Knowledge graph construction method oriented to policy text semantic understanding

Abstract

The invention discloses a knowledge graph construction method oriented to policy text semantic understanding, and relates to the technical field of knowledge graph construction, and the method comprises the following steps: collecting and preprocessing a policy text, extracting a structural text, and screening out sentences containing fuzzy keywords; identifying policy entities and relationships, extracting subjects, operations and fuzzy modification entities, and forming a triple; constructing a localized quantitative index library, accessing a regional economic database, and extracting historical data and distribution of key indexes; dynamically generating a quantization rule, and automatically quantizing the fuzzy expression into a specific threshold rule according to regional data; fusing and storing to a knowledge graph, and storing the triad and a quantization rule node into a graph database together; and providing semantic query and reasoning services, automatically triggering a quantization rule in response to query, and returning an explainable specific result. According to the method, abstract and fuzzy policy terms can be converted into specific and executable quantitative indexes.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph construction technology, and in particular relates to a method for constructing knowledge graphs for semantic understanding of policy texts. Background Technology

[0002] Existing methods typically rely on rule-based template matching or statistical machine learning models to extract and define entities and their relationships. However, these methods have inherent limitations when faced with numerous vague and poorly defined expressions in policy clauses (such as "appropriate support" and "significant growth"). The semantics of these vague terms are highly dependent on the specific implementation context and historical conventions of the issuing region, while traditional technologies can only handle their literal or static general meanings, failing to provide precise and locally relevant quantitative interpretations. Therefore, while knowledge graphs built based on such methods can present the surface relationships between entities in policy texts, they struggle to carry and express the deep, actionable semantic information of these key vague clauses. This results in their inability to provide clear and unambiguous decision-making basis for automated approval or intelligent consultation in practical applications, severely limiting the practical value of knowledge graphs in precise policy services. To address these issues, the following solutions are proposed. Summary of the Invention

[0003] The purpose of this invention is to provide a knowledge graph construction method for semantic understanding of policy texts. By constructing a dynamic localized quantitative rule generator, it automatically extracts ambiguous expressions in policy texts and embeds them into a regionally specific quantitative indicator library, generating computable quantitative rule nodes and associating them with the knowledge graph. This method can transform abstract and ambiguous policy clauses into specific and actionable quantitative indicators, solving the problem that existing technologies cannot interpret ambiguous expressions that rely on local context and historical data, resulting in insufficient semantic depth of policy knowledge graphs and a lack of actionable decision-making basis.

[0004] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0005] This invention provides a knowledge graph construction method for semantic understanding of policy texts, which specifically includes the following steps:

[0006] Step S1: Policy text collection and preprocessing: Collect policy text and preprocess it, extract the structured text and filter out sentences containing fuzzy keywords;

[0007] Step S2, Policy Entity and Relationship Extraction: Identify policy entities and relationships, extract subjects, operations, and fuzzy modified entities, and form triples;

[0008] Step S3: Construction of a localized quantitative indicator library: Construct a localized quantitative indicator library, connect to the regional economic database, and extract historical data and distribution of key indicators;

[0009] Step S4: Dynamic Quantization Rule Generation: Dynamically generate quantization rules, automatically quantizing fuzzy expressions into specific threshold rules based on regional data;

[0010] Step S5, Knowledge Graph Fusion and Storage: Fusion and storage to the knowledge graph, storing triples and quantized rule nodes together in the graph database;

[0011] Step S6, Semantic Query and Reasoning Service: Provides semantic query and reasoning service, automatically triggers quantification rules and returns interpretable specific results when responding to queries.

[0012] Furthermore, step S1, policy text collection and preprocessing, specifically includes the following steps:

[0013] Step S11: Collect the original text of industry support policies (usually in PDF or HTML format) from the public databases of provincial and municipal governments, and use the PDFMiner tool to parse it and extract the plain text content. At the same time, pay attention to preserving the original paragraph divisions and heading hierarchy structure to lay the foundation for subsequent structured analysis.

[0014] Step S12: The extracted text is segmented into sentences and dependency parsing is performed using the SPACY tool to identify sentences containing vague expressions. The selection of sentences relies on a dynamically expanding keyword library containing words such as "appropriate", "reasonable", and "significant". Sentences containing such keywords are included in the candidate range.

