Generative large language model-based official document sensitive information examination system and method

Abstract

The invention provides an official document sensitive information examination system and method based on a generative large language model, and relates to the technical field of information, and the method comprises the steps: obtaining historical official document data based on an official document database, preprocessing the historical official document data to generate target official document data, and obtaining training official document data and verification official document data after classification; a recurrent neural network is adopted to train a generation type large language model used for sensitive information examination based on the training official document data, the sensitive information identification effect of the generation type large language model used for sensitive information examination is tested based on the verification official document data, adjustment and optimization are carried out, and current official document data uploaded by a user side are received; the method comprises the following steps: preprocessing current document data, inputting the preprocessed document data into a model, identifying current document sensitive information corresponding to the current document data, recording position information, outputting a document sensitive information examination result, and displaying the result to a user side through a user interaction interface, so that the document content can be intelligently examined by utilizing a generative large language model; and the sensitive information of the official document is efficiently and accurately identified and marked.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a system and method for reviewing sensitive information in official documents based on a generative large language model. Background Art

[0002] With the rapid development of information technology, official document processing has gradually become electronic and automated, and sensitive information review is an important part of ensuring information security in the official document processing process.

[0003] In the prior art, traditional sensitive information review methods mainly rely on manual review and keyword matching. Specifically, the accuracy of sensitive information review is easily limited by the sensitive word library, so it is necessary to continuously update and maintain the sensitive word library to cope with newly emerging sensitive words. Due to the inability to accurately handle semantic variants and metaphors, this often leads to the omission or misjudgment of certain variants or metaphorical forms of sensitive words. At the same time, the design and maintenance of existing sensitive information review rules are relatively complex, requiring detailed analysis and rule writing of the characteristics and context of sensitive words. In addition, methods based on keyword matching and rule matching need to traverse the text to be reviewed and the sensitive word library one by one for matching, which may lead to low efficiency in large-scale text review.

[0004] Therefore, it is necessary to provide a sensitive information review system and method for official documents based on a generative large language model to solve the above technical problems. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a system and method for reviewing sensitive information of official documents based on a generative large language model to solve the problems that traditional sensitive information review methods often rely on manual review and keyword matching, and have low review efficiency, low accuracy, and difficulty in dealing with new sensitive words.

[0006] The present invention provides a document sensitive information review system based on a generative large language model, the review system comprising:

[0007] The official document data acquisition module is used to acquire historical official document data based on the official document database;

[0008] The document data preprocessing module is used to perform preprocessing operations on the historical document data to generate corresponding target document data;

[0009] An official document data classification module, used to classify the target official document data to generate corresponding training official document data and verification official document data;

[0010] A model training and testing module, used to train a generative large language model for sensitive information review based on the training document data using a recurrent neural network, and to test the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and to adjust and optimize the generative large language model for sensitive information review according to the test results;

[0011] The result output module is used to receive the current official document data uploaded by the user terminal, input it into the generative large language model for sensitive information review after preprocessing, identify and mark the current official document sensitive information corresponding to the current official document data, record the location information of the current official document sensitive information in the current official document data, output the corresponding official document sensitive information review result and display it to the user terminal through the user interaction interface.

[0012] Preferably, the document data preprocessing module is used to perform preprocessing operations on the historical document data to generate corresponding target document data, specifically including:

[0013] A word segmentation unit is used to perform word segmentation processing on the historical official document content by using a word segmentation algorithm, to cut the continuous text in the historical official document content into independent words and / or phrases and perform stem extraction processing to generate the corresponding first official document content;

[0014] A removal unit, configured to remove stop words in the first official document content based on a preset stop word list, and generate corresponding second official document content;

[0015] The standardization unit is used to standardize the content of the second official document to generate the corresponding target official document data.

