A multi-modal collection and evolution graph analysis method and system of a policy text

Abstract

The application belongs to the technical field of natural language processing and knowledge graph, and discloses a kind of multi-modal acquisition and evolution graph analysis method and system of policy text. In view of the passive decision problem caused by the rigidity of existing policy analysis, semantic understanding and lack of dynamic correlation, the application adopts visual and DOM feature weighted fusion and reinforcement learning agent to realize multi-modal adaptive acquisition and joint alignment in anti-crawling environment;Use large language model for semantic deconstruction, and convert implicit logic into computer-readable explicit logic expression through thought chain reasoning;Construct a policy evolution knowledge graph with time and validity level attributes, perform satisfiability analysis to output conflict detection results;Map the target object portrait to the graph and the logic expression to evaluate item by item, and output intelligent decision deduction results combined with quantitative dimensions. The application realizes the whole life cycle intelligent analysis of policy data acquisition, deep logic analysis and active deduction.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and knowledge graph technology, specifically to a method and system for multimodal acquisition and evolutionary graph analysis of policy texts. Background Technology

[0002] With the advancement of digital government construction, government departments at all levels have issued a large number of policy documents. For enterprises, research institutions, and government departments themselves, timely and accurate access to and understanding of policy content, applicable conditions, and supporting requirements is crucial. However, while existing policy information acquisition and analysis technologies have improved information detection efficiency to some extent, they still have many limitations in practical applications: (1) Traditional web crawlers rely heavily on fixed HTML DOM structures (such as XPath and CSS selectors) for data extraction. When government websites undergo page redesign, adopt dynamic loading and rendering, or introduce anti-crawling mechanisms such as behavioral CAPTCHAs and IP frequency restrictions, traditional crawlers are prone to failure and require frequent manual maintenance. At the same time, existing crawlers often have difficulty effectively identifying and parsing key information in unstructured attachments such as scanned PDFs and images, resulting in incomplete data acquisition.

[0003] (2) Existing systems often only grasp the semantics of policy texts at a superficial level, making it difficult to extract complex logical constraints. Many policy analysis systems primarily employ keyword matching, topic classification, or simple rule extraction, failing to adequately support deeper logical elements implicit in policy clauses, such as "mutually exclusive conditions for applicable subjects," "pre-approval requirements," "quantitative indicator calculation rules," and "clause citation relationships." While existing structured mining-based methods can achieve topic clustering, they lack the computational modeling capability for complex logical constraints, making it difficult to meet the needs of automated analysis.

[0004] (3) Existing technologies generally lack a dynamic evolution perspective, and the relationships between policy documents are relatively weak. Policy documents are mostly stored as independent documents, making it difficult to identify the "revision, repeal, refinement, and supporting" relationships across departments, levels, and time periods in a timely manner. Existing knowledge graphs tend to be statically organized and displayed, lacking the ability to uniformly model the effective time, level of effectiveness, and cross-version evolution relationships of policy provisions, and have failed to form an automated conflict detection mechanism for policy provisions from different sources.

[0005] (4) Most existing decision support systems are in a passive response state, mainly relying on user input of keywords for retrieval queries. They lack the ability to automatically push potential applicable policies in combination with user profiles, and cannot combine macro indicators to intelligently predict future policy trends.

[0006] Therefore, there is an urgent need for a technical solution that remains highly adaptable in complex anti-crawling environments, can perform deep semantic deconstruction of policy texts, and supports policy lifecycle evolution analysis and proactive decision-making simulation. Summary of the Invention

[0007] To address the aforementioned problems, this invention presents a method and system for multimodal acquisition and evolutionary graph analysis of policy texts, the specific technical solution of which is as follows: A method for multimodal data collection and evolutionary mapping analysis of policy texts includes the following steps: Obtain the page status information of the target policy website, perform weighted fusion positioning of the visual recognition results of the page status information and DOM structure features to determine the target content area, and dynamically execute page interaction actions based on reinforcement learning interaction agent to trigger page loading and anti-crawling verification, thereby obtaining the policy text and multimodal attachments; The structured information of the multimodal attachments is extracted to generate structured attachment data. The semantic vector similarity between the structured attachment data and paragraphs of the policy text is calculated. The structured attachment data and the policy text are then jointly aligned using existing positional anchors to generate a unified policy corpus. The policy corpus is input into a large language model for fine-tuning in the policy domain for semantic deconstruction, extracting multi-dimensional policy feature objects containing subject, action, quantification, and logic dimensions. For implicit logical expressions in the policy corpus, a thought chain reasoning method is used to generate computer-readable explicit logical expressions, which are then stored in the logical dimension of the multi-dimensional policy feature objects. Based on these multi-dimensional policy feature objects, a policy evolution knowledge graph is constructed, using policy documents, policy clauses, issuing departments, and time points as nodes, and evolutionary relationships including revision, replacement, refinement, matching, basis, and conflict as edges. This graph possesses an effective / invalidation time attribute and a validity level attribute determined by the issuing department. In the policy evolution knowledge graph, based on the specified subject, matter and time window as the constraint boundary, the set of related policy clause nodes is traversed and matched, and the satisfiability analysis and consistency verification of the quantitative constraints are performed on the logical expression corresponding to the set of related policy clause nodes, and the conflict detection results between policy clauses are output. The system receives target object profile information from external input and vectorizes it. It obtains a set of candidate policies by performing similarity retrieval in the policy evolution knowledge graph. It maps the target object profile information to the logical dimension variables of the multi-dimensional policy feature objects. It outputs the applicability judgment result by evaluating the logical expressions in the candidate policy set item by item, and calculates the matching degree in combination with the quantitative dimension. Based on the matching degree, it generates intelligent decision inference results for the target object.

