An incremental policy analysis and dynamic model updating method and system

Abstract

The present application discloses an incremental policy parsing and dynamic model updating method and system, and the present application belongs to the field of natural language processing technology. The method includes: obtaining the old and new versions of the policy text, and converting each clause into a vector representation. Calculating incremental changes based on the vector, and determining the importance weight for each clause in combination with the preset weight rules. Calculating and updating the model parameters based on the model initial parameters, importance weights and incremental update formulas. Input the old and new versions of the policy text into the updated policy parsing model to obtain policy parsing data. This solution improves the efficiency of policy change analysis and enhances the accuracy and meticulousness of the impact of policy changes on different subjects. Users can understand policy changes more quickly and accurately, reduce compliance risks, improve the scientificity and effectiveness of decision-making, improve work efficiency, and reduce manual intervention.

Description

Technical Field

[0001] This application belongs to the field of natural language processing technology, and specifically relates to an incremental policy parsing and dynamic model updating method and system. Background Art

[0002] Policy texts, as the core carrier of social governance, carry laws, regulations and strategic decisions. Their updates present the triple challenges of high frequency, complexity and strong correlation: high frequency is reflected in the need to quickly respond to social changes and technological innovations, with the annual policy update rate in some areas reaching more than 30%; complexity stems from a three-dimensional network with multi-level structures, multi-field intersections and multiple interest games, and a single policy may trigger multi-dimensional impacts such as industrial chains and people's livelihood security; strong correlation is manifested in the linkage effect between policies, such as the need for simultaneous matching of labor market and fiscal support measures for adjustments to education policies.

[0003] Traditional manual policy analysis adopts a three-stage refined process: experts first manually align the new and old policy paragraphs, and establish a comparison benchmark by marking metadata such as version numbers and chapter numbers; then a double-blind cross-verification mechanism is adopted, with two groups of experts independently reading the text sentence by sentence, marking added / deleted / modified clauses with green / red / blue colors, and reaching a consensus on ambiguous statements through joint meetings; finally, combined with the historical policy case library and the field regulatory map, a three-level qualitative analysis of the different clauses is carried out, capturing explicit changes such as wording adjustments on the surface, inferring policy-oriented transformations at the middle level, and predicting potential impacts on industrial structure, fiscal pressure, etc. in depth, forming a complete interpretation chain of "form-content-effect".

[0004] However, traditional manual policy analysis is inefficient and unable to cope with the high-frequency policy update needs; differences in experts' subjective experience lead to divergent interpretation results, affecting the consistency of analysis; manual review can only identify explicit text changes and lacks awareness of deep changes such as semantic fine-tuning and clause reorganization; and this model requires continuous investment of a large amount of expert resources, is costly and difficult to expand to large-scale dynamic policy network analysis, making it difficult to adapt to the needs of complex governance scenarios. Summary of the Invention

[0005] The embodiments of the present application provide an incremental policy analysis and dynamic model updating method and system to address the problems of low efficiency of traditional manual policy analysis and its difficulty in meeting high-frequency policy update needs; differences in expert subjective experience lead to divergent interpretation results, affecting analysis consistency; manual review can only identify explicit text changes and lacks awareness of deep changes such as semantic fine-tuning and clause reorganization; and this model requires continuous investment of a large amount of expert resources, is costly and difficult to expand to large-scale dynamic policy network analysis, and is difficult to adapt to the needs of complex governance scenarios.

[0006] In a first aspect, an embodiment of the present application provides an incremental policy parsing and dynamic model updating method, the method comprising:

[0007] Obtaining an old version of the policy text and a new version of the policy text, determining a first vector for each clause of the old version of the policy text, and determining a second vector for each clause of the new version of the policy text;

[0008] Calculate the incremental changes of each clause in the new policy text based on the first vector, the second vector, and the preset incremental change calculation formula;

[0009] Determining the importance weight of each clause based on the incremental change and a preset weight determination rule, obtaining initial model parameters, calculating model update parameters based on the initial model parameters, the importance weight, the first vector, the second vector, and a preset incremental update formula, and incrementally updating the preset policy parsing model based on the model update parameters;

[0010] The old version of the policy text and the new version of the policy text are input into the updated preset policy parsing model to obtain policy parsing data.

[0011] Furthermore, the preset incremental change calculation formula is:

[0012]

[0013] in, is the incremental change; i is the clause index; is the importance weight of each clause; is the second vector; is the first vector; is the preset smoothing factor.

[0014] Furthermore, the preset incremental update formula is:

[0015]

[0016] in, Update parameters for the model; are the initial parameters of the model; is the preset learning rate; N is the total number of clauses; i is the clause index; is the importance weight; is a loss function used to measure the differences between the new and old policy texts; The loss function is the initial parameters of the model The gradient of , represents the influence of each term on the final model when updating the model parameters.

[0017] Furthermore, after obtaining the policy analysis data, the method further includes:

[0018] Obtain historical policy text data and historical social media data, and use fine-grained sentiment analysis models to determine the historical sentiment tendency data of historical policy texts;

[0019] Use graph neural networks to analyze historical social media data to obtain historical public opinion dissemination paths and historical hot spot analysis data;

[0020] A policy impact analysis model is constructed based on historical policy text data, historical social media data, historical sentiment tendency data, historical public opinion dissemination paths, and historical hot spot analysis data.

[0021] Furthermore, after constructing a policy impact analysis model based on historical policy text data, historical social media data, sentiment trend data, public opinion dissemination paths, and hot spot analysis data, the method further includes:

[0022] Obtain real-time social media data after the release of the new policy text, input the new policy text data and real-time social media data into the policy impact analysis model, and obtain real-time sentiment tendency data, real-time public opinion dissemination path and real-time hot spot analysis data.

[0023] Furthermore, after obtaining the policy analysis data, the method further includes:

[0024] Determine the components of the policy parsing data, and determine the encryption level of each component of the policy parsing data according to a preset encryption strategy;

[0025] Determining the sensitivity and update frequency of each component of the policy parsed data, and determining the encryption strength of each component based on the sensitivity and update frequency;

[0026] Determining contextual relationships among components of the policy parsed data, and determining dynamic encryption strategies for the components based on the contextual relationships;

[0027] Obtain access rights information for different user roles, and set access control rights for different user roles based on the access rights information;

[0028] The policy parsing data is encrypted according to the encryption level, encryption strength, dynamic encryption strategy and the access control authority to obtain encrypted data packets corresponding to different user roles, and the encrypted data packets are transmitted to the control center.

[0029] Furthermore, before transmitting the encrypted data packet to the control center, the method further includes:

[0030] Use quantum key distribution technology to generate a temporary key, perform a hash operation on the temporary key, and generate a hash value of the temporary key;

[0031] Sign the hash value to obtain a hash signature, and encrypt the temporary key using the preset key protection method;

[0032] Forming a transmission encrypted data packet according to the encrypted temporary key, hash value, hash signature and encrypted data packet;

[0033] Accordingly, transmitting the encrypted data packet to the control center includes:

[0034] The encrypted data packet is transmitted to the control center via a TLS / SSL encrypted channel.

