Government official document fair competition examination method and system based on multi-modal artificial intelligence

Abstract

The invention discloses a government official document fair competition examination method and system based on multi-modal artificial intelligence, and relates to the technical field of data processing, and the method comprises the following steps: carrying out clause division and risk preliminary judgment on an official document text; key features based on model judgment are extracted to serve as explanation information; performing semantic matching on the risk terms and a pre-established law and regulation database to obtain a corresponding law and regulation basis; an examination report containing risk terms, regulation basis and suggested measures is automatically generated; and review rule and model parameter optimization based on manual feedback is supported. According to the fair competition examination method and system for the government official documents, through fusion of frontier technologies such as multi-modal learning, a knowledge graph, interpretable AI and NLG, full-chain innovation from'perception 'to'decision' to'output 'is realized. And a plurality of pain points such as image-text disjunction, insufficient regulation matching, poor interpretability, low report generation efficiency and the like in fair competition examination of the government are uniformly solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for fair competition review of government documents based on multimodal artificial intelligence. Background Technology

[0002] Since 2016, my country has fully implemented a fair competition review system. All levels of government and their departments must conduct fair competition reviews when issuing administrative documents and policy measures concerning the economic activities of market entities to prevent clauses that exclude or restrict competition. This system has played a crucial role in ensuring equal participation in competition for market entities and optimizing the business environment. Traditionally, reviews were primarily conducted manually. Reviewers compared the content of documents clause by clause according to regulations such as the "Detailed Rules for the Implementation of the Fair Competition Review System" and the "Anti-Monopoly Law" to determine whether any circumstances hindered competition. The advantage of manual methods lies in combining the professional judgment and contextual understanding of reviewers. However, it also has inherent drawbacks: when dealing with a large number of policy documents and concentrated review tasks, manual review is time-consuming, inefficient, and susceptible to subjective influences, with different personnel potentially yielding inconsistent judgments.

[0003] To improve review efficiency, some institutions and research teams have begun to explore the introduction of Natural Language Processing (NLP) technology and machine learning models to automate the analysis of policy texts. For example, machine learning-based text classification methods can determine the "compliance / non-compliance" of official document clauses. These methods typically rely on large-scale labeled samples, using algorithms to train models capable of recognizing risk statements. However, these methods have two significant problems: First, legal and policy texts are often structurally complex, use standardized terminology, and have diverse semantics, making it difficult for models to achieve high accuracy in the absence of sufficient domain-labeled data. Second, while deep learning models may outperform rule matching in classification accuracy, they are essentially "black box" systems; their internal reasoning processes are opaque, and their output conclusions lack interpretability. This is particularly unacceptable in legal and regulatory scenarios, as regulators need to clearly understand "why" a clause is deemed non-compliant.

[0004] In recent years, the rise of Explainable Artificial Intelligence (XAI) technology has offered a possibility for solving this problem. Explainable AI methods (such as LIME, SHAP, and attention weight visualization) can reveal the contribution of each input feature in the model's decision-making, thus providing an explanation for the model's output. For example, LIME approximates the decision boundary of a deep model in a local area by generating perturbation samples around the input and fitting an interpretable linear model, thereby indicating the keywords that the model focuses on; SHAP, based on the Shapley value in game theory, decomposes the model's prediction into the contribution and bias of each feature, quantifying the impact of each input word on the prediction result.

[0005] Beyond model interpretation, another crucial aspect of actual supervision is connecting detection results with legal and regulatory basis. Supervisory personnel not only need to know that the model has determined a clause to be risky, but also need to quickly identify the legal provisions that the clause may violate. However, most current AI review systems only provide risk warnings and do not establish a "clause-evidence" chain, requiring manual searching of relevant regulations and reducing efficiency. A few studies have attempted to build a regulatory knowledge base and combine it with text matching methods to associate risky clauses with legal provisions, but these still fall short in terms of matching accuracy and automation.

[0006] Furthermore, existing intelligent review tools often lack the ability to automatically generate complete review reports. Even if the model detects risky clauses, the output is often just a list, while the final review report used for archiving and transmission still needs to be written manually, including the original text of the risky clause, the type of risk, the legal basis, and modification suggestions. This not only increases the workload but may also introduce new subjective biases.

