Bidding document risk identification system and method

Abstract

The invention discloses a bidding file risk identification system and method, and the method comprises the steps: extracting key information of a bidding file through a file analysis module, and achieving the automatic risk identification and quantitative analysis through combining with a machine learning model and a risk rule library which are built by a data training module; potential risks in the bidding document can be efficiently and accurately identified, and the bidding document can be classified into high risks, medium risks and low risks according to the severity of the risks and the influence on the project. According to the system, the standardized risk report is generated to provide data support and decision basis for a bid invitation unit, so that the efficiency and accuracy of bid document auditing are remarkably improved, and the possibility of human errors is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of bidding document risk identification, and in particular, relates to a bidding document risk identification system and method. Background Art

[0002] The purpose of risk identification in bidding documents is to conduct a comprehensive and accurate analysis of potential risks in bidding documents through systematic and intelligent means, so as to ensure the fairness and scientific nature of the bidding process and the smooth implementation of the project. Specifically, the process aims to identify high-risk factors in bidding documents, quantify and grade risks, and provide a clear risk assessment report for the bidding unit. Through automated risk identification and assessment, the efficiency and accuracy of document review are improved, the omissions and deviations caused by manual operations are reduced, and the bidding unit is helped to effectively avoid potential risks in bidding decisions and reduce the uncertainty and possibility of legal disputes in the subsequent implementation of the project.

[0003] In the traditional bidding document review process, risk identification mainly relies on the reviewers’ experience and subjective judgment. The process is not only time-consuming and laborious, but also difficult to standardize. Especially when the number of bidding documents is large and the content is complex, the efficiency is even lower. At the same time, the manual identification method is limited by the professional level and working status of the reviewers, which is prone to information omissions, misjudgments or evaluation biases, resulting in key risk points being ignored, affecting the scientificity and fairness of the bidding. The existing technology lacks systematic and automated risk identification tools and methods, and cannot conduct a comprehensive and accurate quantitative analysis of various potential risks in the bidding documents. It also lacks a flexible response mechanism for different project needs, making it difficult to dynamically adjust risk rules according to the special requirements of the project type, reducing the adaptability and accuracy of the review process.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a bidding document risk identification system and method, which solves the problems raised in the above-mentioned background technology.

[0006] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0007] A bidding document risk identification system includes: a document parsing module for parsing the bidding document and extracting key information of the bidding document;

[0008] Data training module, used to build and train machine learning models using historical bidding data to predict potential risk types and levels;

[0009] A risk rule library, used to store risk rules related to bidding documents, wherein the risk rule library supports customization and updating based on different project types;

[0010] A risk identification module, connected to the document parsing module, the data training module and the risk rule library, for comparing the parsed key information with the risk rules to identify the risks in the bidding documents;

[0011] The risk assessment module is used to quantify the identified risks and classify them into high risk, medium risk and low risk according to their severity and impact on the project;

[0012] The risk report generation module is used to organize and analyze the identified risks, generate risk reports, and provide decision-making basis for the bidding unit.

[0013] Optionally, when parsing the bidding documents and extracting key information from the bidding documents, first, load the bidding documents from the system and convert the PDF, Word, and Excel formats into plain text or table data. Next, remove unnecessary symbols, format errors, and blank lines in the files, and divide the files into different chapters according to the directory and title. Subsequently, scan the contents line by line in each chapter to extract the quotation amount, construction period, and contact information, parse the data in the table, and convert all the contents into structured information. Then, verify the extracted data to check whether the text is consistent with the contents in the table. Finally, store the extracted key information in the database in a structured format.

[0014] Optionally, the machine learning model is a convolutional neural network model, wherein the neural network model includes three convolutional layers and three pooling layers, and two fully connected layers. Each convolutional layer uses a different number of convolution kernels, which are 32, 64, and 128 respectively. The convolution kernel size is 3×3, the step size is 1, and the padding is kept as "same" to ensure that the output size remains unchanged. Each convolutional layer is followed by a pooling layer and a ReLU activation function to stabilize the training process and introduce nonlinearity. The pooling layer uses a maximum pooling of 2×2 with a step size of 2 to reduce the number of parameters. The fully connected layer has 256 and 64 nodes, and the final output layer uses a Softmax activation function to convert the output into a probability distribution of each risk level.