[0035] In a second aspect, an embodiment of the present application provides an incremental policy parsing and dynamic model updating system, the system comprising:

[0036] A policy text vectorization module is used to obtain an old version of the policy text and a new version of the policy text, determine a first vector for each clause of the old version of the policy text, and determine a second vector for each clause of the new version of the policy text;

[0037] An incremental change calculation module, used to calculate the incremental change of each clause of the new version of the policy text based on the first vector, the second vector and a preset incremental change calculation formula;

[0038] an incremental update module, configured to determine the importance weight of each clause based on the incremental change and a preset weight determination rule, obtain initial model parameters, calculate model update parameters based on the initial model parameters, the importance weight, the first vector, the second vector, and a preset incremental update formula, and incrementally update the preset policy parsing model based on the model update parameters;

[0039] The policy parsing module is used to input the old version of the policy text and the new version of the policy text into the updated preset policy parsing model to obtain policy parsing data.

[0040] In a third aspect, an embodiment of the present application provides an electronic device comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the first aspect.

[0041] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0042] In an embodiment of the present application, the old version of the policy text and the new version of the policy text are obtained, the first vector of each clause of the old version of the policy text is determined, and the second vector of each clause of the new version of the policy text is determined; the incremental change of each clause of the new version of the policy text is calculated according to the first vector, the second vector and the preset incremental change calculation formula; the importance weight of each clause is determined according to the incremental change and the preset weight determination rule, the model initial parameters are obtained, the model update parameters are calculated according to the model initial parameters, the importance weight, the first vector, the second vector and the preset incremental update formula, and the preset policy parsing model is incrementally updated according to the model update parameters; the old version of the policy text and the new version of the policy text are input into the updated preset policy parsing model to obtain policy parsing data. Through the above-mentioned incremental policy parsing and dynamic model updating method, the old version and the new version of the policy text are input into the updated preset policy parsing model, and the policy parsing data obtained after the incremental update not only improves the efficiency of policy change analysis, but also enhances the accuracy and meticulousness of the impact of policy changes on different subjects. Users can understand policy changes more quickly and accurately, reduce compliance risks, improve the scientificity and effectiveness of decision-making, improve work efficiency, and reduce manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flowchart of the incremental policy parsing and dynamic model updating method provided in Example 1 of the present application;

[0044] Figure 2 This is a flowchart of the incremental policy analysis and dynamic model updating method provided in Example 2 of the present application;

[0045] Figure 3 This is a schematic diagram of the structure of the incremental policy analysis and dynamic model updating system provided in Example 3 of the present application;

[0046] Figure 4 This is a schematic diagram of the structure of the electronic device provided in Example 4 of the present application. DETAILED DESCRIPTION

[0047] To further clarify the objectives, technical solutions, and advantages of this application, specific embodiments of the present application are described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are intended only to illustrate this application and are not intended to limit it. It should also be noted that, for ease of description, the drawings only illustrate portions relevant to this application, not all of them. Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict the various operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process may terminate upon completion of its operations, but may also include additional steps not shown in the accompanying drawings. The process may correspond to a method, function, procedure, subroutine, subprogram, or the like.

[0048] The following will be combined with the accompanying drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of this application.

[0049] The terms "first," "second," and the like in the specification and claims of this application are used to distinguish similar objects, and are not used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of this application can be implemented in an order other than that illustrated or described herein, and that the objects distinguished by "first," "second," and the like are generally of the same type, and do not limit the number of objects; for example, the first object can be one or more. In addition, the term "and / or" in the specification and claims refers to at least one of the connected objects, and the character " / " generally indicates that the objects connected are in an "or" relationship.

[0050] Below, in conjunction with the accompanying drawings, an RSMC chip, a chip multi-stage startup method, and a Beidou communication and navigation device provided in the embodiments of the present application are described in detail through specific embodiments and their application scenarios.

[0051] Example 1

[0052] Figure 1 This is a flow chart of the incremental policy analysis and dynamic model updating method provided in Example 1 of this application. Figure 1 As shown, the specific steps include:

[0053] S101 , obtaining an old version of the policy text and a new version of the policy text, determining a first vector of each clause of the old version of the policy text, and determining a second vector of each clause of the new version of the policy text.

[0054] The old version of the policy text may refer to a policy document released before a certain point in time, which includes the specific content, terms, regulations and other information of the policy.

[0055] The new policy text may refer to a revised or updated policy document containing new or revised terms, which may include additions, deletions, modifications, etc. compared to the old policy.

[0056] A clause is a specific provision or statement within a policy document, typically involving specific behavioral requirements, conditions, obligations, or restrictions. Each clause may consist of a paragraph or a few sentences and typically specifies certain behaviors.

[0057] The first vector can be the result of converting each clause in the "old policy text" into a mathematical representation (vectorization). The vectorization process converts the clause into a digital form, making it easy to process and compare in computers. The vector representation of each clause can be generated using technologies such as word embedding (such as Word2Vec and GloVe), TF-IDF, and BERT.

[0058] The second vector can be the result of vectorizing each clause of the "new policy text" using the same method. The process here is similar to the "first vector", but is applied to the clauses in the new policy text.

[0059] The old and new policy texts can be read from the database, and individual clauses can be extracted from each version. Each clause is vectorized using the following methods: Bag-of-Words: Constructs a vector based on the frequency of words appearing in the clause. TF-IDF: Calculates the frequency and importance of words in each clause. Pre-trained language models (such as BERT and Word2Vec) convert each clause into a high-dimensional vector representation, typically using sentence- or paragraph-level models. The resulting vector representations for each clause form the first vector (old policy clause) and the second vector (new policy clause).

[0060] S102, calculating the incremental change of each clause of the new version of the policy text based on the first vector, the second vector and a preset incremental change calculation formula.

[0061] Incremental change can be a metric that quantifies the changes between policy text versions. It is based on the vector difference of clauses and calculates the change of each clause by weight, reflecting the degree of difference between the old and new versions of the policy text.

[0062] The first vector and the second vector can be substituted into the preset incremental change calculation formula to calculate the incremental changes of each clause in the new version of the policy text.

[0063] Based on the above technical solution, an optional, preset incremental change calculation formula is:

[0064]

[0065] in, is the incremental change; i is the clause index; is the importance weight of each clause; is the second vector; is the first vector; is the preset smoothing factor.

[0066] In this plan, The main purpose of is to avoid division by zero errors during calculations. If the denominator is very close to zero, it may become very small, resulting in numerical instability during the calculation process. As a small constant, it ensures that the formula can be calculated smoothly even in this case. In different application scenarios, It can be adjusted according to the experimental results. If it is too small, it may still cause numerical instability, and if it is too large, it may affect the accuracy of the calculation. Through debugging and optimization, choose a suitable It can make the calculation process more stable and have better accuracy.

[0067] S103, determine the importance weight of each clause according to the incremental change and the preset weight determination rules, obtain the model initial parameters, calculate the model update parameters according to the model initial parameters, the importance weight, the first vector, the second vector and the preset incremental update formula, and incrementally update the preset policy parsing model according to the model update parameters.

[0068] Preset weighting rules can determine the importance of each clause based on its incremental change (i.e., the degree of change between clauses). The size of the incremental change is combined with the preset weighting rules to determine the weight of the clause. For example, clauses with larger incremental changes will be assigned higher weights, indicating that these clauses have a greater impact on policy changes and deserve priority consideration.

[0069] Importance weights are numerical values ​​used to measure the relative importance of each clause within the entire policy text. They are often used to calculate weighted sums when calculating incremental changes to highlight the impact of key clauses.

[0070] Initial model parameters can be the parameter state of the policy parsing model before the update, representing the model's knowledge state at a certain historical moment. Specifically, they can include term embedding vectors: the model's understanding of the previous version of the term; attention weights (such as the self-attention parameter in the Transformer model); rule or logic weights (such as the decision threshold in a rule-based parsing system); and model hyperparameters (such as the learning rate and regularization coefficient). These initial parameters serve as the baseline for incremental updates and are used to calculate the new parameters after the update.

