NLP-based energy policy information intelligent collection and classification arrangement method, system and device, and medium

Abstract

The invention discloses an NLP-based energy policy information intelligent collection and classification arrangement method, system and device and a medium. The method comprises the steps of collecting energy policy data of multiple information sources and performing first processing; performing semantic segmentation and structured recombination on the data based on a first adaptive segmentation algorithm, and identifying a starting point and a hierarchical relationship of policy terms; extracting first key information from the structured policy text; performing subject classification and multi-label assignment on the policy text, and classifying the policy into a predefined multi-layer label system by using a first fusion algorithm; and storing the classified, extracted and vectorized policy data into a database, constructing a dual-mode storage structure, constructing an intelligent retrieval interface supporting keyword retrieval and semantic retrieval, and providing policy association analysis and content generation services. According to the invention, the collection, arrangement, management and utilization level of energy policy information in the power grid field is improved, and a solid guarantee and innovation path is provided for industry digital transformation and efficient policy response.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system, device, and medium for intelligent collection and classification of energy policy information based on NLP. Background Technology

[0002] In the power grid industry, with the increasing penetration rate of new energy sources and the rapid development of smart grid technology, the policy environment is becoming increasingly complex and volatile. National and local governments frequently release a series of policy documents covering power system reform, energy supply-side structural adjustment, power grid planning and construction, energy storage and distributed energy integration, power market mechanism innovation, new energy consumption, and intelligent operation and maintenance. These documents are not only found on the official websites of government ministries and energy regulatory agencies, but are also widely distributed in industry associations, research institutions, mainstream media, and professional forums, often appearing in diverse forms such as announcements, technical specifications, development plans, industry reports, meeting minutes, and public opinion.

[0003] Currently, power grid companies and industry research institutions primarily rely on manual, periodic searches and fragmented attention to obtain energy policy information. This approach is not only labor-intensive and time-consuming but also prone to omissions, making it difficult to grasp the latest policy dynamics and adjustment trends in a timely manner. The diverse sources, varied forms of expression, and complex document structures of policy information further increase the technical difficulty of policy content analysis and thematic summarization. Especially under the backdrop of "dual carbon" goals and the energy internet, the accurate collection, semantic understanding, and intelligent classification of massive amounts of policy information have become crucial prerequisites for power grid companies' scientific decision-making and forward-looking deployment of core resource allocation. At the same time, traditional information systems lack the ability to automatically process policy text content and lack intelligent functions such as cross-source data aggregation, semantic tagging, in-depth thematic analysis, and real-time push notifications and alerts, thus failing to support business innovation and efficiently support the needs of digital transformation.

[0004] Therefore, there is an urgent need for a method for intelligent collection and classification of energy policy information based on natural language processing (NLP) in the power grid sector. This method can comprehensively improve the breadth of policy information collection, the depth of processing, and the level of intelligent application through multi-source automatic collection, full-text intelligent parsing, tag-based aggregation, and customized aggregation analysis. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a method, system, device, and medium for intelligent collection and classification of energy policy information based on NLP, which solves the problem that existing technologies are unable to efficiently collect and accurately classify massive, multi-source, and heterogeneous energy policy information in the power grid field in an automated, intelligent, and structured manner.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, this invention provides a method for intelligent collection and classification of energy policy information based on NLP, including: Collect energy policy data from multiple information sources and perform initial processing; Based on the first adaptive segmentation algorithm, the processed data is semantically segmented and structurally reorganized to identify the starting point and hierarchical relationship of policy clauses; Extract the first key information from the structured policy text; Based on the first key information, the policy text is classified into themes and assigned multiple tags. The first fusion algorithm is used to classify the policy into a predefined multi-level tag system. The policy data, after being classified, extracted, and vectorized, is stored in a database to construct a dual-mode storage structure. Based on the aforementioned dual-mode storage structure, an intelligent retrieval interface supporting keyword retrieval and semantic retrieval is constructed, and policy correlation analysis and content generation services are provided.

[0008] As a preferred embodiment of the NLP-based intelligent collection and classification method for energy policy information described in this invention, the step of semantic segmentation and structured reorganization of the first processed data based on a first adaptive segmentation algorithm includes: Construct a primary policy structure knowledge base containing typical policy provisions; Traverse the first processed policy text and calculate the semantic similarity between the policy text and the clause opening template, the semantic similarity between the transition sentence template and the coupling degree of the distribution of upper and lower level policy entities in the first policy structure knowledge base. Based on the weighted score of semantic similarity and coupling degree, the starting point and hierarchical nodes of policy clauses are identified. Combined with the serial number and indentation logic mark in the text, a policy fragment with a hierarchical structure is output.

[0009] As a preferred embodiment of the NLP-based intelligent collection and classification method for energy policy information described in this invention, the step of extracting the first key information from the structured policy text includes: Build structured prompt word templates for each type of key information element to be extracted; The structured policy fragments are combined with the corresponding prompt word templates and input into the first processing model. The first processing model outputs the structured first key information through semantic understanding and contextual reasoning. Based on predefined rules and a domain entity dictionary, the first key information is subjected to format validation and terminology standardization.

[0010] As a preferred embodiment of the NLP-based intelligent collection and classification method for energy policy information described in this invention, the step of classifying policy texts by topic and assigning multiple tags based on the first key information includes: The vectorized representation of the full policy text and the first key information are input into the large language model. The model is then fine-tuned by instructions to generate a preliminary set of topic tags and confidence scores. The similarity between the policy text vector and the topic node vectors in the predefined multi-level tagging system is calculated using a semantic encoding model to obtain the initial content-based weights. The set of topic tags is transformed into a prior augmentation vector, which is then linearly fused with the initial weights. Based on the first threshold mechanism, multi-dimensional policy labels are automatically assigned to several thematic nodes with the highest fusion weight.

[0011] The beneficial effects of this preferred technical solution are: it can automatically classify and tag complex policy texts by topic, improve the scientificity and consistency of policy archiving, and reduce the errors and workload of traditional manual sorting.

