Government official document fair competition examination method and system based on causal inference

Abstract

A government official document fair competition review method and system based on causal inference relates to the technical field of data processing, and comprises the following steps: policy text analysis and causal element extraction, causal graph construction, causal effect estimation, anti-fact simulation and index evaluation, and review conclusion generation. According to the fair competition examination method and system for the government official documents, through a method of combining natural language processing, a causal graph model, quantitative effect estimation and anti-fact simulation, automatic, quantitative and explainable intelligent examination on the fair competition risk of the government official documents is realized, and the potential market effect of policy modification can be predicted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a government document fair competition review method and system based on causal inference. BACKGROUND

[0002] When government agencies formulate administrative documents (such as policy documents and regulatory documents) related to the economic activities of market entities, fair competition review needs to be conducted to ensure that policy measures do not hinder market fair competition. In current actual work, the fair competition review of government documents still mainly relies on manual methods: the personnel of the document issuing unit reviews the document content item by item against the review rules to determine whether there are situations that hinder market access, limit the free circulation of goods, affect operating costs, or exclude equal competition rights. This purely manual comparison method has clear guidance at the institutional level, but has obvious shortcomings in practice. On the one hand, manual research and judgment rely on the experience and subjective understanding of the auditors, and different personnel may have different understandings of the policy wording, leading to inconsistent application of the review standard. On the other hand, manual review can only compare clauses statically and lacks overall analysis of the impact on market structure and industry dynamics: the current manual method cannot fully obtain information such as changes in market competition patterns after the implementation of the policy, upstream and downstream impact chains, etc. Especially in grassroots or units with limited professional capabilities, manual review is time-consuming and inefficient, making it difficult to handle a large number of documents in a timely manner. In addition, the traditional method lacks interactive simulation tools and cannot dynamically evaluate the behavior response of different market entities and the evolution of competition situation after the policy is introduced, nor can it answer the hypothetical question of "how will the market effect change if the policy is adjusted".

[0003] To improve the efficiency and accuracy of document review, several AI-assisted review attempts have emerged in recent years. For example, some systems incorporate Natural Language Processing (NLP) technology, using keyword matching and shallow semantic analysis to help identify obvious unfair competitive expressions in official documents. This alleviates the burden on human readers to some extent, but such rule-based or keyword-based methods are limited by predefined lexicons and often fail to address implicit expressions or clauses that indirectly restrict competition. Furthermore, NLP models typically rely on limited labeled data and static rule templates, making them prone to misjudgments or omissions when faced with the complex and varied expressions in actual official documents. Other research attempts to construct multi-agent systems (MAS) to simulate the game effects of policies, assessing the competitive impact of certain policies through the interaction of virtual market participants. However, these multi-agent simulation methods are still in the exploratory stage and have not yet formed mature systems in the field of government document review. Their construction complexity is high, and they lack transparent causal reasoning chain outputs. In addition, the latest Large Language Model (LLM) technology is being attempted for document review, using the semantic understanding capabilities of pre-trained models to identify potential risks and combining them with knowledge graphs to provide legal basis. This approach improves the identification rate of implicit risks and provides a degree of interpretability. However, even with the integration of external knowledge, existing technologies still fail to address the issues of causal interpretation and scenario prediction in review results: model outputs often only indicate the existence of a risk, but cannot quantify the specific impact of that clause on market competition, nor can they simulate how market indicators would change if the clause were modified or removed. In other words, current NLP+rule-based methods and intelligent review tools lack the ability to characterize the "causal chain" between policy measures and competitive outcomes, cannot conduct simulations under "counterfactual" scenarios, and review conclusions remain insufficient in terms of transparency and decision support.

[0004] In summary, existing government document fair competition review techniques have the following main defects and gaps: difficulty in extracting deep causal relationships; lack of quantitative assessment and "hypothetical scenario" analysis; insufficient utilization of external knowledge and historical data; lack of interpretability and guidance in review conclusions; and lack of dynamic learning and self-improvement. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for reviewing fair competition in government documents based on causal inference, in order to overcome the deficiencies in current review technologies, realize automated, quantitative and interpretable intelligent review of fair competition risks in government documents, and predict the potential market effects of policy modifications.