[0015] Optionally, the steps of building and training a convolutional neural network model using bidding history data to predict potential risk types and levels are:

[0016] Collect historical bidding data, including quotations, project response time, supplier ratings, and contract performance. Define high and low risk levels based on actual conditions to ensure that labels correspond to feature data one by one. Then convert the historical bidding data into a standard format and normalize it to form an input feature vector of size 32×32×3.

[0017] The feature data is divided into a training set and a test set. Then, the classification model is established using historical data, and the initial weight parameter of the classification model is set to 0.01, the bias term is initialized to 0, and all parameters are initialized to small non-zero values. Then the data set is divided into 80% as a training set and 20% as a test set.

[0018] The input feature vector enters the CNN model, and after the calculation of convolution, pooling, and fully connected layers, the probability distribution of risk type and risk level is generated. Then, the cross entropy loss function is used to calculate the gap between the model prediction and the actual label, which is expressed as: Among them, N represents the number of samples, C represents the number of categories, and y ij is the true label, is the predicted probability;

[0019] By back-propagating and calculating the gradient of the loss relative to each parameter, the Adam optimizer is used to update the model parameters, adjust the convolution kernel and weights, minimize the loss function and accelerate convergence. The expression is: Among them, θ t is a parameter, η is the learning rate, m t and v t are the first and second order momentum respectively;

[0020] The trained model is then tested on the validation set, and the accuracy and recall are recorded to check whether the model is overfitting or underfitting. Next, the model performance is evaluated on the test set, and a confusion matrix and classification report are generated to determine the accuracy and robustness of the model in practical applications. The trained model is then deployed to the system. The model regularly detects the prediction results of the input data and compares them with the actual results. When the model prediction accuracy is lower than 80%, the retraining process is triggered.

[0021] Optionally, when collecting historical bidding data, a clustering algorithm needs to be used to analyze abnormal patterns in the historical data to warn of hidden risks. The steps are as follows:

[0022] Each bidding document is represented as a feature vector x i =(x i1 , x i2 , dots, x id ), where d represents the dimension of the feature. Then, the K-means algorithm is used to divide the data points into K clusters, each cluster is composed of a centroid μ kIt means that, then, the distance between the data point and the centroid is calculated, and it is judged whether it is an outlier and the regularization deviation term is introduced to capture the outlier. Then, the outlier is marked and a risk warning is generated. If the outlier does not match the risk in the rule base, the user is prompted to update the rule base. Finally, the new outlier and its characteristic pattern are added to the data set, and the risk rule base and model parameters are updated. The expression is: Among them, L is the clustering objective function, which is used to minimize the distance within the cluster and the degree of deviation of the outliers. K is the number of clusters, and the data points are divided into K groups. C k is the set of the Kth cluster, containing the data points belonging to this cluster, x i is the characteristic vector of the bidding document, representing the characteristic value of the i-th data point, μ k is the centroid of the kth cluster, that is, the average vector of the data points in the cluster; |x i -μ k | is the Euclidean distance, calculate the data point x i and its cluster centroid μ k The distance between k is the standard deviation vector within the kth cluster, which is used to measure the distribution of data within the cluster.

[0023] Optionally, the steps of comparing the parsed key information with the risk rules and identifying the risks in the bidding documents are:

[0024] The extracted key information is input into the data training module, and the deep learning model trained with historical data is used to generate preliminary risk prediction results to obtain the probability values ​​of risk types and levels. The parsed information is compared with the rules in the risk rule base to identify whether the key information meets the bidding requirements and industry standards. The results of the data training model prediction and the matching results of the risk rule base are summarized. Then, the real-time data information is compared with the standard values ​​in the historical data to detect potential hidden risks. By comparing the differences between the two results, the risk points in the key information are further verified. Risks confirmed by both model prediction and rule base matching are marked as high priority risks. Finally, a structured report is generated for all risks found in the comparison, listing the risk types, specific fields and violated rules, and the comparison results and analysis data are stored in the database. At the same time, the rule base is optimized and updated according to the newly identified risks to ensure that the rule base maintains dynamic adaptability.