[0012] As a preferred embodiment of the NLP-based intelligent collection and classification method for energy policy information described in this invention, the construction of the dual-mode storage structure includes: The policy structured attribute data, after classification and key information extraction, will be stored in a relational database, and a multi-dimensional index will be established. The core policy text content is converted into a high-dimensional vector representation using a semantic embedding model; the high-dimensional vector representation is then associated with the corresponding structured policy data and stored in batches in a vector database. Based on the relational database and vector database, a dual-mode data storage structure with collaborative query capabilities is constructed.

[0013] The beneficial effects of this preferred technical solution are: ensuring high reliability in the process of policy information collection, processing, and dissemination, and improving data security and system operation stability.

[0014] As a preferred embodiment of the NLP-based intelligent collection and classification method for energy policy information described in this invention, the energy policy data collected from multiple information sources includes: Information sources are classified according to interface specifications and content dynamism, and differentiated collection strategies are formulated accordingly; Policy data from government websites and industry platforms with open APIs is collected via API calls; policy content from news and information websites and information disclosure platforms with stable page structures but no open APIs is collected via web crawling. Extract policy text and metadata from offline reports and documents using file parsing methods; Through browser interaction driven by a large language model agent, policy information requiring login, dynamic rendering, or complex interaction can be collected.

[0015] As a preferred embodiment of the NLP-based intelligent collection and classification method for energy policy information described in this invention, the first process includes: The energy policy data will undergo standardized encoding format and character encoding conversion. The text after unified encoding is segmented into words and deduplicated based on semantics. A semantic noise classifier based on a pre-trained language model identifies and filters non-policy text content such as advertisements, navigation menus, and copyright notices. Key terms in policy texts are standardized based on a domain terminology mapping library.

[0016] Secondly, this invention provides an NLP-based intelligent collection and classification system for energy policy information, including: The data processing module is used to collect energy policy data from multiple information sources and perform initial processing. The segmentation and reorganization module is used to perform semantic segmentation and structured reorganization on the first processed data based on the first adaptive segmentation algorithm, and to identify the starting point and hierarchical relationship of policy clauses. The information extraction module is used to extract the first key information from the structured policy text; The classification module is used to classify policy texts by theme and assign multiple labels based on the first key information, and to use the first fusion algorithm to classify the policies into a predefined multi-level label system. The data storage module is used to store the classified, extracted and vectorized policy data into the database, constructing a dual-mode storage structure. The retrieval interface construction module is used to build an intelligent retrieval interface that supports keyword retrieval and semantic retrieval based on the dual-mode storage structure, and to provide policy correlation analysis and content generation services.

[0017] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor executes the computer-executable instructions to implement the steps of an NLP-based intelligent collection and classification method for energy policy information.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of an NLP-based method for intelligent collection and classification of energy policy information.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention utilizes a multi-source automated collection method to efficiently and in real-time collect policy information data from various sources, including government websites, industry platforms, and media information, greatly reducing manual intervention and enabling rapid perception and timely updates of policy dynamics. By leveraging intelligent technologies such as NLP to clean, structure, and standardize policy texts, it effectively solves problems such as diverse policy content formats and complex structures, improving data quality and subsequent analysis capabilities. This invention employs deep semantic analysis and intelligent classification methods to automatically classify and tag complex policy texts by topic, improving the scientific rigor and consistency of policy archiving and reducing errors and workload associated with traditional manual sorting. Furthermore, it supports multi-dimensional policy aggregation display and intelligent retrieval, facilitating users to quickly find and utilize the required policy content from different perspectives, significantly optimizing the user experience. Through end-to-end log tracking and automatic abnormal data completion mechanisms, this invention ensures high reliability in the policy information collection, processing, and push processes, enhancing data security and system stability. It achieves automation and intelligence in large-scale policy data collection and processing, significantly reducing the cost of manual collection, organization, and maintenance, and improving overall work efficiency and system scalability. It provides strong support for the power grid industry to grasp policy changes and development trends in a timely, comprehensive, and structured manner, and promotes the scientific and intelligent development of policy research, risk assessment, and corporate decision-making. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0021] Figure 1 This is a schematic diagram illustrating the overall process logic of an NLP-based intelligent collection and classification method for energy policy information, provided in one embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0023] Example 1, referring to Figure 1 As one embodiment of the present invention, a method for intelligent collection and classification of energy policy information based on NLP is provided, such as... Figure 1The specific steps shown are as follows: S100: Collects energy policy data from multiple information sources and performs initial processing; S200: Based on the first adaptive segmentation algorithm, the processed data is semantically segmented and structurally reorganized to identify the starting point and hierarchical relationship of policy clauses; S300: Extracting the first key information from the structured policy text; S400: Based on the first key information, the policy text is classified into themes and assigned multiple labels. The first fusion algorithm is used to classify the policy into a predefined multi-level label system. S500: Stores the classified, extracted, and vectorized policy data into a database, constructs a dual-mode storage structure, and builds an intelligent retrieval interface that supports keyword retrieval and semantic retrieval based on the dual-mode storage structure, and provides policy correlation analysis and content generation services.

[0024] It should be noted that, to address the challenge of existing technologies in efficiently and accurately collecting and classifying the massive, multi-source, and heterogeneous energy policy information in the power grid sector through automated, intelligent, and structured methods, steps S100-S500 utilize a multi-source automated collection method. This method efficiently and in real-time collects policy information data from various sources, including government websites, industry platforms, and media reports, significantly reducing manual intervention and enabling rapid perception and timely updates of policy dynamics. By leveraging intelligent technologies such as NLP to clean, structure, and standardize policy texts, the invention effectively solves problems related to diverse policy content formats and complex structures, improving data quality and subsequent analysis capabilities. This invention employs deep semantic analysis and intelligent classification methods to automatically classify and tag complex policy texts by topic, enhancing the scientific rigor and consistency of policy archiving and reducing errors and workload associated with traditional manual sorting. Furthermore, it supports multi-dimensional policy aggregation display and intelligent retrieval, facilitating users to quickly find and utilize the required policy content from different perspectives, significantly optimizing the user experience. This invention ensures high reliability in the policy information collection, processing, and dissemination process through end-to-end log tracking and automatic anomaly data completion mechanisms, enhancing data security and system stability. It automates and intelligently processes large-scale policy data collection and processing, significantly reducing the costs of manual collection, organization, and maintenance, and improving overall work efficiency and system scalability. This provides strong support for the power grid industry to grasp policy changes and development trends in a timely, comprehensive, and structured manner, promoting the scientific and intelligent development of policy research, risk assessment, and corporate decision-making.