[0007] Therefore, the current technical challenge lies not only in accurately identifying risky clauses, but also in providing transparent and interpretable judgment criteria while maintaining model accuracy, and automatically generating a complete review report containing a chain of legal evidence. This requires the organic integration of technologies such as NLP, interpretable AI, legal knowledge graphs, and natural language generation to truly achieve a fully automated review process from risk detection to actionable implementation. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for fair competition review of government documents based on multimodal artificial intelligence. By constructing a multimodal artificial intelligence system that can simultaneously understand the semantics of official document text and charts, and automatically associate legal basis with and generate an interpretable chain of evidence, it achieves a significant improvement in the efficiency and accuracy of fair competition review, while ensuring that the review process is transparent, the conclusions are credible, and the results are directly credible.

[0009] The embodiments of the present invention are implemented as follows: A method for fair competition review of government documents based on multimodal artificial intelligence, comprising: S1: Parse the input government document, divide the document content into a set of structured text clauses, and extract the set of charts and images in the document; S2: Encode the text terms and chart images, and fuse them through a cross-modal attention mechanism to obtain a joint feature vector representing the semantic relationship between the terms and the chart; S3: Based on the joint feature vector, determine the probability that the text clause belongs to a predefined fair competition risk category, and automatically retrieve the most relevant legal provisions from the legal knowledge base based on semantic similarity to generate an evidence chain between the risk category and the legal provisions; S4: Based on the judgment results, generate interpretability evidence in both text and chart dimensions, and perform counterfactual stability verification on the key evidence in the charts; S5: Based on the judgment results, matching legal provisions, and interpretable evidence, automatically generate a structured review report; S6: Receive feedback from personnel on the review results and incrementally update the judgment model and regulatory knowledge base based on the feedback data.

[0010] Furthermore, in other preferred embodiments of the present invention, the processing of the chart in step S1 specifically includes: The chart structure encoder identifies and extracts visual objects from chart images, including bars, lines, legends, and axis labels, and generates chart semantic vectors containing numerical and semantic information.

[0011] Furthermore, in other preferred embodiments of the present invention, the fusion process in step S2 is optimized through comparative learning: We use an InfoNCE-based loss function and a hard-to-bear sample mining strategy to bring the semantic distance between the matching terms and the chart closer, while pushing away the distance between mismatched samples that are semantically related but have different legal meanings.

[0012] Furthermore, in other preferred embodiments of the present invention, the process of generating a chain of evidence in step S3 further includes: An ontology graph of fair competition behavior is introduced to impose logical consistency constraints on the probabilities of the determined risk categories, ensuring that the combination relationships of risk categories conform to the predefined regulatory logic.

[0013] Furthermore, in other preferred embodiments of the present invention, generating interpretability evidence in step S4 specifically includes: The integral gradient method is applied to calculate the contribution of specific terms in the text clauses to the risk assessment result; and, The application class activation mapping method generates heatmaps of chart images to identify visual areas that play a key role in risk assessment.

[0014] Furthermore, in other preferred embodiments of the present invention, the counterfactual stability verification in step S4 specifically includes: The values ​​or labels identified as key evidence in the chart are perturbed to generate a perturbed chart. The change in the probability of the risk category before and after the perturbing is calculated. If the change exceeds a preset threshold, the validity of the key evidence is verified.

[0015] Furthermore, in other preferred embodiments of the present invention, the review report generated in step S5 includes at least: The original risk clauses, corresponding charts and images, predicted risk categories and confidence levels, matching regulatory provisions, explanatory evidence of the text and charts, and proposed amendments based on regulations.

[0016] A government document fair competition review system based on multimodal artificial intelligence, used to implement the aforementioned government document fair competition review method, includes: The data processing module is used to execute step S1 and complete the structured parsing of the official document; The multimodal fusion module is used to execute step S2, completing the encoding and semantic alignment of text and charts; The risk assessment and regulatory matching module is used to execute step S3 and complete the risk assessment and evidence chain generation. The interpretability analysis module is used to perform step S4, generate interpretable evidence, and perform counterfactual verification; The report generation module is used to execute step S5 and automatically generate a review report; The human-machine collaboration optimization module is used to execute step S6 to enable the system to continuously learn and update.