[0071] The model update parameters can be new parameters calculated based on the model's initial parameters, importance weights, first vector, second vector, and incremental update formula. These parameters are used to adjust the policy parsing model to adapt it to the new version of the policy text.

[0072] The pre-set policy parsing model can be a system model based on natural language processing (NLP), machine learning, a rules engine, or deep learning technology. It is used to parse, understand, and evaluate the content and changes of policy clauses within the document. This model can perform structured parsing of policy texts and conduct quantitative and qualitative analysis of the content, changes, and significance of policy clauses. Specifically, it can automatically identify and annotate differences between the old and new versions of the policy, including additions, deletions, and revisions. This feature analyzes policy changes, automatically compares clauses, identifies deletions, revisions, or additions, and clearly annotates the changes. It extracts key concepts, terms, and definitions from the policy text and provides detailed explanations. This feature automatically identifies key terms and concepts from the policy text, generates explanations or definitions for these terms, and provides contextual information to facilitate understanding. It analyzes the impact of policy changes on different entities (such as businesses, individuals, and government agencies) and assesses the extent of the impact. By simulating different scenarios, this feature analyzes the specific consequences of policy changes on affected entities and provides an estimate of the intensity and scope of the impact. Specific compliance recommendations and risk warnings are provided to affected entities. After comparing and analyzing policy changes, this feature provides specific compliance action recommendations to businesses, individuals, or institutions to help them comply with the new policies. It also identifies potential compliance risks and alerts relevant parties.

[0073] Based on the incremental changes of each clause, combined with preset weight determination rules (for example, the greater the incremental change, the more important the clause), each clause can be assigned an importance weight. The model initial parameters are read, and the model initial parameters, importance weight, first vector, and second vector are substituted into the preset incremental update formula to obtain the model update parameters. Once the model update parameters are calculated, these parameters can be used to incrementally update the preset policy parsing model. Specifically, the internal parameters of the model can be adjusted based on the model update parameters. Based on the initial training, the updated model parameters can be used for fine-tuning or retraining to make the model better adaptable to the new policy text.

[0074] The training process of the preset policy parsing model is:

[0075] 1. Data preparation and preprocessing

[0076] Data Collection: Collect a large amount of old and new policy texts as training data. This text data should cover a wide range of policies, including different clauses, change types (additions, deletions, and revisions), relevant industry backgrounds, and impacted entities.

[0077] Text Annotation and Notes: Annotate the clauses in the policy text, indicating the clause content, changes (such as additions, modifications, or deletions), and their specific textual content. In addition, highlight key terms, definitions, and concepts in the text.

[0078] Word segmentation and tagging: Perform NLP processing on policy texts, including word segmentation, part-of-speech tagging, and named entity recognition (NER). This extracts entities, relationships, and key terms from the text.

[0079] Text structuring: Convert unstructured policy text into structured data for further analysis. This may include the division of clauses, definition of terms, and types of clause changes.

[0080] 2. Model Design

[0081] Depending on the requirements of the parsing task, the model can be designed based on one or a combination of the following techniques:

[0082] Rule-based engines: These engines process explicit rules or logical relationships within policy text. For example, they match specific policy terms or variations of clauses, performing pattern matching.

[0083] Natural language processing (NLP) technologies: For example, pre-trained language models based on the Transformer architecture (such as BERT and GPT) are used to extract semantic representations and contextual information of terms, allowing comparison and contrast between them.

[0084] Deep learning techniques: For example, using RNN, LSTM, or Transformer to learn the representation of policy terms, combined with attention mechanisms to analyze the importance and changes of different terms.

[0085] Machine learning models: such as classification models and regression models, which train models based on labeled data to predict and analyze the impact of clause additions, deletions, and modifications.

[0086] 3. Training Data Labeling and Enhancement

[0087] Annotation of clause additions, deletions, and modifications: Utilize pre-annotated text data to classify each policy clause, identify which clauses have been modified, which clauses have been added, and which clauses have been deleted, and annotate the specific changes.

[0088] Terminology and Definition Extraction: Extract key terms and definitions from policy texts and generate contextual explanations. Annotate and use these terms for contextual analysis.

[0089] Impact analysis and annotation: Based on policy changes, mark the affected objects (such as enterprises, individuals, and government agencies) and evaluate their impact level.

[0090] 4. Model Training

[0091] Feature extraction: Extract features from text data, such as word vectors, context vectors, grammatical structure information, and change information. Use NLP techniques (such as Word2Vec, GloVe, and BERT) to convert text into numerical vector representations.

[0092] Training the model: Use supervised or unsupervised learning techniques to train the model. For supervised learning, the model can be trained using a labeled dataset to learn tasks such as adding, deleting, and modifying terms, explaining terminology, and the impact of changes.

[0093] DD / Chang detection: Train the model to identify DD / Chang changes between clauses and calculate the size, type, and specific impact of the changes.

[0094] Terminology interpretation and definition generation: Generate definitions of terms based on context and infer changes in terms.

[0095] Impact analysis model: Based on the characteristics of different objects (such as enterprises and individuals), analyze the specific impact of policy changes on these objects.

[0096] Compliance analysis and risk warnings: The training model predicts the compliance risks that policy changes may bring to the object and generates recommendations.

[0097] 5. Model Evaluation and Tuning

[0098] Model evaluation: Use evaluation metrics (such as precision, recall, and F1 score) to evaluate the trained model. In particular, evaluate the model's performance on different tasks, such as clause addition, deletion, and modification detection, term parsing, and impact analysis.

[0099] Adjust parameters: Based on the evaluation results, adjust the model's hyperparameters or training strategy. This may require re-adjusting the training set, optimization algorithm (such as the Adam optimizer), and regularization strategy.

[0100] Model validation: Use cross-validation or an independent validation set to test the generalization ability of the model.

[0101] 6. Incremental Learning and Updates

[0102] Incremental Updates: As policy text is updated, the model needs to be incrementally updated to reflect the new policy version. Based on the aforementioned incremental change calculation formula and importance weight model, the model can perform incremental learning, gradually improving its ability to parse and evaluate new policies.

[0103] Feedback mechanism: Based on user feedback, continue to adjust and optimize the model to enhance its accuracy and reliability.

[0104] Deploy the trained policy parsing model to the production environment to support real-time parsing, comparison, impact analysis, compliance recommendations, and other functions of new policy texts.

[0105] Based on the above technical solution, an optional, preset incremental update formula is:

[0106]

[0107] in, Update parameters for the model; are the initial parameters of the model; is the preset learning rate; N is the total number of clauses; i is the clause index; is the importance weight; is a loss function used to measure the differences between the new and old policy texts; The loss function is the initial parameters of the model The gradient of , represents the influence of each term on the final model when updating the model parameters.

[0108] In this plan, The choice is based on previous experimental experience or the value recommended in the literature. For example, the learning rate is often set to a small positive number. For deep learning models, the learning rate is a key hyperparameter. A larger learning rate may lead to unstable training process, while a smaller learning rate may lead to slow convergence. In some cases, It will be adjusted dynamically during the training process. For example, using learning rate decay or adaptive learning rate algorithms (such as Adam, RMSprop, etc.) to automatically adjust the learning rate according to the changes in the gradient to improve training efficiency and performance. When , the optimal learning rate can be selected by cross-validation method. The performance of the model under different learning rates can be evaluated by the performance on the training set and the validation set, and then the most appropriate value.