[0006] The embodiments of the present invention are implemented as follows: A method for fair competition review of government documents based on causal inference, comprising: S1. Perform natural language processing on the input policy document text to extract semantic elements describing policy measures, mediating variables, and market outcome indicators, and represent them in a structured form as a set of multiple variables; S2. Based on the extracted set of variables, construct a causal relationship graph model with policy intervention variables as the starting point, market outcome indicator variables as the ending point, and including mediating variables; S3. Based on the causal relationship diagram model, the causal inference algorithm is used to quantitatively calculate the causal effect value of each policy intervention variable on the market outcome indicator variable; S4. Simulate a counterfactual scenario of policy intervention variable change on a causal relationship graph model, calculate and compare the changes in market competition indicators under the counterfactual scenario and the current scenario; S5. Based on causal effect values ​​and counterfactual simulation results, automatically generate a review report that includes risk clause identification, causal path explanation, quantitative impact assessment, and policy optimization recommendations.

[0007] Furthermore, in other preferred embodiments of the present invention, step S1 includes: By using dependency parsing and named entity recognition technology, policy documents are decomposed into semantic triples in the form of "subject-verb-object"; Each semantic triple is labeled with policy attributes using a pre-trained classification model to obtain the corresponding policy intervention variable label. Based on the semantic analysis results, the output is categorized into sets of policy measures variables, sets of intermediary variables, and sets of market outcome indicator variables.

[0008] Furthermore, in other preferred embodiments of the present invention, step S2 includes: Based on a predefined base of fair competition review rules, establish basic causal relationship edges between variable nodes; Calculate the semantic similarity between variable nodes and expand the basic causal relationship edges; When encountering new policy concepts that are not covered, the Bayesian structural learning method is used to dynamically update the node and edge structure of the causal graph model.

[0009] Furthermore, in other preferred embodiments of the present invention, step S3 includes: Parameter estimation was performed using structural equation modeling. The causal effect values ​​included the direct effect of policy intervention variables on market outcome indicator variables, the indirect effect generated through mediating variables, and the total effect.

[0010] Furthermore, in other preferred embodiments of the present invention, step S3 further includes: When historical data is available, propensity score matching is used to screen comparable samples in order to control the impact of confounding variables on the estimation of causal effects. When historical data is insufficient, expert knowledge is introduced as a prior distribution, and the causal effect parameters are calibrated using a Bayesian update method.

[0011] Furthermore, in other preferred embodiments of the present invention, step S4 includes: The do operator is used to intervene in the policy intervention variables in the causal graph model to set up a counterfactual scenario; The probability distribution of market outcome indicator values ​​under counterfactual scenarios is generated using the Monte Carlo sampling method. Calculate and output the difference between the market outcome indicators under the counterfactual scenario and the current scenario. The market outcome indicators include at least one of the following: market concentration HHI index, SME exit rate, and market entry rate.

[0012] Furthermore, in other preferred embodiments of the present invention, step S4 also supports a gradual policy adjustment simulation, by continuously changing the intensity of policy intervention variables, plotting the relationship curve between policy intensity and market outcome indicators, and calculating the marginal effect of policy adjustment.

[0013] A government document fair competition review system based on causal inference, used to implement the aforementioned government document fair competition review method, includes: The text parsing and feature extraction module is used to process policy document texts and output a structured set of variables; The causal graph construction module, connected to the text parsing and feature extraction module, is used to receive a set of variables and build a causal graph model. The causal effect calculation module, connected to the causal graph construction module, is used to execute causal inference algorithms based on the causal relationship graph model and output causal effect values. The counterfactual simulation engine, connected to the causal graph construction module and the causal effect calculation module, is used to perform simulations and evaluations of counterfactual scenarios; The report generation module, connected to the causal effect calculation module and the counterfactual simulation engine, is used to automatically generate structured review reports.

[0014] Furthermore, in other preferred embodiments of the present invention, the text parsing and feature extraction module integrates a BERT-BiLSTM model and a dependency parser to extract and classify semantic triples.

[0015] Furthermore, in other preferred embodiments of the present invention, the cause-effect graph construction module includes: The rule base unit stores causal rules based on the fair competition review guidelines; Semantic similarity calculation unit, used to expand causal relationship edges; The structure learning unit is used to dynamically update the causal graph structure when encountering new policy concepts.