[0008] In a preferred implementation, the step of weightedly fusing the visual recognition results of the page state information with DOM structure features to determine the target content area, and dynamically executing page interaction actions based on a reinforcement learning interaction agent, specifically includes: Obtain screenshots and DOM structure information of the target page, identify page blocks through the page visual understanding model, and extract tag hierarchy and link relationship features from the DOM tree; The visual recognition results of the page blocks are weighted and fused with the DOM structure features to determine the precise boundaries of the target content area; A reinforcement learning interaction agent trained based on the PPO algorithm is used. The current page state and historical operation sequence are taken as input, and the corresponding scrolling, clicking, waiting or input action sequence is output to complete the interaction verification. Extract the semantic vector of the target content region and compare it with the semantic vector of the historical version. When the vector similarity is lower than the preset update threshold, it is determined to be a substantial update and the collection task is triggered.

[0009] In a preferred implementation, the step of extracting structured information from the multimodal attachments to generate structured attachment data, calculating semantic vector similarity between the structured attachment data and paragraphs of the policy text, and jointly aligning the structured attachment data with the policy text using existing positional anchors, specifically includes: When the multimodal attachment is an image or scanned document, optical character recognition is performed to extract the text, and a table structure restoration algorithm is used to restore the header relationship and cell merging relationship to generate the structured data of the attachment; The field content in the structured data of the attachments and each paragraph of the policy text are encoded into semantic vectors respectively; When the policy text contains clause numbers and chapter titles, the clause numbers and chapter titles are extracted as location anchors, and the similarity is calculated in combination with the semantic vector to establish a corresponding mapping relationship between the policy text and the structured data in the appendices.

[0010] In a preferred implementation, the policy corpus is input into a large language model fine-tuned for the policy domain for semantic deconstruction, extracting multi-dimensional policy feature objects including subject dimension, action dimension, quantitative dimension, and logical dimension; wherein, for the implicit logical expressions existing in the policy corpus, a thought chain reasoning method is used to generate computer-readable, explicit logical expressions, specifically including: The policy corpus is segmented into paragraphs and input into the large language model to extract the subject dimension information of applicable objects, qualifications, industries and regions, action dimension information of support or restriction, and quantitative dimension information including amount, proportion and time limit. Identify implicit logical trigger words containing preconditions, exclusion conditions, and mutually exclusive conditions in the policy corpus to locate the implicit logical expression; For the implicit logical expression, the natural language long sentence is decomposed and converted into a computer-readable IF-THEN conditional rule format through the thought chain reasoning of the large language model, and the explicit logical expression is generated. If there are different versions of the same policy in the corpus, the corresponding clauses are identified based on the semantic similarity comparison at the clause level, and the differential results containing the addition, deletion or modification identifiers are written into the multidimensional policy feature object. The large language model for fine-tuning in the policy domain is a large language model specifically designed for extracting four-dimensional features of policy texts and generating explicit logical expressions. It is obtained by using a specialized dataset that includes policy text parsing, clause logic decomposition, quantitative indicator extraction, and policy conflict rule learning, and is fine-tuned through domain adaptation.

[0011] In a preferred implementation, the step of constructing a policy evolution knowledge graph based on the multi-dimensional policy feature object, using policy documents, policy clauses, issuing departments, and time points as nodes, and evolutionary relationships including revision, replacement, refinement, supporting, basis, and conflict as edges, with an effective / invalidation time attribute and a validity level attribute determined by the issuing department, specifically includes: In the graph, the policy documents, policy provisions, issuing departments, and time points are instantiated as graph nodes, and applicable industry nodes and applicable region nodes are added. Based on the information extracted from the multidimensional policy feature objects, establish relationship edges between each node, including revision, replacement, refinement, matching, basis, and conflict; Configure corresponding effective timestamps and expiration timestamps for each node, and configure credibility scores for the relationship edges to generate a policy evolution knowledge graph that evolves dynamically over time.

[0012] In a preferred implementation, the step of performing satisfiability analysis and consistency verification of quantitative constraints on the logical expression corresponding to the set of associated policy clause nodes specifically includes: The logical expression corresponding to the set of related policy clause nodes is parsed, a logical dependency tree is constructed, and a constraint solver is used to detect whether there are mutually exclusive applicable conditions between each logical branch; Extract the amount, proportion, and time limit features for the same matter within the quantitative dimensions described in different policy clauses, calculate the intersection of their threshold intervals, and if the intersection is empty, it is determined that there is a numerical inconsistency in the quantitative constraints. Clause pairs that are logically mutually exclusive or numerically inconsistent are marked as potential conflicts, and a conflict detection report is generated that includes the conflict type, the conflict association path in the graph, and the confidence level.