[0016] Preferably, the adopting of a recurrent neural network to train a generative large language model for sensitive information review based on the training document data specifically includes:

[0017] Based on the recurrent neural network, the training document data is input into the generative large language model for sensitive information review, and the corresponding calculation formula for the hidden layer state update is as follows:

[0018] h t =σ(W hh h t-1 +W xh x t +b h )

[0019] Among them, h t represents the hidden layer state at the current time step t; σ represents the activation function; W hh Represents the weight matrix from the hidden layer of the previous time step to the hidden layer of the current time step; h t-1represents the hidden layer state at the previous time step t-1; W xh Represents the weight matrix from the input layer of the current time step to the hidden layer of the current time step; x t represents the training document data input at the current time step t; b h Represents the bias term of the hidden layer.

[0020] Preferably, the hidden layer state is updated iteratively, and the training document sensitive information corresponding to the training document data is generated through the output layer. The calculation formula of the output layer is as follows:

[0021] g t =W hy h t +b y

[0022] In the formula, g t W represents the sensitive information of the training document corresponding to the training document data output at the current time step t; hy Represents the weight matrix from the hidden layer of the current time step to the output layer of the current time step; h t represents the hidden layer state at the current time step t; b y Represents the bias term of the output layer;

[0023] Based on a recurrent neural network, the training document data is input into the generative large language model for sensitive information review, the hidden layer state is updated through iteration, and the training document sensitive information corresponding to the training document data is generated through the output layer, thereby completing the training of the generative large language model for sensitive information review.

[0024] Preferably, the actual document sensitive information corresponding to the training document data is obtained, and combined with the training document sensitive information, a loss function corresponding to the generative large language model for sensitive information review is constructed, and the corresponding calculation formula is:

[0025]

[0026] Where L represents the loss function corresponding to the generative large language model for sensitive information review; N represents the total number of training document data; g u represents the sensitive information of the training document corresponding to the u-th training document data; represents the actual sensitive information of the official document corresponding to the u-th training official document data; log() represents the logarithmic function; ∑ represents the summation symbol.

[0027] Preferably, a gradient descent algorithm is used to iteratively update the model parameters of the generative large language model for sensitive information review to minimize the loss function. The corresponding calculation formula is as follows:

[0028]

[0029] In the formula, θ t+1 represents the model parameters of the generative large language model for sensitive information review at the t+1th iteration; θ t represents the model parameters of the generative large language model for sensitive information review at the tth iteration; η represents the learning rate, which is used to control the update step size of the model parameters; L represents the loss function corresponding to the generative large language model for sensitive information review.

[0030] Preferably, the step of testing the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and adjusting and optimizing the generative large language model for sensitive information review according to the test result, specifically includes:

[0031] Testing the performance of the generative large language model for sensitive information review based on the verification document data and outputting a corresponding confusion matrix, and evaluating the sensitive information recognition effect of the generative large language model for sensitive information review according to the confusion matrix;

[0032] The elements of the confusion matrix include true positive examples, false positive examples, true negative examples and false negative examples, wherein the true positive examples are correct identification of sensitive information; the false positive examples are incorrect identification of sensitive information; the true negative examples are correct identification of non-sensitive information; and the false negative examples are incorrect identification of non-sensitive information;

[0033] Based on the confusion matrix, the sensitive information identification index corresponding to the generative large language model for sensitive information review is calculated, wherein the sensitive information identification index includes accuracy, precision, recall, F1 score and false positive rate;

[0034] The calculation formula of the accuracy is:

[0035]

[0036] In the formula, ZQL represents the accuracy of the generative large language model used for sensitive information review; TP represents true positive examples, that is, correctly identifying sensitive information; TN represents true negative examples, that is, correctly identifying non-sensitive information; FP represents false positive examples, that is, incorrectly identifying sensitive information; FN represents false negative examples, that is, incorrectly identifying non-sensitive information;

[0037] The calculation formula of the accuracy is:

[0038]

[0039] Where JQL represents the precision of the generative large language model used for sensitive information review; TP represents true positives, i.e., correctly identifying sensitive information; FP represents false positives, i.e., incorrectly identifying sensitive information;