[0025] In this embodiment of the invention, step S100 includes the following sub-steps A1 and A2: In A1: Energy policy data are collected from multiple information sources; detailed steps include: Information sources are classified according to interface specifications and content dynamism, and differentiated collection strategies are formulated accordingly; Policy data from government websites and industry platforms with open APIs is collected via API calls; policy content from news and information websites and information disclosure platforms with stable page structures but no open APIs is collected via web crawling. Extract policy text and metadata from offline reports and documents using file parsing methods; Through browser interaction driven by a large language model agent, policy information requiring login, dynamic rendering, or complex interaction can be collected.

[0026] Specifically, in the process of multi-source collection of energy policy information, the first step is to thoroughly sort and classify the target information sources, and formulate differentiated collection plans based on their technical interface specifications, content dynamism, and interaction complexity. For sources such as government websites and industry reporting platforms that provide open API interfaces and have standardized data structures, priority is given to efficient and stable data retrieval via APIs, and incremental synchronization is achieved using update flags provided by the interfaces. For news and information websites and information disclosure platforms that do not have open APIs but have relatively stable page structures, customized web crawlers are developed. By parsing the HTML DOM structure and setting reasonable crawling frequency and depth strategies, anti-crawling mechanisms are avoided and collection compliance is ensured. At the same time, for offline reports and documents in formats such as PDF, Word, and Excel, text content and metadata are automatically extracted using a file reading and document parsing library (Apache POI + PDFBox).

[0027] Specifically, to overcome the technical bottlenecks faced by traditional web crawlers and static methods when dealing with websites requiring login, dynamic JavaScript rendering, complex pagination, or form queries, a novel large model agent (LLM Agent) is introduced as a highly adaptive data collection method. After receiving collection tasks from a unified scheduling system, this agent can simulate user behavior and automatically control the browser (Selenium) to access the target page. Its core lies in leveraging the powerful natural language understanding and contextual reasoning capabilities of the large language model to analyze the overall layout, text semantics, and interactive elements of the loaded page in real time. Based on its instant understanding of the page content, the agent autonomously makes decisions and executes a series of operational instructions, such as: identifying and triggering pagination or content loading controls to obtain a complete list; entering relevant policy keywords in the search box and executing a query; handling basic verification mechanisms such as graphical CAPTCHAs; and even simulating the login process to enter restricted policy query areas. Through this closed loop of "perception-decision-execution," the agent can effectively acquire dynamic interactive content that traditional web crawlers cannot collect.

[0028] It should be noted that, to ensure the timeliness and completeness of the collected data, a unified scheduled task scheduling and event-driven monitoring module was configured in the collection system. Regular information sources are periodically polled and crawled; for frequently updated news and social media content, incremental monitoring is implemented using RSS subscriptions, event stream monitoring, or keyword alerting mechanisms. All collection tasks, whether traditional or agent-driven, are subject to end-to-end monitoring. The system records detailed collection logs and implements anomaly detection and automatic retry strategies to ensure the stability of the collection process and the completion of missing data. Finally, all collected raw data is uniformly appended with timestamps, source identifiers, collection methods (such as API / crawler / agent), and batch information, providing a solid foundation for subsequent data cleaning, tracing, and fusion. The entire multi-source collection process, through modularization, intelligence, and adaptive methods, achieves significant improvements in the breadth, depth, and real-time synchronization of energy policy information.

[0029] In A2: the energy policy data collected from multiple information sources undergoes initial processing; detailed steps include: Standardize the encoding format and convert character encodings for energy policy data; The text after unified encoding is segmented into words and deduplicated based on semantics. A semantic noise classifier based on a pre-trained language model identifies and filters non-policy text content such as advertisements, navigation menus, and copyright notices. Key terms in policy texts are standardized based on a domain terminology mapping library.

[0030] Specifically, the first step is to uniformly schedule and process the collected multi-source heterogeneous original policy texts in parallel. The primary task is to solve the character encoding consistency problem. The system automatically detects the encoding format of the original data stream and uniformly converts it to UTF-8 encoding to avoid processing anomalies caused by inconsistent character encoding.

[0031] Specifically, for Chinese text, efficient word segmentation tools are used for basic word segmentation, and a dictionary of energy policy is combined to discover new words and eliminate ambiguities, improving the accuracy of word segmentation. Subsequently, algorithms such as SimHash and MinHash are used to deduplicate the text, effectively removing reprinted and duplicate content to ensure the uniqueness of information.

[0032] Specifically, to overcome the limitations of traditional rule-based methods (such as regular expressions and fixed noise dictionaries) in identifying irrelevant content in non-fixed positions and multi-sample expressions, this invention introduces a semantic noise classifier based on a pre-trained language model. This classifier has been fine-tuned on a large-scale corpus primarily consisting of policy texts, enabling it to deeply understand the writing style and content characteristics of policy documents. When processing texts that have undergone encoding standardization and basic preprocessing, the system inputs the paragraphs or sentences to be cleaned into the model. Based on their semantic representation, the model accurately determines whether the segment belongs to the noise that needs to be removed and outputs its classification confidence score. Compared to regular expressions that rely on surface string matching, this semantic model can effectively identify various types of interfering information that are not in the preset dictionary but whose semantics are irrelevant to the policy text, such as "Welcome to follow our WeChat official account XXX" or "Related reading: Product recommendations from a certain company," achieving high-precision identification and filtering of unstructured noise.

[0033] Specifically, after intelligent noise reduction, in-depth domain standardization processing is performed. Addressing the diversity of terminology in energy policy, the system incorporates a domain terminology standardization mapping library. Based on a thesaurus and industry standards, it normalizes key terms in the text to ensure consistency in subsequent analysis. Furthermore, for potential mixing of simplified and traditional Chinese characters, the system automatically performs simplified-traditional conversion to unify the text style.