[0017] An electronic device comprising: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement the above-mentioned fair competition review method for government documents.

[0018] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for fair competition review of government documents.

[0019] The beneficial effects of the embodiments of the present invention are: This invention provides a method and system for fair competition review of government documents based on multimodal artificial intelligence. The method includes the following steps: segmenting the document text into clauses and making a preliminary risk assessment; extracting key features used by the model for assessment as explanatory information; semantically matching the risk clauses with a pre-established legal database to obtain the corresponding legal basis; automatically generating a review report containing the risk clauses, legal basis, and recommended measures; and supporting optimization of review rules and model parameters based on human feedback. This method and system for fair competition review of government documents achieves end-to-end innovation from "perception" to "decision-making" and then to "output" by integrating cutting-edge technologies such as multimodal learning, knowledge graphs, explainable AI, and NLG. It systematically solves several pain points in government fair competition review, such as the disconnect between text and graphics, insufficient legal matching, poor explainability, and low report generation efficiency. Detailed Implementation

[0020] 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.

[0021] The abbreviations and key terms used in the following embodiments are defined as follows: Example

[0022] This embodiment provides a method for fair competition review of government documents based on multimodal artificial intelligence, which includes: S1: Parse the input government document, divide the document content into a set of structured text clauses, and extract the set of charts and images in the document.

[0023] First, the system performs structured parsing on the input government documents. Documents can be in PDF, Word, or HTML formats, and typically consist of multiple chapters, clauses, paragraphs, and accompanying tables and figures. The system uses a combination of layout recognition models and rule templates to automatically split the document, dividing the main text into independent clauses. Each clause... d i Represented as: , in, d i Indicates the first i The text of each clause, N This represents the total number of clauses.

[0024] Subsequently, through layout parser detection and natural language anchor point recognition (e.g., "see Figure 1" and "as shown in Table 2"), the system extracts the set of charts corresponding to the clauses in the official document: , in, I j For the first j A chart image, M This represents the total number of charts extracted. Charts can include bar charts, line charts, pie charts, or a combination of statistical charts.

[0025] For each chart I j The system uses a self-developed chart-aware encoder to identify and extract key objects (bars, lines, legends, axes, text labels, etc.) from the image into a set of objects. , Each object o jk This includes the chart's category (e.g., bar, line, label), location coordinates, corresponding numerical values, and text descriptions. Through a combination of object modeling and OCR extraction, the system generates a structured semantic vector for the chart. , in, f chart (·) is a special coding function for charts; I j Original icon image; C j A collection of structured objects; Represents the semantic vector of the chart; d v is the dimension of the visual feature space.

[0026] To enhance the regulatory relevance of the text, this invention also designs a regulation-aware prompting text encoding mechanism, which automatically inserts relevant legal concept prompts before the clause text, such as "This clause involves competitive conduct; please determine whether there is any restriction on the trading partner." The input following the prompt is denoted as: , Among them, Reg_Keywords represents key legal phrases extracted from the legal knowledge base.

[0027] This step can explicitly and structurally represent the text and charts of official documents, providing a unified foundation for subsequent multimodal semantic alignment, enabling the model to identify the inherent logical relationships between clauses and charts at the data level.

[0028] Furthermore, the fair competition review method for government documents based on multimodal artificial intelligence in this embodiment also includes: S2: Encode the text terms and chart images, and fuse them through a cross-modal attention mechanism to obtain a joint feature vector representing the semantic relationship between the terms and the chart.

[0029] Encoding text terms is achieved using a text encoder: , in,f T (·) represents the Transformer-based text encoding function; The text of the clauses to include regulatory notices; The semantic vector representation of the clause; d t The dimension is the text vector.

[0030] Encoding icon images is achieved using a visual encoder: , in, f V (·) represents a visual feature extraction function based on ViT or ResNet backbone; This is the visual feature vector of the chart.