[0040] The calculation formula of the recall rate is:

[0041]

[0042] Where ZHL represents the recall rate corresponding to the generative large language model for sensitive information review; TP represents true positive examples, i.e., correctly identifying sensitive information; FN represents false negative examples, i.e., incorrectly identifying non-sensitive information;

[0043] The calculation formula of the F1 score is:

[0044]

[0045] In the formula, F1 represents the F1 score corresponding to the generative large language model for sensitive information review; JQL represents the precision corresponding to the generative large language model for sensitive information review; ZHL represents the recall corresponding to the generative large language model for sensitive information review;

[0046] The calculation formula of the false positive rate is:

[0047]

[0048] Where JZL represents the false positive rate corresponding to the generative large language model for sensitive information review; FP represents false positive examples, i.e., incorrectly identifying sensitive information; TN represents true negative examples, i.e., correctly identifying non-sensitive information;

[0049] The sensitive information recognition indicators are analyzed, and the precision and recall are weighed, the sensitive information recognition effect of the generative large language model for sensitive information review is evaluated, and the generative large language model for sensitive information review is adjusted and optimized.

[0050] Preferably, the position information of the current official document sensitive information in the current official document data includes paragraph number, line number, character start position and character end position.

[0051] A method for reviewing sensitive information in official documents based on a generative large language model, the method comprising:

[0052] Obtain historical official document data based on the official document database;

[0053] Performing preprocessing operations on the historical official document data to generate corresponding target official document data;

[0054] Classifying the target official document data to generate corresponding training official document data and verification official document data;

[0055] Using a recurrent neural network to train a generative large language model for sensitive information review based on the training document data, and testing the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and adjusting and optimizing the generative large language model for sensitive information review according to the test results;

[0056] The current official document data uploaded by the user end is received, and after preprocessing, is input into the generative large language model for sensitive information review, the current official document sensitive information corresponding to the current official document data is identified and marked, the location information of the current official document sensitive information in the current official document data is recorded, the corresponding official document sensitive information review result is output and displayed to the user end through the user interaction interface.

[0057] Compared with related technologies, the system and method for reviewing sensitive information of official documents based on a generative large language model provided by the present invention have the following beneficial effects:

[0058] The present invention can obtain historical official document data based on an official document database through an official document data acquisition module, an official document data preprocessing module, an official document data classification module, a model training test module and a result output module; perform preprocessing operations on the historical official document data to generate corresponding target official document data; classify the target official document data to generate corresponding training official document data and verification official document data; use a recurrent neural network to train a generative large language model for sensitive information review based on the training official document data, and test the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification official document data, and adjust and optimize the generative large language model for sensitive information review according to the test results; receive the current official document data uploaded by the user end, and input it into the model for sensitive information review after preprocessing. In the generative large language model of the search, the current official document sensitive information corresponding to the current official document data is identified and marked, the position information of the current official document sensitive information in the current official document data is recorded, the corresponding official document sensitive information review result is output and displayed to the user end through the user interaction interface, so that the generative large language model can be used to perform intelligent review of the official document sensitive information, which greatly improves the review efficiency and reduces the workload of manual review; the generative large language model of the present invention can identify and mark the sensitive information in the official document, including sensitive words expressed in variants and metaphors, thereby improving the accuracy of the review; the system of the present invention supports real-time updating of the sensitive information database, can cope with newly emerging sensitive words and phrases, and ensure the timeliness of the review; and can adapt to a variety of review needs and be applied to different types of official document reviews. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a system block diagram of the sensitive information review system of official documents based on the generative large language model of the present invention;

[0060] Figure 2 A schematic diagram showing the results of the sensitive information review of official documents of the present invention;

[0061] Figure 3 It is a structural block diagram of the official document data preprocessing module of the official document sensitive information review system based on the generative large language model of the present invention;

