Government policy analysis and public opinion intelligent early warning system based on large model

Abstract

The application provides a government policy analysis and public opinion intelligent early warning system based on a large model, relates to the technical field of intelligent government affairs, collects government data and public opinion data, analyzes the government data to obtain a policy information table, constructs an index system of influence research and judgment dimensions and predicts the research and judgment weight trend of the corresponding indexes, analyzes the cause-effect relationship of the indexes based on a cause-effect diagram, performs risk mining based on the government exclusive large model, generates an interpretation file according to a template, monitors the public opinion data in real time, classifies, sentiment analyzes, trajectory trend analyzes and displays the public opinion data, obtains public opinion monitoring results, identifies and warns risks according to the interpretation file and the public opinion monitoring results, generates a risk report and pushes the risk report to relevant departments. The large model is used to analyze the whole life cycle of the reform policy, the risk is judged in combination with the real-time monitored public opinion data, sudden events can be quickly disposed, the satisfaction of residents is improved, and the pressure of government work is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent government affairs, and more particularly to a government policy analysis and public opinion intelligent early warning system based on a large model. Background Technology

[0002] In current government work, policy formulation and implementation must balance scientific rigor, rationality, and social acceptance. However, traditional policy analysis relies on manual interpretation, resulting in low efficiency, narrow coverage, and insufficient prediction of potential impacts. Meanwhile, online public opinion spreads rapidly, utilizes diverse platforms, and exhibits strong emotional characteristics, making it easy for various emergencies and negative public opinion to escalate quickly, posing significant challenges to government emergency response. Particularly in the context of old residential community renovation, numerous problems have caused considerable inconvenience to the normal operation of government work. For example: the compatibility between provincial policies and the actual conditions of the old residential communities is unclear; the financial pressure of renovation is unknown; the renovation process involves diverse resident demands (disagreements over elevator installation, disputes over renovation standards, etc.), easily triggering negative public opinion; some communities have construction safety hazards and delayed renovation progress, lacking real-time early warning mechanisms; and untimely handling of emergencies (such as construction disturbances and mass resident complaints) can easily amplify their impact.

[0003] Therefore, the technical problem this invention aims to solve is how to use a large model to conduct a full-process analysis of the renovation policy for old residential communities, combine real-time monitoring of public opinion related to the renovation to analyze potential risks, quickly handle emergencies, ensure the smooth implementation of the policy, improve residents' satisfaction, and reduce the workload of government officials. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a government policy analysis and public opinion intelligent early warning system based on a large-scale model. By conducting full lifecycle analysis of reform policies through a large-scale model and combining it with real-time monitored public opinion data for risk assessment, the system can quickly handle emergencies, provide efficient and reliable technical support for government decision-making, and help improve the modernization level of government governance.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a government policy analysis and public opinion intelligent early warning system based on a large model, including a communication connection: Data acquisition and processing module: used to collect government data and public opinion data, and to preprocess the collected data; Policy Analysis Module: It has a built-in government-specific large model for parsing government data to obtain policy information tables. It constructs an indicator system that affects the assessment dimensions and predicts the assessment weight trends of the corresponding indicators. It analyzes the causal relationships of the indicators by combining causal graphs. It performs risk mining based on the government-specific large model and generates interpretation documents according to the template. Public opinion monitoring module: used to monitor public opinion data in real time, and to classify, perform sentiment analysis, trajectory trend analysis and display the public opinion data to obtain public opinion monitoring results; Risk identification and early warning module: used to identify and warn of risks based on the interpretation documents and public opinion monitoring results, generate risk reports and push them to relevant departments.

[0006] Preferably, the policy analysis module includes a policy text parsing unit, a policy impact assessment unit, and a policy interpretation generation unit; Policy text parsing unit: used to construct government affairs extraction tasks using Prompt templates, extract core information from government affairs data based on joint information extraction algorithms, and obtain structured policy information tables; Policy Impact Assessment Unit: This unit is used to construct an indicator system for impact assessment dimensions and predict the weight change trends of the indicators. By analyzing the causal transmission links between different assessment dimensions, it performs implicit risk mining on each impact assessment dimension based on the government-specific big data model, and finally generates structured assessment results. Policy Interpretation Generation Unit: This unit generates interpretation documents based on structured policy information tables, suitability scores, and analysis results, combined with preset templates. It also generates short video scripts for government publicity.

[0007] Preferably, the execution process of the policy impact assessment unit includes: Based on the structured policy information table, an indicator system for impact assessment dimensions is constructed, which includes fiscal pressure dimension, implementation dimension, residents' conflict dimension, and derivative impact dimension. Calculate the fit score between the structured policy information table and the local actual situation data, assign judgment weights to the indicators based on the fit score, and dynamically adjust the judgment weights through an attention mechanism; A prediction model is constructed based on the LSTM algorithm to deduce the trend of changes in the judgment weights of the indicators at different stages after policy implementation; By constructing a policy-impact causal graph using causal inference algorithms, we can explore the causal relationship between policy provisions and local impacts, and sort out the causal transmission links between different indicators. The analysis integrates the trends in weight changes and causal transmission links across dimensions, and performs hidden risk mining based on the government-specific big data model, ultimately generating structured analysis results.