[0031] After obtaining the semantic vector and the visual feature vector of the chart, the association between the clause text and the corresponding chart is calculated through a cross-modal attention mechanism, specifically as follows: , in, CrossAttn (·) represents the cross-modal attention layer; For the fused multimodal representation, d z To fuse feature dimensions.

[0032] Furthermore, the fusion process in step S2 is optimized through contrastive learning: We use an InfoNCE-based loss function and a hard-to-bear sample mining strategy to bring the semantic distance between the matching terms and the chart closer, while pushing away the distance between mismatched samples that are semantically related but have different legal meanings.

[0033] Specifically, text and chart representations are projected onto a shared semantic space after a linear transformation: , in, For text projection matrix; For the chart projection matrix, This is a representation within a shared space.

[0034] Subsequently, the improved InfoNCE contrast loss optimization terms - chart pairing were used: , Where B is the batch size; cos(·) is the cosine similarity function; τ is the temperature coefficient, which controls the smoothness of the distribution; Semantic vectors paired with the same clause-chart.

[0035] During training, the model receives both paired inputs (clause-chart) and mismatched inputs (clause-other chart), learning alignment relationships through multi-head attention and cosine distance. The system employs Hard Negative Mining, prioritizing samples with similar semantics but different legal meanings (e.g., "support local businesses" versus "limit to foreign businesses") to ensure the model is sensitive to subtle semantic differences. After this stage, clauses and charts are bound together in a shared semantic space. Through cross-modal semantic alignment learning, the model can understand the implicit correspondence between bar heights, line slopes, and scale differences in charts and the wording of clauses.

[0036] Furthermore, the fair competition review method for government documents based on multimodal artificial intelligence in this embodiment also includes: S3: Based on the joint feature vector, determine the probability that the text clause belongs to a predefined fair competition risk category, and automatically retrieve the most relevant legal provisions from the legal knowledge base based on semantic similarity to generate an evidence chain between the risk category and the legal provisions.

[0037] Specifically, the joint feature obtained by merging the clauses and the chart is: , Where Pool(·) is the fusion operation (which can be a weighted average or attention convergence); output This is a comprehensive semantic vector.

[0038] Based on this joint feature vector, the risk class probability is output through a fully connected layer: , in, This is the classification weight matrix; The bias is σ(·), which is the Sigmoid activation operation. K Number of risk categories; For belonging to the first k The probability of a class of risks.

[0039] Its loss function is: , in, y ik This is a real label; The probability output by the model.

[0040] The system will further output vector u i Vector of each legal provision in the RegK library r m Perform similarity calculation: , The Top-K regulations with the highest similarity are selected as the matching results.

[0041] The system first outputs a multi-label probability vector during the inference phase. Each dimension represents a risk category (such as "exclusivity agreement," "discriminatory treatment," "unreasonable subsidies," etc.). Next, similarity is calculated based on semantic embedding in the RegK regulatory database, returning the legal provisions most relevant to the clause's content, such as Article 6 of the Anti-Monopoly Law or Article 4 of the Implementation Rules for Fair Competition Review. If a clause has a similar match with multiple regulations, the system uses regulatory ontology relationships (priority, subordination, prohibited clauses) to rearrange them, ensuring the output regulation has the highest interpretative authority. The final output includes: a category probability vector, a Top-K list of regulatory matches, and their similarity scores.

[0042] Optionally, to ensure that the output conforms to the legal logic, this embodiment introduces a legal ontology diagram. And apply consistency regularization during inference.

[0043] Specifically, the ontology consistency constraint adopts the following function: , in, The set of edges in the legal entity; w pq Relationship weights; α p For predefined dependency coefficients; This represents the predicted probability for the corresponding category.

[0044] The evidence consistency score is calculated using the following function: , in, l 1, l 2 represents the fusion weight; For a set of candidate regulations; S ev A comprehensive credibility score is given.

[0045] The system verifies the consistency of model predictions based on the logical structure of the regulatory diagram. For example, if the model predicts both "local protectionism" and "restricted trading objects," then the two should have a strong correlation. w pq If the confidence level is too high, the confidence level should be lowered. Conversely, if both "differential treatment" and "price discrimination" are predicted simultaneously, the system must ensure that the logical direction is correct. Finally, the constraints are jointly optimized using a comprehensive loss analysis.