[0025] Optionally, the steps of generating preliminary risk prediction results using a convolutional neural network model trained with historical data to obtain probability values ​​of risk types and levels are as follows:

[0026] Extract and clean various features from the historical bidding document dataset, including quotation, technical matching, and performance record, and convert each bidding document into a standardized feature matrix;

[0027] The pre-processed bidding document feature matrix is ​​input into the trained CNN model. Through layer-by-layer calculation of convolutional layer, activation layer, pooling layer and fully connected layer, the deep features of the input data are extracted and the risk feature vector is gradually formed.

[0028] After multiple layers of convolution and pooling operations, the model generates multi-dimensional feature maps and maps these feature maps into a risk feature space through a fully connected layer to form a comprehensive risk representation of the bid document.

[0029] In the output layer, the Softmax activation function is used to calculate the probability of each risk type. Through the Softmax operation, the risk features output by the fully connected layer are mapped to probability values ​​for different risk types, each value is between 0 and 11;

[0030] Output the probability value of each risk type and its level, sort them by probability, and mark the risk type and risk level with the highest probability as the preliminary risk prediction result.

[0031] Optionally, the steps to quantify the identified risks and classify them into high risk, medium risk and low risk according to their severity and impact on the project are as follows:

[0032] Summarize the identified risk information, including risk type, violated rules, frequency of occurrence, and project links involved. Then, set specific quantitative indicators for each risk type, and then use the formula based on the quantitative indicators of each risk. Calculate the comprehensive score, divide the risk into high risk, medium risk and low risk according to the comprehensive score, and classify each risk into the corresponding level category. Finally, organize the quantified risk data to generate an analysis report, which lists the comprehensive score, level classification and impact description of each risk in detail, and stores the results in the database. Among them, S is the comprehensive risk score of the bidding document, ranging from 0 to 5. The higher the value, the lower the risk. i is the weight of risk type i, indicating the importance of each risk type in the comprehensive score, n i is the number of quantitative indicators under risk type i, R ij is the score of the jth quantitative indicator of risk type i, ranging from 0 to 5, indicating the indicator score, I ij is the importance coefficient of the j-th quantitative indicator of risk type i, which is used to indicate the weight of the indicator in this risk type.

[0033] Optionally, the steps to organize and analyze the identified risks and generate a risk report to provide a basis for decision-making for the tendering unit are as follows:

[0034] Collect all identified risk data, classify the data according to risk type, and group risk points of the same risk type together. Then, sort each type of risk by severity and display high risks more intuitively at the front end of the report. Calculate the overall risk score based on the comprehensive score obtained by the risk assessment module. By comparing historical data, mark the risk differences with similar bids in the past to help the bidding unit quickly determine the relative risk level of the bid. Then generate charts to show the proportion and score of different risk types, intuitively display the distribution and proportion of risks, generate standardized report documents, including risk type and risk level, and export them in PDF or Excel format.

[0035] A method for identifying risks in bidding documents comprises the following steps:

[0036] Parse the bidding documents and extract key information from them;

[0037] Use historical bidding data to build and train machine learning models to predict potential risk types and levels;

[0038] Storing risk rules related to bidding documents, wherein the risk rule library supports customization and updating based on different project types;

[0039] Compare the parsed key information with the risk rules to identify risks in the bidding documents;

[0040] Conduct quantitative analysis on the identified risks and classify them into high risk, medium risk and low risk according to the severity of the risks and their impact on the project;

[0041] The identified risks will be sorted and analyzed to generate a risk report to provide a basis for decision-making for the tendering unit.

[0042] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0043] The key information of the bidding documents is extracted through the document parsing module, and the machine learning model and risk rule library established by the data training module are combined to realize automated risk identification and quantitative analysis. It can not only efficiently and accurately identify the potential risks in the bidding documents, but also classify them into high, medium and low risks according to the severity of the risks and the impact on the project. The system generates standardized risk reports to provide data support and decision-making basis for the bidding units, thereby significantly improving the efficiency and accuracy of bidding document review and reducing the possibility of human errors.