[0109] S104: Input the old version of the policy text and the new version of the policy text into the updated preset policy analysis model to obtain policy analysis data.

[0110] Policy parsing data may refer to the structured and quantitative output data about policy changes generated by the model after the old and new policy texts are input into the updated preset policy parsing model. These data include the multi-dimensional analysis results of the policy text to help users understand and respond to the specific content, impact and compliance of policy changes. Specifically, it may include clause difference analysis: addition, deletion and modification annotation: automatic identification and annotation of changes in policy clauses, including which clauses have been added, deleted or modified. For modified clauses, the specific content before and after the modification is clearly listed. The nature of the clause change: indicate the type of change (addition, deletion, modification) of each clause, as well as the scope of the change and the specific content of the modification.

[0111] Extraction of key concepts, terms, and definitions: Term interpretation: The model extracts key terms from policy texts and automatically generates definitions or interpretations of the terms. Contextual information: Provides context for each term to help understand its application and meaning in specific clauses.

[0112] Impact Analysis: Affected Targets: Based on the policy changes, analyze the impact of the policy on different entities (such as businesses, individuals, and government agencies). Impact Intensity and Scope: Assess the specific consequences of the policy changes on these entities, providing an impact intensity (such as high, medium, and low) and the possible scope of impact (such as industry and region). Impact Scenario Simulation: Simulate the specific impact of the policy changes on the entities under different scenarios to provide a more comprehensive assessment.

[0113] Compliance Advice and Risk Warnings: Compliance Advice: Provides specific compliance advice to affected entities (such as businesses, individuals, or institutions) to help them comply with the new policy. Risk Warnings: Identifies potential compliance risks, alerts relevant parties to potential issues, and enables them to prepare for them in advance.

[0114] By feeding the old and new policy texts into the updated pre-set policy parsing model, the model will generate policy parsing data through multiple steps, including text preprocessing, vectorization, clause comparison and change calculation, semantic parsing, and impact analysis. These data packages provide users with clear and structured information on policy changes, helping them better understand and respond to them.

[0115] In an embodiment of the present application, the old version of the policy text and the new version of the policy text are obtained, the first vector of each clause of the old version of the policy text is determined, and the second vector of each clause of the new version of the policy text is determined; the incremental change of each clause of the new version of the policy text is calculated according to the first vector, the second vector and the preset incremental change calculation formula; the importance weight of each clause is determined according to the incremental change and the preset weight determination rule, the model initial parameters are obtained, the model update parameters are calculated according to the model initial parameters, the importance weight, the first vector, the second vector and the preset incremental update formula, and the preset policy parsing model is incrementally updated according to the model update parameters; the old version of the policy text and the new version of the policy text are input into the updated preset policy parsing model to obtain policy parsing data. Through the above-mentioned incremental policy parsing and dynamic model updating method, the old version and the new version of the policy text are input into the updated preset policy parsing model, and the policy parsing data obtained after the incremental update not only improves the efficiency of policy change analysis, but also enhances the accuracy and meticulousness of the impact of policy changes on different subjects. Users can understand policy changes more quickly and accurately, reduce compliance risks, improve the scientificity and effectiveness of decision-making, improve work efficiency, and reduce manual intervention.

[0116] Based on the above technical solution, optionally, after obtaining the policy analysis data, the method further includes:

[0117] Determine the components of the policy parsing data, and determine the encryption level of each component of the policy parsing data according to a preset encryption strategy;

[0118] Determining the sensitivity and update frequency of each component of the policy parsed data, and determining the encryption strength of each component based on the sensitivity and update frequency;

[0119] Determining contextual relationships among components of the policy parsed data, and determining dynamic encryption strategies for the components based on the contextual relationships;

[0120] Obtain access rights information for different user roles, and set access control rights for different user roles based on the access rights information;

[0121] The policy parsing data is encrypted according to the encryption level, encryption strength, dynamic encryption strategy and the access control authority to obtain encrypted data packets corresponding to different user roles, and the encrypted data packets are transmitted to the control center.

[0122] In this solution, the components can be the core content units of policy analysis data, such as policy clauses, scope of application, implementation details, relevant legal basis, historical modification records, implementation impact assessment, etc.

[0123] Pre-set encryption policies can be pre-defined encryption rules based on data sensitivity and access requirements. For example, AES (symmetric encryption) can be used to protect policy text and clauses. Homomorphic encryption can be used to protect numerical data (such as fine amounts), ensuring that they can still be calculated even in an encrypted state. A hash algorithm (SHA-256) can be used to ensure data integrity and prevent tampering. Differential privacy can be used to protect personal privacy during data analysis.

[0124] Encryption levels can be set based on data sensitivity. For example, high sensitivity refers to the core policy terms and fine amounts, which require strong encryption. Medium sensitivity refers to the execution time and applicable industries, which require basic encryption. Low sensitivity refers to the policy summary, which can be stored in plain text or with weak encryption.

[0125] Sensitivity can be determined by the importance and security requirements of the policy data. For example, the amount of a fine is highly sensitive data and should be strongly encrypted. Public policy summaries are less sensitive data and can be weakly encrypted.

[0126] The update frequency can be the frequency of changes in different data parts. For example, some regulatory provisions are stable for a long time, while the implementation details may be adjusted frequently.

[0127] Encryption strength can be determined based on sensitivity and update frequency. For fine amounts (high sensitivity, high volatility), strong symmetric encryption plus homomorphic encryption is used. For policy terms (high sensitivity, low volatility), strong AES-256 encryption is used. For implementation time (low sensitivity, medium volatility), basic AES-128 encryption is used.

[0128] Contextual relationships can be dependencies between components. For example, fines and emission reduction requirements are closely linked and must be consistently encrypted to ensure uniform access rights. Policy terms and execution time are primary and secondary and can be managed separately.

[0129] Dynamic encryption policies can be dynamically adjusted based on context. Specifically, they can include clause dependencies: certain policy clauses are interdependent and must be encrypted and decrypted in a specific order. For example, the "fine amount" depends on the "emission reduction requirement" because the applicable conditions for the fine amount depend on whether the emission reduction requirement has been violated. A strategy of encrypting the fine amount first and then the emission reduction requirement ensures that the fine amount is not visible until the emission reduction requirement is decrypted. Implementation: Step 1: Use homomorphic encryption to first encrypt the fine amount, ensuring that even in the event of a data leak, the fine amount cannot be directly read. Step 2: Use AES encryption to encrypt the emission reduction requirement. The fine amount can only be calculated after the clause is decrypted.

[0130] Access rights information can be defined for different users, including the terms and data scope of access, whether decryption and editing are allowed, and whether additional authentication (such as two-factor authentication) is required.

[0131] User roles can include: Government staff: Access to complete policy analysis data; Enterprise users: Access only policy requirements and implementation time, with fine amounts encrypted; Public users: View only policy summaries, with all sensitive information encrypted.

[0132] Access control permissions can refer to the permissions for different user roles to access, decrypt, and manipulate encrypted data. It ensures that sensitive data can only be viewed by authorized users, while restricting unauthorized users from accessing critical information.

[0133] The encrypted data packet may refer to a data packet generated for different users after policy parsing data is separately encrypted based on access control permissions, ensuring that different roles can only decrypt their own data.