[0034] In an optional embodiment, the first processing step may also be to perform text structure pre-identification and redundant information removal. Based on layout features and fixed tags, the macro structure of the text is pre-identified, such as automatically identifying and separating the policy text from fixed-format supplementary information such as the reporting unit at the beginning of the text and the copying agency at the end of the text.

[0035] In another optional embodiment, the first processing step can also be to perform sentence-level semantic integrity verification and fragmented sentence repair. That is, for fragmented text caused by errors in collection or parsing, such as missing sentence components or disordered word order, the system uses the contextual understanding capability of the language model to evaluate the semantic integrity and logical fluency of the sentences. For sentences identified as semantically incomplete or logically broken, the model will attempt to perform reasonable semantic completion or structural reorganization based on the context.

[0036] In this embodiment of the invention, step S200, which performs semantic segmentation and structured reorganization on the first processed data based on a first adaptive segmentation algorithm to identify the starting point and hierarchical relationship of policy clauses, includes the following sub-steps B1 to B3: In B1: Construct a first policy structure knowledge base containing typical policy provisions; Specifically, by conducting statistical analysis on a large number of policy texts, a first policy structure knowledge base is constructed, consisting of opening templates for typical policy clauses, common transitional sentence patterns, and unique logical markers.

[0037] Specifically, the construction of this policy structure knowledge base begins with the systematic collection and cleaning of massive amounts of historical policy texts, forming a high-quality domain corpus foundation. Based on this, the structural features of policy texts are automatically extracted through a combination of rule import and statistical mining: First, strong rule patterns are defined based on writing norms and sequence identifiers, and first-sentence frequency statistics and cluster analysis are used to summarize various typical clause opening semantic templates; second, by analyzing discourse structure and syntax, common transitional sentence patterns used to connect preceding and following clauses are identified and statistically analyzed; simultaneously, according to the characteristics of the energy policy domain, a systematic dictionary of logical markers including categories such as obligatory, authorizing, and conditional is established. Subsequently, the mined patterns, sentence patterns, and vocabulary are standardized and transformed into computable vector representations using a vectorization model, and initial weights are assigned based on their frequency of occurrence in the corpus, completing the quantitative construction of the knowledge base. Finally, this knowledge base is integrated into a segmentation algorithm, and the templates and weights are periodically fine-tuned by continuously collecting user feedback, forming a closed-loop system with continuous optimization capabilities, thus providing deep semantic and logical support for intelligent segmentation.

[0038] In an optional embodiment, the construction of the first policy structure knowledge base can also be achieved through a construction method based on deep semantic clustering and automatic template generation. This method utilizes a sentence embedding model trained on an energy policy corpus to convert the opening and transition sentences of massive policy clauses into high-dimensional semantic vectors. Subsequently, unsupervised deep clustering is performed on these vectors to automatically discover and summarize semantically similar clause opening and transition sentence patterns, forming cluster centers. Finally, the core syntactic-semantic pattern represented by each cluster center is extracted as a reusable structural template, and weights are assigned to it according to the cluster size, thereby realizing the automatic emergence and summarization of the knowledge base from the data.

[0039] In another optional embodiment, the construction of the first policy structure knowledge base can also be achieved through a construction method based on instruction mining and knowledge distillation of a large language model. Using a general large language model as the knowledge source, the model generates a large number of structural templates and logical markers that conform to the style of policy language through carefully designed instructions. Subsequently, the candidate knowledge generated by these models is verified and denoised with real policy corpora to select high-frequency and effective patterns. Finally, through knowledge distillation technology, these selected patterns are solidified into a lightweight knowledge base dedicated to policy structure analysis.

[0040] In B2: Traverse the policy text after the first processing, and calculate the semantic similarity between the policy text and the clause opening template, the semantic similarity between the transition sentence template and the coupling degree of the distribution of upper and lower level policy entities in the first policy structure knowledge base. Specifically, the cleaned policy text is traversed sentence by sentence or paragraph by paragraph, and for each sentence... Calculate the similarity between each clause and the database of typical clause beginnings. Similarity to common transition sentences and the degree of coupling with the distribution of policy entities at different levels. ; Specifically, the similarity between sentence beginnings and transition sentences can be vectorized, such as using word embeddings (BGE) to analyze the sentence structure. and template collection (Similarity to the beginning of a typical clause) (Similarity with common transition sentences) is expressed as an embedding vector, and calculated using the cosine similarity formula: in, Let S represent the semantic vector of the sentence to be analyzed, a typical policy clause opening template, and a common policy transition sentence template, respectively. The highest similarity between the current sentence and each type of segmentation template is obtained, while the coupling degree... This measures the frequency of occurrence and semantic relevance of sentences with entities in the surrounding policy domain.

[0041] For example, entity association weights can be calculated based on the statistical co-occurrence probability of entities using a sliding window or dependency parsing. In one window, the central sentence... The phrase "subsidy standards will be formulated separately" appears in sentences before and after it, containing entities such as "scope of application," "photovoltaic projects," and "subsidy recipients." Through the above calculations, it was found that... The co-occurrence probability with "subsidy recipients" is much higher than the random expectation, therefore Weight( "Subsidy recipients" will receive a higher positive weight. Meanwhile, The entity appearing in the window, labeled "Penalties," may have no co-occurrence with the target, and its weight may be close to zero or negative. Ultimately, of The score will be increased by its association with high-weight entities such as "subsidy recipients" and "scope of application." This high-score signal indicates to the algorithm that the sentence is in a close semantic context discussing "subsidy" policies and is likely an important policy clause, thus increasing the probability that it will be identified as the starting point of the segment.

[0042] In B3: Based on the weighted score of semantic similarity and coupling, the starting point and hierarchical nodes of policy clauses are identified. Combined with the serial number and indentation logic mark in the text, policy fragments with hierarchical structure are output. Specifically, the overall piecewise scoring function can be expressed as: in, The weighting coefficients are adjusted according to the actual policy text structure to accommodate policy documents of different types and sources.

[0043] Specifically, dynamically scan the entire text, when Exceeding the preset threshold The system automatically identifies policy clauses as starting, transitional, or hierarchical nodes. Combining regular expression logic, it further subdivides the text structure, outputting segmentation markers and hierarchical paths to achieve structured reorganization of the original policy text. This process effectively improves the adaptability to processing energy policy texts across formats and styles, laying a solid data foundation for subsequent key information extraction and multi-label classification.