[0046] , in, βThe ontology constraint weighting factor; , , These three items correspond to the goals of alignment, classification, and regulatory consistency, respectively.

[0047] The output results satisfy both semantic consistency and legal logic consistency, making the model conclusions more interpretable and legally credible.

[0048] Furthermore, the fair competition review method for government documents based on multimodal artificial intelligence in this embodiment also includes: S4: Based on the judgment results, generate interpretability evidence in both text and chart dimensions, and perform counterfactual stability verification on the key evidence in the charts.

[0049] Furthermore, the generation of interpretable evidence in step S4 specifically includes: The integral gradient method is applied to calculate the contribution of specific terms in the text clauses to the risk assessment result. α w ;as well as, The application class activation mapping method generates heatmaps of chart images to identify visual areas that play a key role in risk assessment.

[0050] Furthermore, the counterfactual stability verification in step S4 specifically involves: The values ​​or labels identified as key evidence in the chart are perturbed to generate a perturbed chart. The change in the probability of the risk category before and after the perturbing is calculated. If the change exceeds a preset threshold, the validity of the key evidence is verified.

[0051] Its functional expression is: , in, I This is the original chart; The chart after perturbation; For category k The predicted probability; Δ P The difference before and after the perturbation is used to quantify the impact of key evidence on the results. If Δ P A larger value indicates that the object makes a significant contribution to the classification result.

[0052] The system simultaneously presents textual explanations and overlay images of heatmaps. If adjusting the chart values ​​to reflect fair conditions results in a significant decrease in the probability of the output category (e.g., the confidence level for "discriminatory treatment" decreases from 0.6 to 0.2), it proves that the original clause does indeed contain discriminatory elements. All explanations are accompanied by stability tests and confidence lower bounds to ensure they are acceptable for manual review.

[0053] Furthermore, the fair competition review method for government documents based on multimodal artificial intelligence in this embodiment also includes: S5: Based on the judgment results, matching legal provisions, and interpretable evidence, automatically generate a structured review report.

[0054] Furthermore, the review report generated in step S5 includes at least: The original risk clauses, corresponding charts and images, predicted risk categories and confidence levels, matching regulatory provisions, explanatory evidence of the text and charts, and proposed amendments based on regulations.

[0055] Furthermore, the fair competition review method for government documents based on multimodal artificial intelligence in this embodiment also includes: S6: Receive feedback from personnel on the review results and incrementally update the judgment model and regulatory knowledge base based on the feedback data.

[0056] Human reviewers can perform "confirm, modify, or reject" operations on the system results. Feedback data will be stored in the database in the form of sample pairs for incremental fine-tuning. The system employs an active learning strategy for samples with high uncertainty (low...). S ev (For models with unstable interpretations, retraining is prioritized.) The RegK regulatory library supports versioning; after updates, the system automatically reconstructs regulatory vectors and ontology edge weights. Continuous optimization of the process maintains the model's adaptability to regulatory changes, ensuring its long-term accurate, interpretable, and reliable review capabilities.

[0057] Furthermore, this embodiment also provides a government document fair competition review system based on multimodal artificial intelligence, used to implement the above-mentioned government document fair competition review method, which includes: The data processing module is used to execute step S1 and complete the structured parsing of the official document; The multimodal fusion module is used to execute step S2, completing the encoding and semantic alignment of text and charts; The risk assessment and regulatory matching module is used to execute step S3 and complete the risk assessment and evidence chain generation. The interpretability analysis module is used to perform step S4, generate interpretable evidence, and perform counterfactual verification; The report generation module is used to execute step S5 and automatically generate a review report; The human-machine collaboration optimization module is used to execute step S6 to enable the system to continuously learn and update.

[0058] Furthermore, this embodiment also provides an electronic device, which includes: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement the above-mentioned fair competition review method for government documents.

[0059] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for fair competition review of government documents.