[0134] The components of policy analysis data can be determined. Policy analysis data typically includes multiple core content units, such as policy clauses, scope of application, implementation details, relevant legal basis, historical revision records, and implementation impact assessments. Each component plays a different role in policy implementation and interpretation, so their classification and data attributes need to be clearly defined. Secondly, based on the pre-set encryption strategy, the encryption level for each component of the policy analysis data should be determined. The encryption level depends on the sensitivity of the data. For example, highly sensitive data requires strong encryption, moderately sensitive data uses basic encryption, and low-sensitivity data can be weakly encrypted or stored in plaintext. The encryption strategy can use symmetric encryption (such as AES-256) to protect core policy content, homomorphic encryption to protect numeric data (ensuring data remains computable even in an encrypted state), hashing algorithms (such as SHA-256) to ensure data integrity, and differential privacy techniques to protect personal privacy during data analysis. Finally, the sensitivity level and update frequency of each component of the policy analysis data should be determined. Sensitivity determines the importance and security requirements of the data, while update frequency influences the choice of encryption method. Highly sensitive and frequently changing data should use stronger encryption mechanisms, such as a combination of symmetric encryption and homomorphic encryption. Highly sensitive but rarely changing data can use strong symmetric encryption. Moderately or low-sensitivity data should use an appropriate encryption method based on the specific situation. The combination of sensitivity and update frequency determines encryption strength, ensuring both secure and efficient data access. Next, determine the contextual relationships between the various components of the policy-parsed data and formulate dynamic encryption policies accordingly. Contextual relationships refer to logical connections between data. For example, some policy clauses depend on the content of other clauses. Therefore, the correctness of these dependencies must be ensured during encryption and decryption. Dynamic encryption policies analyze data dependencies and adjust encryption order and policies to ensure information flow and consistency during decryption. For example, if the decryption of a piece of data depends on another piece of data, the latter should be encrypted before the former. Next, obtain access rights information for different user roles and set access control permissions based on this access rights information. Access control permissions determine which data different users can access and whether they can decrypt or edit specific information. Access control mechanisms can be managed based on roles (RBAC) or attributes (ABAC), ensuring that only authorized users can decrypt specific data, while unauthorized users can only access filtered and encrypted data. Finally, based on the encryption level, encryption strength, dynamic encryption policy, and access control permissions, the policy parsing data is encrypted, and encrypted data packets corresponding to different user roles are generated. The encrypted data packets contain data adjusted according to user permissions, ensuring that different users can only access information within their authorized scope. After generation, the encrypted data packets are transmitted to the control center via wireless communication technology. For example, an environmental protection policy may include multiple components such as "emission reduction requirements," "implementation details," "fine amounts," and "scope of application.""Fine amount" is highly sensitive data and requires AES-256 encryption combined with homomorphic encryption. "Implementation details" are moderately sensitive data and require AES-128 encryption. "Scope of application" is low-sensitivity data and can be stored using weak encryption or plaintext. Fine amounts involve direct economic impact, are highly sensitive, and may change with policy adjustments, so strong encryption is required. Historical modification records are moderately sensitive data, but change infrequently, so basic encryption is used. The policy summary is low-sensitivity data and rarely changes, so it can be stored in plaintext. The applicable conditions for fines depend on emission reduction requirements. Therefore, encryption must be performed first, followed by the emission reduction requirements, to ensure that the fine amount is not visible without decrypting the emission reduction requirements. Government officials can decrypt all clauses. Enterprise users can only decrypt the policy requirements, scope of application, and implementation details; the fine amount remains encrypted. Public users can only view the policy summary; all sensitive data remains encrypted. A fully decrypted data packet is generated for government officials. A partially decrypted data packet is generated for enterprise users, with the fine amount remaining encrypted. Generates highly encrypted data packets for public users, leaving only the policy summary readable.

[0135] This solution implements appropriate encryption measures for data of varying sensitivity levels, ensuring that sensitive information (such as fine amounts and policy terms) is highly encrypted. Non-sensitive information can be stored using basic encryption or plaintext, mitigating the risk of data leakage. Role-based permission management ensures that government officials, corporate users, and the public can access only the data within their authorized scope, preventing unauthorized access and improving data security and compliance.

[0136] Based on the above technical solution, optionally, before transmitting the encrypted data packet to the control center, the method further includes:

[0137] Use quantum key distribution technology to generate a temporary key, perform a hash operation on the temporary key, and generate a hash value of the temporary key;

[0138] Sign the hash value to obtain a hash signature, and encrypt the temporary key using the preset key protection method;

[0139] Forming a transmission encrypted data packet according to the encrypted temporary key, hash value, hash signature and encrypted data packet;

[0140] Accordingly, transmitting the encrypted data packet to the control center includes:

[0141] The encrypted data packet is transmitted to the control center via a TLS / SSL encrypted channel.

[0142] In this scenario, quantum key distribution (QKD) is a secure key exchange method based on the principles of quantum mechanics. It utilizes quantum states (such as the polarization state of photons) to transmit keys, relying on the cloning and non-copyability of quantum measurements to prevent eavesdropping. If a third party attempts to intercept the quantum key, the quantum state will change, making it detectable to the legitimate communicating party, ensuring the security of key transmission.

[0143] A temporary key is a short-term encryption key, typically valid for a single communication or a period of time. It is used to encrypt data transmission, preventing the leakage of long-term keys and improving security. Temporary keys are typically generated dynamically by the security protocol and destroyed after the session ends.

[0144] A hash value is a unique, fixed-length summary of input data (such as a key or message) generated by applying a hash function. It is used to verify data integrity, ensuring it has not been tampered with during transmission or storage. For example, SHA-256 is a commonly used hash algorithm that converts data of any length into a fixed-length 256-bit hash value.

[0145] A hash signature is the result of digitally signing a hash value. Typically, a hash signature is encrypted with a private key, ensuring that only the corresponding public key can verify its authenticity. The purpose of a hash signature is to prevent the hash value from being tampered with, thereby ensuring data integrity and source authenticity.

[0146] Pre-defined key protection methods can include encryption technologies for storing and protecting keys, such as: Key encryption key: Temporary keys are encrypted using a strong master key. Hardware security module: Utilizes dedicated secure hardware to store keys and prevent them from being stolen. Trusted execution environment: Storing and processing keys in a protected computing environment. Sharded storage: Splits keys into multiple fragments so that the key can only be recovered when a sufficient number of fragments are combined.

[0147] A transport encryption data packet is an encapsulated data packet containing encrypted data, used to securely transmit information between different systems or devices. It can include a temporary encryption key (to prevent key leakage), a hash value (for integrity verification), a hash signature (to ensure the authenticity of the hash value), and encrypted data content (to prevent unauthorized access).

[0148] TLS (Transport Layer Security) and SSL (Secure Sockets Layer) are cryptographic protocols used to protect internet communications. They establish a secure connection between client and server, ensuring that data cannot be eavesdropped or tampered with during transmission. TLS, an upgraded version of SSL, offers enhanced security and is commonly used in applications such as HTTPS, VPNs, and email encryption.

