry knowledge features include standardized economic industry classification codes, the enterprise's position in the industrial chain (upstream raw material supply, midstream manufacturing and processing, etc.), and other relevant information. Downstream sales and services) and industry policy direction characteristics (encouraged development, restricted development, and outdated industries are eliminated); the specific policy orientation corresponding to the industry is queried through the industry policy knowledge base, including key supported strategic emerging industries, traditional advantageous industries, modern service industries, etc., to form policy orientation labels; the industry environment characteristics such as the average technology level, market competition level, and development prospect expectation of the industry are obtained through the industry statistical database to provide macro background information for the enterprise industry profile; the result of industry semantic expansion is that the enterprise industry characteristics are expanded from simple classification labels to multi-dimensional industry knowledge characteristics that include classification level, industrial chain position, policy orientation, and industry environment.

[0039] Based on the registered capital in basic business information, the enterprise size level is inferred, and a comprehensive knowledge feature of enterprise size is constructed. Enterprise size inference is a process of comprehensively evaluating the enterprise size level based on multi-dimensional size indicators. The inference process first extracts the registered capital value from basic business registration information. Registered capital is an important indicator of enterprise size, but not the only standard. Referring to the classification standards set for different industries in existing technologies, a size inference model based on registered capital is used for estimation. The inference model establishes a regression relationship between registered capital and other size indicators based on historical enterprise data. The size level corresponding to the target enterprise is determined. The size level is divided into four basic categories: micro-enterprises, small enterprises, medium-sized enterprises, and large enterprises. Each category corresponds to different policy support priorities and application conditions. The size knowledge characteristics of the corresponding enterprise are obtained. The size knowledge characteristics include asset size range estimation based on statistical analysis, financing capacity rating based on industry data (three levels: strong, medium, and weak), risk resistance label based on enterprise life cycle theory (high risk, medium risk, and low risk), and competitiveness label based on market position analysis (industry leading, industry average, and industry lagging). The financing capacity rating comprehensively considers factors such as enterprise size, industry characteristics, and historical performance. The risk resistance label is based on the matching analysis of enterprise size and industry risk characteristics. The competitiveness label is evaluated based on the relative position of enterprise size within the industry.

[0040] Regional economic attributes are analyzed based on the enterprise address in basic business registration information to obtain detailed regional knowledge characteristics; regional economic attribute analysis is a geographic information processing process that converts enterprise geographical location information into policy-related regional feature labels. The parsing process first geocodes the registered address of the enterprise in the basic business registration information. Using an address resolution service, the text address is converted into a standardized administrative region code to obtain the enterprise's administrative region affiliation. Next, regional economic database matching is used to obtain regional knowledge features, including the specific administrative region to which the enterprise belongs and the corresponding economic development level classification (developed region, moderately developed region, underdeveloped region, impoverished region). Economic development level is classified based on comprehensive indicators such as GDP, fiscal revenue, and industrial structure. Regional industrial structure features are then obtained, classified based on the proportion of the primary, secondary, and tertiary industries, categorized into manufacturing-led, service-led, agriculture-led, and balanced development types. Regional policy incentive type features are also obtained, including policy preference identifiers for special economic zones, high-tech industrial development zones, western regions, old industrial bases in Northeast China, and impoverished regions. Finally, policy geographic data is used to query the various policy incentives currently enjoyed by the region, forming a regional policy environment label. The acquisition of regional knowledge features transforms the enterprise's geographical location information into an important basis for judging policy applicability.

[0041] Based on the main products or business scope in the basic business information, a technology field mapping is performed to construct product knowledge characteristics; the technology field mapping is a product analysis process that converts enterprise product information into technical attributes and innovation characteristic tags. The mapping process first extracts the main product information of enterprises from the business scope field of basic business registration information. The business scope usually exists in the form of standardized business description text. The core product or service content is identified through keyword extraction and semantic analysis. Next, a product technology classification knowledge base is used for technology field mapping. This knowledge base is built based on technology field classification standards and industry technology development trends, including tags for technologies such as artificial intelligence, new materials, new energy, and modern agriculture. Product knowledge features are then acquired, including technology field tags corresponding to the main products (which can be a combination of multiple technology fields), innovation attribute tags based on technology maturity analysis (traditional mature technologies, improved and optimized technologies, emerging development technologies, and cutting-edge exploration technologies), and application scenario tags based on target market analysis (B2B enterprise market, B2C consumer market, B2G government market, and cross-border integration market). The determination of innovation attribute tags is based on technology development cycle theory, and the determination of application scenario tags is based on the matching analysis of product characteristics and market models. Macro-environmental information such as policy support priorities, development prospect assessments, and competitive landscape analysis for this technology field is queried through an industry innovation database to provide policy environment background for product characteristics. The construction of product knowledge features provides technical dimension support for judging product-type and industry-type conditions in policy matching.