[0016] The beneficial effects of the embodiments of the present invention are: This invention provides a method and system for reviewing fair competition in government documents based on causal inference, comprising the following steps: policy text parsing and causal element extraction, causal graph construction, causal effect estimation, counterfactual simulation and indicator evaluation, and review conclusion generation. This method and system for reviewing fair competition in government documents achieves automated, quantifiable, and interpretable intelligent review of fair competition risks in government documents by combining natural language processing, causal graph models, quantitative effect estimation, and counterfactual simulation, and can predict the potential market effects of policy modifications. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The abbreviations and key terms used in the following embodiments are defined as follows: Example

[0019] This embodiment provides a method for fair competition review of government documents based on causal inference, which includes: S1. Perform natural language processing on the input policy document text to extract semantic elements describing policy measures, mediating variables, and market outcome indicators, and represent them in a structured form as a set of multiple variables.

[0020] Specifically, natural language processing techniques (including BERT-BiLSTM dependency parsing and named entity recognition) are first used to parse the syntactic structure of the official document, decomposing the text into semantic triples of "subject-verb-object".

[0021] For example, the clause "collecting security deposits from out-of-town enterprises" can be analyzed as follows: (Subject: Out-of-town enterprises; Measures: Collection of security deposits; Impact: Increased entry costs).

[0022] In this model, “measures” corresponds to the causal variable X, and “impact” corresponds to the outcome variable Y.

[0023] Each semantic triple is labeled with policy attributes using a pre-trained classification model to obtain the corresponding policy intervention variable label. Furthermore, the system uses a classifier f cls Label each triple with policy attributes:

[0024] Based on the semantic analysis results, the output is categorized into three sets: policy measure variables, mediating variables, and market outcome indicator variables. These are represented as follows: Collection of policy measures: ; Set of mediator variables: ; Set of outcome metrics: .

[0025] These sets will serve as node inputs for subsequent causal graph modeling. Compared to traditional keyword matching methods, this step, through deep semantic extraction, can identify implicitly expressed policy intentions (such as "supporting local enterprises" implying "exclusivity"), significantly improving the comprehensiveness and accuracy of policy element identification.

[0026] Furthermore, the fair competition review method for government documents based on causal inference provided in this embodiment also includes: S2. Based on the extracted set of variables, construct a causal relationship graph model with policy intervention variables as the starting point, market outcome indicator variables as the ending point, and mediating variables.

[0027] Specifically, the causal relationship diagram model is as follows: G =( V , E ),in, V For a set of variable nodes, E Let be a set of directed edges representing causal relationships. This model can lay the mathematical foundation for subsequent causal inferences.

[0028] Furthermore, the set of variable nodes V Represented as , in, As a variable for policy intervention; Mediating variables (such as firm behavior, market concentration); For outcome indicators (such as HHI, entry rate).

[0029] Furthermore, based on a predefined base of fair competition review rules, basic causal relationship edges are established between variable nodes; edge setE Generated jointly by a rule base and a semantic similarity model. If a policy provision involves "restricting access," the system generates: .

[0030] Calculate the semantic similarity between variable nodes and expand the basic causal relationship edges. Optionally, first construct a basic rule set based on the items of the "Implementation Rules for Fair Competition Review" (such as "restriction on market access → increase in market concentration"), and then use BERT vector similarity to expand the connection edges of semantically similar clauses.

[0031] When encountering new policy concepts that are not covered, the Bayesian structural learning (BIC scoring) method is used to dynamically update the node and edge structure of the causal graph model.

[0032] This causal graph maps semantic text to a formal model, making policy impact paths visible and interpretable. Unlike traditional NLP classification, this method provides a structured representation of causal chains, offering a computable framework for subsequent interventions and counterfactual reasoning.

[0033] Furthermore, the fair competition review method for government documents based on causal inference provided in this embodiment also includes: S3. Based on the causal relationship diagram model, the causal inference algorithm is used to quantitatively calculate the causal effect value of each policy intervention variable on the market outcome indicator variable.

[0034] This step aims to calculate the true causal impact of each policy intervention on competition indicators, i.e., the average treatment effect (ATE). Unlike traditional NLP review models that can only determine "there may be a risk," this invention, through quantitative causal estimation, can answer key questions such as "how much did the policy reduce competition?", "is it significant?", and "through which path does it work?"

[0035] Specifically, structural equation modeling is used for parameter estimation. The causal effect values ​​include the direct effect of policy intervention variables on market outcome indicator variables, the indirect effect generated through mediating variables, and the total effect.

[0036] Assume the structural equation model is as follows: Y = α + βX + γM + ε , in, Y For outcome variables (such as market concentration HHI or entry rate); X As a variable for policy intervention; M As a mediating variable; α For constant terms; β The direct effect coefficient represents the policy's direct effect coefficient.X right Y The direct impact on strength; γ The mediation coefficient reflects the mediating effect. X pass M Indirect impact Y The degree; ε This is the random error term.