[0013] In a preferred implementation, the step of calculating the matching degree in conjunction with the quantification dimension and generating an intelligent decision-making inference result for the target object based on the matching degree specifically includes: Extract target policies with similarity reaching a preset threshold from the candidate policy set, and substitute the target object profile information into the logical expression and quantitative dimension of the target policy to calculate the expected return value; Historical policy data for specific business areas are extracted from the policy evolution knowledge graph to quantify the regulatory intensity of policies over the years and form a time series of historical policy intensity. The system calls an external database to obtain macroeconomic indicators related to the target object profile information as covariates, uses a time series forecasting model to fit the historical policy intensity time series, predicts the probability of policy adjustment and the direction of evolution in this field within a preset time period, and integrates the results with the applicability judgment to output a trend analysis report.

[0014] A multimodal data acquisition and evolutionary mapping system for policy texts includes: The data acquisition and positioning module is used to acquire page status information of the target policy website, perform weighted fusion positioning of the visual recognition results of the page status information and DOM structure features to determine the target content area, and dynamically execute page interaction actions based on reinforcement learning interaction agent to trigger page loading and anti-crawling verification, thereby obtaining the policy text and multimodal attachments. The multimodal alignment module is used to extract structured information from the multimodal attachments, generate structured attachment data, calculate the semantic vector similarity between the structured attachment data and paragraphs of the policy text, and combine the existing positional anchors to jointly align the structured attachment data with the policy text to generate a unified policy corpus. The semantic deconstruction module is used to input the policy corpus into a large language model for fine-tuning in the policy domain for semantic deconstruction, and extract multi-dimensional policy feature objects containing subject dimension, action dimension, quantitative dimension and logical dimension; wherein, for the implicit logical expressions in the policy corpus, the module uses thought chain reasoning to generate computer-readable explicit logical expressions, and stores the logical expressions in the logical dimension of the multi-dimensional policy feature objects; The graph construction module is used to construct a policy evolution knowledge graph based on the multi-dimensional policy feature objects, with policy documents, policy clauses, issuing departments and time points as nodes, and evolutionary relationships including revision, replacement, refinement, supporting, basis and conflict as edges, which has an effective / invalid time attribute and an effectiveness level attribute determined by the issuing department. The conflict detection module is used to traverse and match the set of related policy clause nodes in the policy evolution knowledge graph based on the specified subject, matter and time window as constraint boundaries, perform satisfiability analysis and consistency verification of the logical expression corresponding to the set of related policy clause nodes, and output the conflict detection results between policy clauses. The decision inference module is used to receive target object profile information from external input and vectorize it, obtain a set of candidate policies by performing similarity retrieval in the policy evolution knowledge graph, map the target object profile information to the logical dimension variables of the multi-dimensional policy feature objects, evaluate the logical expressions in the candidate policy set item by item to output the applicability judgment result, calculate the matching degree in combination with the quantitative dimension, and generate intelligent decision inference results for the target object based on the matching degree.

[0015] An electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when executed by the processor, the computer program implements the steps of the multimodal acquisition and evolutionary mapping analysis method for policy texts as described in any of the preceding claims.

[0016] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the multimodal acquisition and evolutionary graph analysis method for policy texts as described in any of the preceding claims.

[0017] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention adopts the weighted fusion positioning of page block visual recognition and DOM structure features, and dynamically executes interactive actions based on reinforcement learning interactive agent, effectively bypassing dynamic rendering and anti-crawling verification; at the same time, it combines the existing position anchor points to jointly align unstructured attachments with the main text, ensuring the multimodal integrity of policy data.

[0018] (2) In view of the problem that existing systems have difficulty in extracting complex logical constraints, this invention not only uses a fine-tuned large language model to extract multi-dimensional features, but also innovatively uses thought chain reasoning to generate computer-readable, explicit IF-THEN logical expressions for implicit logical expressions in policy texts, successfully transforming the fuzzy boundaries and mutually exclusive conditions in human language into a rule engine that can be directly executed and evaluated by machines.

[0019] (3) To address the issues of fragmented and weakly connected policy documents, this invention constructs a dynamic evolutionary knowledge graph with time and hierarchical attributes. Based on this, by performing satisfiability analysis on the explicit logical expressions of related clauses and verifying the consistency of interval intersections of quantitative constraints, it is possible to accurately identify logical mutual exclusion and numerical conflicts between policy clauses across departments and levels, thereby significantly improving the rigor of policy analysis.

[0020] (4) In view of the shortcomings of the existing system's passive and lagging decision support, the present invention vectorizes the target object profile information and maps it to the map and logical dimension variables. By evaluating the logical expression item by item, it can not only automatically output the applicability judgment result with high accuracy, but also combine historical data sequence to conduct macro trend analysis, providing the target object with full-link, interpretable intelligent decision inference service. Attached Figure Description

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] Figure 1 A flowchart illustrating a method for multimodal acquisition and evolutionary graph analysis of policy texts provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the data processing flow of the anthropomorphic adaptive acquisition engine provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure for policy multidimensional semantic feature extraction and logical reasoning provided in an embodiment of the present invention; Figure 4 A logical schematic diagram of the policy evolution knowledge graph topology and conflict detection provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the data flow for intelligent decision-making inference provided in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described below through specific embodiments, but this is not a limitation of the present invention. Those skilled in the art can make various modifications or improvements based on the basic idea of ​​the present invention, but as long as they do not depart from the basic idea of ​​the present invention, they are all within the protection scope of the present invention.