[0042] The extracted industry knowledge features, scale knowledge features, regional knowledge features, and product knowledge features undergo knowledge consistency verification to construct consistent enterprise extended information. Knowledge consistency verification is a quality control process that ensures the logical consistency and information reliability of multi-source knowledge features. The verification process uses the enterprise in the basic business registration information as the central entity node and the industry knowledge features, scale knowledge features, regional knowledge features, and product knowledge features as attribute nodes. It establishes association edges between the central entity node and attribute nodes to construct an enterprise feature knowledge graph. Semantic disambiguation is performed on the attribute nodes in the enterprise feature knowledge graph to identify conflicting attribute values ​​from different knowledge sources under the same attribute dimension, such as differences in the industry classification of the same enterprise in different knowledge sources. For each conflicting attribute value, its corresponding knowledge source is traced to obtain metadata such as the publishing institution, data update time, and data coverage of the corresponding knowledge source. Based on the authority, timeliness, and coverage of the knowledge source, an authority score is calculated. The formula for calculating the authority score is: Authority score = α0 × Institutional Authority + β0 × Timeliness Score + γ0 × Coverage Score; where α0, β0, and γ0 are weighting coefficients, institutional authority is assessed based on the nature of the publishing institution, timeliness is assessed based on data update frequency, and coverage is assessed based on data completeness. For conflicting attribute dimensions, the authority scores of the knowledge sources corresponding to each conflicting attribute value are compared, and the attribute value corresponding to the knowledge source with the highest authority score is selected as the primary attribute value, while other conflicting attribute values ​​are retained as candidate attribute values. A consistency check score is calculated between the primary attribute value and the basic business information. Consistency verification is evaluated based on the logical compatibility and semantic similarity between attribute values ​​and basic information. When the consistency verification score is less than the preset consistency threshold, the main attribute value and candidate attribute value are weighted and fused to obtain the fused attribute value as the enterprise feature information after conflict resolution. The confidence of each knowledge feature in the enterprise extended information is analyzed. The confidence of the knowledge feature takes into account factors such as the reliability of the knowledge source, the completeness of information acquisition, and the consistency verification results. Knowledge features with confidence scores lower than the preset confidence threshold are marked as features to be verified. Features to be verified will be discounted in subsequent matching confidence calculations.

[0043] In an embodiment of the present invention, the process of constructing a condition information matching library includes: The process iterates through the types of criteria used to determine atomic conditions and constructs specific matching rules for each type. Matching rule construction is the core step in achieving accurate matching; it establishes a mapping relationship between policy conditions and enterprise information to form executable matching logic.