[0149] The sender and receiver establish a secure channel using a quantum key distribution (QKD) protocol (such as BB84). The sender generates a set of random quantum bits (qubits) and encodes them in different bases (such as rectangular and diagonal bases). The sender sends these qubits to the receiver, who measures them and negotiates a final shared key with the sender. Through information coordination and privacy amplification, bits that could be eavesdropped on are removed, ultimately generating a symmetric ephemeral key. The generated ephemeral key is hashed using a secure hash algorithm (such as SHA-256) to produce a fixed-length hash value. This hash value is used to verify the key's integrity and ensure it has not been tampered with. The hash value is digitally signed using the sender's private key (such as RSA or ECDSA) to generate a hash signature. The recipient can verify this signature using the corresponding public key, ensuring the key's origin is authentic and has not been tampered with. A pre-defined key protection method is used to encrypt the ephemeral key with a more secure master key to prevent it from being eavesdropped on. A hardware security module (HSM) or trusted execution environment (TEE) can be used to store and encrypt keys for enhanced security. The above information is then combined and encapsulated into an encrypted data packet for transmission. A secure connection is then established using TLS 1.3 or SSL protocols, ensuring that data cannot be eavesdropped or tampered with during transmission. The sender sends the encrypted data packet via a secure channel to the control center, which decrypts and verifies it. The control center verifies the hash signature to ensure the packet's integrity and decrypts the temporary key using the master key. The decrypted key is then used to further decrypt the packet to restore the original data.

[0150] In this solution, the security of temporary keys is ensured through quantum key distribution, and the key integrity is protected by combining hash signatures and key encryption keys. TLS / SSL encrypted channels are then used for secure transmission, ultimately ensuring that data is not eavesdropped, tampered with, or forged during the entire communication process.

[0151] Example 2

[0152] Figure 2 This is a flow chart of the incremental policy analysis and dynamic model updating method provided in Example 2 of this application. Figure 2 As shown, the specific steps include:

[0153] S201, obtain historical policy text data and historical social media data, and use a fine-grained sentiment analysis model to determine the historical sentiment tendency data of the historical policy text.

[0154] Historical policy text data can be published policy documents, bills, government notices, etc. that have historical background. They are usually official documents issued by governments or organizations. These documents can cover various areas, such as economic, social, educational, and environmental policies.

[0155] Historical social media data can be publicly available user-generated content published in the past on social platforms such as Twitter, Facebook, and Weibo. This content includes comments, posts, reposts, likes, etc. User discussions and feedback on policies or events are key elements of social media data.

[0156] Fine-grained sentiment analysis is an advanced sentiment analysis technique that not only determines the overall sentiment of a text (e.g., positive, negative, or neutral) but also deeply analyzes the various details and emotional levels within the text. Fine-grained sentiment analysis identifies the polarity, intensity, and target of sentiment within different parts of a text (e.g., sentences, paragraphs, or keywords). It can distinguish subtle sentiment differences. Specifically, it can identify multiple sentiment entities within a text, such as "Government policy reforms bring better opportunities (positive)" and "Tax increases place a heavy burden on businesses (negative)." It analyzes not only the polarity of sentiment (positive, negative, neutral) but also quantifies its intensity (e.g., "strongly supportive" vs. "somewhat supportive"). It can also analyze the specific targets of sentiment, such as policy reforms, taxes, and businesses.

[0157] Historical sentiment data is generated through sentiment analysis models and represents the sentiment, polarity, and intensity of text related to a specific period or event. For example, when analyzing historical policy texts or social media comments, sentiment data might include the sentiment label (positive, negative, neutral) and intensity (strong, moderate, weak) for each piece of historical policy text.

[0158] All relevant policy text documents, government announcements, and legal documents can be collected. User comments and posts related to relevant policies or events can be captured from social media platforms (such as Twitter, Facebook, and Weibo). This can be done through APIs or web scraping. The captured historical social media data is denoised to remove irrelevant content (such as advertisements and meaningless symbols) and normalized (e.g., through lemmatization). Historical policy text and social media data are split into individual sentences, paragraphs, or comments for sentence-by-sentence sentiment analysis. Transformer-based models such as BERT and RoBERTa are then used for fine-grained sentiment analysis. Each policy clause or paragraph is input into the model to obtain its sentiment orientation (positive, negative, neutral) and intensity. The model analyzes the sentiment target of each clause. For example, the sentiment target of "tax rate increase" is likely "tax policy," and its sentiment polarity is likely negative. The intensity of the sentiment is further analyzed, such as whether it expresses strong support or strong opposition. The analysis results are then aggregated to obtain sentiment data for each policy clause in each historical policy text data set. For example: The sentiment tendency towards a certain policy clause is "negative" and the sentiment intensity is "medium".

[0159] S202: Analyze historical social media data using graph neural networks to obtain historical public opinion dissemination paths and historical hot spot analysis data.

[0160] Graph neural networks (GNNs) are a type of neural network model specifically designed for processing graph data. Graph data consists of nodes (vertices) and edges (edges), with nodes representing entities and edges representing relationships between entities. In social media data, nodes can represent users, posts, comments, and so on, while edges represent interactions between them, such as comments, reposts, and likes. GNNs can iteratively learn the relationships between nodes and edges, processing and learning information from graph-structured data. Their primary goal is to leverage the structural information of graphs to learn node and edge representations and use these representations to accomplish tasks such as node classification, graph classification, and link prediction. The relationships between nodes and edges are a crucial input for GNNs. GNNs are able to effectively learn and represent this structured data. In GNNs, nodes exchange information with their neighbors through edges to update their representations. The final representation of each node is determined by the characteristics of both itself and its neighbors.

[0161] Historical public opinion diffusion paths can be traced back to how discussions, comments, and reposts about a topic or event spread among different users on social media. They illustrate the flow of information within the network, revealing which nodes (users) play a significant role in its dissemination, and the speed, breadth, and depth of its spread. For example, a comment about a new policy might be spread to other users through one user's repost, and then further spread through comments and likes. GNN analysis can identify which users or posts played a significant role in this dissemination process.

[0162] Historical hotspot analysis data can be data on the level of attention and discussion surrounding specific events or topics on social media over a specific time period. These hotspots can be specific policies, social events, or popular topics, reflecting public sentiment and focus. For example, when a policy is released, data such as the volume of discussion, likes, and reposts surrounding it reflects its popularity. If discussion of a policy is frequent and intense over a specific period of time, it may indicate that the policy has become a hot topic on social media.

[0163] Social media data typically contains information such as posts, comments, and user interactions (likes, reposts, and follows). This data is inherently unstructured, so to analyze it using graph neural networks (GNNs), it must first be converted into a graph structure. In a graph, nodes represent users or posts, and edges represent interactions between users or connections between posts. Specifically, user nodes represent each social media user, representing their social behavior. Post nodes represent each social media post (including original posts, reposts, and comments), representing a piece of information. User-user edges create an edge if a user follows another user or has frequent interactions with a user (such as reposts or comments). User-post edges create an edge if a user publishes a post, comments on it, or likes it. Post-post edges create an edge if a post is a repost or comment on another post, or if two posts have similar content. This creates a social network graph that includes users, posts, and their relationships. Posts and comments in social media data primarily consist of text, which needs to be converted into computable numerical representations for input into graph neural networks for learning. Text feature representation methods include TF-IDF, which calculates the importance of keywords in each post and converts text into feature vectors. Word2Vec / GloVe, based on word embedding methods, converts each post into a distributed representation of words. BERT / Transformer, which uses more advanced pre-trained language models to capture the deeper semantic information of text and make the relationships between posts more explicit. Ultimately, each post is converted into a numerical vector, which serves as the node features of the graph.