[0024] See Figure 1 This invention provides a method for multimodal data acquisition and evolutionary graph analysis of policy texts. This method can be executed by computer equipment, a cloud server, or a corresponding execution engine. The method specifically includes the following steps: Obtain the page status information of the target policy website, perform weighted fusion of the visual recognition results of the page status information and DOM structure features to locate the target content area, and dynamically execute page interaction actions based on reinforcement learning interaction agent to trigger page loading and anti-crawling verification, and obtain the policy text and multimodal attachments; In a preferred implementation, the specific methods include: Obtain screenshots and DOM structure information of the target page, identify page blocks through the page visual understanding model, and extract tag hierarchy and link relationship features from the DOM tree; The visual recognition results of page blocks are weighted and fused with DOM structural features to determine the precise boundaries of the target content area; A reinforcement learning interaction agent trained based on the PPO algorithm is used. The current page state and historical operation sequence are taken as input, and the corresponding scrolling, clicking, waiting or input action sequence is output to complete the interaction verification. Extract the semantic vector of the target content region and compare it with the semantic vector of the historical version. When the vector similarity is lower than the preset update threshold, it is determined to be a substantial update and the collection task is triggered.

[0025] For details, please refer to Figure 2 In this embodiment, we take providing policy interpretation services to a new energy engineering design institute as an example. First, a data source set is configured, including international news websites, industry news websites, and policy release websites of national / provincial / municipal competent authorities. The URLs of the category pages of each site are written into the URL queue as seed URLs. The page visual understanding model uses the YOLOv8n network to perform visual object detection on page blocks (list area, body text area, attachment area, and CAPTCHA area). It then combines this with the structure parsing module to extract elements such as title, publication time, source, body text links, and attachment links for weighted fusion positioning.

[0026] To address common dynamic loading and anti-scraping scenarios encountered in new energy policy implementation (such as scrolling loading, pop-up prompts, slider / click verification codes, etc.), a PPO (Proximal Policy Optimization) reinforcement learning algorithm is employed. Its action space is defined as: {scroll, click, input, wait, back, refresh}; the state space includes: page screenshot features, DOM tree summary features, and historical action sequences. Its core reward function is set as follows: successful extraction of the main text block +2, successful acquisition of attachment links +1, successful verification code verification and entry into the main text page +2, action timeout or getting stuck in a repetitive loop -1, being blocked or redirected to an abnormal page -2. The successful collection criteria for this interactive agent are: the page contains a predefined main text DOM pattern and the number of characters in the main text exceeds a set threshold (e.g., greater than 500 characters), and at least two of the following are detected: publication time, source, and attachment links. Through this mechanism, the agent can reliably bypass anti-scraping mechanisms and download policy text and attachments (such as PDFs, Word documents, images, etc.).

[0027] Structured information is extracted from multimodal attachments to generate structured attachment data. Semantic vector similarity is calculated between the structured attachment data and paragraphs of the policy text. The structured attachment data and the policy text are then jointly aligned using existing positional anchors to generate a unified policy corpus. In a preferred implementation, the specific methods include: When the multimodal attachment is an image or scanned document, optical character recognition is performed to extract the text, and a table structure restoration algorithm is used to restore the header relationship and cell merging relationship to generate structured attachment data; Encode the field content in the structured data of the attachments and each paragraph of the policy text into semantic vectors respectively; When the policy text contains clause numbers and chapter titles, the clause numbers and chapter titles are extracted as location anchors. Similarity is calculated by combining semantic vectors to establish a corresponding mapping relationship between the policy text and the structured data in the appendices.

[0028] Specifically, in this embodiment, after downloading the attachment, OCR recognition and table structure restoration are performed on the scanned document or image to output structured fields. In the text redirection and joint alignment stage, the system uses the bge-large-zh-v1.5 model to encode the text into a 1024-dimensional semantic vector and uses cosine similarity as the metric. The specific logic of joint alignment is as follows: first, coarse alignment is performed based on clause numbers or chapter titles, followed by fine alignment based on semantic similarity. When the cosine similarity... When it is, it is directly determined as strong alignment; when When this occurs, it is included in the candidate set, and a secondary judgment is performed based on the anchor point score. The comprehensive scoring formula is set as follows: When the overall score The time frame is used as the final matching threshold to determine successful alignment. This method significantly improves the machine-readable accuracy of tables such as subsidy standards and energy storage ratios in policy appendices. It solves the problem that charts such as subsidy standards and energy storage ratios in previous new energy policy appendices were difficult to read effectively by machines.

[0029] The policy corpus is input into a large language model fine-tuned for the policy domain for semantic deconstruction, extracting multi-dimensional policy feature objects containing subject, action, quantification, and logic dimensions. For implicit logical expressions in the policy corpus, a thought chain reasoning method is used to generate computer-readable explicit logical expressions, which are then stored in the logical dimensions of the multi-dimensional policy feature objects. Based on these multi-dimensional policy feature objects, a policy evolution knowledge graph is constructed, using policy documents, policy clauses, issuing departments, and time points as nodes, and evolutionary relationships including revision, replacement, refinement, supporting elements, basis, and conflict as edges. This graph possesses both effective / expiration time attributes and a validity level attribute determined by the issuing department. In a preferred implementation, the specific methods include: After segmenting the policy corpus into paragraphs, it was input into a large language model to extract the subject dimension information of applicable objects, qualifications, industries and regions, action dimension information of support or restriction, and quantitative dimension information including amount, proportion and time limit. Identify implicit logical trigger words in policy corpora containing preconditions, exclusion conditions, and mutually exclusive conditions in order to locate implicit logical expressions; For implicit logical expressions, the natural language long sentences are decomposed and converted into computer-readable IF-THEN conditional rule format through the thought chain reasoning of the large language model, generating explicit logical expressions. If there are different versions of the same policy in the corpus, the corresponding clauses are identified based on the semantic similarity comparison at the clause level, and the differential results containing the addition, deletion or modification identifiers are written into the multidimensional policy feature object. The large language model for fine-tuning in the policy domain is a large language model specifically designed for extracting four-dimensional features of policy texts and generating explicit logical expressions. It is obtained by using a specialized dataset that includes policy text parsing, clause logic decomposition, quantitative indicator extraction, and policy conflict rule learning, and is fine-tuned through domain adaptation.