[0015] Step S13: Store all the selected sentences into a set called Candidate Fuzzy Sentences. In the structure, each Each represents a separate candidate fuzzy sentence, and this set will serve as the direct input for subsequent deep semantic processing.

[0016] Furthermore, step S2, the extraction of policy entities and relationships, specifically includes the following steps:

[0017] Step S21: Use the BiLSTM-CRF model to perform named entity recognition on the policy text, aiming to accurately extract three types of key entities: first, subject entities, such as enterprise types like "high-tech enterprises" or "micro and small enterprises" and project types like "R&D projects"; second, operational entities, such as subsidy forms like "funding subsidies" or "tax reductions" and approval conditions like "passing expert review"; and third, fuzzy modifier entities, such as vague expressions like "significant" or "appropriate" and the quantitative dimensions they modify, such as "output value" or "R&D investment," thereby laying a data foundation for subsequent semantic understanding.

[0018] Step S22: Based on the Policy-BERT language model pre-trained on the policy corpus, deeply analyze the semantic relationships and contextual dependencies between these identified entities, automatically extract the various relationships between them, and thus form a structured set of triples. ( For the head entity, For tail entities, To connect entities and entity (The relationship between them), such as generating knowledge assertions like (high-tech enterprises enjoy, financial subsidies) or (financial subsidies, the condition is, significant R&D investment) to clearly express the provisions and constraints in the policy clauses.

[0019] Furthermore, step S3, the construction of the localized quantitative indicator library, specifically includes the following steps:

[0020] Step S31: The system automatically accesses the economic database of the target region. These databases include, but are not limited to, annual enterprise reports published by the statistics bureau and records of policy subsidies retained internally by the government over the years. Key quantitative indicators are then extracted from this unstructured or semi-structured data, specifically including the distribution of enterprise annual output growth rates (its mathematical representation is...). That is, the indicator follows the mean. variance is (Normal distribution) and the fluctuation range of the proportion of R&D investment It will also extract the range of historical subsidy amounts. and the percentage of companies receiving subsidies out of the total. Key threshold data

[0021] Step S32: After completing the indicator extraction, the system will assign a corresponding regional code attribute to each extracted data indicator. Then, these structured indicator data with regional labels will be persistently stored to jointly construct a standardized localized quantitative indicator library. Logically, this library can be viewed as a set of tuples. This set of data provides solid and specialized data support for the generation of dynamic rules in subsequent steps.

[0022] Furthermore, step S4, the dynamic quantization rule generation, specifically includes the following steps:

[0023] Step S41: For each sentence containing ambiguous expressions selected after preprocessing, perform deep semantic analysis to accurately identify the specific quantitative dimension pointed to by the ambiguous modifier (such as "significant") (for example, whether the word modifies "output growth rate" or "R&D investment ratio"), and record this dimension as d, which will serve as the core object for subsequent quantitative calculations.

[0024] Step S42: Based on the issuing region code of the current policy text, retrieve historical data statistics matching dimension d in real time from the constructed localized quantitative indicator library; when d is a continuous dimension, call the formula. Perform the calculation, where, This represents the average value of historical data for dimension d in this region. Its standard deviation, This is the adjustment coefficient obtained by fitting the distribution of historical subsidy cases, which ultimately yields a specific threshold value; when When the dimension is proportional, the quantiles in the historical records are extracted directly (e.g., the 80th quantile). () as the benchmark value;

[0025] Step S43: Transform the specific numerical results obtained from the above calculations or extractions into a structured, machine-readable quantification rule (e.g., "output growth rate ≥ 12%)", and create an independent quantification rule node in the knowledge graph. Finally, establish a clear "quantified as" relationship between this node and the fuzzy modified entity extracted in step S2, thereby completing the mapping from fuzzy semantics to precise rules.

[0026] Furthermore, step S5, knowledge graph fusion and storage, specifically includes the following steps:

[0027] Step S51: Extract the policy entity relationship triplet set from step S2. With the quantization rule nodes dynamically generated in step S4 The nodes are integrated to form an extended knowledge graph. This graph is stored using the Neo4j graph database, which defines various node types, including policy clauses, entities (such as principal entities and operational entities), and innovative quantitative rule nodes. It also defines various relationship types, such as conditional relationships, quantitative relationships, and membership relationships, to accurately describe the logical and quantitative rule dependencies between entities.