[0062] Figure 4 is a schematic diagram of confusion matrix elements of the present invention;

[0063] Figure 5 The present invention is a flow chart of the method for reviewing sensitive information in official documents based on a generative large language model. DETAILED DESCRIPTION

[0064] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0065] Embodiment 1

[0066] like Figure 1 As shown, a document sensitive information review system based on a generative large language model includes:

[0067] The official document data acquisition module is used to acquire historical official document data based on the official document database;

[0068] The document data preprocessing module is used to perform preprocessing operations on the historical document data to generate corresponding target document data;

[0069] An official document data classification module, used to classify the target official document data to generate corresponding training official document data and verification official document data;

[0070] A model training and testing module, used to train a generative large language model for sensitive information review based on the training document data using a recurrent neural network, and to test the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and to adjust and optimize the generative large language model for sensitive information review according to the test results;

[0071] The result output module is used to receive the current official document data uploaded by the user terminal, input it into the generative large language model for sensitive information review after preprocessing, identify and mark the current official document sensitive information corresponding to the current official document data, record the location information of the current official document sensitive information in the current official document data, output the corresponding official document sensitive information review result and display it to the user terminal through the user interaction interface.

[0072] Among them, the document data acquisition module can automatically collect historical document data from the document database through an integrated data interface. The historical document data covers a wide range of document types and can ensure the timeliness and representativeness of the data.

[0073] The official document data preprocessing module uses natural language processing (NLP) technology to perform deep cleaning, word segmentation, part-of-speech tagging, and stop word removal on historical official document data, generating target official document data with a unified format and optimized quality, thereby improving subsequent model training efficiency and recognition accuracy.

[0074] The official document data classification module can subdivide the target official document data into two parts, namely training document data and verification document data, based on preset classification standards, such as content nature, confidentiality level, etc., thereby helping the model to learn more extensive and representative features in the subsequent training process.

[0075] The model training and testing module can use a recurrent neural network (RNN) to train a generative large language model based on training document data. After training, the model can be tested using verification document data to evaluate its accuracy and robustness in identifying sensitive information. Based on the test results, the model performance can be continuously optimized through parameter tuning, model architecture adjustment, and other means.

[0076] Finally, the result output module can receive the current document data uploaded by the user, and input it into the optimized generative large language model after preprocessing. This model can perform sensitive information identification tasks, accurately mark the sensitive information in the current document, and record the specific location of this information in the document. Figure 2 As shown, the results of the sensitive information review of official documents are output in a structured form and intuitively displayed to the user through the user interaction interface on the terminal device, so that the user can quickly view and take corresponding processing measures.

[0077] Through the above method, a generative large language model can be used to conduct intelligent review of sensitive information in official documents, identify and mark sensitive information in official documents, greatly improve the efficiency and accuracy of the review, reduce the workload of manual review, and support real-time updating of the sensitive information database to ensure the timeliness of the review. It can also adapt to a variety of review needs and be applied to different types of official document reviews.

[0078] In the specific implementation process, Figure 3 As shown, the document data preprocessing module 300 is used to perform preprocessing operations on the historical document data to generate corresponding target document data, specifically including:

[0079] The word segmentation unit 301 is used to perform word segmentation processing on the historical official document content by using a word segmentation algorithm, to divide the continuous text in the historical official document content into independent words and / or phrases and perform stem extraction processing to generate the corresponding first official document content;

[0080] A removal unit 302, configured to remove stop words in the first official document content based on a preset stop word list to generate corresponding second official document content;

[0081] The standardization unit 303 is used to perform standardization processing on the second official document content to generate the corresponding target official document data.

[0082] It can be understood that the word segmentation unit can use a word segmentation algorithm to perform word segmentation processing on the input historical document content, effectively divide the continuous text content into independent words and / or phrases, and further perform stem extraction operations to eliminate differences in lexical morphology, thereby generating a first document content that is consistent and comparable.