[0030] For details, please refer to Figure 3In this embodiment, the base model of the large language model adopts the Qwen-7B network. To achieve accurate deconstruction of policy features, the system constructed a special dataset of approximately 52,000 policy-related data points for supervised fine-tuning. The sample format covers "original policy text -> structured four-dimensional features + logical rule expressions," encompassing policies, notices, implementation details, and appendix tables at all levels from the central government to provincial and municipal governments. The fine-tuning method employs LoRA fine-tuning technology, with a learning rate set to 2e-5 and LoRArank set to 16.

[0031] To clearly illustrate the computer-readable structure of the multidimensional policy characteristic object (policy DNA), consider the following real policy text as an example: "Applicant enterprises must be legally registered and have been operating continuously in Hunan Province for more than one year. Those that are high-tech enterprises will receive a subsidy of 20% of the project investment, with a maximum of 500,000 yuan per project." The structured data in JSON format output by the fine-tuned large language model is shown below: { "policy_id": "HN-2025-NE-001", "clause_id": "C3.2", The applicant company must be legally registered and have been operating continuously in Hunan Province for more than one year. For companies classified as high-tech enterprises, a subsidy of 20% of the project investment will be granted, with a maximum of 500,000 yuan per project. "policy_dna": { "subject_dimension": { "region": ["Hunan Province"], "entity_type": ["Enterprise"], "qualification": ["High-tech Enterprise"], "operation_years_min": 1 }, "action_dimension": { "action_type": "subsidy", "action_target": "Project investment amount", "calculation_rule": "Subsidize according to the proportion of investment amount and set a cap." }, "quantitative_dimension": { "subsidy_ratio": 0.20, "subsidy_cap_cny": 500000, "currency": "CNY" }, "logic_dimension": { "if_then_rules": [ { "rule_id": "R-C3.2-01", "if": [ "registered_region == 'Hunan Province'", "operation_years>= 1", The qualification includes 'High-tech Enterprise'. ], "then": [ "eligible = true", "subsidy = min(project_investment * 0.20, 500000)" ], "else": [ "eligible = false" ] } ], "logic_expr": "(region=HN AND years>=1 AND is_hightech=1) =>subsidy=min(invest*0.2,500000)" } } } The above parsing process not only completes the structuring of the four dimensions, but also transforms the implicit conditions into an if_then_rules list and a logic_expr explicit logical expression through thought chain reasoning.

[0032] The multidimensional semantic deconstruction center utilizes a large language model to perform clause-level deconstruction of the aforementioned policy corpus. When extracting logical dimensions, for common policy expressions with strong implicit logic such as "except...", "in principle", and "implemented with reference to...", the model uses thought chain reasoning to generate machine-readable IF-THEN conditional rules for subsequent automatic verification. Furthermore, for news articles, in addition to extracting structured key points, the large language model also generates interpretation points specifically for new energy engineering design institutes (e.g., clearly indicating the policy's impact on specific business aspects such as project bidding, grid connection, subsidies, carbon trading, and energy storage infrastructure), and retains traceable source links.

[0033] In a preferred implementation, the specific methods include: In the graph, policy documents, policy provisions, issuing departments, and time points are instantiated as graph nodes, and applicable industry nodes and applicable region nodes are added for instantiation; Based on the information extracted from the multidimensional policy feature objects, establish relationship edges between each node, including revision, replacement, refinement, matching, basis, and conflict; Configure corresponding effective and ineffective timestamps for each node, and configure credibility scores for relationship edges to generate a policy evolution knowledge graph that evolves dynamically over time.

[0034] For details, please refer to Figure 4 In this embodiment, for multiple releases / revisions of the same policy topic (e.g., "grid connection management of new energy projects," "energy storage configuration requirements," "subsidies / tax incentives," etc.) or the same policy document, the system performs semantic differential monitoring between newly collected versions and historical versions. Based on clause-level alignment, it outputs the added, deleted, and modified content and writes it back into the policy DNA. Subsequently, these entities are added to the policy evolution knowledge graph to construct a policy evolution tree, recording the effective time, expiration time, and effectiveness level to support full lifecycle evolution tracking.