[0008] As a preferred approach, a policy-impact causal graph is constructed using a causal inference algorithm to uncover the causal relationship between policy provisions and local impacts, and to trace the causal transmission links between different indicators, including: An undirected graph is constructed by identifying the relationships between variables using a constraint-based causal discovery algorithm. Based on the causal graph skeleton of the undirected graph, the causal direction between variables is determined by combining a scoring-based causal discovery algorithm. The identified causal relationships are logically traced through the transmission chain from policy intervention to direct impact, indirect impact, and final result, and a visual causal diagram is generated.

[0009] Preferably, the cross-dimensional integration of the judgment weight change trend and causal transmission link includes: The judgment weights of the predicted indicators at each stage are transformed into a three-dimensional structured matrix with time dimension, indicator dimension, and weight value, and the slope and inflection point of the weight change are marked. Transform the policies, indicators, and causal relationships in the causal graph into structured data of nodes, directed edges, and effect strengths; Using the policy implementation stage as the timeline, and based on the three-dimensional structured matrix, the judgment weight values ​​of indicators at each stage are associated with the corresponding nodes in the causal chain to generate a fusion result.

[0010] As a preferred approach, when mining latent risks based on the aforementioned government-specific big data model, latent risk mining is achieved through three layers of reasoning. The first layer extends the reasoning of causal links, that is, it mines latent risks at the end of the fused causal links. The second layer performs risk reasoning on the inflection points of weight trends, that is, it mines stage-type latent risks from the inflection points of weight changes. The third layer performs reasoning based on the migration of historical cases, that is, it combines similar cases to mine scenario-type latent risks not covered by local data.

[0011] As a preferred approach, when identifying hidden risks, multiple random sampling simulations based on the Monte Carlo simulation algorithm are conducted to predict the probability and extent of potential negative impacts regarding uncertainties in policy implementation.

[0012] Preferably, the public opinion monitoring module includes a public opinion classification and sentiment analysis unit, a public opinion trajectory trend analysis unit, and a public opinion visualization display unit; Public opinion classification and sentiment analysis unit: used to perform semantic and sentiment analysis on public opinion data, classify according to the semantic analysis results, and score sentiment according to the sentiment analysis results; The Public Opinion Trajectory Trend Analysis Unit is used to prioritize public opinion based on classification results and sentiment scores using the SHP algorithm. It also analyzes the source location, propagation trajectory, and scope of influence of high-priority public opinion by constructing a public opinion propagation map, and generates a public opinion trend map. The public opinion visualization display unit is used to display the public opinion monitoring results in the form of visual charts, including public opinion classification charts, public opinion sentiment distribution charts, public opinion priority ranking, dissemination channel charts, public opinion trend charts, and public opinion heat maps.

[0013] Preferably, the execution process of the public opinion trajectory trend analysis unit includes: Features are extracted based on real-time monitored public opinion data, including classification tags, sentiment scores, dissemination speed, number of people involved, and sensitive keywords; The SHP algorithm is used to mine risk shape patterns under different feature combinations, and all public opinion is quantified and scored according to risk pattern matching. The public opinion is then sorted from high to low scores to determine the priority of public opinion. For high-priority public opinion, a dissemination map is constructed for correlation analysis to locate the source, analyze the dissemination trajectory and the scope of impact.

[0014] As a preferred option, it also includes an emergency response module and a human-computer interaction module; The emergency response module is used to automatically identify emergencies based on public opinion data, trace the source of the emergencies and predict their trends, and generate targeted response plans by calling a preset solution template library according to the type and level of the emergencies. The human-computer interaction module provides customized operations for visual charts, allowing staff to filter and adjust the displayed content as needed. By clicking on any data point in the visual icon, users can directly jump to view detailed information about public opinion and trigger the handling process directly based on public opinion. The module also supports the export and collaborative sharing of charts.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the renovation of an old residential community by applying a large-scale model-based government policy analysis and intelligent public opinion early warning system to this scenario. The large-scale model performs full-process analysis of relevant policies, combines real-time monitored public opinion data for risk assessment, and evaluates the impact of policy implementation. This enables rapid response to emergencies and solves the problems of data silos across departments, inaccurate policy risk assessment, and inaccurate public opinion sentiment analysis. It provides efficient and reliable technical support for government decision-making and helps improve the modernization level of government governance. Attached Figure Description

[0016] Figure 1 This is a block diagram of the government policy analysis and public opinion intelligent early warning system based on a large model, as described in this invention. Figure 2 This is a flowchart of the execution process of the policy impact assessment unit of this invention. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0018] Please see Figure 1As shown, for the scenario of renovating old residential communities, this invention provides a government policy analysis and public opinion intelligent early warning system based on a large model, including communication connections: Data acquisition and processing module: used to collect government data and public opinion data, and to preprocess the collected data; Government data: Through federated learning and data sandbox technology, the data barriers between the Municipal Housing and Construction Bureau (old residential area ledgers, renovation standards, and construction unit information), the Municipal Finance Bureau (renovation fund budget and disbursement details), the Municipal Civil Affairs Bureau (distribution of special groups such as low-income households and the elderly), the Municipal 12345 hotline (complaints and inquiries related to old residential areas), and the street offices of various districts / counties (resident request records) are broken down, and government data is collected from multiple dimensions to achieve cross-departmental data interoperability.