[0083] Based on a pre-constructed stop word list, the removal unit can accurately identify and remove stop words in the first official document content that contribute little to the text meaning or have no practical meaning, and generate refined second official document content, thereby effectively reducing data noise and improving the efficiency and accuracy of subsequent processing.

[0084] The standardization unit can perform comprehensive standardization processing on the content of the second official document, including standardized mapping of vocabulary, spelling correction, and format unification, etc., to generate target official document data that meets specific standards and requirements, thereby ensuring the consistency of all official document data in format and content and improving data availability.

[0085] The adopting of a recurrent neural network to train a generative large language model for sensitive information review based on the training document data specifically includes:

[0086] Based on the recurrent neural network, the training document data is input into the generative large language model for sensitive information review, and the corresponding calculation formula for the hidden layer state update is as follows:

[0087] h t =σ(W hh h t-1 +W xh x t +b h )

[0088] Among them, h t represents the hidden layer state at the current time step t; σ represents the activation function; W hh Represents the weight matrix from the hidden layer of the previous time step to the hidden layer of the current time step; h t-1represents the hidden layer state at the previous time step t-1; W xh Represents the weight matrix from the input layer of the current time step to the hidden layer of the current time step; x t represents the training document data input at the current time step t; b h Represents the bias term of the hidden layer.

[0089] By iteratively updating the hidden layer state, and generating the training document sensitive information corresponding to the training document data through the output layer, the calculation formula of the output layer is as follows:

[0090] g t =W hy h t +b y

[0091] In the formula, g t W represents the sensitive information of the training document corresponding to the training document data output at the current time step t; hy Represents the weight matrix from the hidden layer of the current time step to the output layer of the current time step; h t represents the hidden layer state at the current time step t; b y Represents the bias term of the output layer;

[0092] Based on a recurrent neural network, the training document data is input into the generative large language model for sensitive information review, the hidden layer state is updated through iteration, and the training document sensitive information corresponding to the training document data is generated through the output layer, thereby completing the training of the generative large language model for sensitive information review.

[0093] It should be noted that, first, the preprocessed and formatted training document data can be input into the generative large language model based on the recurrent neural network. In this process, the hidden layer state inside the model can be dynamically updated according to the input document data. Specifically, the hidden layer state of the current time step is determined by the hidden layer state of the previous time step, the input data of the current time step, the weight matrix and the bias term, and the activation function can ensure the nonlinear transformation of the hidden layer state and enhance the expressiveness of the model.

[0094] By continuously iteratively updating the hidden layer state, the temporal dependency of text data can be simulated, allowing the model to capture contextual information in the official document data. The output layer can generate sensitive information corresponding to the training document data based on the latest hidden layer state, thus completing the training process of the generative large language model for sensitive information review.

[0095] The actual document sensitive information corresponding to the training document data is obtained, and combined with the training document sensitive information, a loss function corresponding to the generative large language model for sensitive information review is constructed. The corresponding calculation formula is:

[0096]

[0097] Where L represents the loss function corresponding to the generative large language model for sensitive information review; N represents the total number of training document data; g u represents the sensitive information of the training document corresponding to the u-th training document data; represents the actual sensitive information of the official document corresponding to the u-th training official document data; log() represents the logarithmic function; ∑ represents the summation symbol.

[0098] The model parameters of the generative large language model for sensitive information review are iteratively updated using a gradient descent algorithm to minimize the loss function. The corresponding calculation formula is as follows:

[0099]

[0100] In the formula, θ t+1 represents the model parameters of the generative large language model for sensitive information review at the t+1th iteration; θ t represents the model parameters of the generative large language model for sensitive information review at the tth iteration; η represents the learning rate, which is used to control the update step size of the model parameters; L represents the loss function corresponding to the generative large language model for sensitive information review.