[0037] right X The mean treatment effect (ATE) is defined as: , This refers to the difference between the expected impact of policy implementation and the expected impact on market indicators.

[0038] The system extracts policy-related economic indicators, such as the number of enterprises, market share, subsidy amount, and price level, from historical data and a review case library. It automatically performs normalization, missing value completion, and time alignment to ensure comparability across samples.

[0039] Furthermore, when historical data is available, propensity score matching (PSM) is used to screen comparable samples to control for the influence of confounding variables on the estimation of causal effects. , in, Z To control for variables (industry, region, time, etc.), the system performs a one-to-one match between the implementing group and the non-implementing group to ensure that the two groups are similar in terms of external policy conditions, thereby isolating the influence of the policy itself.

[0040] Calculate the average difference in the matched samples: , Furthermore, when historical data is insufficient, expert knowledge is introduced as a prior distribution. β 0, and calibrate the causal effect parameters using a Bayesian update method: , in, To observe the noise variance, This represents the prior variance.

[0041] Output the direct effect, indirect effect, and total effect, and calculate the p-value and confidence interval to determine the strength and direction of the influence.

[0042] The system maps the results onto a causal path graph, using different colors to mark positive or negative effects, line thickness to represent effect strength, and node labels to display the numerical value of the effect (e.g., ...). β =0.032 indicates a 3.2% increase in concentration.

[0043] The system calculates the causal strength sequentially based on each edge relationship in the causal graph, solves for each parameter using simultaneous equations via SEM, and performs intervention simulations using do-Calculus. If potential confounding variables (such as the economic cycle) exist, propensity score is introduced to eliminate bias. The output is a table of effect values ​​corresponding to the policy provisions, including... β , R 2. Significance level p .

[0044] This step effectively solves the problem of not being able to quantify the relationship between policy provisions and market competition; it supports robust estimation based on limited samples; and the output causal coefficient can be directly interpreted as the "intensity of policy impact," improving the readability of the conclusions.

[0045] Furthermore, the fair competition review method for government documents based on causal inference provided in this embodiment also includes: S4. Simulate a counterfactual scenario of policy intervention variable change on a causal graph model, calculate and compare the changes in market competition indicators under the counterfactual scenario and the current scenario.

[0046] After obtaining the causal effects of policies, the system further uses counterfactual reasoning techniques to simulate "how market indicators would change if the policy were canceled or adjusted," providing a predictive basis for policy revisions. This transforms the system from a simple ex-post review to "predictable policy simulation."

[0047] Further, step S4 includes: The do operator is used to intervene in the policy intervention variables in the causal graph model, setting up a counterfactual scenario.

[0048] Specifically, the counterfactual calculation model is as follows: When policy intervention variables X When the condition changes from 1 (implementation) to 0 (cancellation), the system generates a counterfactual output based on the causal equation: , Compare observations under actual policies Y real Counterfactual Y cf The policy impact difference was quantified as follows: , Where, if Δ Y A value greater than 0 indicates that the policy has increased market concentration (limited competition); if Δ Y If the value is less than 0, it indicates that the policy has promoted competition.

[0049] Market outcome indicators include at least one of the following: market concentration HHI index, SME exit rate, and market entry rate.

[0050] Market concentration (Herfindahl–Hirschman Index, HHI) is defined as: , in, s i For the first i Market share of home enterprises N For the number of enterprises.

[0051] The policy effect is assessed as follows: , If ΔHHI>0, it means that the policy has led to an increase in market concentration, i.e., reduced competition.

[0052] Simultaneously define: SME exit rate: ; Percentage of potential entrants: .

[0053] The probability distribution of market outcome index values ​​under counterfactual scenarios is generated using the Monte Carlo sampling method.

[0054] Specifically, all non-policy variables are first frozen, keeping the macroeconomic environment, industry cycles, and other conditions unchanged, and only the value of the intervention variable X is modified (e.g., from 1 to 0) to ensure that the comparison results reflect the impact of the policy itself, rather than external changes.

[0055] Based on the estimated causality coefficient α , β , γ With noise distribution ε Monte Carlo sampling is used to generate multiple sets of counterfactual samples (typically 10,000 times) to form... Y cf The distribution interval is given. From this, the counterfactual mean, variance, and 95% confidence interval can be obtained.

[0056] Calculate and output the difference between the market outcome index values ​​under the counterfactual scenario and the current scenario.