[0044] In this embodiment of the invention, the first adaptive segmentation algorithm is an adaptive segmentation algorithm for energy policy content scenarios. It combines the distribution of entities in the context domain, logical flow signals, and special policy terminology patterns to automatically mark the start and end points of clauses and hierarchical relationships, improving the ability to process cross-format and diverse texts. For each sentence, its similarity to the beginning and transition sentences of typical policy clauses, as well as its coupling degree with upper and lower level policy entities, are calculated, and policy segments and levels are dynamically divided accordingly.

[0045] In an optional embodiment, the first adaptive segmentation algorithm can also be a segmentation algorithm based on multi-scale sliding windows and semantic boundary detection. It uses a multi-scale sliding window to scan the policy text and identifies potential paragraph boundaries by calculating abrupt changes in semantic coherence within the window. By comprehensively utilizing changes in lexical distribution, syntactic structure features, and the local minimum cut value of semantic vectors, it can adaptively discover semantic units of different granularities in the policy text.

[0046] In another alternative embodiment, the first adaptive segmentation algorithm can also be a sequence labeling segmentation algorithm based on pointer networks, which models the segmentation problem of policy text as a sequence labeling task and uses the pointer network architecture to directly predict the probability that each sentence is the start point, middle point or end point of a clause.

[0047] In this embodiment of the invention, step S300, which involves extracting the first key information from the structured policy text, includes the following sub-steps C1 to C3: In C1: Build a structured prompt word template for each type of key information element to be extracted; Specifically, to overcome the inherent shortcomings of traditional rule-based and template-based methods, such as weak generalization ability and high maintenance costs when faced with new policy texts, complex sentence structures, or implicit information, a large language model fine-tuned in the energy policy domain is introduced as the core engine for information extraction. Specifically, the system constructs structured prompts for each type of primary key information to be extracted (such as "issuing unit," "policy name," and "applicable objects"). These prompts clearly define the information extraction task, output format requirements, and provide a few examples.

[0048] For example, the system constructed and used the following Prompt template to extract the key information of "incentives": You are an energy policy analysis expert. From the given policy text snippet, please extract the specific details related to "incentive measures". The output must be in strict JSON format and include the following fields: "measure_type" (measure type, such as subsidies, tax breaks, loan interest subsidies, etc.), "measure_description" (detailed description of the measure), "beneficiary" (beneficiary), and "condition" (applicable conditions; fill in "none" if none).

[0049] Example of result 1: Enter the text: "For distributed photovoltaic projects included in the subsidy catalog, a subsidy of 0.1 yuan / kWh will be given to the project investment company based on the amount of electricity generated, and the subsidy period is tentatively set at 5 years." Output: { "measure_type": "funding subsidy", "measure_description": "A subsidy of 0.1 yuan / kWh will be provided based on the amount of electricity generated." "beneficiary": "project investment enterprise", "condition": "Must be included in the subsidy catalog; subsidy period: 5 years" } Example 2 of the result: Enter the text: "Financial institutions are encouraged to offer preferential loan interest rates to eligible energy storage projects, in principle, by reducing the rate by 50 basis points from the Loan Prime Rate (LPR) for the same period." Output: { "measure_type": "Loan interest subsidy", "measure_description": "Offers preferential loan interest rates, reducing the rate by 50 basis points from the prevailing LPR." "beneficiary": "Investor in energy storage projects", "condition": "The project must meet the conditions stipulated by the financial institution." } In C2: The structured policy fragments are combined with the corresponding prompt word templates and input into the first processing model. The first processing model outputs the structured first key information through semantic understanding and contextual reasoning. Specifically, the first processing model receives input consisting of prompt words and policy text fragments to be analyzed, and directly outputs structured key information through its powerful semantic understanding and contextual reasoning capabilities.

[0050] For example, traditional methods for extracting "issuing units" rely on a pre-defined list of unit names for string matching, which is prone to omissions due to incomplete or abbreviated names. Larger models, however, can perform semantic parsing based on context. Even when encountering new institutions or joint issuing units not in the list, they can accurately identify and standardize a list of unit names based on semantics (e.g., "with the consent of the XX Provincial People's Government, jointly issued by the XX Department and XX Bureau"). For unstructured information such as "clause summaries," the model can perform global semantic analysis on clause paragraphs, accurately summarize their core points, and automatically extract specific elements such as "subsidy standards," "application conditions," and "implementation period."

[0051] In an optional embodiment, the first processing model can also be an information extraction model based on a pre-trained sequence labeling model. This model uses an architecture pre-trained on a large amount of text as its foundation, with fine-tuning performed at its top layer specifically for policy information extraction tasks. The model takes policy text fragments as input and can directly and efficiently output the information category label for each word without relying on complex prompting engineering.

[0052] In another alternative embodiment, the first processing model can also be a joint parsing model based on predefined rules and semantic role labeling. First, it uses predefined syntactic rules and domain dictionaries to match key components in the text. Then, through semantic role labeling technology, it identifies the "predicate-argument" structure related to policy elements in the sentence, thereby transforming the unstructured policy description into a structured, standardized information output of "who-to whom-what-under what conditions".

[0053] In C3: Based on predefined rules and a domain entity dictionary, the first key information is formatted and standardized in terms of terminology. Specifically, the initial extraction results from the large model will be sent to a validation process based on rules and a domain entity dictionary. For example, for the extracted "publication time," regular expressions will be used for secondary validation and standardization of the date format; for the extracted "domains involved," they will be matched and semantically aligned with the constructed energy domain ontology to ensure the standardization of terminology. This collaborative paradigm of "LLM main extraction + rule-based post-validation" leverages the powerful generalization and reasoning capabilities of the large model while using rules to ensure the accuracy and controllability of key information, forming a complementary advantage.

[0054] It should be noted that the entire extraction process strictly records the original text of each field, the prediction results generated by the large model, the confidence score, and the post-processing status. All identified policy elements and their metadata are systematically archived into a structured database or knowledge graph, achieving accurate and semantic organization and efficient storage of energy policy information, providing high-quality and reliable data support for subsequent topic classification, multi-tag assignment, and intelligent retrieval.