[0101] First, the actual sensitive information of the official documents corresponding to the training official document data can be obtained, and the actual sensitive information can be compared with the sensitive information marked in the training official document data to construct the loss function of the generative large language model for sensitive information review. In order to minimize the loss function, the gradient descent algorithm can be used to iteratively update the model parameters of the generative large language model for sensitive information review to gradually approach the optimal model parameters and achieve accurate review of sensitive information.

[0102] The step of testing the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and adjusting and optimizing the generative large language model for sensitive information review according to the test result, specifically includes:

[0103] Testing the performance of the generative large language model for sensitive information review based on the verification document data and outputting a corresponding confusion matrix, and evaluating the sensitive information recognition effect of the generative large language model for sensitive information review according to the confusion matrix;

[0104] The elements of the confusion matrix include true positive examples, false positive examples, true negative examples and false negative examples, wherein the true positive examples are correct identification of sensitive information; the false positive examples are incorrect identification of sensitive information; the true negative examples are correct identification of non-sensitive information; and the false negative examples are incorrect identification of non-sensitive information;

[0105] Based on the confusion matrix, the sensitive information identification index corresponding to the generative large language model for sensitive information review is calculated, wherein the sensitive information identification index includes accuracy, precision, recall, F1 score and false positive rate;

[0106] The calculation formula of the accuracy is:

[0107]

[0108] In the formula, ZQL represents the accuracy of the generative large language model used for sensitive information review; TP represents true positive examples, that is, correctly identifying sensitive information; TN represents true negative examples, that is, correctly identifying non-sensitive information; FP represents false positive examples, that is, incorrectly identifying sensitive information; FN represents false negative examples, that is, incorrectly identifying non-sensitive information;

[0109] The calculation formula of the accuracy is:

[0110]

[0111] Where JQL represents the precision of the generative large language model used for sensitive information review; TP represents true positives, i.e., correctly identifying sensitive information; FP represents false positives, i.e., incorrectly identifying sensitive information;

[0112] The calculation formula of the recall rate is:

[0113]

[0114] Where ZHL represents the recall rate corresponding to the generative large language model for sensitive information review; TP represents true positive examples, i.e., correctly identifying sensitive information; FN represents false negative examples, i.e., incorrectly identifying non-sensitive information;

[0115] The calculation formula of the F1 score is:

[0116]

[0117] In the formula, F1 represents the F1 score corresponding to the generative large language model for sensitive information review; JQL represents the precision corresponding to the generative large language model for sensitive information review; ZHL represents the recall corresponding to the generative large language model for sensitive information review;

[0118] The calculation formula of the false positive rate is:

[0119]

[0120] Where JZL represents the false positive rate corresponding to the generative large language model for sensitive information review; FP represents false positive examples, i.e., incorrectly identifying sensitive information; TN represents true negative examples, i.e., correctly identifying non-sensitive information;

[0121] The sensitive information recognition indicators are analyzed, and the precision and recall are weighed, the sensitive information recognition effect of the generative large language model for sensitive information review is evaluated, and the generative large language model for sensitive information review is adjusted and optimized.

[0122] In practical applications, we can first use the verification document data to test the performance of the generative large language model used for sensitive information review and generate the corresponding confusion matrix as the basis for evaluation. Figure 4 As shown in the figure, the confusion matrix is ​​a key tool for evaluating the performance of classification models. Its elements can be divided into true positive examples TP, false positive examples FP, true negative examples TN, and false negative examples FN. Among them, true positive examples are instances where the model correctly identifies sensitive information; false positive examples refer to instances where the model incorrectly identifies sensitive information; true negative examples refer to instances where the model correctly identifies non-sensitive information; and false negative examples refer to instances where the model incorrectly identifies non-sensitive information.