[0044] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0046] In the figure:

[0047] Figure 1 Flowchart of the bidding document risk identification system.

[0048] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0049] The present invention will now be described in further detail with reference to the accompanying drawings.

[0050] See also Figure 1 As shown, in this embodiment, a bidding document risk identification system is provided, including:

[0051] The document parsing module is used to parse the bidding documents and extract key information of the bidding documents. Key information includes but is not limited to bidding unit information, project demand response status and quotation information;

[0052] Data training module, used to build and train machine learning models using historical bidding data to predict potential risk types and levels;

[0053] The risk rule library is used to store risk rules related to the bidding documents. The risk rules include quotation risk rules, construction period risk rules, technical solution risk rules, contract clause risk rules, bidder qualification risk rules, environment and project adaptability risk rules, legal and compliance risk rules, and project implementation risk rules. The risk rule library supports customization and updating based on different project types.

[0054] A risk identification module, connected to the document parsing module, the data training module and the risk rule library, is used to compare the parsed key information with the risk rules to identify the risks in the bidding documents; the risks include but are not limited to the risks of non-qualification, non-standard technology and unreasonable quotation;

[0055] The risk assessment module is used to quantify the identified risks and classify them into high risk, medium risk and low risk according to their severity and impact on the project;

[0056] The risk report generation module is used to organize and analyze the identified risks, generate risk reports, and provide decision-making basis for the bidding unit.

[0057] In this embodiment, when parsing the bidding document and extracting the key information of the bidding document, first, load the bidding document from the system, convert the PDF, Word or Excel format into plain text or table data, then remove the redundant symbols, format errors and blank lines in the file, and divide the file into different chapters according to the directory and title, then scan the content line by line in each chapter, extract key fields such as quotation amount, construction period, contact information, etc., parse the data in the table, convert all the content into structured information, then verify the extracted data to check whether the text is consistent with the content in the table, and finally store the extracted key information in the database in a structured format.

[0058] In this embodiment, the machine learning model is a convolutional neural network model, wherein the neural network model includes three convolutional layers and three pooling layers, and two fully connected layers. Each convolutional layer uses a different number of convolution kernels, 32, 64, and 128 respectively, and the convolution kernel size is 3×3, the step size is 1, and the padding is kept as "same" to ensure that the output size remains unchanged. Each convolutional layer is followed by a pooling layer and a ReLU activation function to stabilize the training process and introduce nonlinearity. The pooling layer uses a maximum pooling of 2×2 with a step size of 2 to reduce the number of parameters. The fully connected layer has 256 and 64 nodes, and the final output layer uses a Softmax activation function to convert the output into a probability distribution of each risk level.

[0059] In this embodiment, the steps of establishing and training a convolutional neural network model using bidding history data to predict potential risk types and levels are as follows:

[0060] Collect historical bidding data, including key features such as quotations, project response time, supplier ratings, and contract performance. Define high and low risk levels based on actual conditions to ensure that labels correspond to feature data one by one. Then convert the historical bidding data into a standard format and normalize it to form an input feature vector of size 32×32×3.

[0061] The feature data is divided into a training set and a test set. Then, the classification model is established using historical data, and the initial weight parameter of the classification model is set to 0.01, the bias term is initialized to 0, and all parameters are initialized to small non-zero values. Then the data set is divided into 80% as a training set and 20% as a test set.

[0062] The input feature vector enters the CNN model, and after the calculation of convolution, pooling, and fully connected layers, the probability distribution of risk type and risk level is generated. Then, the cross entropy loss function is used to calculate the gap between the model prediction and the actual label, which is expressed as: Among them, N represents the number of samples, C represents the number of categories, and y ij is the true label, is the predicted probability;

[0063] By back-propagating and calculating the gradient of the loss relative to each parameter, the Adam optimizer is used to update the model parameters, adjust the convolution kernel and weights, minimize the loss function and accelerate convergence. The expression is: Among them, θ t is a parameter, η is the learning rate, m t and v t are the first and second order momentum respectively;

[0064] The trained model is then tested on the validation set, and indicators such as accuracy and recall are recorded to check whether the model is overfitting or underfitting. Next, the model performance is evaluated on the test set, and a confusion matrix and classification report are generated to determine the accuracy and robustness of the model in practical applications. The trained model is then deployed to the system. The model regularly detects the prediction results of the input data and compares them with the actual results. When the model prediction accuracy is lower than 80%, the retraining process is triggered.