[0055] In an optional embodiment, the first key information may also be information on policy effectiveness and scope of application. For example, from the text “This Measures shall be implemented in the Yangtze River Delta Ecological Green Integrated Development Demonstration Zone from January 1, 2025”, the following can be extracted: the level of effectiveness is “local level”, the effective date is “2025-01-01”, and the scope of application is “Yangtze River Delta Ecological Green Integrated Development Demonstration Zone”.

[0056] In another alternative embodiment, the first key information may also be information on policy tools and implementation mechanisms. For example, from "establishing a cross-departmental coordination mechanism led by the Provincial Energy Bureau and scheduling and reporting on project progress on a quarterly basis," the policy tool can be extracted as "administrative coordination," the implementing agency as "Provincial Energy Bureau," and the supervision mechanism as "quarterly scheduling and reporting."

[0057] It should be noted that step S300 above parses out key information of energy policies based on segmented and standardized text. An innovative approach combining intelligent parsing based on a large language model with rule-based and domain entity dictionary-based verification and correction is employed for accurate extraction, effectively addressing the challenges of diverse policy text expressions and varied structures, and significantly reducing erroneous and missed extractions.

[0058] In this embodiment of the invention, step S400, which involves classifying policy texts by topic and assigning multiple tags based on the first key information, and using the first fusion algorithm to categorize the policy into a predefined multi-level tagging system, includes: The vectorized representation of the full policy text and the first key information are input into the large language model. The model is then fine-tuned through instructions to generate a preliminary set of topic tags and confidence scores. The similarity between the policy text vector and the topic node vectors in the predefined multi-level tagging system is calculated using a semantic encoding model to obtain the initial content-based weights. The set of topic tags is transformed into a prior augmentation vector, which is then linearly fused with the initial weights; Based on the first threshold mechanism, multi-dimensional policy labels are automatically assigned to several thematic nodes with the highest fusion weight.

[0059] It should be noted that LLM is used to perform deep semantic understanding and potential topic mining of policy documents. Then, the domain prior knowledge, clustering features and LLM inference results are integrated through classification algorithms. By dynamically adjusting the label weights, high-accuracy, multi-granularity and interpretable automatic archiving of policy multi-labels can be achieved. Combined with policy tools, a multi-level label classification system architecture as shown in Table 1 is constructed.

[0060] Table 1: Multi-level label classification system architecture.

[0061] Specifically, the system first inputs the structured policy text into a large language model that has been fine-tuned by instructions. Through carefully designed prompt engineering, the system guides the model to perform two core tasks: first, to perform zero-sample / few-sample topic identification, requiring the model to directly generate a series of relevant topic tags based on its deep understanding of the policy content; and second, to perform fine-grained semantic association analysis of clauses, determining the relevance strength between specific paragraphs or clauses and various nodes in the predefined topic system.

[0062] Specifically, for a policy document, LLM generates a preliminary set of topic tags and their confidence levels based on semantic reasoning. For example, for a text about "rural distributed photovoltaic subsidy standards," the model can generate potential topic tags such as "distributed energy," "renewable energy subsidies," and "rural revitalization energy policy." This process fully utilizes LLM's powerful zero-shot learning capabilities and broad prior knowledge, effectively identifying emerging, overlapping, or undefined topic concepts. Subsequently, the system deeply integrates the LLM's inference results with the traditional "multidimensional progressive topic fusion classification algorithm." The algorithm converts the policy text into a vector representation using a semantic encoding model (BGE), calculates its similarity to predefined topic centers, and obtains initial content-based weights. Meanwhile, the set of topic tags generated by LLM is transformed into a prior augmentation vector. This vector reflects the strength of the model's recommendation for each predefined topic.

[0063] Specifically, the final multi-label fusion weight is calculated using the following formula: in, The fusion coefficient is adaptively adjustable. This represents the set of topics that have a strong co-occurrence with topic k. These are the elements in the co-occurrence correlation matrix.

[0064] Specifically, a dynamic threshold mechanism is used to adjust the final score. The top few topics are automatically assigned tags.

[0065] In an optional embodiment, the first fusion algorithm can also be a fusion algorithm based on topic graph propagation, which constructs the label system into a graph and uses graph neural networks to collaboratively propagate and fuse policy semantic features with topic priors generated by LLM in the graph, thereby enhancing the semantic consistency of multi-label classification through the association relationship between nodes.

[0066] In another optional embodiment, the first fusion algorithm can also be a dynamically weighted fusion algorithm that integrates the results of multiple classifiers such as semantic vector matching, LLM inference and keyword rules, and dynamically allocates voting weights according to the real-time performance of each classifier, and determines the final multi-label set through a weighted majority voting mechanism.

[0067] It should be noted that the introduction of step S400 LLM in the above-mentioned process acts like providing crucial semantic reasoning priors, equivalent to policy researchers with professional analytical capabilities, effectively complementing statistical clustering and co-occurrence methods. This not only significantly improves the classification accuracy of complex, cross-domain policies but also endows the system with the ability to dynamically discover and adapt to new policy themes. The entire process, through continuous collection of user feedback on the classification results, can reverse-optimize the LLM's prompting strategy and the weight parameters of the fusion algorithm, forming an intelligent classification system with a continuous optimization mechanism.

[0068] In this embodiment of the invention, step S500 stores the classified, extracted, and vectorized policy data into a database, constructing a dual-mode storage structure. Based on the dual-mode storage structure, an intelligent retrieval interface supporting keyword retrieval and semantic retrieval is constructed, and policy association analysis and content generation services are provided, including: Specifically, constructing a dual-mode storage structure includes: The policy structured attribute data, after classification and key information extraction, will be stored in a relational database, and a multi-dimensional index will be established. The core policy text content is converted into a high-dimensional vector representation using a semantic embedding model; the high-dimensional vector representation is then associated with the corresponding structured policy data and stored in batches in a vector database. Based on relational databases and vector databases, a dual-mode data storage structure with collaborative query capabilities is constructed.