[0035] In the policy evolution knowledge graph, based on the specified subject, matter and time window as the constraint boundary, the set of related policy clause nodes is traversed and matched. Satisfaction analysis and consistency verification of quantitative constraints are performed on the logical expressions corresponding to the set of related policy clause nodes, and the conflict detection results between policy clauses are output. In a preferred implementation, the specific methods include: The logical expressions corresponding to the set of related policy clause nodes are parsed, a logical dependency tree is constructed, and a constraint solver is used to detect whether there are mutually exclusive applicable conditions between each logical branch. Extract the amount, proportion, and time limit characteristics of the same matter within the quantitative dimension of different policy clauses, calculate the intersection of their threshold intervals, and if the intersection is empty, it is determined that there is numerical inconsistency in the quantitative constraints. Clause pairs that are logically mutually exclusive or numerically inconsistent are marked as potential conflicts, and a conflict detection report is generated that includes the conflict type, the conflict association path in the graph, and the confidence level.

[0036] Specifically, in combination Figure 4 In this embodiment, the conflict detection mechanism adopts a joint architecture of rule engine and constraint solver. At the rule layer, the system compiles the pre-extracted IF-THEN logical expression into an executable predicate; at the solution layer, it calls the Z3 SMT Solver (constraint solver) to perform a rigorous satisfiability analysis.

[0037] The judgment criteria are as follows: Under the same subject, the same matter, and the same time window, if two rules cannot be satisfied (UNSAT) after being jointly input into the constraint solver, they are judged as logically mutually exclusive hard conflicts. For the quantitative dimension, amounts, proportions, etc., are normalized into numerical ranges and their intersections are calculated. For example, policy A requires a reserve allocation ratio of ≥10% (range [10%, 100%]), and policy B requires a reserve allocation ratio of ≤8% (range [0%, 8%]); if the intersection of the two ranges is empty, the system directly judges it as a numerical conflict (hard conflict); if the intersection is not empty but the range is significantly narrowed (e.g., becomes [12%, 15%]), the system judges it as a rule tightening and marks it as a "potential conflict," thereby avoiding failure of engineering design or enterprise application due to policy contradictions.

[0038] The system receives target object profile information from external input and vectorizes it. It obtains a set of candidate policies by performing similarity retrieval in the policy evolution knowledge graph. It maps the target object profile information to the logical dimension variables of multi-dimensional policy feature objects. It outputs the applicability judgment result by evaluating the logical expressions in the candidate policy set item by item, and calculates the matching degree in combination with the quantitative dimension. Based on the matching degree, it generates intelligent decision inference results for the target object.

[0039] In a preferred implementation, the specific methods include: Extract target policies from the candidate policy set whose similarity reaches a preset threshold, and substitute the target object profile information into the logical expression and quantitative dimension of the target policy to calculate the expected return value; Extract historical policy data from specific business areas in the policy evolution knowledge graph, quantify the regulatory intensity of policies over the years, and form a time series of historical policy intensity. The system calls upon external databases to obtain macroeconomic indicators related to the target object's profile information as covariates, uses a time series forecasting model to fit the historical policy intensity time series, predicts the probability and evolution direction of policy adjustments in this field within a preset time period, and integrates the results with the applicability judgment to output a trend analysis report.

[0040] For details, please refer to Figure 5In this embodiment, the system provides an interactive interface (such as a web interface). After the user inputs user profile information, including the company's industry code, region, size, qualifications, and historical application status, the system vectorizes this information and automatically maps it to the subject_dimension (e.g., region, qualification) of the aforementioned policy JSON object for logical evaluation. If logic_expr is determined to be True (applicable), the system will automatically calculate and display expected benefits such as "maximum subsidy of 500,000 for a single project" and the application success rate based on quantitative_dimension. Furthermore, based on historical policy strength and macroeconomic indicators, the system plots a policy trend prediction curve in the form of a time-strength line graph, providing companies with an intuitive intelligent decision-making simulation report for forward-looking project bidding and planning. In addition, users can annotate the interpretation results (e.g., "useful," "requires manual review") for subsequent model and rule iteration optimization.

[0041] Based on the same inventive concept, this invention also provides a multimodal acquisition and evolutionary graph analysis system for policy texts, including: The data acquisition and positioning module is used to obtain the page status information of the target policy website, perform weighted fusion positioning of the visual recognition results of the page status information and DOM structure features to determine the target content area, and dynamically execute page interaction actions based on reinforcement learning interaction agent to trigger page loading and anti-crawling verification, and obtain the policy text and multimodal attachments. The multimodal alignment module is used to extract structured information from multimodal attachments, generate structured attachment data, calculate semantic vector similarity between the structured attachment data and paragraphs of the policy text, and combine the existing positional anchors to jointly align the structured attachment data with the policy text to generate a unified policy corpus. The semantic deconstruction module is used to input policy corpora into a large language model that is fine-tuned for the policy domain for semantic deconstruction, extracting multi-dimensional policy feature objects containing subject dimension, action dimension, quantitative dimension and logical dimension; among them, for the implicit logical expressions in the policy corpora, the module uses thought chain reasoning to generate computer-readable explicit logical expressions, and stores the logical expressions in the logical dimension of the multi-dimensional policy feature objects; The graph construction module is used to construct a policy evolution knowledge graph based on multi-dimensional policy feature objects, with policy documents, policy clauses, issuing departments and time points as nodes, and evolutionary relationships including revision, replacement, refinement, supporting, basis and conflict as edges, which has the attributes of effective / invalid time and the level of effectiveness determined by the issuing department. The conflict detection module is used to traverse and match the set of related policy clause nodes in the policy evolution knowledge graph based on the specified subject, matter and time window as the constraint boundary, perform satisfiability analysis and consistency verification of the logical expression corresponding to the set of related policy clause nodes, and output the conflict detection results between policy clauses. The decision inference module receives target object profile information from external input and vectorizes it. It obtains a set of candidate policies by performing similarity retrieval in the policy evolution knowledge graph, maps the target object profile information to the logical dimension variables of multi-dimensional policy feature objects, evaluates the logical expressions in the candidate policy set item by item to output the applicability judgment result, and calculates the matching degree in combination with the quantitative dimension. Based on the matching degree, it generates intelligent decision inference results for the target object.