[0028] Step S52: Attach the necessary metadata attributes to each quantification rule node, including its applicable region code, the data source on which the rule was generated, and the effective time of the rule, so as to ensure the traceability and timeliness of knowledge in the graph.

[0029] Furthermore, step S6, the semantic query and reasoning service, specifically includes the following steps:

[0030] Step S61: After receiving the query request input by the user in natural language, the system first parses the query statement, identifies the key entities (such as "City A", "high-tech enterprise", "R&D subsidy") and intent, and then matches them in the graph database to locate the policy clause nodes containing these entities.

[0031] Step S62: If the matched policy clause contains a vague expression (e.g., "significant R&D investment"), the inference engine will automatically traverse along the "quantification relationship" edges in the knowledge graph, triggering the associated dynamic quantification rule nodes, and storing the specific numerical threshold in the rule (e.g., the rule may contain a formula). The calculated result of "R&D investment ratio ≥ 8%" is returned to the user as the final condition;

[0032] Step S63: When the system returns the answer to the user, it will also provide a complete and traceable explanation of the reasoning path, clearly indicating which policy clause and what vague expression the final specific threshold is derived from, and indicating the local data source (such as "2023 A City Statistics Bureau Report") and other metadata such as the regional code on which the quantification rule is based, thereby ensuring the transparency and interpretability of the results.

[0033] The present invention has the following beneficial effects:

[0034] 1. This invention uses a dynamic localized quantitative rule generator to transform abstract and vague expressions in policy texts into specific and calculable quantitative thresholds based on actual regional data. This process improves the accuracy of machine understanding of the deep semantics and implicit implementation standards of policies, overcoming the limitations of traditional text analysis methods that can only process explicit and structured information. The resulting knowledge graph is no longer just a network of terminology relationships, but a decision support system that carries specific implementation rules. This enables enterprises or approval departments to obtain clear and unambiguous indicator requirements, enhances the operability of policy provisions, and provides a reliable semantic foundation for automated approval and intelligent consultation.

[0035] 2. This invention automatically generates quantitative rules that conform to the economic development level and policy implementation history of different regions by accessing and learning from localized quantitative indicator databases of different regions, rather than adopting a nationally unified static standard. This mechanism solves the problem of inconsistent policy implementation standards caused by regional differences, ensuring that the same ambiguous clause can be interpreted fairly and reasonably in different regions in accordance with their local realities. This respects the particularity of local policies and, from a technical perspective, ensures that the evaluation standards faced by enterprises in different regions when applying for support are matched with their environment, thus promoting substantive fairness in policy implementation.

[0036] 3. This invention derives quantitative rules from structured local data through an automated process, achieving a fundamental shift in the way knowledge is acquired. The system can autonomously discover statistical patterns and historical execution conventions in the data and transform them into usable knowledge, thereby reducing the continuous reliance on manual annotation and expert intervention. This not only reduces the time and economic costs of system construction and subsequent updates, but also makes the maintenance of the knowledge graph more efficient and sustainable, making it suitable for application scenarios with frequent policy updates and extensive coverage areas.

[0037] 4. The knowledge graph constructed by this invention clearly records the complete generation path and data source from fuzzy expressions to specific quantitative rules; when outputting a reasoning conclusion, it can provide a clear and traceable explanation chain; this deep interpretability surpasses the method of only providing the original text of policy provisions, allowing users not only to know the conclusion, but also to understand the logic and basis for the conclusion; this enhances the information transparency between the government and enterprises, improves the credibility of automated decision-making systems, and also helps to audit and supervise the implementation of policies.

[0038] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the knowledge graph construction method for semantic understanding of policy texts according to the present invention. Detailed Implementation

[0041] 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.

[0042] Please see Figure 1 As shown, this invention is a knowledge graph construction method for semantic understanding of policy texts. The construction method specifically includes the following steps:

[0043] Step S1: Policy text collection and preprocessing: Collect policy text and preprocess it, extract the structured text and filter out sentences containing fuzzy keywords;

[0044] Step S1, policy text collection and preprocessing, specifically includes the following steps:

[0045] Step S11: Collect the original text of industry support policies (usually in PDF or HTML format) from the public databases of provincial and municipal governments, and use the PDFMiner tool to parse it and extract the plain text content. At the same time, pay attention to preserving the original paragraph divisions and heading hierarchy structure to lay the foundation for subsequent structured analysis.