[0123] Based on the confusion matrix, we can further calculate sensitive information recognition indicators, including accuracy, precision, recall, F1 score, and false positive rate, to comprehensively quantify the recognition performance of the model. Specifically, accuracy reflects the proportion of all test samples that are correctly classified by the model; precision reflects the proportion of instances that are actually sensitive information among those identified as sensitive information by the model; and recall measures the proportion of all actual sensitive information that is correctly identified by the model. The F1 score, as the harmonic mean of precision and recall, provides a comprehensive performance indicator; the false positive rate reflects the proportion of non-sensitive information that is incorrectly identified as sensitive information by the model.

[0124] Furthermore, these sensitive information identification indicators can be analyzed in depth. By weighing the relationship between precision and recall, the performance of the model in sensitive information review can be comprehensively evaluated. According to the evaluation results, the model can be adjusted and optimized accordingly, including parameter tuning, algorithm improvement, or the introduction of new feature engineering strategies, to improve the overall performance of the model in sensitive information identification.

[0125] The position information of the current document sensitive information in the current document data includes paragraph number, line number, character start position and character end position.

[0126] It should be noted that the location information of the current sensitive information in the current document data set includes multiple dimensions such as paragraph number, line number, character start position and character end position. Specifically, the paragraph number is used to identify the document paragraph where the sensitive information is located; the line number is used to determine the line where the sensitive information is located; the character start position and character end position respectively indicate the specific start and end boundaries of the sensitive information in the line, so that the sensitive information can be comprehensively and accurately located in the document data.

[0127] Embodiment 2

[0128] like Figure 5 As shown, a method for reviewing sensitive information in official documents based on a generative large language model includes:

[0129] S1, obtaining historical official document data based on the official document database;

[0130] S2, performing a preprocessing operation on the historical official document data to generate corresponding target official document data;

[0131] S3, classifying the target official document data to generate corresponding training official document data and verification official document data;

[0132] S4, using a recurrent neural network to train a generative large language model for sensitive information review based on the training document data, and testing the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and adjusting and optimizing the generative large language model for sensitive information review according to the test results;

[0133] S5, receiving the current official document data uploaded by the user terminal, inputting it into the generative large language model for sensitive information review after preprocessing, identifying and marking the current official document sensitive information corresponding to the current official document data, recording the location information of the current official document sensitive information in the current official document data, outputting the corresponding official document sensitive information review result and displaying it to the user terminal through the user interaction interface.

[0134] Through the introduction of the above embodiments, the present invention uses a system and method for reviewing sensitive information of official documents based on a generative large language model, through an official document data acquisition module, an official document data preprocessing module, an official document data classification module, a model training test module and a result output module, can obtain historical official document data based on an official document database; preprocess the historical official document data to generate corresponding target official document data; classify the target official document data to generate corresponding training official document data and verification official document data; use a recurrent neural network to train a generative large language model for sensitive information review based on the training official document data, and test the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification official document data, and adjust and optimize the generative large language model for sensitive information review according to the test results; receive the current official document uploaded by the user end The document data is pre-processed and input into a generative large language model for sensitive information review, the current document sensitive information corresponding to the current document data is identified and marked, the position information of the current document sensitive information in the current document data is recorded, the corresponding document sensitive information review result is output and displayed to the user end through a user interaction interface, so that the generative large language model can be used to perform intelligent review of the document sensitive information, which greatly improves the review efficiency and reduces the workload of manual review; the generative large language model of the present invention can identify and mark sensitive information in the document, including sensitive words expressed in variants and metaphors, thereby improving the accuracy of the review; the system of the present invention supports real-time updating of the sensitive information database, can cope with newly emerging sensitive words and phrases, and ensure the timeliness of the review; and can adapt to a variety of review needs and be applied to different types of document reviews.