[0044] For regional conditions, regional scope identifiers within the judgment threshold are extracted, including specific regional names or special economic zone names; these regional identifiers are matched with regional knowledge features in the enterprise's extended information, and a multi-level matching strategy is used to generate regional matching rules; the regional matching rules include: extracting the regional scope description from the atomic judgment conditions, standardizing the regional scope description into administrative division codes or latitude and longitude ranges; parsing the enterprise address into administrative division codes or latitude and longitude coordinates; determining whether the enterprise's administrative division code belongs to the set of administrative division codes required by policy, or determining whether the enterprise's latitude and longitude coordinates fall within the latitude and longitude range required by policy, and if so, it is determined as a regional condition match; For scale-related conditions, the numerical range and judgment operator within the judgment threshold are extracted and compared numerically with the scale knowledge features in the enterprise's extended information. Based on the judgment operator, corresponding comparison operations are performed to generate scale matching rules. The scale matching rules include: identifying the scale indicator names and indicator thresholds in the atomic judgment conditions; the scale indicator names include registered capital, annual operating revenue, number of employees, etc.; establishing a mapping relationship between scale indicators and basic business registration information, wherein registered capital can be directly obtained from basic business registration information, and annual operating revenue and number of employees are estimated by interval estimation using the registered capital and the scale level; comparing the enterprise's scale indicator value or estimated range with the indicator threshold, and if the threshold requirement is met, the scale condition is determined to be matched. For industry-specific conditions, extract industry description text within the judgment threshold, such as high-tech enterprises, manufacturing, software and information technology services, etc.; perform semantic similarity calculation between the industry description and industry knowledge features in the enterprise's extended information to generate industry matching rules, using word vector-based similarity calculation with a similarity range of [0,1]. The industry matching rules include: extracting industry requirement descriptions from the atomic judgment conditions, mapping the industry requirement descriptions to a set of industry codes in the industry classification standard system, the mapping process being implemented through industry keyword matching and industry synonym expansion; obtaining the industry code corresponding to the enterprise's industry category; determining whether the enterprise's industry code belongs to the set of industry codes required by policy, if so, determining it as a direct industry match; if the enterprise's industry code is not in the set, further calculating the semantic distance between the enterprise's industry code and each industry code in the set in the industry classification tree, if the minimum semantic distance is less than the preset industry similarity threshold, determining it as an industry similarity match; For product-related conditions, product feature requirements are extracted from the judgment threshold, such as artificial intelligence products, environmental protection products, and products with independent intellectual property rights. These product requirements are then matched with product knowledge features in the company's extended information using multiple tags to generate product matching rules. These matching rules include: extracting product requirement descriptions from the atomic judgment conditions, including product technical field requirements, product innovation level requirements, and product quality certification requirements; matching the product technical field requirements with the technical field tags of the company's main products to calculate the overlap in technical fields; matching the product innovation level requirements with the innovation attribute types of the company's main products to determine whether the innovation level requirements are met; and calculating the matching score for the product condition based on the combined results of the technical field overlap and innovation level matching. All matching rules undergo conflict detection and priority ranking, and are organized into a conditional information matching rule base. Conflict detection is a crucial step in ensuring the consistency of the rule base, avoiding matching errors caused by logical contradictions. The detection process traverses all matching rule pairs, identifying conflicting rule pairs with logical contradictions, such as different matching rules for the same condition yielding opposite conclusions (one condition satisfies the condition, while the other does not). Based on the core admission chain in the logical dependency graph, conflicting rule pairs are prioritized, with rules corresponding to conditions in the core admission chain and rules generated by authoritative data sources having higher priority. High-priority rules are retained, while low-priority rules are deleted or downgraded to candidate rules. All non-conflicting rules and the optimized rules are organized into a conditional information matching rule base, stored in the standard format of the rule engine, supporting efficient rule retrieval and execution.

[0045] The matching rules undergo type validation and weight assignment. Rule type validation is a crucial step in distinguishing between deterministic and fuzzy matching, as different rule types are handled differently in confidence calculation. The validation process determines whether the matching logic of each rule is deterministic: precise numerical comparisons of regional and scale conditions are deterministic matching rules, with a clear true or false result; semantic similarity calculations for industry and product conditions are fuzzy matching rules, with continuous numerical scores. Rule weights are assigned to each matching rule based on the dependency strength of the edges in the logical dependency graph; conditions corresponding to edges with higher dependency strength have higher matching rule weights. This ultimately forms a complete matching rule library containing rule types and weights, providing a refined rule foundation for subsequent confidence calculations.

[0046] In an embodiment of the present invention, the process of calculating the matching confidence score includes: The matching confidence level between each policy clause and the corresponding enterprise is obtained based on the conditional information matching rule base. The matching confidence level is a quantitative expression of the degree of matching between the policy clause and the enterprise characteristics, comprehensively reflecting the likelihood of the policy's applicability. The calculation process first obtains the completeness of basic business information and assigns initial weights based on the completeness. The completeness is calculated based on the filling ratio of key fields and the information quality score. The key fields include basic business information items such as enterprise name and enterprise address. The filling ratio of key fields is obtained by dividing the number of filled fields by the total number of fields. The information quality score is based on the standardization, consistency, and validity of the field content, using multi-dimensional quality checks, including data format checks, data consistency checks, and data validity checks. Initial weights are set according to the completeness, and the initial weights are assigned using a non-linear mapping method. The higher the completeness, the faster the initial weights increase.