[0019] Public opinion data is collected from multiple channels using a combination of distributed real-time web crawlers and API interfaces, including mainstream media (central and local news websites, Party media platforms), social media platforms (WeChat official accounts, Weibo, Douyin, Kuaishou, Xiaohongshu, Bilibili, Zhihu), forums and online communities, government comments, and other channels (short video comment sections, news posts, self-media platforms). Collected content includes the title, body, publication time, publishing account, number of likes, number of comments, number of reposts, geographical location, and user avatar / nickname. Pre-set filtering rules automatically filter irrelevant public opinion.

[0020] All collected data undergoes preprocessing, including data cleaning to remove duplicates and invalid data, and standardization of data format; anonymization of residents' personal information; and tiered and categorized management of sensitive data such as fund details and construction plans. Policy texts, public opinion data, and risk cases are labeled for fine-tuning of the large model. A distributed database is deployed on a government private cloud to achieve real-time data synchronization and backup, with a retention period of 5 years (compliant with government archive management requirements). Data watermarking and operation log traceability are provided to prevent data leakage and tampering.

[0021] Policy Analysis Module: It has a built-in government-specific large model for parsing government data to obtain policy information tables. It constructs an indicator system that affects the assessment dimensions and predicts the assessment weight trends of the corresponding indicators. It analyzes the causal relationships of the indicators by combining causal graphs. It performs risk mining based on the government-specific large model and generates interpretation documents according to the template. The policy analysis module includes a policy text parsing unit, a policy impact assessment unit, and a policy interpretation generation unit; Policy text parsing unit: used to construct government affairs extraction tasks using Prompt templates, extract core information from government affairs data based on the Joint Information Extraction (UIE) algorithm, and obtain a structured policy information table; To address the core extraction requirements of policies for the renovation of old residential communities, two types of Prompt templates were designed: entity extraction and relationship extraction. These templates transform vague extraction requirements into instructions that the UIE algorithm can understand. When designing Prompt templates, it is necessary to clearly define the extraction objects and align them with government service scenarios. The system automatically matches templates based on policy types and supports users to add custom templates.

[0022] Open-source general-purpose information extraction models (such as Baidu ERNIE-UIE and PaddleNLP UIE) were selected as the base model, with a preference for lightweight versions (such as UIE-small) to meet the computing power requirements of government private clouds. A small-sample fine-tuning strategy was employed, using a pre-labeled government policy dataset containing 1000 provincial / municipal policies on the renovation of old residential communities. The annotations were "entity-relationship" triples, and the fine-tuning goal was to optimize the model's understanding of government terminology. The pre-processed policy segmented text was concatenated with the corresponding Prompt template to form the model input format. This concatenated input text was then fed into the fine-tuned UIE model, which output the extraction results. The model's output extraction results were integrated, and errors were eliminated through rule validation. The validated extraction results were then organized according to the standard format of government data to generate a structured policy information table.

[0023] For the renovation of old residential communities, the pre-set fields of the structured policy information table include: policy number, issuing department, issuing time, renovation scope, subsidy ratio, responsible department, approval time limit, implementation period, and supporting measures.

[0024] Policy Impact Assessment Unit: This unit is used to construct an indicator system for impact assessment dimensions and predict the weight change trends of the indicators. By analyzing the causal transmission links between different assessment dimensions, it performs implicit risk mining on each impact assessment dimension based on the government-specific big data model, and finally generates structured assessment results. The execution process of the policy impact assessment unit includes: Based on the structured policy information table, an indicator system for impact assessment dimensions is constructed, which includes fiscal pressure dimension, implementation dimension, residents' conflict dimension, and derivative impact dimension. Based on the core fields of the structured policy information table (scope of renovation, subsidy ratio, approval time limit, etc.) and local actual data (finance, residents, construction, etc.), it is broken down into 4 primary dimensions and 12 quantifiable sub-indicators. The fiscal pressure dimension includes three indicators: the proportion of local subsidies, the renovation funding gap rate, and the pace of fund disbursement, used to predict the short-term / long-term pressure of policy implementation on local finances; the implementation dimension includes three indicators: the matching degree of construction unit capacity, the sufficiency rate of building material supply, and the complexity of the approval process, used to predict the implementation risks such as construction delays and non-compliance with standards during policy implementation; the resident conflict dimension includes three indicators: the disagreement rate on elevator installation, the dispute rate on renovation standards, and the resident survey support rate, used to predict resident conflicts and negative public opinion caused by policy implementation; the derivative impact dimension includes three indicators: the rate of complaints about construction disturbance, the efficiency of surrounding traffic, and the degree of temporary environmental pollution, used to predict the derivative negative impacts on surrounding residents and the environment during the policy construction process.