[0065] In this embodiment, when collecting historical bidding data, it is necessary to use a clustering algorithm to analyze abnormal patterns in the historical data and issue an early warning for possible hidden risks. The steps are as follows:

[0066] Each bidding document is represented as a feature vector x i =(x i1 , x i2 , dots, x id ), where d represents the dimension of the feature. Then, the K-means algorithm is used to divide the data points into K clusters, each cluster is composed of a centroid μ k It means that, then, the distance between the data point and the centroid is calculated, and it is judged whether it is an outlier and the regularization deviation term is introduced to capture the outlier. Then, the outlier is marked and a risk warning is generated. If the outlier does not match the risk in the rule base, the user is prompted to update the rule base. Finally, the new outlier and its characteristic pattern are added to the data set, and the risk rule base and model parameters are updated. The expression is: Among them, L is the clustering objective function, which is used to minimize the distance within the cluster and the degree of deviation of the outliers. K is the number of clusters, and the data points are divided into K groups. C k is the set of the Kth cluster, containing the data points belonging to this cluster, x i is the characteristic vector of the bidding document, representing the characteristic value of the i-th data point, μ k is the centroid of the kth cluster, that is, the average vector of the data points in the cluster: Among them, |C k |, is the number of points in the kth cluster;

[0067] |x i -μk | is the Euclidean distance, calculate the data point x i and its cluster centroid μ k Distance between: Where d is the dimension of the feature vector, Δx i =x i -μ k For data point x i Deviation vector from the centroid, used to assess how far a data point is from the cluster center.

[0068] σ k is the standard deviation vector within the kth cluster, which is used to measure the distribution of data within the cluster: is a regularization parameter used to adjust the weight of the deviation term and balance the influence of normal points and abnormal points.

[0069] In this embodiment, the steps of comparing the parsed key information with the risk rules and identifying the risks in the bidding documents are as follows:

[0070] The extracted key information is input into the data training module, and the deep learning model trained with historical data is used to generate preliminary risk prediction results to obtain the probability values ​​of risk types and levels. The parsed information is compared with the rules in the risk rule base to identify whether the key information meets the bidding requirements and industry standards. The results of the data training model prediction and the matching results of the risk rule base are summarized. Then, the real-time data information is compared with the standard values ​​in the historical data to detect potential hidden risks. By comparing the differences between the two results, the risk points in the key information are further verified. Risks confirmed by both model prediction and rule base matching are marked as high priority risks. Finally, a structured report is generated for all risks found in the comparison, listing the risk types, specific fields and violated rules, and the comparison results and analysis data are stored in the database. At the same time, the rule base is optimized and updated according to the newly identified risks to ensure that the rule base maintains dynamic adaptability.

[0071] In this embodiment, the steps of generating preliminary risk prediction results using the convolutional neural network model trained with historical data and obtaining the probability values ​​of risk types and levels are as follows:

[0072] Extract and clean various features from the historical bidding document dataset, including quotation, technical matching, performance record, etc., and convert each bidding document into a standardized feature matrix;

[0073] The pre-processed bidding document feature matrix is ​​input into the trained CNN model. Through layer-by-layer calculation of convolutional layer, activation layer, pooling layer and fully connected layer, the deep features of the input data are extracted and the risk feature vector is gradually formed.

[0074] After multiple layers of convolution and pooling operations, the model generates multi-dimensional feature maps and maps these feature maps into a risk feature space through a fully connected layer to form a comprehensive risk representation of the bid document.

[0075] In the output layer, the Softmax activation function is used to calculate the probability of each risk type. Through the Softmax operation, the risk features output by the fully connected layer are mapped to probability values ​​for different risk types, each value is between 0 and 11;

[0076] Output the probability value of each risk type and its level, sort them by probability, and mark the risk type and risk level with the highest probability as the preliminary risk prediction result.