[0047] The atomic decision conditions of each policy clause are validated using the matching rules in the conditional information matching rule base. Each atomic decision condition obtains an initial validation score based on the execution result of its matching rule. The validation score ranges from [0,1], where 1 represents a complete match and 0 represents a complete mismatch. In the clause validation process, for deterministic matching rules, the validation score is a discrete value of 0 or 1; for fuzzy matching rules, the validation score is a continuous value in the interval from 0 to 1, and the score is determined according to the degree of matching.

[0048] The initial validation scores are corrected through logical propagation based on the logical relation edges in the logical dependency graph. A logical AND relation requires all connected conditional node validation scores to reach a threshold; a logical OR relation requires at least one connected conditional node validation score to reach a threshold; and a logical NOT relation requires connected conditional node validation scores to be low, resulting in a corrected validation score. The corrected validation score is then multiplied by the rule weight of the corresponding matching rule to obtain a weighted validation score. The correction process starts from the leaf nodes of the logical dependency graph and uses a bottom-up propagation strategy, correcting the validation scores according to the type of logical relation edge. For node groups with logical AND relations, the corrected validation score is calculated using a weighted minimum function, considering both the minimum validation score among the nodes and the influence of the weighted average of all child nodes through a smoothing factor. The smoothing factor is typically set to 0.3 to ensure logical rigor while avoiding excessive penalty for a single low-scoring node. The system employs a multi-level scoring method. For logically OR nodes, the corrected verification score is calculated using a weighted maximum function, which considers both the highest verification score within the node and the influence of the weighted average of all child nodes through a smoothing factor. For logically NOT nodes, the corrected verification score is calculated using a reverse function, which subtracts the original verification score from 1. For complex logical relationships, a recursive propagation algorithm is used, starting from the lowest logical unit and propagating upwards layer by layer until the root node is reached. During propagation, the dependency strength of logical relationship edges is considered; edges with higher dependency strength have a greater impact on upper-level nodes. For nodes with multiple incoming edges, weighted merging is performed based on the dependency strength of each incoming edge to avoid duplicate condition calculations and weight imbalances. Through logical propagation correction, the verification score is ensured to reflect both the satisfaction of each atomic decision condition and accurately express the logical combination relationship between conditions, providing a logically consistent score basis for subsequent weighted calculations.

[0049] Then, the weighted verification scores of all atomic judgment conditions within the same policy clause are summed to obtain the original matching score. The original matching score is a comprehensive quantitative expression of the initial matching degree between the policy clause and the enterprise characteristics, laying the foundation for subsequent information discounting and confidence calculation.

[0050] The number of knowledge features marked as features to be verified within the enterprise's extended information is obtained. The information discount coefficient is calculated based on the proportion of features to be verified, and the information discount coefficient is multiplied by the original matching score to obtain the discounted matching score. The discount coefficient is calculated as follows: Information discount coefficient = 1 - (number of features to be verified ÷ total number of features) × discount penalty factor. In the formula, the discount penalty factor controls the degree of influence of information uncertainty on the matching result. It is usually set to 0.3, which means that the matching score will be discounted by up to 30%.

[0051] The satisfaction level of the core access chain is calculated by first extracting the core access chain from the logical dependency graph; obtaining the corrected verification scores of all condition nodes on the core access chain and the path weights of each condition node in the core access chain, and then calculating the satisfaction level of the core access chain based on these scores. The satisfaction level is the weighted average of the verification scores of all condition nodes in the core access chain. When the satisfaction level is lower than a preset satisfaction threshold, even if the satisfaction levels of other non-core conditions are high, the overall applicability of the policy will be severely limited. The satisfaction level of the core access chain, as a key indicator for judging the applicability of the policy, together with the discount matching score, constitutes a core component of the matching confidence calculation.

[0052] The core admission chain satisfaction, discount matching score, and initial weight are multiplied together, and the result is normalized to obtain the final matching confidence.