[0042] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the steps of the multimodal acquisition and evolutionary graph analysis method for policy texts as described above.

[0043] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the multimodal acquisition and evolutionary graph analysis method for policy texts as described above.

[0044] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for multimodal acquisition and evolutionary graph analysis of policy texts, characterized in that, Includes the following steps: Obtain the page status information of the target policy website, perform weighted fusion positioning of the visual recognition results of the page status information and DOM structure features to determine the target content area, and dynamically execute page interaction actions based on reinforcement learning interaction agent to trigger page loading and anti-crawling verification, thereby obtaining the policy text and multimodal attachments; The structured information of the multimodal attachments is extracted to generate structured attachment data. The semantic vector similarity between the structured attachment data and paragraphs of the policy text is calculated. The structured attachment data and the policy text are then jointly aligned using existing positional anchors to generate a unified policy corpus. The policy corpus is input into a large language model for fine-tuning in the policy domain for semantic deconstruction, extracting multi-dimensional policy feature objects containing subject, action, quantification, and logic dimensions. For implicit logical expressions in the policy corpus, a thought chain reasoning method is used to generate computer-readable explicit logical expressions, which are then stored in the logical dimensions of the multi-dimensional policy feature objects. Based on these multi-dimensional policy feature objects, a policy evolution knowledge graph is constructed, using policy documents, policy clauses, issuing departments, and time points as nodes, and evolutionary relationships including revision, replacement, refinement, matching, basis, and conflict as edges. This graph possesses effective time attributes, expiration time attributes, and a validity level attribute determined by the issuing department. In the policy evolution knowledge graph, based on the specified subject, matter and time window as the constraint boundary, the set of related policy clause nodes is traversed and matched, and the satisfiability analysis and consistency verification of the quantitative constraints are performed on the logical expression corresponding to the set of related policy clause nodes, and the conflict detection results between policy clauses are output. The system receives target object profile information from external input and vectorizes it. It obtains a set of candidate policies by performing similarity retrieval in the policy evolution knowledge graph. It maps the target object profile information to the logical dimension variables of the multi-dimensional policy feature objects. It outputs the applicability judgment result by evaluating the logical expressions in the candidate policy set item by item, and calculates the matching degree in combination with the quantitative dimension. Based on the matching degree, it generates intelligent decision inference results for the target object.

2. The method for multimodal acquisition and evolutionary graph analysis of policy texts according to claim 1, characterized in that, The step of weightedly fusing the visual recognition results of the page state information with DOM structure features to locate the target content area, and dynamically executing page interaction actions based on a reinforcement learning interaction agent, specifically includes: Obtain screenshots and DOM structure information of the target page, identify page blocks through the page visual understanding model, and extract tag hierarchy and link relationship features from the DOM tree; The visual recognition results of the page blocks are weighted and fused with the DOM structure features to determine the precise boundaries of the target content area; A reinforcement learning interaction agent trained based on the PPO algorithm is used. The current page state and historical operation sequence are taken as input, and the corresponding scrolling, clicking, waiting or input action sequence is output to complete the interaction verification. Extract the semantic vector of the target content region and compare it with the semantic vector of the historical version. When the vector similarity is lower than the preset update threshold, it is determined to be a substantial update and the collection task is triggered.

3. The method for multimodal acquisition and evolutionary graph analysis of policy texts according to claim 1, characterized in that, The step of extracting structured information from the multimodal attachments to generate structured attachment data, calculating semantic vector similarity between the structured attachment data and paragraphs of the policy text, and jointly aligning the structured attachment data with the policy text using existing positional anchors, specifically includes: When the multimodal attachment is an image or scanned document, optical character recognition is performed to extract the text, and a table structure restoration algorithm is used to restore the header relationship and cell merging relationship to generate the structured data of the attachment; The field content in the structured data of the attachments and each paragraph of the policy text are encoded into semantic vectors respectively; When the policy text contains clause numbers and chapter titles, the clause numbers and chapter titles are extracted as location anchors, and the similarity is calculated in combination with the semantic vector to establish a corresponding mapping relationship between the policy text and the structured data in the appendices.

4. The method for multimodal acquisition and evolutionary graph analysis of policy texts according to claim 1, characterized in that, The policy corpus is input into a large language model fine-tuned for the policy domain for semantic deconstruction, extracting multi-dimensional policy feature objects including subject, action, quantification, and logic dimensions; wherein, for implicit logical expressions in the policy corpus, a thought chain reasoning method is used to generate computer-readable explicit logical expressions, specifically including: The policy corpus is segmented into paragraphs and input into the large language model to extract the subject dimension information of applicable objects, qualifications, industries and regions, action dimension information of support or restriction, and quantitative dimension information including amount, proportion and time limit. Identify implicit logical trigger words containing preconditions, exclusion conditions, and mutually exclusive conditions in the policy corpus to locate the implicit logical expression; For the implicit logical expression, the natural language long sentence is decomposed and converted into a computer-readable IF-THEN conditional rule format through the thought chain reasoning of the large language model, and the explicit logical expression is generated. If there are different versions of the same policy in the corpus, the corresponding clauses are identified based on the semantic similarity comparison at the clause level, and the differential results containing the addition, deletion or modification identifiers are written into the multidimensional policy feature object.