[0057] For multiple indicators such as market concentration HHI, entry rate, and exit rate, the system simultaneously predicts their counterfactual changes, establishing a multi-dimensional profile of policy impact. The system also calculates a comprehensive competition score. , in w 1, w 2, w 3 is the weighting coefficient, which is set by the regulatory authorities based on the importance of the policy.

[0058] Furthermore, step S4 also supports a gradual policy adjustment simulation, which plots the relationship curve between policy intensity and market outcome indicators by continuously changing the intensity of policy intervention variables, and calculates the marginal effect of policy adjustment.

[0059] For example, the policy intensity can be mapped by gradually reducing "strong local protectionism" to "complete elimination". x With market concentration HHI ( x The curve is used to calculate marginal revenue: , This allows policy optimization to move beyond relying on single-point judgments and instead form a continuous adjustment strategy.

[0060] Furthermore, the simulation results can be displayed graphically, with the red line representing the current policy status and the blue line representing the counterfactual scenario. If the blue line is significantly lower than the red line, it indicates that removing the policy would significantly enhance competition. Explanatory descriptions are also provided below the chart, such as: "If local protection policies are removed, the predicted market HHI will decrease by 0.032, the entry rate will increase by 1.8%, and the exit rate of SMEs will decrease by 7.5%, with a confidence level of 95%." This step enables verifiable "what if policy changes" reasoning; it also supports multi-indicator linked evaluation (HHI, price, entry rate, etc.); the counterfactual results obtained can serve as a basis for policy optimization suggestions. This fills the gap in previous systems' inability to perform "hypothetical scenario analysis," giving the review results predictive and decision-support capabilities.

[0061] Furthermore, the fair competition review method for government documents based on causal inference provided in this embodiment also includes: S5. Based on causal effect values ​​and counterfactual simulation results, automatically generate a review report that includes risk clause identification, causal path explanation, quantitative impact assessment, and policy optimization recommendations.

[0062] Specifically, based on causal inferences and counterfactual results, the system generates a comprehensive risk score: , in, w i The indicator weights (set by the regulatory authorities); Δ Y i This represents the difference in the policy impact of each indicator.

[0063] Optionally, the report input includes: 1. The suspected clause number and original text; 2. Risk type and level (high / medium / low); 3. Visualization of the causal chain (e.g., "Clause A → Increased entry barriers → Increased HHI"); 4. Quantified effect value and significance results; and modification suggestions (e.g., "Removing Clause A will decrease the HHI by 0.03").

[0064] This step automates the conversion from causal inference results to natural language reports, making review conclusions transparent and traceable. Through visualized causal chains and numerical tables, decision-makers can intuitively understand the potential risks and optimization opportunities for each clause. Combined with expert feedback mechanisms, the system can also continuously adjust the structure and parameters of the causal graph, improving the accuracy of subsequent reviews.

[0065] This embodiment also provides a government document fair competition review system based on causal inference, used to implement the above-mentioned government document fair competition review method, which includes: The text parsing and feature extraction module is used to process policy document texts and output a structured set of variables; The causal graph construction module, connected to the text parsing and feature extraction module, is used to receive a set of variables and build a causal graph model. The causal effect calculation module, connected to the causal graph construction module, is used to execute causal inference algorithms based on the causal relationship graph model and output causal effect values. The counterfactual simulation engine, connected to the causal graph construction module and the causal effect calculation module, is used to perform simulations and evaluations of counterfactual scenarios; The report generation module, connected to the causal effect calculation module and the counterfactual simulation engine, is used to automatically generate structured review reports.

[0066] Furthermore, the text parsing and feature extraction module integrates a BERT-BiLSTM model and a dependency parser to extract and classify semantic triples.

[0067] Furthermore, the causal graph construction module includes: The rule base unit stores causal rules based on the fair competition review guidelines; Semantic similarity calculation unit, used to expand causal relationship edges; The structure learning unit is used to dynamically update the causal graph structure when encountering new policy concepts.