[0053] In an embodiment of the present invention, the process of performing differentiated push notifications includes: Based on matching confidence levels, relevant policy clauses are screened to determine a candidate policy set. This candidate policy screening process, based on quantitative evaluation results, prioritizes policies and provides enterprises with the most applicable and actionable policy recommendations. The screening process first marks policy clauses with matching confidence levels greater than a pre-set confidence threshold as candidate policies. Next, duplicate policy clauses are checked to identify repeated policy clauses on the same policy topic. The policy clause with the highest matching confidence level is retained, while redundant policy clauses are eliminated to avoid repetitive interference in the pushed content. Finally, the timeliness of the candidate policies is verified by checking the validity period, application deadline, and other time constraints. Expired or soon-to-expire policy clauses are eliminated to ensure the timeliness and actionability of the pushed policies, resulting in a candidate policy set. This set provides high-quality policy resources for subsequent tiered processing and personalized content generation.

[0054] Based on matching confidence and core access chain satisfaction, candidate push policies are categorized into high-confidence, medium-confidence, and low-confidence policies. Policy grading is a process of precise classification based on confidence assessment results, enabling differentiated push strategies for policies with different levels of applicability. The grading process establishes a two-dimensional grading standard: the first dimension is matching confidence, reflecting the overall degree of matching between policy conditions and enterprise characteristics; the second dimension is core access chain satisfaction, reflecting the degree to which the enterprise meets the key conditions of the policy. Candidate push policies are then classified one by one according to the grading standard, forming policy subsets with three different confidence levels, providing a classification basis for generating targeted push content subsequently.

[0055] For policies with high confidence, we generate directly actionable application guidelines. These guidelines are operational documents that provide detailed application instructions for highly relevant policies, helping companies quickly initiate the policy application process. The generation process first extracts the condition nodes that have passed the clause verification in the logical dependency graph. These condition nodes correspond to the policy conditions that the enterprise has already met. The verified atomic judgment conditions are then converted into descriptions of the access elements that the enterprise has already met, transforming abstract judgment conditions into concrete descriptions of the enterprise's advantages. Key information is extracted from the application requirements of the policy text data, including the policy title (e.g., the SME Technology Innovation Fund Support Plan), the core support content of the policy (e.g., providing technology R&D funding), a list of conditions that the enterprise has already met (generated based on condition nodes with a clause verification result of 1), a brief description of the application process (including key steps such as application material preparation, submission channels, and review process), and the application deadline (clearly defining the specific application time window). Based on the extracted information, an application guidance document is constructed. The document adopts a structured content organization method, highlighting the application advantages that the enterprise already possesses, emphasizing the direct operability of the policy, and providing clear action guidance so that the enterprise can immediately begin policy application preparation. The language style of the application guidance document adopts an affirmative and guiding expression to enhance the enterprise's confidence and motivation in applying.

[0056] The document provides supplementary guidance on policies with a medium confidence level. This supplementary guidance document is a guiding document that provides information improvement instructions for some matching degree policies, and can help enterprises improve the applicability of policies through information supplementation. The generation process analyzes information gaps in policy conditions for enterprises based on logical dependency graphs, identifying key factors leading to insufficient confidence, such as incomplete information, partial fulfillment of conditions, and pending verification. It then generates supplementary information guidance, including the policy title, core policy support content, possible conditions met by the enterprise (based on verification results with moderate confidence), a list of missing or pending information items (clearly outlining key missing or pending information), and methods for supplementing information (providing information acquisition channels, material preparation guidance, and verification method suggestions). The information item list is prioritized, placing the items most impactful on policy applicability at the top, providing enterprises with targeted information supplementation paths. The supplementary information methods offer specific operational guidance, including practical suggestions on obtaining authoritative information from which channels, how to prepare relevant supporting materials, and how to conduct self-checks. The language style of the guidance is suggestive and persuasive, pointing out areas for improvement without placing excessive pressure on enterprises, encouraging them to improve policy applicability through their efforts.

[0057] For policies with low confidence levels, we generate policy awareness expansion documents. These documents are educational materials that provide guidance for learning and capacity building regarding policies with low matching scores, helping companies understand policy requirements and develop long-term development plans. The generation process analyzes the main gaps between the company and the policy conditions, identifying shortcomings in areas such as scale, technology, management, and market. The generated policy awareness expansion documents include the policy title, core policy support content, a description of the main conditions the company currently does not meet (based on the low confidence verification results, clearly pointing out specific gaps), and suggestions for improvement (providing targeted development advice and capacity building plans). The description of the main conditions uses a comparative analysis approach, comparing policy requirements with the company's current situation to help the company clearly understand the gaps. The suggestions for improvement, based on the gap analysis results, provide phased and actionable development recommendations, including aspects that can be improved in the short term and directions that need long-term effort. The recommendations cover multiple dimensions such as improving technological R&D capabilities, expanding business scale, improving management systems, and expanding markets, providing a reference for companies to formulate development plans. The language style of the documents adopts an educational and motivational approach, objectively pointing out gaps while actively encouraging company development, reflecting the long-term value of policy services.