[0046] Step S12: The extracted text is segmented into sentences and dependency parsing is performed using the SPACY tool to identify sentences containing vague expressions. The selection of sentences relies on a dynamically expanding keyword library containing words such as "appropriate", "reasonable", and "significant". Sentences containing such keywords are included in the candidate range.

[0047] Step S13: Store all the selected sentences into a set called Candidate Fuzzy Sentences. In the structure, each Each represents a separate candidate fuzzy sentence, and this set will serve as the direct input for subsequent deep semantic processing.

[0048] Step S2, Policy Entity and Relationship Extraction: Identify policy entities and relationships, extract subjects, operations, and fuzzy modified entities, and form triples;

[0049] Step S2, the extraction of policy entities and relationships, specifically includes the following steps:

[0050] Step S21: Use the BiLSTM-CRF model to perform named entity recognition on the policy text, aiming to accurately extract three types of key entities: first, subject entities, such as enterprise types like "high-tech enterprises" or "micro and small enterprises" and project types like "R&D projects"; second, operational entities, such as subsidy forms like "funding subsidies" or "tax reductions" and approval conditions like "passing expert review"; and third, fuzzy modifier entities, such as vague expressions like "significant" or "appropriate" and the quantitative dimensions they modify, such as "output value" or "R&D investment," thereby laying a data foundation for subsequent semantic understanding.

[0051] Step S22: Based on the Policy-BERT language model pre-trained on the policy corpus, deeply analyze the semantic relationships and contextual dependencies between these identified entities, automatically extract the various relationships between them, and thus form a structured set of triples. ( For the head entity, For tail entities, To connect entities and entity (The relationship between them), such as generating knowledge assertions like (high-tech enterprises enjoy, financial subsidies) or (financial subsidies, the condition is, significant R&D investment) to clearly express the provisions and constraints in the policy clauses.

[0052] Step S3: Construction of a localized quantitative indicator library: Construct a localized quantitative indicator library, connect to the regional economic database, and extract historical data and distribution of key indicators;

[0053] Step S3, the construction of the localized quantitative indicator library specifically includes the following steps:

[0054] Step S31: The system automatically accesses the economic database of the target region. These databases include, but are not limited to, annual enterprise reports published by the statistics bureau and records of policy subsidies retained internally by the government over the years. Key quantitative indicators are then extracted from this unstructured or semi-structured data, specifically including the distribution of enterprise annual output growth rates (its mathematical representation is...). That is, the indicator follows the mean. variance is (Normal distribution) and the fluctuation range of the proportion of R&D investment It will also extract the range of historical subsidy amounts. and the percentage of companies receiving subsidies out of the total. Key threshold data

[0055] Step S32: After completing the indicator extraction, the system will assign a corresponding regional code attribute to each extracted data indicator. Then, these structured indicator data with regional labels will be persistently stored to jointly construct a standardized localized quantitative indicator library. Logically, this library can be viewed as a set of tuples. This set of data provides solid and specialized data support for the generation of dynamic rules in subsequent steps.

[0056] Step S4: Dynamic Quantization Rule Generation: Dynamically generate quantization rules, automatically quantizing fuzzy expressions into specific threshold rules based on regional data;

[0057] Step S4, the dynamic quantization rule generation specifically includes the following steps:

[0058] Step S41: For each sentence containing ambiguous expressions selected after preprocessing, perform deep semantic analysis to accurately identify the specific quantitative dimension pointed to by the ambiguous modifier (such as "significant") (for example, whether the word modifies "output growth rate" or "R&D investment ratio"), and record this dimension as d, which will serve as the core object for subsequent quantitative calculations.

[0059] Step S42: Based on the issuing region code of the current policy text, retrieve historical data statistics matching dimension d in real time from the constructed localized quantitative indicator library; when d is a continuous dimension, call the formula. Perform the calculation, where, This represents the average value of historical data for dimension d in this region. Its standard deviation, This is the adjustment coefficient obtained by fitting the distribution of historical subsidy cases, which ultimately yields a specific threshold value; when When the dimension is proportional, the quantiles in the historical records are extracted directly (e.g., the 80th quantile). () as the benchmark value;

[0060] Step S43: Transform the specific numerical results obtained from the above calculations or extractions into a structured, machine-readable quantification rule (e.g., "output growth rate ≥ 12%)", and create an independent quantification rule node in the knowledge graph. Finally, establish a clear "quantified as" relationship between this node and the fuzzy modified entity extracted in step S2, thereby completing the mapping from fuzzy semantics to precise rules.