[0069] Specifically, a policy vector knowledge base is constructed: using a semantic embedding model (BGE) optimized with corpus in the energy policy domain, the core content of the policy, including policy titles, key clause texts, and summaries, is converted into high-dimensional vector representations, which are then associated with the corresponding structured data and stored in batches into a high-performance vector database.

[0070] Furthermore, an integrated intelligent retrieval and service interface is constructed. When a user initiates a query, the system executes traditional keyword retrieval and semantic retrieval based on vector similarity in parallel. Keyword retrieval relies on the inverted index of the relational database to quickly locate policies containing explicit terminology; semantic retrieval, on the other hand, calculates the cosine similarity between the vector of the query statement and the vector of the policy content to recall the most semantically relevant policy documents from the vector database that may not contain the user's input keywords. The results of both are then weighted and rearranged based on relevance before being presented to the user in a unified manner, effectively solving the omission problem caused by keyword mismatch in traditional retrieval.

[0071] Furthermore, to enhance the depth and convenience of policy knowledge utilization, an innovative retrieval-enhanced generation framework was introduced. When a user poses a complex natural language question (such as "compare the differences between the photovoltaic subsidy policies issued by Jiangsu Province and Zhejiang Province in 2023"), the system first converts the question into a retrieval query, accurately retrieving several of the most relevant policy text fragments from the vector library as authoritative context. Subsequently, these fragments, along with the user's original question, are submitted to the large language model. Carefully designed instructions prompt the large language model to strictly summarize, compare, or answer questions based on the provided policy texts, requiring the policy source to be cited in the generated answer. This process ensures the accuracy and credibility of the generated content, effectively preventing the large model from generating content inconsistent with the facts, and providing users with decision-making references that possess analytical depth and data support.

[0072] Specifically, the retrieval interface supports a variety of services, including multi-condition combination filtering, semantic search, intelligent question answering, policy association graph jump, and multi-dimensional visualization aggregation analysis. It also ensures data security through hierarchical permission management, providing an efficient, intelligent, and reliable one-stop knowledge service platform for policy research, market analysis, and strategic planning in the power grid industry.

[0073] It should be noted that step S500 above classifies, extracts, and vectorizes the energy policy data, storing it in parallel into a structured relational database with multidimensional indexes and a vector database dedicated to high-dimensional vector retrieval, thus constructing a collaborative dual-mode data foundation. Based on this, a new generation of intelligent retrieval interface is designed and implemented, integrating traditional keyword matching with advanced semantic vector retrieval, and incorporating retrieval enhancement generation technology to provide users with accurate semantic queries, multi-dimensional aggregation analysis, policy correlation exploration, and intelligent question answering and content summary generation functions based on authoritative policy texts.

[0074] Example 2, based on the previous example, provides a specific implementation method for the intelligent collection and classification of energy policy information based on NLP, to illustrate the technical means used in this method.

[0075] This embodiment uses a real-world business scenario as an example to demonstrate how the present invention can automatically and accurately capture, analyze, and categorize a newly issued policy from multi-source information. On November 15, 2023, a provincial development and reform commission and energy bureau jointly issued a notice entitled "Several Measures on Further Promoting the High-Quality Development of Distributed Photovoltaic Power Generation in the Province" (hereinafter referred to as the "Several Measures"). This policy document was published in PDF format on the official website of the provincial development and reform commission and was also reprinted and reported by several industry media outlets.

[0076] The system captured the PDF file within two hours of the policy's release using a crawler configured for the provincial government's official website. Simultaneously, a media news crawler retrieved relevant news reports from several core industry media websites. The system added metadata such as source and timestamp to all collected raw data. The system automatically parsed the PDF file, extracting the plain text content. Subsequently, a BERT-based semantic noise classifier successfully filtered out the headers and footers ("Document from the XX Provincial Development and Reform Commission"), as well as advertising links and editor comments from reprinted news articles, resulting in a clean policy text. The "Energy Policy Content Scenario Adaptive Segmentation Algorithm" processed the main text of the "Several Measures." The algorithm successfully identified chapter starting points marked by "I. General Requirements," "II. Main Objectives," and "III. Key Tasks," and divided the specific clauses under "Key Tasks" ("(I) For distributed photovoltaic projects connected to the grid during 2024-2026, a provincial subsidy of 0.12 yuan / kWh will be given to the investing enterprise based on power generation, with a subsidy period of 5 years.") into independent, hierarchical policy segments. The system invoked a finely tuned DeepSeek-V3 model to accurately extract information from the aforementioned independent subsidy clause fragments. Based on the preset Prompt template, the model successfully output the following structured JSON data: { "policy_name": "Several Measures on Further Promoting the High-Quality Development of Distributed Photovoltaic Power Generation in the Province", "publishing_units": ["XX Province Development and Reform Commission", "XX Province Energy Bureau"], "publish_time": "2023-11-15", "involved_field": "Distributed photovoltaics", "clause_summary": "It is clarified that distributed photovoltaic projects connected to the grid between 2024 and 2026 are eligible for provincial electricity subsidies." "applicable_objects": "Distributed photovoltaic project investment enterprises", "incentive_measures":[ { "measure_type": "Electricity price and grid connection subsidy", "measure_description": "A provincial subsidy of 0.12 yuan / kWh will be provided based on the amount of electricity generated." "beneficiary": "Distributed photovoltaic project investment enterprise", "Condition": "The project must be connected to the grid between 2024 and 2026; the subsidy period is 5 years." } ] } This structured data was then accurately archived into the database. The system fused the vector representation of the entire policy document with the semantic reasoning results of LLM. Ultimately, the system automatically assigned multiple precise tags to the policy with high confidence, including Incentives and Subsidies -> Electricity Prices and Grid Connection Subsidies, Planning and Guidance -> Special Guidance Opinions, and Energy Technology Innovation -> Demonstration Projects and Pilot Management.