[0077] In this embodiment, the steps of quantitatively analyzing the identified risks and classifying them into high risk, medium risk and low risk according to the severity of the risks and the impact on the project are as follows:

[0078] Summarize the identified risk information, including risk type, violated rules, frequency of occurrence, and project links involved. Then, set specific quantitative indicators for each risk type, and then use the formula based on the quantitative indicators of each risk. Calculate the comprehensive score, divide the risk into high risk, medium risk and low risk according to the comprehensive score, and classify each risk into the corresponding level category. Finally, organize the quantified risk data to generate an analysis report, which lists the comprehensive score, level classification and impact description of each risk in detail, and stores the results in the database. Among them, S is the comprehensive risk score of the bidding document, ranging from 0 to 5. The higher the value, the lower the risk. i is the weight of risk type i, indicating the importance of each risk type in the comprehensive score.

[0079] Example: If a project has a high requirement for the rationality of the quotation, the weight w3 of "unreasonable quotation risk" can be set higher, n i is the number of quantitative indicators under risk type i. For example, there are two indicators under “qualification failure risk”, that is, n1 = 2. ij is the score of the jth quantitative indicator of risk type i, ranging from 0 to 5, indicating the indicator score, I ij is the importance coefficient of the j-th quantitative indicator of risk type i, which is used to indicate the weight of the indicator in this risk type.

[0080] In this embodiment, the steps of sorting and analyzing the identified risks, generating a risk report, and providing a decision basis for the tendering unit are as follows:

[0081] Collect all identified risk data, classify the data according to risk type, and group risk points of the same risk type together. Then, sort each type of risk by severity and display high risks more intuitively at the front end of the report. Calculate the overall risk score based on the comprehensive score obtained by the risk assessment module. By comparing historical data, mark the risk differences with similar bids in the past to help the bidding unit quickly determine the relative risk level of the bid. Then generate charts to show the proportion and score of different risk types, intuitively display the distribution and proportion of risks, generate standardized report documents, including risk type and risk level, and export them in PDF or Excel format.

[0082] A method for identifying risks in bidding documents comprises the following steps:

[0083] Parse the bidding documents and extract key information from them;

[0084] Use historical bidding data to build and train machine learning models to predict potential risk types and levels;

[0085] Storing risk rules related to bidding documents, wherein the risk rule library supports customization and updating based on different project types;

[0086] Compare the parsed key information with the risk rules to identify risks in the bidding documents;

[0087] Conduct quantitative analysis on the identified risks and classify them into high risk, medium risk and low risk according to the severity of the risks and their impact on the project;

[0088] The identified risks will be sorted and analyzed to generate a risk report to provide a basis for decision-making for the tendering unit.

[0089] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A bidding document risk identification system, characterized in that: include: The file parsing module is used to parse the bidding documents and extract key information of the bidding documents; Data training module, used to build and train machine learning models using historical bidding data to predict potential risk types and levels; A risk rule library, used to store risk rules related to bidding documents, wherein the risk rule library supports customization and updating based on different project types; A risk identification module, connected to the document parsing module, the data training module and the risk rule library, for comparing the parsed key information with the risk rules to identify the risks in the bidding documents; The risk assessment module is used to quantify the identified risks and classify them into high risk, medium risk and low risk according to their severity and impact on the project; The risk report generation module is used to organize and analyze the identified risks, generate risk reports, and provide decision-making basis for the bidding unit.

2. A bidding document risk identification system according to claim 1, characterized in that: When parsing the bidding documents and extracting key information from them, first, load the bidding documents from the system and convert PDF, Word, and Excel formats into plain text or table data. Next, remove redundant symbols, format errors, and blank lines in the files, and divide the files into different chapters according to the directory and title. Then, scan the content line by line in each chapter to extract the quotation amount, construction period, and contact information, parse the data in the table, and convert all content into structured information. Then, verify the extracted data to check whether the text is consistent with the content in the table. Finally, store the extracted key information in the database in a structured format.