[0164] The primary goal of analyzing public opinion diffusion paths is to reveal how information spreads on social media, identify information propagation paths, and identify key nodes (i.e., important users or content) and bottlenecks in information dissemination. Specifically, graph neural networks (GNNs) analyze the graph structure of social networks to compute an embedding vector for each node (user or post). These embedding vectors represent the characteristics of the node within the social network. By calculating these embedding vectors, GNNs can assess the influence of each node and identify which users or posts are core nodes in information dissemination. Influence can be further quantified using node centrality metrics (such as PageRank and Betweenness Centrality), which help identify users or posts with the most forwarding and interactions. GNNs can trace the path of information dissemination within social networks, starting from the initial poster of a post. The dissemination path shows how information gradually spreads through user interactions (such as forwarding, commenting, and likes). GNNs learn the connections within the graph structure and, based on user interactions and the chain of information forwarding, reveal the order and pattern of information dissemination, helping to identify information diffusion paths. During the information dissemination process, some nodes (users or posts) may become bottlenecks in the chain. GNN can identify key nodes that play a restrictive or promoting role in the propagation process by learning the graph structure.

[0165] Historical hotspot analysis aims to identify trending topics within a specific time period, determine their spread within social networks, and analyze the connections between different hot topics. GNNs cluster topics by grouping posts with similar semantics. Using technologies such as graph convolutional networks (GCNs), GNNs can classify social media posts, grouping posts discussing similar topics into the same category. Cluster analysis can help identify hot topics within social media and identify topics that have attracted significant attention within a specific time period. Analyzing the relationships between hot topics can further reveal underlying connections and diffusion patterns between different topics. GNNs can not only analyze static relationships between nodes but also model temporal changes. Temporal graph neural networks (GNNs) can analyze trends in topics over time. Time series analysis helps reveal which topics suddenly become popular at a specific point in time and which topics persist over a longer period. By analyzing the temporal spread of topics, short-term and long-term hot topics can be identified, thereby revealing the evolutionary patterns of hot topics within social media. GNNs can calculate the influence metric for each node in a social network—that is, the posts or topics that are discussed, forwarded, or interacted with by the most users. Influence analysis helps identify which posts have the greatest diffusion effect on social media. By analyzing the propagation paths and interaction frequencies between nodes, we can identify the most viral content and topics on social media. Graph neural networks can be used to analyze historical social media data to obtain historical opinion propagation paths and historical hotspot analysis data. Public opinion propagation path analysis reveals how information spreads within social networks, identifying the most influential nodes, propagation paths, and potential bottlenecks. Historical hotspot analysis helps identify and analyze trending topics on social media, revealing relationships between topics, their temporal evolution, and the distribution of their influence. This approach provides a deep understanding of the mechanisms by which information and topics spread on social media, providing powerful support for public opinion monitoring, trend forecasting, and social network optimization.

[0166] S203, constructing a policy impact analysis model based on historical policy text data, historical social media data, historical sentiment tendency data, historical public opinion dissemination paths, and historical hot spot analysis data.

[0167] A policy impact analysis model can be a comprehensive model that combines historical policy text data, historical social media data, historical sentiment data, historical public opinion diffusion data, and historical hotspot analysis data to conduct multi-dimensional public opinion analysis. This model can help understand the public response to policies, assess the impact of policies on social opinion, and predict the hot topics and sentiment trends that policies may trigger.

[0168] A dataset can be formed from historical policy text data and social media data. Historical sentiment data, historical public opinion diffusion paths, and historical hot topic analysis data are used as labels. Natural Language Processing (NLP) techniques are used for preprocessing, including cleaning, tokenization, and stop word removal. Word embeddings (such as Word2Vec and BERT) are used to convert text into vector representations that are understandable by machine learning models. Social media data is cleaned to remove irrelevant content. User interaction data, comments, and forwarding content are extracted, sorted by time, and the sentiment of each piece of social media data is analyzed. Feature extraction using TF-IDF, topic modeling (such as LDA), or BERT is used to convert policy text into numerical feature vectors. Behavioral features (such as likes, forwarding, and comment counts), social interaction features (such as user influence and interaction network structure), and user sentiment features (using sentiment analysis to extract positive and negative sentiment indicators) are extracted from social media data. Social media data has temporal properties, and the model can analyze the time of posting, commenting, and the temporal evolution of hot topics. Sentiment analysis tools are used to classify each piece of social media data, obtaining a sentiment label (positive, negative, or neutral) and quantifying the intensity of the sentiment. By constructing a social network graph, centrality metrics (such as PageRank and Betweenness Centrality) are calculated along the propagation path to identify key nodes and transmission chains. Cluster analysis and time series analysis are used to identify hot topics, and the spread intensity and relevance of each topic at different time points are output. Based on the characteristics of the data, appropriate machine learning or deep learning models are selected for training. Common models include: Sentiment Analysis Models: Deep learning models such as LSTM and BERT can be used to analyze sentiment on social media data, outputting sentiment labels (positive, negative, neutral) and sentiment intensity. Public Opinion Propagation Path Analysis: Graph Neural Networks (GNNs) are well-suited for processing social network graph data, capturing information propagation paths and key nodes. Models can be trained to identify bottlenecks, key users, or optimal information propagation paths. Hot Topic Analysis: Clustering algorithms (such as K-means) and time series analysis (such as LSTM and Transformer) can identify the evolution of hot topics and their spread trends. Given the diverse target tasks (e.g., sentiment orientation, public opinion propagation paths, hot topics), multi-task learning (MTL) can be used to train a joint model to simultaneously optimize the objectives of multiple tasks. By sharing some network parameters, the model can handle multiple tasks (for example, sentiment analysis, communication path analysis, and hot topic identification) during the same training process. By then splitting the historical dataset into training and validation sets, a multi-task loss function can be designed to calculate losses for each task based on the true labels and optimize the model by weightedly combining these losses.For sentiment analysis tasks, you can use cross-entropy loss, for propagation path tasks, you can use graph analysis loss functions, and for hotspot tasks, you can use clustering error. Use an appropriate optimization algorithm (such as Adam) to train the model and adjust the model parameters to achieve good performance on multiple tasks.

[0169] In this embodiment, since the model is trained based on historical data, after learning, it can be used to predict the emotional trends, public opinion dissemination paths and hot topics after the release of new policies, discover possible public opinion risks in advance, and help governments or companies formulate better communication strategies.

[0170] Based on the above technical solution, optionally, after constructing a policy impact analysis model based on historical policy text data, historical social media data, sentiment tendency data, public opinion dissemination paths, and hot spot analysis data, the method further includes:

[0171] Obtain real-time social media data after the release of the new policy text, input the new policy text data and real-time social media data into the policy impact analysis model, and obtain real-time sentiment tendency data, real-time public opinion dissemination path and real-time hot spot analysis data.

[0172] In this solution, real-time social media data can be user interaction data related to the policy on social media platforms after the new policy is released. Specifically, it can include comments, posts, reposts, likes, etc. Users' discussions and feedback on policies or events are the key content of social media data.

[0173] Real-time sentiment data can be the result of analyzing social media data to measure public attitudes towards policies.

[0174] The real-time public opinion dissemination path can refer to the dissemination trajectory of new policy-related information on social media.

[0175] Real-time hotspot analysis data can be used to identify the most popular policy-related topics on social media.

[0176] It can crawl policy-related content on major social media platforms, obtain real-time social media data, and input the new policy text data and real-time social media data into the policy impact analysis model. The model uses sentiment analysis to identify the public's real-time emotional tendencies towards policies, tracks the information dissemination path in social networks through public opinion dissemination analysis, and extracts hot topics related to policies and their development trends through hot spot analysis. It ultimately outputs real-time emotional tendency data, real-time public opinion dissemination paths, and real-time hot spot analysis data to help optimize policy dissemination and decision-making.