[0025] Calculate the fit score between the structured policy information table and the local actual situation data, assign judgment weights to the indicators based on the fit score, and dynamically adjust the judgment weights through an attention mechanism; When calculating the suitability score, the requirements of the structured policy information form are matched one by one with the actual local data, and the score is calculated according to the scoring rules of the sub-indicators. For example, if the policy requires a "30% subsidy ratio for elevator installation" and the actual local subsidy ratio is 25%, the corresponding sub-indicator "local subsidy ratio" is calculated, and the score for this sub-indicator is calculated as (25% / 30%) × 100 ≈ 83 points. Using the Analytic Hierarchy Process (AHP), experts from housing and construction, finance, and community sectors are invited to conduct pairwise importance comparisons of the four primary dimensions, construct a judgment matrix, calculate the dimension weights, first calculate the average score for each primary dimension, and then weight them according to the dimension weights to obtain the comprehensive suitability score. For example, if the average score for fiscal carrying capacity is 80 points, implementation is 85 points, resident acceptance is 75 points, and derivative impact suitability is 90 points, then the comprehensive score is 80 × 0.3 + 85 × 0.25 + 75 × 0.3 + 90 × 0.15 = 80.25 points (highly suitable).

[0026] Specifically, the process of assigning assessment weights to the indicators based on the suitability score, and dynamically adjusting the assessment weights through an attention mechanism, includes: Based on the suitability score, initial weights are assigned to the indicators using a score-back mapping method. Based on the calculated suitability scores of the four assessment dimensions, a score-back mapping method is used to allocate weights. Dimensions with lower suitability scores indicate greater risk / impact and should be assigned higher assessment weights. Specifically, first, the inverse value of each dimension's score is calculated (inverse value = 100 - dimension score). Then, the proportion of inverse values ​​for each dimension is calculated as the initial weight (initial weight = inverse value of that dimension / sum of inverse values ​​of all dimensions). For each sub-indicator under each dimension, the initial weight is allocated based on the corresponding suitability sub-score using the same logic, ensuring that the weights are distributed across the specific assessment indicators.

[0027] Construct an indicator feature matrix, capture the dynamic correlation between indicators based on the attention mechanism, and calculate the attention weights; The quantitative data of each indicator is transformed into feature vectors to form an indicator feature matrix (dimension: number of indicators × feature dimension). The importance of the correlation between each indicator is calculated using a self-attention mechanism to obtain the attention weight.

[0028] The initial weights and attention weights are weighted and fused to obtain the final dynamic weights.

[0029] A prediction model is constructed based on the LSTM algorithm to deduce the trend of changes in the judgment weights of the indicators at different stages after policy implementation; The policy implementation cycle is divided into five stages: preparation, initial construction, mid-construction, completion, and operation. The prediction model is trained using indicator data from historical policy cases at each stage and the corresponding assessment weights. The model structure consists of four layers: the input layer receives the indicator relationships for each stage (data dimension: number of time steps × number of indicator features); 1-2 layers of LSTM units are used to capture long-term dependencies in the time series data; a fully connected layer maps the output of the LSTM layer to the weight values ​​of each assessment indicator; the output layer outputs the weights of all assessment indicators for the corresponding stage, and a Sigmoid activation function is used to scale the output to the 0-1 range. The predicted weight changes are displayed as a line graph, visually showing the trend of each indicator weight changing with the policy implementation stage.

[0030] A policy-impact causal graph is constructed using a causal inference algorithm to uncover the causal relationship between policy provisions and local impacts, and to trace the causal transmission links between different indicators. The steps include: Undirected graphs are constructed by identifying relationships between variables using constraint-based causal discovery algorithms (such as PC algorithms). First, identify the variables to be analyzed (policy-related), including policy intervention variables, local impact indicators, and confounding variables. Then, use the PC algorithm to identify the relationships between variables. This involves using conditional independence tests to progressively eliminate cases where two variables are not related after controlling for other variables, ultimately retaining the truly related variable pairs. Connect these retained variable pairs with undirected edges to form a preliminary association network (undirected graph). For example, if the elevator installation subsidy ratio and the number of construction companies involved are still significantly related after controlling for residents' income, then this variable pair is retained; if the policy release time and residents' complaint rate are not related after controlling for building material prices, then this variable pair is excluded.

[0031] Based on the causal graph skeleton of the undirected graph, the causal direction between variables is determined by combining a scoring-based causal discovery algorithm (such as the GES algorithm). This embodiment uses the GES algorithm to calculate the fit score of the causal direction. By assigning different directions to undirected edges (such as "A→B" or "B→A"), and calculating the data fit score under each direction, the higher the score, the more it conforms to the true causal relationship. Time series data is then incorporated for correction, adding directions to the undirected edges on the undirected graph to form directed relationships of "policy → indicator → indicator".

[0032] The identified causal relationships are logically traced through the transmission chain from policy intervention to direct impact, indirect impact, and final result, and a visual causal diagram is generated (the chain is presented in the form of "nodes (variables) + directed edges (causal direction)" and the strength of the causal effect is marked).

[0033] The analysis integrates the trends in weight changes and causal transmission links across dimensions, and performs hidden risk mining based on the government-specific big data model, ultimately generating structured analysis results.