[0058] Personalize and integrate the push notification texts of all candidate push policies to generate comprehensive push content. The integration process prioritizes policies based on high, medium, and low confidence levels, placing the most actionable content in the most prominent positions to ensure businesses focus on the most suitable policy opportunities. Each policy document is refined to avoid redundancy and repetitive expression, highlighting its core value and action requirements. A clear content structure and layout are designed, enhancing readability and comprehension through hierarchical headings, highlighting key points, and extracting essential information. The content is personalized based on the company's industry characteristics, size, and geographic location, reflecting individual preferences in expression, case studies, and key recommendations. Structured, comprehensive push content is generated, including an implementation summary (outlining policy opportunities and key actions), high-priority policy recommendations (detailed application guidelines), medium-priority policy opportunities (supplementary information), policies of long-term focus (knowledge expansion and capacity building), and policy service contact information. The comprehensive push content undergoes format optimization and readability adjustments to ensure clear, logical, and focused expression. This optimized, personalized content is then precisely delivered to target companies through a policy push platform, achieving accurate matching and effective delivery of policy information to business needs. Example

[0059] Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A system for precisely targeting policy services for SMEs is provided, including: The data acquisition module is used to obtain basic business registration information and policy documents of the target company. The data parsing module performs clause-level atomic decomposition on the policy clauses within the policy text data to obtain several atomic judgment conditions. It extracts the judgment basis type and judgment threshold from the atomic judgment conditions and simultaneously constructs a logical dependency graph between the atomic judgment conditions. The judgment basis type includes regional conditions, scale conditions, industry conditions, and product conditions. The knowledge expansion module expands the basic business information into multi-dimensional knowledge to obtain enterprise extended information; and constructs a conditional information matching rule base based on the enterprise extended information and the judgment basis type, and obtains the matching confidence of each policy clause with the corresponding enterprise according to the conditional information matching rule base. The data push module filters out suitable policy clauses based on the matching confidence level and logical dependency relationship graph, and generates personalized policy push content for the corresponding enterprise based on them. The personalized policy push content is then pushed in a differentiated manner based on the matching confidence level.

[0060] The modules are connected via wired and / or wireless means to enable data transmission between them.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0062] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0063] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for precisely delivering policy services to SMEs, characterized in that: include: Obtain basic business registration information and policy documents of the target company; The policy clauses within the policy text data are decomposed into clause-level atomic decompositions to obtain several atomic judgment conditions. The judgment basis type and judgment threshold in the atomic judgment conditions are extracted, and a logical dependency graph between the atomic judgment conditions is constructed simultaneously. The judgment basis type includes regional conditions, scale conditions, industry conditions, and product conditions. The basic business information is expanded into multiple dimensions to obtain enterprise extended information; and a conditional information matching rule base is constructed based on the enterprise extended information and the judgment criteria type. The matching confidence of each policy clause with the corresponding enterprise is obtained according to the conditional information matching rule base. Based on the matching confidence level and logical dependency graph, suitable policy clauses are selected, and personalized policy push content is generated for the corresponding enterprises. The personalized policy push content is then pushed in a differentiated manner based on the matching confidence level.

2. The method for precisely delivering policy services to SMEs according to claim 1, characterized in that, The process of performing clause-level atomization decomposition includes: Syntactic dependency analysis is performed on the policy clauses in the policy text data to identify the subject-verb-object semantic units in each policy clause. By extracting the modification and dominance relationships between the subject-verb-object semantic units, a clause semantic dependency tree is obtained. The policy clauses are separated into conditional and conclusive clauses to obtain the corresponding conditional clauses. The limiting and constraining components in the conditional clauses are then extracted based on the semantic dependency tree of the clauses. Traverse the logical connector nodes in the semantic dependency tree of the clause, determine the semantic boundary of the atomic decision condition based on the position of the logical connector node, and divide each policy clause into several semantically independent atomic decision conditions based on it. For each atomic decision condition, the components are extracted to obtain the decision object, decision dimension, decision operator and decision value in the atomic decision condition; The judgment criteria type for the corresponding atomic judgment conditions is determined based on the judgment dimension, judgment object, and constraint components, and constraint strength is assigned to each atomic judgment condition. The decision operator and the decision value are combined to form the decision threshold, and a structured representation is constructed for each atomic decision condition.