3. A bidding document risk identification system according to claim 1, characterized in that: The machine learning model is a convolutional neural network model, which includes three convolutional layers, three pooling layers, and two fully connected layers. Each convolutional layer uses a different number of convolution kernels, namely 32, 64, and 128. The convolution kernel size is 3×3, the step size is 1, and the padding is kept "same" to ensure that the output size remains unchanged. Each convolutional layer is followed by a pooling layer and a ReLU activation function to stabilize the training process and introduce nonlinearity. The pooling layer uses a maximum pooling of 2×2 with a step size of 2 to reduce the number of parameters. The fully connected layer has 256 and 64 nodes. The final output layer uses a Softmax activation function to convert the output into a probability distribution of each risk level.

4. A bidding document risk identification system according to claim 1, characterized in that: The steps to build and train a convolutional neural network model using historical bidding data to predict potential risk types and levels are: Collect historical bidding data, including quotations, project response time, supplier ratings, and contract performance. Define high and low risk levels based on actual conditions to ensure that labels correspond to feature data one by one. Then convert the historical bidding data into a standard format and normalize it to form an input feature vector of size 32×32×3. The feature data is divided into a training set and a test set. Then, the classification model is established using historical data, and the initial weight parameters of the classification model are set to 0.01, the bias term is initialized to 0, all parameters are initialized to small non-zero values, and then the data set is divided into 80% as a training set and 20% as a test set. The input feature vector enters the CNN model, and after the calculation of convolution, pooling, and fully connected layers, the probability distribution of risk type and risk level is generated. Then, the cross entropy loss function is used to calculate the gap between the model prediction and the actual label, which is expressed as: Among them, N represents the number of samples, C represents the number of categories, and y ij is the true label, is the predicted probability; By back-propagating and calculating the gradient of the loss relative to each parameter, the Adam optimizer is used to update the model parameters, adjust the convolution kernel and weights, minimize the loss function and accelerate convergence. The expression is: Among them, θ t is a parameter, η is the learning rate, m t and v t are the first and second order momentum respectively; The trained model is then tested on the validation set, and the accuracy and recall are recorded to check whether the model is overfitting or underfitting. Next, the model performance is evaluated on the test set, and a confusion matrix and classification report are generated to determine the accuracy and robustness of the model in practical applications. The trained model is then deployed to the system. The model regularly detects the prediction results of the input data and compares them with the actual results. When the model prediction accuracy is lower than 80%, the retraining process is triggered.

5. A bidding document risk identification system according to claim 1, characterized in that: When collecting historical bidding data, clustering algorithms need to be used to analyze abnormal patterns in historical data and warn of hidden risks. The steps are as follows: Each bidding document is represented as a feature vector x i =(x i1 , x i2 ,dots,x id ), where d represents the dimension of the feature. Then, the K-means algorithm is used to divide the data points into K clusters, each cluster is composed of a centroid μ k It means that, then, the distance between the data point and the centroid is calculated, and it is judged whether it is an outlier and the regularization deviation term is introduced to capture the outlier. Then, the outlier is marked and a risk warning is generated. If the outlier does not match the risk in the rule base, the user is prompted to update the rule base. Finally, the new outlier and its characteristic pattern are added to the data set, and the risk rule base and model parameters are updated. The expression is: Among them, L is the clustering objective function, which is used to minimize the distance within the cluster and the degree of deviation of the outliers. K is the number of clusters, and the data points are divided into K groups. C k is the set of the Kth cluster, containing the data points belonging to this cluster, x i is the characteristic vector of the bidding document, representing the characteristic value of the i-th data point, μ k is the centroid of the kth cluster, that is, the average vector of the data points in the cluster; |x i -μ k | is the Euclidean distance, calculate the data point x i and its cluster centroid μ k The distance between k is the standard deviation vector within the kth cluster, which is used to measure the distribution of data within the cluster.