[0177] This plan can timely identify policy-related hot topics, grasp the focus of public attention, quickly respond to social concerns, and improve policy communication effectiveness.

[0178] Example 3

[0179] Figure 3 This is a schematic diagram of the structure of the incremental policy analysis and dynamic model update system provided in Example 3 of this application. Figure 3 As shown, specifically including:

[0180] The policy text vectorization module 301 is used to obtain the old version of the policy text and the new version of the policy text, determine the first vector of each clause of the old version of the policy text, and determine the second vector of each clause of the new version of the policy text;

[0181] The incremental change calculation module 302 is used to calculate the incremental change of each clause of the new version of the policy text based on the first vector, the second vector and a preset incremental change calculation formula;

[0182] Incremental update module 303, configured to determine the importance weight of each clause based on the incremental change and a preset weight determination rule, obtain initial model parameters, calculate model update parameters based on the initial model parameters, the importance weight, the first vector, the second vector, and a preset incremental update formula, and incrementally update the preset policy parsing model based on the model update parameters;

[0183] The policy parsing module 304 is used to input the old version of the policy text and the new version of the policy text into the updated preset policy parsing model to obtain policy parsing data.

[0184] The incremental policy analysis and dynamic model updating system provided by the embodiment of the present application can achieve Figure 1 To avoid repetition, the various processes implemented in the method embodiment are not described here.

[0185] Example 4

[0186] like Figure 4 As shown, an embodiment of the present application also provides an electronic device 400, including a processor 401, a memory 402, and a program or instruction stored in the memory 402 and executable on the processor 401. When the program or instruction is executed by the processor 401, each process of the above-mentioned incremental policy parsing and dynamic model updating method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0187] It should be noted that the electronic devices in the embodiments of the present application include the mobile electronic devices and non-mobile electronic devices mentioned above.

[0188] Example 5

[0189] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned cable installation process based on the tension adaptive control system embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0190] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk.

[0191] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or system comprising the element. In addition, it should be noted that the scope of the methods and systems in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved. For example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0192] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of this application, or the part that contributes to the existing technology, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of this application.

[0193] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

[0194] The above are only preferred embodiments of the present application and the technical principles employed. The present application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that are possible for those skilled in the art will not depart from the scope of protection of the present application. Therefore, although the present application has been described in more detail through the above embodiments, the present application is not limited to the above embodiments and may include more other equivalent embodiments without departing from the concept of the present application. The scope of the present application is determined by the scope of the claims.

Claims

1. An incremental policy analysis and dynamic model updating method, characterized by: The method comprises: Obtaining an old version of the policy text and a new version of the policy text, determining a first vector for each clause of the old version of the policy text, and determining a second vector for each clause of the new version of the policy text; Based on the first vector, the second vector, and the preset incremental change calculation formula, the incremental change of each clause in the new policy text is calculated; wherein the preset incremental change calculation formula is: in, is the incremental change; i is the clause index; is the importance weight of each clause; is the second vector; is the first vector; is the preset smoothing factor; The importance weight of each clause is determined based on the incremental change and a preset weight determination rule, the initial model parameters are obtained, the model update parameters are calculated based on the initial model parameters, the importance weight, the first vector, the second vector, and a preset incremental update formula, and the preset policy parsing model is incrementally updated based on the model update parameters; wherein the preset incremental update formula is: in, Update parameters for the model; are the initial parameters of the model; is the preset learning rate; N is the total number of clauses; i is the clause index; is the importance weight; is a loss function used to measure the differences between the new and old policy texts; The loss function is the initial parameters of the model The gradient of , which represents the impact of each term on the final model when updating the model parameters; The old version of the policy text and the new version of the policy text are input into the updated preset policy parsing model to obtain policy parsing data.

2. The method according to claim 1, characterized in that in, After obtaining the policy analysis data, the method further includes: Obtain historical policy text data and historical social media data, and use fine-grained sentiment analysis models to determine the historical sentiment tendency data of historical policy texts; Use graph neural networks to analyze historical social media data to obtain historical public opinion dissemination paths and historical hot spot analysis data; A policy impact analysis model is constructed based on historical policy text data, historical social media data, historical sentiment tendency data, historical public opinion dissemination paths, and historical hot spot analysis data.

3. The method according to claim 2, characterized in that in, After constructing a policy impact analysis model based on historical policy text data, historical social media data, sentiment trend data, public opinion dissemination paths, and hot spot analysis data, the method further includes: Obtain real-time social media data after the release of the new policy text, input the new policy text data and real-time social media data into the policy impact analysis model, and obtain real-time sentiment tendency data, real-time public opinion dissemination path and real-time hot spot analysis data.

4. The method according to claim 1, wherein in, After obtaining the policy analysis data, the method further includes: Determine the components of the policy parsing data, and determine the encryption level of each component of the policy parsing data according to a preset encryption strategy; Determining the sensitivity and update frequency of each component of the policy parsed data, and determining the encryption strength of each component based on the sensitivity and update frequency; Determining contextual relationships among components of the policy parsed data, and determining dynamic encryption strategies for the components based on the contextual relationships; Obtain access rights information for different user roles, and set access control rights for different user roles based on the access rights information; The policy parsing data is encrypted according to the encryption level, encryption strength, dynamic encryption strategy and the access control authority to obtain encrypted data packets corresponding to different user roles, and the encrypted data packets are transmitted to the control center.

5. The method according to claim 4, characterized in that in, Before transmitting the encrypted data packet to the control center, the method further includes: Use quantum key distribution technology to generate a temporary key, perform a hash operation on the temporary key, and generate a hash value of the temporary key; Sign the hash value to obtain a hash signature, and encrypt the temporary key using the preset key protection method; Forming a transmission encrypted data packet according to the encrypted temporary key, hash value, hash signature and encrypted data packet; Accordingly, transmitting the encrypted data packet to the control center includes: The encrypted data packet is transmitted to the control center via a TLS / SSL encrypted channel.

6. An incremental policy analysis and dynamic model updating system, characterized by: The system comprises: A policy text vectorization module is used to obtain an old version of the policy text and a new version of the policy text, determine a first vector for each clause of the old version of the policy text, and determine a second vector for each clause of the new version of the policy text; The incremental change calculation module is used to calculate the incremental change of each clause of the new policy text based on the first vector, the second vector and a preset incremental change calculation formula; wherein the preset incremental change calculation formula is: in, is the incremental change; i is the clause index; is the importance weight of each clause; is the second vector; is the first vector; is the preset smoothing factor; The incremental update module is configured to determine the importance weight of each clause based on the incremental change and a preset weight determination rule, obtain the initial model parameters, calculate the model update parameters based on the initial model parameters, the importance weight, the first vector, the second vector, and a preset incremental update formula, and incrementally update the preset policy parsing model based on the model update parameters; wherein the preset incremental update formula is: in, Update parameters for the model; are the initial parameters of the model; is the preset learning rate; N is the total number of clauses; i is the clause index; is the importance weight; is a loss function used to measure the differences between the new and old policy texts; The loss function is the initial parameters of the model The gradient of , which represents the impact of each term on the final model when updating the model parameters; The policy parsing module is used to input the old version of the policy text and the new version of the policy text into the updated preset policy parsing model to obtain policy parsing data.

7. An electronic device, characterized in that: The method comprises a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the incremental policy parsing and dynamic model updating method as described in any one of claims 1 to 5.

8. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the incremental policy parsing and dynamic model updating method according to any one of claims 1 to 5 are implemented.

Images