[0061] Step S5, Knowledge Graph Fusion and Storage: Fusion and storage to the knowledge graph, storing triples and quantized rule nodes together in the graph database;

[0062] Step S5, knowledge graph fusion and storage specifically includes the following steps:

[0063] Step S51: Extract the policy entity relationship triplet set from step S2. With the quantization rule nodes dynamically generated in step S4 The nodes are integrated to form an extended knowledge graph. This graph is stored using the Neo4j graph database, which defines various node types, including policy clauses, entities (such as principal entities and operational entities), and innovative quantitative rule nodes. It also defines various relationship types, such as conditional relationships, quantitative relationships, and membership relationships, to accurately describe the logical and quantitative rule dependencies between entities.

[0064] Step S52: Attach the necessary metadata attributes to each quantification rule node, including its applicable region code, the data source on which the rule was generated, and the effective time of the rule, so as to ensure the traceability and timeliness of knowledge in the graph.

[0065] Step S6, Semantic Query and Reasoning Service: Provides semantic query and reasoning service, automatically triggers quantification rules and returns interpretable specific results when responding to queries.

[0066] Step S6, the semantic query and reasoning service specifically includes the following steps:

[0067] Step S61: After receiving the query request input by the user in natural language, the system first parses the query statement, identifies the key entities (such as "City A", "high-tech enterprise", "R&D subsidy") and intent, and then matches them in the graph database to locate the policy clause nodes containing these entities.

[0068] Step S62: If the matched policy clause contains a vague expression (e.g., "significant R&D investment"), the inference engine will automatically traverse along the "quantification relationship" edges in the knowledge graph, triggering the associated dynamic quantification rule nodes, and storing the specific numerical threshold in the rule (e.g., the rule may contain a formula). The calculated result of "R&D investment ratio ≥ 8%" is returned to the user as the final condition;

[0069] Step S63: When the system returns the answer to the user, it will also provide a complete and traceable explanation of the reasoning path, clearly indicating which policy clause and what vague expression the final specific threshold is derived from, and indicating the local data source (such as "2023 A City Statistics Bureau Report") and other metadata such as the regional code on which the quantification rule is based, thereby ensuring the transparency and interpretability of the results.

[0070] One specific application of this embodiment is:

[0071] Background: City A issued the "Measures for Supporting High-tech Enterprises," which includes the clause: "Enterprises with significant R&D investment will be given appropriate financial subsidies." However, this clause lacks quantifiable standards for "significant" and "appropriate," leading to frequent disputes during the application process. This implementation uses a Dynamic Localized Quantitative Rule Generator (LQRG) to accurately parse this clause.

[0072] Implementation steps:

[0073] Step S1: Policy Text Collection and Preprocessing

[0074] Obtain the PDF file of "Measures for Supporting High-tech Enterprises" from the official website of Municipal Government A, use PDFMiner to extract the text content, and retain the chapter structure (such as "Chapter 3 Subsidy Standards").

[0075] Sentences containing ambiguous expressions are identified through dependency parsing, and target sentences are selected:

[0076] =“Appropriate financial subsidies will be provided to enterprises that make significant investments in research and development”;

[0077] The sentence contains the keywords "significant" and "appropriate," and is added to the candidate fuzzy sentence set. .

[0078] Step S2: Extraction of Policy Entities and Relationships

[0079] Entity extraction using the BiLSTM-CRF model:

[0080] Main entity: High-tech enterprise (Type: Corporate entity);

[0081] Operating entity: Financial subsidies (Type: Subsidy form);

[0082] Fuzzy modifiers for entities: significant (modification dimension: R&D investment), appropriate (modification dimension: subsidy amount);

[0083] Extracting relation triples using the Policy-BERT model:

[0084] High-tech enterprises are eligible for financial subsidies; the condition for receiving financial subsidies is significant investment in research and development.

[0085] Financial subsidies, a modifier, appropriate.