6. A bidding document risk identification system according to claim 1, characterized in that: Compare the parsed key information with the risk rules and identify the risks in the bidding documents as follows: The extracted key information is input into the data training module, and the deep learning model trained with historical data is used to generate preliminary risk prediction results to obtain the probability values ​​of risk types and levels. The parsed information is compared with the rules in the risk rule base to identify whether the key information meets the bidding requirements and industry standards. The results of the data training model prediction and the matching results of the risk rule base are summarized. Then, the real-time data information is compared with the standard values ​​in the historical data to detect potential hidden risks. By comparing the differences between the two results, the risk points in the key information are further verified. Risks confirmed by both model prediction and rule base matching are marked as high priority risks. Finally, a structured report is generated for all risks found in the comparison, listing the risk types, specific fields and violated rules, and the comparison results and analysis data are stored in the database. At the same time, the rule base is optimized and updated according to the newly identified risks to ensure that the rule base maintains dynamic adaptability.

7. A bidding document risk identification system according to claim 1, characterized in that: The steps to generate preliminary risk prediction results using the convolutional neural network model trained with historical data and obtain the probability values ​​of risk types and levels are as follows: Extract and clean various features from the historical bidding document dataset, including quotation, technical matching, and performance record, and convert each bidding document into a standardized feature matrix; The pre-processed bidding document feature matrix is ​​input into the trained CNN model. Through layer-by-layer calculation of convolutional layer, activation layer, pooling layer and fully connected layer, the deep features of the input data are extracted and the risk feature vector is gradually formed. After multiple layers of convolution and pooling operations, the model generates multi-dimensional feature maps and maps these feature maps into a risk feature space through a fully connected layer to form a comprehensive risk representation of the bid document. In the output layer, the Softmax activation function is used to calculate the probability of each risk type. Through the Softmax operation, the risk features output by the fully connected layer are mapped to probability values ​​of different risk types, each value is between 0 and 11; Output the probability value of each risk type and its level, sort them by probability, and mark the risk type and risk level with the highest probability as the preliminary risk prediction result.

8. A bidding document risk identification system according to claim 1, characterized in that: The steps to quantify the identified risks and classify them into high risk, medium risk and low risk according to their severity and impact on the project are as follows: Summarize the identified risk information, including risk type, violated rules, frequency of occurrence, and project links involved. Then, set specific quantitative indicators for each risk type, and then use the formula based on the quantitative indicators of each risk. Calculate the comprehensive score, divide the risk into high risk, medium risk and low risk according to the comprehensive score, and classify each risk into the corresponding level category. Finally, organize the quantified risk data to generate an analysis report, which lists the comprehensive score, level classification and impact description of each risk in detail, and stores the results in the database. S is the comprehensive risk score of the bidding document, ranging from 0 to 5. The higher the value, the lower the risk. i is the weight of risk type i, indicating the importance of each risk type in the comprehensive score, n i is the number of quantitative indicators under risk type i, R ij is the score of the jth quantitative indicator of risk type i, ranging from 0 to 5, indicating the indicator score, I ij is the importance coefficient of the j-th quantitative indicator of risk type i, which is used to indicate the weight of the indicator in this risk type.

9. A bidding document risk identification system according to claim 1, characterized in that: The steps to organize and analyze the identified risks, generate a risk report, and provide a basis for decision-making for the tendering unit are as follows: Collect all identified risk data, classify the data according to risk type, and group risk points of the same risk type together. Then, sort each type of risk by severity and display high risks more intuitively at the front end of the report. Calculate the overall risk score based on the comprehensive score obtained by the risk assessment module. By comparing historical data, mark the risk differences with similar bids in the past to help the bidding unit quickly determine the relative risk level of the bid. Then generate charts to show the proportion and score of different risk types, intuitively display the distribution and proportion of risks, generate standardized report documents, including risk type and risk level, and export them in PDF or Excel format.

10. A method for identifying risks in bidding documents, characterized in that: The following steps are involved: Parse the bidding documents and extract key information from them; Use historical bidding data to build and train machine learning models to predict potential risk types and levels; Storing risk rules related to bidding documents, wherein the risk rule library supports customization and updating based on different project types; Compare the parsed key information with the risk rules to identify risks in the bidding documents; Conduct quantitative analysis on the identified risks and classify them into high risk, medium risk and low risk according to the severity of the risks and their impact on the project; The identified risks will be sorted and analyzed to generate a risk report to provide a basis for decision-making for the tendering unit.

Images