5. The method for multimodal acquisition and evolutionary graph analysis of policy texts according to claim 1, characterized in that, Based on the multidimensional policy feature objects, a policy evolution knowledge graph is constructed, using policy documents, policy clauses, issuing departments, and time points as nodes, and evolutionary relationships including revision, replacement, refinement, supporting, basis, and conflict as edges. This graph possesses effective time attributes, expiration time attributes, and a validity level attribute determined by the issuing department. Specifically, this includes: In the graph, the policy documents, policy provisions, issuing departments, and time points are instantiated as graph nodes, and applicable industry nodes and applicable region nodes are added. Based on the information extracted from the multidimensional policy feature objects, establish relationship edges between each node, including revision, replacement, refinement, matching, basis, and conflict; Configure corresponding effective timestamps and expiration timestamps for each node, and configure credibility scores for the relationship edges to generate a policy evolution knowledge graph that evolves dynamically over time.

6. The method for multimodal acquisition and evolutionary graph analysis of policy texts according to claim 1, characterized in that, The process of performing satisfiability analysis and consistency verification of quantitative constraints on the logical expressions corresponding to the set of associated policy clause nodes specifically includes: The logical expression corresponding to the set of related policy clause nodes is parsed, a logical dependency tree is constructed, and a constraint solver is used to detect whether there are mutually exclusive applicable conditions between each logical branch; Extract the amount, proportion, and time limit features for the same matter within the quantitative dimensions described in different policy clauses, calculate the intersection of their threshold intervals, and if the intersection is empty, it is determined that there is a numerical inconsistency in the quantitative constraints. Clause pairs that are logically mutually exclusive or numerically inconsistent are marked as potential conflicts, and a conflict detection report is generated that includes the conflict type, the conflict association path in the graph, and the confidence level.

7. The method for multimodal acquisition and evolutionary graph analysis of policy texts according to claim 1, characterized in that, The step of calculating the matching degree by combining the quantification dimensions and generating intelligent decision-making inference results for the target object based on the matching degree specifically includes: Extract target policies with similarity reaching a preset threshold from the candidate policy set, and substitute the target object profile information into the logical expression and quantitative dimension of the target policy to calculate the expected return value; Historical policy data for specific business areas are extracted from the policy evolution knowledge graph to quantify the regulatory intensity of policies over the years and form a time series of historical policy intensity. The system calls an external database to obtain macroeconomic indicators related to the target object profile information as covariates, uses a time series forecasting model to fit the historical policy intensity time series, predicts the probability of policy adjustment and the direction of evolution in this field within a preset time period, and integrates the results with the applicability judgment to output a trend analysis report.

8. A multimodal acquisition and evolutionary graph analysis system for policy texts, characterized in that, include: The data acquisition and positioning module is used to acquire page status information of the target policy website, perform weighted fusion positioning of the visual recognition results of the page status information and DOM structure features to determine the target content area, and dynamically execute page interaction actions based on reinforcement learning interaction agent to trigger page loading and anti-crawling verification, thereby obtaining the policy text and multimodal attachments. The multimodal alignment module is used to extract structured information from the multimodal attachments, generate structured attachment data, calculate the semantic vector similarity between the structured attachment data and paragraphs of the policy text, and combine the existing positional anchors to jointly align the structured attachment data with the policy text to generate a unified policy corpus. The semantic deconstruction module is used to input the policy corpus into a large language model for fine-tuning in the policy domain for semantic deconstruction, and extract multi-dimensional policy feature objects containing subject dimension, action dimension, quantitative dimension and logical dimension; wherein, for the implicit logical expressions in the policy corpus, the module uses thought chain reasoning to generate computer-readable explicit logical expressions, and stores the logical expressions in the logical dimension of the multi-dimensional policy feature objects; The graph construction module is used to construct a policy evolution knowledge graph based on the multi-dimensional policy feature objects, with policy documents, policy clauses, issuing departments and time points as nodes, and evolutionary relationships including revision, replacement, refinement, supporting, basis and conflict as edges, which has effective time attributes, expiration time attributes and effectiveness level attributes determined by the issuing department. The conflict detection module is used to traverse and match the set of related policy clause nodes in the policy evolution knowledge graph based on the specified subject, matter and time window as constraint boundaries, perform satisfiability analysis and consistency verification of the logical expression corresponding to the set of related policy clause nodes, and output the conflict detection results between policy clauses. The decision inference module is used to receive target object profile information from external input and vectorize it, obtain a set of candidate policies by performing similarity retrieval in the policy evolution knowledge graph, map the target object profile information to the logical dimension variables of the multi-dimensional policy feature objects, evaluate the logical expressions in the candidate policy set item by item to output the applicability judgment result, calculate the matching degree in combination with the quantitative dimension, and generate intelligent decision inference results for the target object based on the matching degree.

9. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for multimodal acquisition and evolutionary mapping analysis of policy texts as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the multimodal acquisition and evolutionary graph analysis method for policy texts as described in any one of claims 1 to 7.

Images