[0086] Step S3: Construction of a localized quantitative indicator library

[0087] Accessing the database of the Statistics Bureau of City A and the records of subsidies for high-tech enterprises from 2018 to 2022, the following indicators were extracted:

[0088] R&D investment as a percentage of total expenditure: Average Standard deviation ;

[0089] Historical subsidy amount range: Ten thousand yuan;

[0090] The number of subsidized enterprises accounts for the top 25% of all enterprises.

[0091] Store as a quantitative indicator library :

[0092] City A, R&D investment ratio, distribution parameters (6.5, 2.1); City A, subsidy amount, range [10, 50].

[0093] Step S4: Dynamic Quantization Rule Generation

[0094] For ambiguous sentences "Significant" in the context:

[0095] Related quantitative dimension: R&D investment ratio (d = R&D_ratio);

[0096] Search the indicator database of City A and calculate the threshold:

[0097] (in Fitting data using historical data: covering the standard normal quantiles of the top 25% of companies);

[0098] Generate rule node: significant, quantified as R&D investment ratio ≥ 8.26%;

[0099] Regarding “appropriate”:

[0100] Related dimension: subsidy amount ( );

[0101] Take the 80th percentile of historical subsidy amounts (corresponding to the upper limit of the amount, since subsidies cover the top 25% of enterprises):

[0102]

[0103] Generate rule nodes: (“Appropriate” is quantified as “subsidy amount ≤ 420,000 yuan”); Step S5, knowledge graph fusion and storage

[0104] The original triplet and quantized rule nodes are merged into an extended graph:

[0105] Node 1: Policy Clause Node (ID: P_Clause_01);

[0106] Node 2: Entity node "High-tech Enterprise";

[0107] Node 3: "Financial subsidies" for physical nodes;

[0108] Node 4: Rule node "R&D investment ≥ 8.26%" (linked to "significant");

[0109] Node 5: Rule node "Subsidy amount ≤ 420,000" (linked to "Appropriate");

[0110] relation:

[0111] (P_Clause_01, contains condition, node 4); (node ​​3, has constraint, node 5).

[0112] Step S6, Semantic Query and Reasoning Service

[0113] A user inquired: "What are the eligibility requirements for high-tech enterprises in City A to apply for R&D subsidies? What is the maximum subsidy amount?"

[0114] Graphical reasoning process:

[0115] Match policy clause P_Clause_01 and associated entity "funding subsidies";

[0116] The condition "significant R&D investment" was detected, triggering rule node 4, and the quantitative condition "R&D investment ratio ≥ 8.26%" was returned;

[0117] The modifier "appropriate" was detected, triggering rule node 5 and returning the quantitative result: "Maximum subsidy amount ≤ 420,000 yuan";

[0118] Return an explainable answer:

[0119] Application requirements: R&D investment accounts for no less than 8.26% of total revenue (calculated based on enterprise data from City A from 2018 to 2022).

[0120] Maximum subsidy: 420,000 yuan (determined based on the 80th percentile of historical subsidy amounts in City A).

[0121] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0122] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A knowledge graph construction method for semantic understanding of policy texts, characterized in that, The construction method specifically includes the following steps: Step S1: Policy text collection and preprocessing: Collect policy text and preprocess it, extract the structured text and filter out sentences containing fuzzy keywords; Step S2, Policy Entity and Relationship Extraction: Identify policy entities and relationships, extract subjects, operations, and fuzzy modified entities, and form triples; Step S3: Construction of a localized quantitative indicator library: Construct a localized quantitative indicator library, connect to the regional economic database, and extract historical data and distribution of key indicators; Step S4: Dynamic Quantization Rule Generation: Dynamically generate quantization rules, automatically quantizing fuzzy expressions into specific threshold rules based on regional data; Step S5, Knowledge Graph Fusion and Storage: Fusion and storage to the knowledge graph, storing triples and quantized rule nodes together in the graph database; Step S6, Semantic Query and Reasoning Service: Provides semantic query and reasoning service, automatically triggers quantification rules and returns interpretable specific results when responding to queries.

2. The knowledge graph construction method for policy text semantic understanding according to claim 1, characterized in that, Step S1, policy text collection and preprocessing, specifically includes the following steps: Step S11: Collect the original text of industry support policies from the public databases of provincial and municipal governments, and use PDFMiner to parse them to extract the plain text content. At the same time, pay attention to preserving the original paragraph divisions and heading hierarchy structure to lay the foundation for subsequent structured analysis. Step S12: The extracted text is segmented into sentences and the SPACY tool is used for dependency parsing to identify sentences containing fuzzy expressions. The selection of sentences depends on a dynamically expanded keyword library. Step S13: Store all the selected sentences into a candidate fuzzy sentence set. In the structure, the set will serve as the direct input for subsequent deep semantic processing, where each This represents an independent candidate fuzzy sentence.