[0060] In summary, this invention provides a method and system for fair competition review of government documents based on multimodal artificial intelligence. The method includes the following steps: segmenting the document text into clauses and making a preliminary risk assessment; extracting key features used by the model for assessment as explanatory information; semantically matching the risk clauses with a pre-established legal database to obtain the corresponding legal basis; automatically generating a review report containing the risk clauses, legal basis, and recommended measures; and supporting optimization of review rules and model parameters based on human feedback. This method and system for fair competition review of government documents achieves end-to-end innovation from "perception" to "decision-making" and then to "output" by integrating cutting-edge technologies such as multimodal learning, knowledge graphs, explainable AI, and NLG. It systematically solves several pain points in government fair competition review, such as the disconnect between text and graphics, insufficient legal matching, poor explainability, and low report generation efficiency.

[0061] 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 multimodal artificial intelligence, characterized in that, include: S1: Parse the input government document, divide the document content into a set of structured text clauses, and extract the set of charts and images in the document; S2: Encode the text terms and chart images, and fuse them through a cross-modal attention mechanism to obtain a joint feature vector representing the semantic relationship between the terms and the chart; S3: Based on the joint feature vector, determine the probability that the text clause belongs to a predefined fair competition risk category, and automatically retrieve the most relevant legal provisions from the legal knowledge base based on semantic similarity to generate an evidence chain between the risk category and the legal provisions; S4: Based on the judgment result, generate interpretability evidence in both text and chart dimensions, and perform counterfactual stability verification on the key evidence in the chart; S5: Based on the judgment result, the matching legal provisions and interpretable evidence, automatically generate a structured review report; S6: Receive feedback from personnel on the review results, and incrementally update the judgment model and regulatory knowledge base based on the feedback data.

2. The method for fair competition review of government documents according to claim 1, characterized in that, The chart processing in step S1 specifically includes: The chart structure encoder identifies and extracts visual objects from the chart image, including bars, lines, legends and axis labels, and generates a chart semantic vector containing numerical and semantic information.

3. The method for fair competition review of government documents according to claim 1, characterized in that, The fusion process in step S2 is optimized through contrastive learning: We use an InfoNCE-based loss function and a hard-to-bear sample mining strategy to bring the semantic distance between the matching terms and the chart closer, while pushing away the distance between mismatched samples that are semantically related but have different legal meanings.

4. The method for fair competition review of government documents according to claim 1, characterized in that, The process of generating the chain of evidence in step S3 further includes: An ontology graph of fair competition behavior is introduced to impose logical consistency constraints on the probabilities of the risk categories determined, ensuring that the combination relationships of risk categories conform to predefined regulatory logic.

5. The method for fair competition review of government documents according to claim 1, characterized in that, The generation of interpretable evidence in step S4 specifically includes: The integral gradient method is applied to calculate the contribution of specific terms in the text clauses to the risk assessment result; and, The application class activation mapping method generates heatmaps of chart images to identify visual areas that play a key role in risk assessment.

6. The method for fair competition review of government documents according to claim 1, characterized in that, The counterfactual stability verification in step S4 specifically involves: The values ​​or labels identified as key evidence in the chart are perturbed to generate a perturbed chart. The change in the probability of the risk category before and after the perturbing is calculated. If the change exceeds a preset threshold, the validity of the key evidence is verified.

7. The method for fair competition review of government documents according to claim 1, characterized in that, The review report generated in step S5 includes at least the following: The original risk clauses, corresponding charts and images, predicted risk categories and confidence levels, matching regulatory provisions, explanatory evidence of the text and charts, and proposed amendments based on regulations.

8. A government document fair competition review system based on multimodal artificial intelligence, used to implement the government document fair competition review method according to any one of claims 1-7, characterized in that, include: The data processing module is used to execute step S1 and complete the structured parsing of the official document; The multimodal fusion module is used to perform step S2, completing the encoding and semantic alignment of text and charts; The risk assessment and regulatory matching module is used to execute step S3 to complete risk assessment and evidence chain generation. The interpretability analysis module is used to perform step S4, generate interpretable evidence and perform counterfactual verification; The report generation module is used to execute step S5 and automatically generate a review report; The human-machine collaboration optimization module is used to execute step S6 to realize the continuous learning and updating of the system.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the fair competition review method for government documents as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the fair competition review method for government documents as described in any one of claims 1-7.