[0068] In summary, this invention provides a method and system for reviewing fair competition in government documents based on causal inference, comprising the following steps: policy text parsing and causal element extraction, causal graph construction, causal effect estimation, counterfactual simulation and indicator evaluation, and review conclusion generation. This method and system for reviewing fair competition in government documents, by combining natural language processing, causal graph models, quantitative effect estimation, and counterfactual simulation, achieves automated, quantifiable, and interpretable intelligent review of fair competition risks in government documents, and can predict the potential market effects of policy modifications.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for fair competition review of government documents based on causal inference, characterized in that, include: S1. Perform natural language processing on the input policy document text to extract semantic elements describing policy measures, mediating variables, and market outcome indicators, and represent them in a structured form as a set of multiple variables; S2. Based on the extracted set of variables, construct a causal relationship graph model with policy intervention variables as the starting point, market outcome indicator variables as the ending point, and including mediating variables; S3. Based on the aforementioned causal relationship diagram model, a causal inference algorithm is used to quantitatively calculate the causal effect values ​​of each policy intervention variable on the market outcome indicator variable; S4. Simulate a counterfactual scenario of policy intervention variable change on the causal relationship diagram model, and calculate and compare the changes in market competition indicators under the counterfactual scenario and the current scenario; S5. Based on the causal effect value and the counterfactual simulation results, automatically generate a review report that includes risk clause identification, causal path explanation, quantitative impact assessment, and policy optimization recommendations.

2. The method for fair competition review of government documents according to claim 1, characterized in that, Step S1 includes: By using dependency parsing and named entity recognition technology, policy document texts are decomposed into semantic triples in the form of "subject-verb-object"; Each semantic triple is labeled with policy attributes using a pre-trained classification model to obtain the corresponding policy intervention variable label. Based on the semantic analysis results, the output is categorized into sets of policy measures variables, sets of intermediary variables, and sets of market outcome indicator variables.

3. The method for fair competition review of government documents according to claim 1, characterized in that, Step S2 includes: Based on a predefined base of fair competition review rules, establish basic causal relationship edges between variable nodes; Calculate the semantic similarity between variable nodes and expand the basic causal relationship edges; When encountering new policy concepts that are not covered, the node and edge structure of the causal graph model is dynamically updated using a Bayesian structural learning method.

4. The method for fair competition review of government documents according to claim 1, characterized in that, Step S3 includes: Parameter estimation is performed using structural equation modeling. The causal effect values ​​include the direct effect of policy intervention variables on market outcome indicator variables, the indirect effect generated through mediating variables, and the total effect.

5. The method for fair competition review of government documents according to claim 4, characterized in that, Step S3 further includes: When historical data is available, propensity score matching is used to screen comparable samples in order to control the impact of confounding variables on the estimation of causal effects. When historical data is insufficient, expert knowledge is introduced as a prior distribution, and the causal effect parameters are calibrated using a Bayesian update method.

6. The method for fair competition review of government documents according to claim 1, characterized in that, Step S4 includes: The do operator is used to intervene in the policy intervention variables in the causal graph model to set up a counterfactual scenario; The probability distribution of market outcome indicator values ​​under counterfactual scenarios is generated using the Monte Carlo sampling method. Calculate and output the difference between the market outcome index values ​​under the counterfactual scenario and the current scenario, wherein the market outcome index includes at least one of the following: market concentration HHI index, SME exit rate and market entry rate.

7. The method for fair competition review of government documents according to claim 6, characterized in that, Step S4 also supports a gradual policy adjustment simulation, which plots the relationship curve between policy intensity and market outcome indicators by continuously changing the intensity of policy intervention variables, and calculates the marginal effect of policy adjustment.

8. A government document fair competition review system based on causal inference, used to implement the government document fair competition review method according to any one of claims 1-7, characterized in that, include: The text parsing and feature extraction module is used to process policy document texts and output a structured set of variables; The causal graph construction module, connected to the text parsing and feature extraction module, is used to receive the variable set and construct a causal graph model. The causal effect calculation module, connected to the causal graph construction module, is used to execute a causal inference algorithm based on the causal relationship graph model and output the causal effect value. The counterfactual simulation engine, connected to the causal graph construction module and the causal effect calculation module, is used to perform simulation and evaluation of counterfactual scenarios; The report generation module, connected to the causal effect calculation module and the counterfactual simulation engine, is used to automatically generate structured review reports.

9. The government document fair competition review system according to claim 8, characterized in that, The text parsing and feature extraction module integrates a BERT-BiLSTM model and a dependency parser to extract and classify semantic triples.

10. The government document fair competition review system according to claim 8, characterized in that, The cause-effect graph construction module includes: The rule base unit stores causal rules based on the fair competition review guidelines; Semantic similarity calculation unit, used to expand causal relationship edges; The structure learning unit is used to dynamically update the causal graph structure when encountering new policy concepts.

Images