3. The knowledge graph construction method for policy text semantic understanding according to claim 1, characterized in that, Step S2, the extraction of policy entities and relationships, specifically includes the following steps: Step S21: Use the BiLSTM-CRF model to perform named entity recognition on the policy text and extract the following three types of key entities: Main entities: Enterprise type and project type; Operational entities: subsidy form and approval conditions; Fuzzy modifiers for entities: fuzzy terms and quantified dimensions; Step S22: Based on the Policy-BERT language model pre-trained on the policy corpus, analyze the semantic relationships and contextual dependencies between the identified entities, automatically extract the existing relationships, and form a structured set of triples.

4. The knowledge graph construction method for policy text semantic understanding according to claim 1, characterized in that, Step S3, the construction of the localized quantitative indicator library, specifically includes the following steps: Step S31: Access the economic database of the target region. The database includes, but is not limited to, the annual reports of enterprises published by the statistics bureau and the records of policy subsidies retained by the government over the years. Extract key quantitative indicators from unstructured or semi-structured data, specifically including the distribution of the annual output growth rate of enterprises and the fluctuation range of the proportion of R&D investment. At the same time, extract key threshold data including the range of historical subsidy amounts and the percentage of enterprises that have received subsidies. Step S32: Assign corresponding regional coding attributes to the extracted data indicators, persistently store the structured indicator data with regional labels, and build a standardized localized quantitative indicator library.

5. The knowledge graph construction method for policy text semantic understanding according to claim 1, characterized in that, Step S4, the dynamic quantization rule generation, specifically includes the following steps: Step S41: For each sentence containing ambiguous expressions selected after preprocessing, perform deep semantic analysis to identify the specific quantization dimension pointed to by the ambiguous modifiers, and denote this dimension as d; Step S42: Based on the issuing region code of the current policy text, retrieve historical data statistics matching dimension d in real time from the constructed localized quantitative indicator library; when d is a continuous dimension, call the formula. Perform the calculation, where, This represents the average value of historical data for dimension d in this region. Its standard deviation, To obtain the adjustment coefficient by fitting the distribution of historical subsidy cases, the specific threshold value is finally calculated; when When the dimension is proportional, the quantiles in the historical records are directly extracted as the baseline value; Step S43: Transform the calculated or extracted numerical results into structured, machine-readable quantization rules, and create independent quantization rule nodes in the knowledge graph. Establish a clear "quantized as" relationship between the nodes and the fuzzy modified entities extracted in step S2, thereby completing the mapping from fuzzy semantics to precise rules.

6. The knowledge graph construction method for policy text semantic understanding according to claim 1, characterized in that, Step S5, knowledge graph fusion and storage, specifically includes the following steps: Step S51: Extract the policy entity relationship triplet set from step S2. With the dynamically generated quantization rule nodes in step S4 The nodes are integrated to form an extended knowledge graph. The graph is stored using the Neo4j graph database, which defines various node types, including policy clauses, entities, and quantitative rule nodes, and also defines various relationship types. Step S52: Attach the necessary metadata attributes to each quantification rule node, including its applicable region code, the data source on which the rule was generated, and the effective time of the rule, so as to ensure the traceability and timeliness of knowledge in the graph.

7. The knowledge graph construction method for policy text semantic understanding according to claim 1, characterized in that, Step S6, the semantic query and reasoning service specifically includes the following steps: Step S61: After receiving the query request input by the user in natural language, the system parses the query statement, identifies key entities and intents, matches them in the graph database, and locates the policy clause node containing the entity. Step S62: When the matched policy clause contains a vague expression, the inference engine will automatically traverse along the "quantification relationship" edge in the knowledge graph, trigger the associated dynamic quantification rule node, and return the specific numerical threshold stored in the rule as the final condition to the user. Step S63: When the system returns the answer to the user, it will also provide a complete and traceable explanation of the reasoning path, clearly indicating which policy clause and what vague expression the final specific threshold is derived from, and indicating the local data source and regional code and other metadata on which the quantification rule is based.

Images