3. The method for precisely delivering policy services to SMEs according to claim 1, characterized in that, The process of constructing a logical dependency graph includes: Atomic decision conditions are mapped to condition nodes, and each condition node is labeled with the decision basis type, decision threshold, and constraint strength; Traverse all atomic decision conditions within the same policy clause and identify logical relationships based on the logical connectors between the atomic decision conditions; A logical relationship dependency graph is constructed using the identified logical relationships as edges and the mapped condition nodes as graph nodes; Calculate the dependency strength of each logical relation edge in the logical relation dependency graph, perform topological sorting on the logical relation dependency graph based on the dependency strength of the logical relation edges, identify the key condition paths in the logical relation dependency graph as core admission chains, and assign path weights to each condition node in the core admission chains.

4. The method for precisely delivering policy services to SMEs according to claim 1, characterized in that, The process of expanding knowledge in multiple dimensions includes: Semantic expansion of the industry categories in the basic business information yields industry knowledge features. Based on the registered capital in the basic business information, the scale level is inferred to obtain the scale knowledge characteristics of the corresponding enterprise; Based on the enterprise addresses in the basic business information, regional economic attributes are analyzed to obtain regional knowledge characteristics; Based on the main products in the basic business information, a technology field mapping is performed to obtain product knowledge characteristics; The industry knowledge characteristics, scale knowledge characteristics, regional knowledge characteristics, and product knowledge characteristics are verified for knowledge consistency. Conflict information between different knowledge sources is detected, and the detected conflict information is resolved to obtain enterprise extended information. Analyze the confidence level of each knowledge feature in the enterprise's extended information, and mark knowledge features with confidence levels below a preset confidence threshold as features to be verified.

5. The method for precisely delivering policy services to SMEs according to claim 1, characterized in that, The process of building a conditional information matching rule base includes: Iterate through the judgment criteria types corresponding to the atomic judgment conditions. If the judgment criteria type is a regional condition, perform text matching between the regional range identifier within the corresponding judgment threshold and the regional knowledge features in the enterprise extended information to obtain the regional matching rules. If the judgment criteria are scale-based conditions, the numerical range within the corresponding judgment threshold is compared with the scale knowledge features in the enterprise's extended information to obtain the scale matching rules. If the judgment criteria are industry-specific conditions, the semantic similarity between the industry description text within the corresponding judgment threshold and the industry knowledge features in the enterprise's extended information is calculated to obtain the industry matching rules. If the judgment criteria are product-type conditions, the product feature requirements in the corresponding judgment threshold are matched with the product knowledge features in the enterprise's extended information using multiple tags to generate product matching rules. The region matching rules, scale matching rules, industry matching rules and product matching rules are subjected to conflict detection to identify conflicting rule pairs with logical contradictions. The conflicting rule pairs are prioritized in combination with the core access chain, and high-priority conflicting rules are selected and organized with non-conflicting rules into a condition information matching rule library. The rule type of each matching rule in the condition information matching rule base is verified, and the rule type includes deterministic matching rules and fuzzy matching rules; and rule weights are assigned to each matching rule according to the dependency strength of each logical relation edge in the logical dependency graph.

6. The method for precisely delivering policy services to SMEs according to claim 1, characterized in that, The process of obtaining matching confidence includes: Obtain the completeness of the collected basic business information and assign initial weights based on it; Based on the matching rules in the conditional information matching rule base, the atomic decision conditions of each policy clause are verified to obtain the initial verification score of each atomic decision condition. The initial verification score is corrected by logical propagation based on the logical relationship edges in the logical relationship dependency graph to obtain the corrected verification score; The modified verification score is multiplied by the rule weight of the corresponding matching rule to obtain the weighted verification score; and the weighted verification scores of all atomic judgment conditions within the same policy clause are summed to obtain the original matching score. Obtain the number of knowledge features marked as features to be verified within the enterprise's extended information, calculate the information discount coefficient based on it, and calculate the discounted matching score by combining the information discount coefficient with the original matching score. The satisfaction of the core admission chain is calculated. The satisfaction of the core admission chain, the discount matching score and the initial weight are multiplied together and the result is normalized to obtain the matching confidence.