[0034] Among them, the cross-dimensional fusion of the judgment weight change trend and causal transmission link includes: The judgment weights of the predicted indicators at each stage are transformed into a three-dimensional structured matrix with time dimension, indicator dimension, and weight value, and the slope and inflection point of the weight change are marked. Each stage of the policy implementation cycle corresponds to a time node, which is used as the row index of a matrix. Twelve sub-indicators are used as column indices, and the weights of each stage's sub-indicators predicted by the LSTM model are filled in to form a three-dimensional structured matrix. By performing a difference operation on the weights of the same indicator in adjacent stages, the weight slope is calculated as (weight of the next stage - weight of the previous stage) / time interval, reflecting the rate of weight change. When the absolute value of the slope of an indicator is ≥15% / month (a preset threshold), or when the slope direction reverses for two consecutive stages (e.g., from a positive slope to a negative slope), it is determined as a weight inflection point. The matrix labeled with the weight slope and inflection points is stored.

[0035] Transform the policy, indicator, and indicator causal relationships in the causal graph into a structured data table with nodes (variables), directed edges (causal directions), and effect strengths (causal coefficients); The nodes include policy intervention nodes (such as a 30% subsidy for elevator installation), indicator nodes, and outcome nodes (such as resident satisfaction). Each node is labeled with a unique ID and type (e.g., ID=001, type=policy intervention). The causal relationship is represented by "starting node ID → ending node ID", which clarifies the transmission direction of "policy → indicator" and "indicator → indicator". The causal effect is calculated through causal inference algorithms (such as causal forest, DID), and "the change in the outcome variable for every 1 unit change in the dependent variable" is expressed as a percentage or numerical value.

[0036] Traverse the policy-impact causal graph, extracting the "node-edge-effect strength" combination one by one. For causal relationships with confounding variables (such as "building material prices" simultaneously affecting "number of construction unit participants" and "funding gap rate"), additionally label the "confounding variable ID and influence method," for example, "confounding variable = building material prices (ID010), affected nodes = ID005 (number of construction unit participants), ID003 (funding gap rate)." Construct a structured data table indexed by "causal link ID," with fields including link ID, starting node ID / name, ending node ID / name, directed edge direction, effect strength, and confounding variable label.

[0037] Using the policy implementation phases (preparation period / initial construction / mid-construction / completion period / operation period) as the timeline, and based on the aforementioned three-dimensional structured matrix, the judgment weight values ​​of indicators for each phase are associated with the corresponding nodes of the causal link, generating a fusion result (a five-dimensional fusion table of "phase-indicator-weight-causal link-effect intensity").

[0038] If the weight of an indicator increases in a certain stage, the causal effect is strengthened according to the formula: "Adjusted effect strength = Original effect strength × (Current stage weight / Baseline stage weight)". A five-dimensional fusion table is formed by integrating five dimensions: "stage, indicator name, indicator weight, causal link (start-end point), and adjusted effect strength". The five-dimensional fusion table is then validated and corrected based on local data.

[0039] By connecting with local government knowledge bases (such as regulations related to the renovation of old residential areas, methods for managing fiscal funds, and norms for handling resident complaints), the reasoning boundaries of the large-scale model are constrained. Using a five-dimensional fusion table resulting from cross-dimensional integration, local actual data, and historical cases of similar policies (including failed / risk cases) as input, the government-specific large-scale model achieves hidden risk discovery through three layers of reasoning. The first layer extends the causal chain, i.e., discovering latent risks at the end of the fused causal chain; the second layer performs risk reasoning on the inflection points of weight trends, i.e., discovering stage-specific hidden risks from the inflection points of weight changes; the third layer performs reasoning based on the migration of historical cases, i.e., combining similar cases to discover scenario-specific hidden risks not covered by local data. Hidden risks are output in a unified format to ensure they can be directly incorporated into the assessment results, for example, outputting them in tabular form, including risk type, risk description triggering conditions, impact stage, related indicators / chains, and degree of impact.

[0040] The cross-dimensional fusion results and the results of hidden risk mining are integrated into standardized and implementable structured judgment results, generating a judgment report that includes basic information, judgment on weight change trends, judgment on causal transmission links, judgment on explicit risks, judgment on hidden risks, and risk response suggestions. The report is then output after manual verification.

[0041] In the process of identifying hidden risks, multiple random sampling simulations are conducted based on the Monte Carlo simulation algorithm for uncertainties in policy implementation (such as the allocation of fiscal funds, the supply of building materials, and the support rate of residents) to predict the probability and degree of potential negative impacts.

[0042] First, identify the uncertainties (variables) in policy implementation and the corresponding potential negative impact indicators (outcomes). Set a realistic probability distribution for each uncertainty and use the Monte Carlo algorithm to conduct multiple random sampling simulations. That is, use a computer program to randomly sample a large number of uncertainties (usually more than 1,000 times). Each sampling yields a set of values ​​for uncertainties. Then, combine the causal transmission links to calculate the corresponding negative impact indicators. Statistically analyze the results of 1,000 simulations and output a quantitative conclusion on the risk.

[0043] Policy Interpretation Generation Unit: This unit generates interpretation documents based on structured policy information tables, suitability scores, and analysis results, combined with preset templates. It also generates short video scripts for government publicity.