[0077] Example 3: This example provides an NLP-based intelligent collection and classification system for energy policy information, including: The data processing module is used to collect energy policy data from multiple information sources and perform initial processing. The segmentation and reorganization module is used to perform semantic segmentation and structured reorganization on the first processed data based on the first adaptive segmentation algorithm, and to identify the starting point and hierarchical relationship of policy clauses. The information extraction module is used to extract the first key information from the structured policy text; The classification module is used to classify policy texts by theme and assign multiple labels based on the first key information, and to use the first fusion algorithm to classify policies into a predefined multi-level label system. The data storage module is used to store the classified, extracted and vectorized policy data into the database, constructing a dual-mode storage structure. The retrieval interface building module is used to build an intelligent retrieval interface that supports keyword retrieval and semantic retrieval based on a dual-mode storage structure, and to provide policy correlation analysis and content generation services.

[0078] It should be noted that the technical solution of the NLP-based intelligent collection and classification system for energy policy information is based on the same concept as the technical solution of the aforementioned NLP-based intelligent collection and classification method for energy policy information. For details not described in detail in the technical solution of the NLP-based intelligent collection and classification system for energy policy information in this embodiment, please refer to the description of the aforementioned technical solution of the NLP-based intelligent collection and classification method for energy policy information.

[0079] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0080] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements an NLP-based intelligent collection and classification method for energy policy information. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0081] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0082] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0083] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent collection and classification of energy policy information based on NLP, characterized in that, include: Collect energy policy data from multiple information sources and perform initial processing; Based on the first adaptive segmentation algorithm, the processed data is semantically segmented and structurally reorganized to identify the starting point and hierarchical relationship of policy clauses; Extract the first key information from the structured policy text; Based on the first key information, the policy text is classified into themes and assigned multiple tags. The first fusion algorithm is used to classify the policy into a predefined multi-level tag system. The policy data, after being classified, extracted, and vectorized, is stored in a database to construct a dual-mode storage structure. Based on the aforementioned dual-mode storage structure, an intelligent retrieval interface supporting keyword retrieval and semantic retrieval is constructed, and policy correlation analysis and content generation services are provided.

2. The method for intelligent collection and classification of energy policy information based on NLP as described in claim 1, characterized in that, The semantic segmentation and structured recombination of the first processed data based on the first adaptive segmentation algorithm includes: Construct a primary policy structure knowledge base containing typical policy provisions; Traverse the first processed policy text and calculate the semantic similarity between the policy text and the clause opening template, the semantic similarity between the transition sentence template and the coupling degree of the distribution of upper and lower level policy entities in the first policy structure knowledge base. Based on the weighted score of semantic similarity and coupling degree, the starting point and hierarchical nodes of policy clauses are identified. Combined with the serial number and indentation logic mark in the text, a policy fragment with a hierarchical structure is output.

3. The method for intelligent collection and classification of energy policy information based on NLP as described in claim 2, characterized in that, The extraction of the first key information from the structured policy text includes: Build structured prompt word templates for each type of key information element to be extracted; The structured policy fragments are combined with the corresponding prompt word templates and input into the first processing model. The first processing model outputs the structured first key information through semantic understanding and contextual reasoning. Based on predefined rules and a domain entity dictionary, the first key information is subjected to format validation and terminology standardization.

4. The method for intelligent collection and classification of energy policy information based on NLP as described in claim 3, characterized in that, The process of classifying policy texts by theme and assigning multiple tags based on the first key information includes: The vectorized representation of the full policy text and the first key information are input into the large language model. The model is then fine-tuned by instructions to generate a preliminary set of topic tags and confidence scores. The similarity between the policy text vector and the topic node vectors in the predefined multi-level tagging system is calculated using a semantic encoding model to obtain the initial content-based weights. The set of topic tags is transformed into a prior augmentation vector, which is then linearly fused with the initial weights. Based on the first threshold mechanism, multi-dimensional policy labels are automatically assigned to several thematic nodes with the highest fusion weight.

5. The method for intelligent collection and classification of energy policy information based on NLP as described in claim 4, characterized in that, The construction of the dual-mode storage structure includes: The policy structured attribute data, after classification and key information extraction, will be stored in a relational database, and a multi-dimensional index will be established. The core policy text content is converted into a high-dimensional vector representation using a semantic embedding model; the high-dimensional vector representation is then associated with the corresponding structured policy data and stored in batches in a vector database. Based on the relational database and vector database, a dual-mode data storage structure with collaborative query capabilities is constructed.

6. The method for intelligent collection and classification of energy policy information based on NLP as described in claim 1, characterized in that, The energy policy data collected from multiple information sources includes: Information sources are classified according to interface specifications and content dynamism, and differentiated collection strategies are formulated accordingly; Policy data from government websites and industry platforms with open APIs is collected via API calls; policy content from news and information websites and information disclosure platforms with stable page structures but no open APIs is collected via web crawling. Extract policy text and metadata from offline reports and documents using file parsing methods; Through browser interaction driven by a large language model agent, policy information requiring login, dynamic rendering, or complex interaction can be collected.

7. The method for intelligent collection and classification of energy policy information based on NLP as described in claim 6, characterized in that, The first process includes: The energy policy data will undergo standardized encoding format and character encoding conversion. The text after unified encoding is segmented into words and deduplicated based on semantics. A semantic noise classifier based on a pre-trained language model identifies and filters non-policy text content such as advertisements, navigation menus, and copyright notices. Key terms in policy texts are standardized based on a domain terminology mapping library.

8. An NLP-based intelligent collection and classification system for energy policy information, employing the NLP-based intelligent collection and classification method for energy policy information as described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to collect energy policy data from multiple information sources and perform initial processing. The segmentation and reorganization module is used to perform semantic segmentation and structured reorganization on the first processed data based on the first adaptive segmentation algorithm, and to identify the starting point and hierarchical relationship of policy clauses. The information extraction module is used to extract the first key information from the structured policy text; The classification module is used to classify policy texts by theme and assign multiple labels based on the first key information, and to use the first fusion algorithm to classify the policies into a predefined multi-level label system. The data storage module is used to store the classified, extracted and vectorized policy data into the database, constructing a dual-mode storage structure. The retrieval interface construction module is used to build an intelligent retrieval interface that supports keyword retrieval and semantic retrieval based on the dual-mode storage structure, and to provide policy correlation analysis and content generation services.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the NLP-based intelligent collection and classification method for energy policy information as described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by the processor, they implement the steps of the NLP-based intelligent collection and classification method for energy policy information as described in any one of claims 1 to 7.

Images