7. The method for precisely delivering policy services to SMEs according to claim 1, characterized in that, The process of providing differentiated push notifications includes: Based on the matching confidence level, relevant policy clauses are filtered, and policy clauses with a matching confidence level greater than a preset confidence threshold are marked as candidate policies to be pushed. Based on the matching confidence and core access chain satisfaction, the candidate push policy is divided into high-confidence policy, medium-confidence policy and low-confidence policy; For high-confidence policies, extract the atomic judgment conditions corresponding to the condition nodes that have passed the clause verification in the logical dependency graph, convert the atomic judgment conditions into descriptions of the access elements that the enterprise has met, and generate application guidance text in combination with policy text data; for medium-confidence policies, generate supplementary information guidance text based on the dependency graph; for low-confidence policies, generate policy understanding expansion text. Personalize and integrate the corresponding texts of all candidate push policies, prioritize them according to high confidence, medium confidence and low confidence policies, generate comprehensive push content, optimize the format of comprehensive push content, and push it to the target enterprise.

8. The method for precisely delivering policy services to SMEs according to claim 2, characterized in that, The process of extracting limiting and constraining components from conditional clauses based on the aforementioned semantic dependency tree includes: Each semantic unit in the semantic dependency tree of the clause is tagged with part of speech to obtain candidate limiting components; the central words used for modification in the candidate limiting components are traversed and extracted, and it is determined whether the corresponding central words are nouns related to enterprise attributes; By using nouns related to enterprise attributes, numerical and range expressions are extracted from candidate bounded components. Candidate bounded components containing numerical or range expressions are marked as bounded components. The numerical content and range boundaries of these bounded components are then extracted and denoted as the quantitative attributes of the bounded components. Pattern matching is performed on the sentence structure in the semantic dependency tree of the clauses to identify and mark the condition markers in the semantic units, and the semantic units containing the condition markers are marked as constraint components; and constraint strength is assigned to the atomic decision conditions modified by the constraint components.

9. The method for precisely delivering policy services to SMEs according to claim 4, characterized in that, The process of performing knowledge consistency verification includes: The enterprise in the basic business information is taken as the central entity node, and the knowledge features in the enterprise extended information are taken as attribute nodes. The association edges between the central entity node and the attribute nodes are established to obtain the enterprise feature knowledge graph. Semantic disambiguation is performed on the attribute nodes in the enterprise feature knowledge graph to identify conflicting attribute values ​​from different knowledge sources under the same attribute dimension; For each conflicting attribute value, trace its corresponding knowledge source, obtain the publishing institution, update time and data coverage of the corresponding knowledge source, and calculate the authority score of the corresponding knowledge source based on it; For conflicting attribute dimensions, compare the authority scores of the knowledge sources corresponding to each conflicting attribute value, select the attribute value corresponding to the knowledge source with the highest authority score as the main attribute value, and retain the other conflicting attribute values ​​as candidate attribute values. Calculate the consistency check score between the main attribute value and the basic business information; when the consistency check score is less than the preset consistency threshold, perform weighted fusion of the main attribute value and the candidate attribute value, and use the fused attribute value as the enterprise feature information after resolving the conflict.

10. A policy service precision delivery system for SMEs, used to implement the policy service precision delivery method for SMEs as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to obtain basic business registration information and policy documents of the target company. The data parsing module performs clause-level atomic decomposition on the policy clauses within the policy text data to obtain several atomic judgment conditions. It extracts the judgment basis type and judgment threshold from the atomic judgment conditions and simultaneously constructs a logical dependency graph between the atomic judgment conditions. The judgment basis type includes regional conditions, scale conditions, industry conditions, and product conditions. The knowledge expansion module expands the basic business information into multi-dimensional knowledge to obtain enterprise extended information; and constructs a conditional information matching rule base based on the enterprise extended information and the judgment basis type, and obtains the matching confidence of each policy clause with the corresponding enterprise according to the conditional information matching rule base. The data push module filters out suitable policy clauses based on the matching confidence level and logical dependency relationship graph, and generates personalized policy push content for the corresponding enterprise based on them. The personalized policy push content is then pushed in a differentiated manner based on the matching confidence level.

Images