[0044] Based on a pre-set government template library, three types of targeted interpretation documents are automatically generated. The templates can be customized and modified. The interpretation document types include the public version (focusing on residents' concerns and operational guidelines, using simple language), the grassroots implementation version (for implementing entities such as street offices and communities, clarifying work processes, division of responsibilities, and time nodes), and the expert / enterprise version (for construction units, social capital, etc., focusing on policy support, access conditions, and cooperation models).

[0045] Public opinion monitoring module: used to monitor public opinion data in real time, and to classify, perform sentiment analysis, trajectory trend analysis and display the public opinion data to obtain public opinion monitoring results; The public opinion monitoring module includes a public opinion classification and sentiment analysis unit, a public opinion trajectory trend analysis unit, and a public opinion visualization display unit; Public opinion classification and sentiment analysis unit: used to perform semantic and sentiment analysis on public opinion data, classify according to the semantic analysis results, and score sentiment according to the sentiment analysis results; The system preprocesses, segments, and filters stop words in real-time monitored public opinion data. Then, it extracts the core semantics of the public opinion using a government-specific large-scale model, identifies the corresponding demands / themes in the text, and matches the semantic analysis results to the corresponding categories by comparing them with pre-defined public opinion classification tags (such as policy consultation, construction complaints, suggestion feedback, negative complaints, etc.). Simultaneously, it identifies the sentiment tendency of the public opinion text and quantifies it into a sentiment score, ultimately outputting a structured result of the public opinion text, classification tags, and sentiment score. For example, based on the text "How to apply for subsidies for community renovation?", the core semantic "policy consultation (subsidy application)" is extracted, classifying it as "policy consultation public opinion".

[0046] The Public Opinion Trajectory Trend Analysis Unit is used to prioritize public opinion based on classification results and sentiment scores using the SHP algorithm. It also analyzes the source location, propagation trajectory, and scope of influence of high-priority public opinion by constructing a public opinion propagation map, and generates a public opinion trend map. The execution process of the public opinion trajectory trend analysis unit includes: Features are extracted based on real-time monitored public opinion data, including classification tags, sentiment scores, dissemination speed, number of people involved, and sensitive keywords; The SHP algorithm is used to mine risk shape patterns under different feature combinations, and all public opinion is quantified and scored according to risk pattern matching. The public opinion is then sorted from high to low scores to determine the priority of public opinion. By combining historical public opinion cases related to the renovation of old residential areas, a risk shape pattern library is constructed by pre-setting high-risk feature combination patterns. The extracted features are matched with the patterns in the risk shape pattern library. The higher the matching degree, the higher the risk score (out of 100). According to the risk score, the top 10% are usually high priority (to be handled within 1 hour), 10%-30% are medium priority (to be handled within 4 hours), and the rest are low priority (regular attention), ensuring that government resources are focused on high-risk public opinion.

[0047] For high-priority public opinion, a dissemination map is constructed for correlation analysis to locate the source, analyze the dissemination trajectory and the scope of impact.

[0048] For high-priority public opinion, we collect data on the initial platform (e.g., local forums, Douyin), posting time, and posting account / IP. Combined with text similarity comparison (e.g., TF-IDF algorithm), we exclude reposted content and pinpoint the initial node. For example, the source of the "construction disturbance post" is "a user posted on a local XX forum at 9:00 AM, the IP address belongs to the street where the community is located." We track the flow of public opinion from the initial node to secondary dissemination nodes (e.g., forwarding accounts, cited media), recording the dissemination time, method (forwarding / commenting / reposting), and account influence (e.g., whether it is a local influential figure). For example, "first posted on the forum → reposted to Douyin by local self-media at 10:00 AM → forwarded in the community residents' group at 10:30 AM → commented on by a local blogger on Weibo at 11:00 AM." We statistically analyze the number of platforms covered by the public opinion (e.g., involving forums, Douyin, and Weibo), geographical scope (e.g., covering two districts in the city), and number of users (e.g., reaching a cumulative total of 12,000 people). We quantify the degree of impact by combining the influence of the dissemination nodes (e.g., a 30% increase in the scope of impact due to a repost by an influential figure). The dissemination path is visualized by using a method of "node (publishing account / platform) + directed edge (direction of dissemination) + label (time / reach of people)", and a public opinion heat trend chart is generated (the number of reposts / comments is counted by hour), which intuitively presents the fermentation trend of public opinion.

[0049] The public opinion visualization display unit is used to display the public opinion monitoring results in the form of visual charts, including public opinion classification charts, public opinion sentiment distribution charts, public opinion priority ranking, dissemination channel charts, public opinion trend charts, and public opinion heat maps.

[0050] Risk identification and early warning module: used to identify and warn of risks based on the interpretation documents and public opinion monitoring results, generate risk reports and push them to relevant departments.

[0051] Based on the assessment results output by the policy impact analysis unit and the public opinion monitoring results output by the public opinion visualization unit, the system identifies and issues early warnings for risks. The system identifies public opinion-related risks every 30 seconds and policy implementation-related risks daily. For example, if a residential community receives six complaints about construction disturbance over two consecutive days, the system automatically identifies this as a "construction disturbance warning risk"; if renovation funds in a district are delayed by 18 days, the system identifies this as a "delayed fund disbursement warning risk." The identified risks are categorized and pushed to relevant departments through multiple channels. After receiving the warnings, relevant departments report their progress through the system. The system automatically tracks the handling process and generates a monthly warning review report, statistically analyzing the number of warnings, the timeliness of handling, and the effectiveness of handling. Based on the review results, the system optimizes the public opinion classification model and risk thresholds, improving the accuracy of subsequent warnings.

[0052] In addition, the system of the present invention also includes an emergency response module and a human-computer interaction module; The emergency response module is used to automatically identify emergencies based on public opinion data, trace the source of the emergencies and predict their trends, and generate targeted response plans by calling a preset solution template library according to the type and level of the emergencies. For example, in an older residential community, a disagreement over elevator installation led to a gathering of over 20 residents at the community entrance, obstructing construction. The related public opinion spread over 500 times on a local forum within one hour. The system, through public opinion monitoring data, automatically identified it as a mass incident and, based on its scope and severity, classified it as a Level III (significant event), requiring intervention from municipal departments. The system quickly located the source of the public opinion (a community residents' group), traced its spread (residents' group → local forum → Douyin), identified key dissemination accounts, and determined the core demand was "opposition to elevator installation, concerns about reduced lighting and property value." Based on historical cases, the system predicted that the public opinion would reach its peak within two hours, and if not addressed promptly, could lead to a larger gathering, simultaneously issuing a warning of the risk of negative spread. The system calls the solution template library and, based on the type of this incident (mass incident) and its level (Level III), automatically generates a personalized handling plan, which clarifies the handling principles (people-oriented, patient communication, and handling in accordance with the law), the division of responsibilities (led by the Municipal Housing and Construction Bureau, with the District Housing and Construction Bureau, the Sub-district Office, and the Community Residents' Committee working together), and the handling steps (on-site reassurance → organizing communication → providing solutions → full-process monitoring).

[0053] The system automatically pushes the generated handling plan to relevant personnel in the Municipal Housing and Construction Bureau, District Housing and Construction Bureau, Sub-district Office, and Community Residents' Committee through multiple channels, including the OA system, government affairs APP, and SMS, requiring responsible personnel to arrive at the scene within 30 minutes. Simultaneously, multi-departmental collaboration interfaces are activated to facilitate real-time communication on handling progress. Relevant departments carry out handling work according to the plan and update the progress in real time through the system. The system also monitors changes in public opinion, tracks the effectiveness of public opinion guidance, and adjusts the warning level after a decrease in negative public opinion. After the incident is handled, the system automatically generates a handling summary report, summarizing the handling process, achievements (residents reached a consensus, construction resumed normally, and public opinion intensity decreased by 80%), existing problems (insufficient initial communication), and proposing improvement suggestions (establishing resident communication groups in each renovated community and regularly reporting progress). At the same time, the optimized handling plan is synchronized to the template library to provide a reference for handling similar incidents in the future.

[0054] The human-computer interaction module provides customized operations for visual charts, allowing staff to filter and adjust the displayed content as needed. By clicking on any data point in the visual icon, users can directly jump to view detailed information about public opinion and trigger the handling process directly based on public opinion. The module also supports the export and collaborative sharing of charts.

[0055] This embodiment takes the implementation of the old residential area renovation policy in this city as a specific scenario. It utilizes a government policy analysis and public opinion intelligent early warning system based on a large model to complete data integration, model adaptation, system deployment, and full-process implementation. It effectively solves core technical problems in this scenario, such as data silos, insufficient model adaptation, delayed public opinion, inaccurate risk warning, and inefficient handling of emergencies. It achieves integrated intelligent support for "policy analysis - public opinion monitoring - risk warning - emergency handling", achieves the expected implementation goals, and verifies the system's practicality, feasibility, and professionalism. It can provide replicable and scalable implementation experience for the implementation of other livelihood-related government policies (such as education, medical care, and employment).

[0056] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A government policy analysis and public opinion intelligent early warning system based on a large model, characterized in that: Including communication connections: Data acquisition and processing module: used to collect government data and public opinion data, and to preprocess the collected data; Policy Analysis Module: It has a built-in government-specific large model for parsing government data to obtain policy information tables. It constructs an indicator system that affects the assessment dimensions and predicts the assessment weight trends of the corresponding indicators. It analyzes the causal relationships of the indicators by combining causal graphs. It performs risk mining based on the government-specific large model and generates interpretation documents according to the template. Public opinion monitoring module: used to monitor public opinion data in real time, and to classify, perform sentiment analysis, trajectory trend analysis and display the public opinion data to obtain public opinion monitoring results; Risk identification and early warning module: used to identify and warn of risks based on the interpretation documents and public opinion monitoring results, generate risk reports and push them to relevant departments.

2. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 1, characterized in that, The policy analysis module includes a policy text parsing unit, a policy impact assessment unit, and a policy interpretation generation unit; Policy text parsing unit: used to construct government affairs extraction tasks using Prompt templates, extract core information from government affairs data based on joint information extraction algorithms, and obtain structured policy information tables; Policy Impact Assessment Unit: This unit is used to construct an indicator system for impact assessment dimensions and predict the weight change trends of the indicators. By analyzing the causal transmission links between different assessment dimensions, it performs implicit risk mining on each impact assessment dimension based on the government-specific big data model, and finally generates structured assessment results. Policy Interpretation Generation Unit: This unit generates interpretation documents based on structured policy information tables, suitability scores, and analysis results, combined with preset templates. It also generates short video scripts for government publicity.

3. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 2, characterized in that, The execution process of the policy impact assessment unit includes: Based on the structured policy information table, an indicator system for impact assessment dimensions is constructed, which includes fiscal pressure dimension, implementation dimension, residents' conflict dimension, and derivative impact dimension. Calculate the fit score between the structured policy information table and the local actual situation data, assign judgment weights to the indicators based on the fit score, and dynamically adjust the judgment weights through an attention mechanism; A prediction model is constructed based on the LSTM algorithm to deduce the trend of changes in the judgment weights of the indicators at different stages after policy implementation; By constructing a policy-impact causal graph using causal inference algorithms, we can explore the causal relationship between policy provisions and local impacts, and sort out the causal transmission links between different indicators. The analysis integrates the trends in weight changes and causal transmission links across dimensions, and performs hidden risk mining based on the government-specific big data model, ultimately generating structured analysis results.

4. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 3, characterized in that, A policy-impact causal graph is constructed using causal inference algorithms to uncover the causal relationship between policy provisions and local impacts, and to trace the causal transmission links between different indicators, including: An undirected graph is constructed by identifying the relationships between variables using a constraint-based causal discovery algorithm. Based on the causal graph skeleton of the undirected graph, the causal direction between variables is determined by combining a scoring-based causal discovery algorithm. The identified causal relationships are logically traced through the transmission chain from policy intervention to direct impact, indirect impact, and final result, and a visual causal diagram is generated.

5. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 3, characterized in that, The cross-dimensional fusion of the aforementioned judgment weight change trends and causal transmission links includes: The judgment weights of the predicted indicators at each stage are transformed into a three-dimensional structured matrix with time dimension, indicator dimension, and weight value, and the slope and inflection point of the weight change are marked. Transform the policies, indicators, and causal relationships in the causal graph into structured data of nodes, directed edges, and effect strengths; Using the policy implementation stage as the timeline, and based on the three-dimensional structured matrix, the judgment weight values ​​of indicators at each stage are associated with the corresponding nodes in the causal chain to generate a fusion result.

6. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 3, characterized in that, When mining hidden risks based on the aforementioned government-specific big data model, the mining of hidden risks is achieved through three layers of reasoning. The first layer extends the reasoning of the causal links, that is, it mines the terminal hidden risks from the fused causal links. The second layer performs risk reasoning on the inflection point of the weight trend, that is, it explores the stage-type hidden risks from the inflection point of weight change; the third layer performs reasoning based on the migration of historical cases, that is, it combines similar cases to explore scenario-type hidden risks not covered by local data.

7. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 3, characterized in that, When identifying hidden risks, multiple random sampling simulations are conducted based on the Monte Carlo simulation algorithm to predict the probability and extent of potential negative impacts regarding uncertainties in policy implementation.

8. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 1, characterized in that, The public opinion monitoring module includes a public opinion classification and sentiment analysis unit, a public opinion trajectory trend analysis unit, and a public opinion visualization display unit; Public opinion classification and sentiment analysis unit: used to perform semantic and sentiment analysis on public opinion data, classify according to the semantic analysis results, and score sentiment according to the sentiment analysis results; The Public Opinion Trajectory Trend Analysis Unit is used to prioritize public opinion based on classification results and sentiment scores using the SHP algorithm. It also analyzes the source location, propagation trajectory, and scope of influence of high-priority public opinion by constructing a public opinion propagation map, and generates a public opinion trend map. The public opinion visualization display unit is used to display the public opinion monitoring results in the form of visual charts, including public opinion classification charts, public opinion sentiment distribution charts, public opinion priority ranking, dissemination channel charts, public opinion trend charts, and public opinion heat maps.

9. The government policy analysis and public opinion intelligent early warning system based on a large model as described in claim 8, characterized in that, The execution process of the public opinion trajectory trend analysis unit includes: Features are extracted based on real-time monitored public opinion data, including classification tags, sentiment scores, dissemination speed, number of people involved, and sensitive keywords; The SHP algorithm is used to mine risk shape patterns under different feature combinations, and all public opinion is quantified and scored according to risk pattern matching. The public opinion is then sorted from high to low scores to determine the priority of public opinion. For high-priority public opinion, a dissemination map is constructed for correlation analysis to locate the source, analyze the dissemination trajectory and the scope of impact.

10. The government policy analysis and public opinion intelligent early warning system based on a large model according to claim 1, characterized in that, It also includes an emergency response module and a human-computer interaction module; The emergency response module is used to automatically identify emergencies based on public opinion data, trace the source of the emergencies and predict their trends, and generate targeted response plans by calling a preset solution template library according to the type and level of the emergencies. The human-computer interaction module provides customized operations for visual charts, allowing staff to filter and adjust the displayed content as needed. By clicking on any data point in the visual icon, users can directly jump to view detailed information about public opinion and trigger the handling process directly based on public opinion. The module also supports the export and collaborative sharing of charts.

Images