[0135] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0136] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0137] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

Claims

1. The sensitive information review system for official documents based on the generative large language model is characterized by: The review system includes: The official document data acquisition module is used to acquire historical official document data based on the official document database; The document data preprocessing module is used to perform preprocessing operations on the historical document data to generate corresponding target document data; An official document data classification module, used to classify the target official document data to generate corresponding training official document data and verification official document data; A model training and testing module, used to train a generative large language model for sensitive information review based on the training document data using a recurrent neural network, and to test the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and to adjust and optimize the generative large language model for sensitive information review according to the test results; The result output module is used to receive the current official document data uploaded by the user terminal, input it into the generative large language model for sensitive information review after preprocessing, identify and mark the current official document sensitive information corresponding to the current official document data, record the location information of the current official document sensitive information in the current official document data, output the corresponding official document sensitive information review result and display it to the user terminal through the user interaction interface.

2. The sensitive information review system for official documents based on a generative large language model according to claim 1 is characterized in that: The document data preprocessing module is used to perform preprocessing operations on the historical document data to generate corresponding target document data, specifically including: A word segmentation unit is used to perform word segmentation processing on the historical official document content by using a word segmentation algorithm, to cut the continuous text in the historical official document content into independent words and / or phrases and perform stem extraction processing to generate the corresponding first official document content; A removal unit, configured to remove stop words in the first official document content based on a preset stop word list, and generate corresponding second official document content; The standardization unit is used to standardize the content of the second official document to generate the corresponding target official document data.

3. The sensitive information review system for official documents based on a generative large language model according to claim 1 is characterized in that: The adopting of a recurrent neural network to train a generative large language model for sensitive information review based on the training document data specifically includes: Based on the recurrent neural network, the training document data is input into the generative large language model for sensitive information review, and the corresponding calculation formula for the hidden layer state update is as follows: h t =σ(W hh h t-1 +W xh x t +b h ) Among them, h t represents the hidden layer state at the current time step t; σ represents the activation function; W hh Represents the weight matrix from the hidden layer of the previous time step to the hidden layer of the current time step; h t-1 represents the hidden layer state at the previous time step t-1; W xh Represents the weight matrix from the input layer of the current time step to the hidden layer of the current time step; x t represents the training document data input at the current time step t; b h Represents the bias term of the hidden layer.

4. The sensitive information review system for official documents based on a generative large language model according to claim 3 is characterized in that: By iteratively updating the hidden layer state, and generating the training document sensitive information corresponding to the training document data through the output layer, the calculation formula of the output layer is as follows: g t =W hy h t +b y In the formula, g t W represents the sensitive information of the training document corresponding to the training document data output at the current time step t; hy Represents the weight matrix from the hidden layer of the current time step to the output layer of the current time step; h t represents the hidden layer state at the current time step t; b y Represents the bias term of the output layer; Based on a recurrent neural network, the training document data is input into the generative large language model for sensitive information review, the hidden layer state is updated through iteration, and the training document sensitive information corresponding to the training document data is generated through the output layer, thereby completing the training of the generative large language model for sensitive information review.

5. The sensitive information review system for official documents based on a generative large language model according to claim 4 is characterized in that: The actual document sensitive information corresponding to the training document data is obtained, and combined with the training document sensitive information, a loss function corresponding to the generative large language model for sensitive information review is constructed. The corresponding calculation formula is: Where L represents the loss function corresponding to the generative large language model for sensitive information review; N represents the total number of training document data; g u represents the sensitive information of the training document corresponding to the u-th training document data; represents the actual sensitive information of the official document corresponding to the u-th training official document data; log() represents the logarithmic function; ∑ represents the summation symbol.

6. The sensitive information review system for official documents based on a generative large language model according to claim 5 is characterized in that: The model parameters of the generative large language model for sensitive information review are iteratively updated using a gradient descent algorithm to minimize the loss function. The corresponding calculation formula is as follows: In the formula, θ t+1 represents the model parameters of the generative large language model for sensitive information review at the t+1th iteration; θ t represents the model parameters of the generative large language model for sensitive information review at the tth iteration; η represents the learning rate, which is used to control the update step size of the model parameters; L represents the loss function corresponding to the generative large language model for sensitive information review.

7. The sensitive information review system for official documents based on a generative large language model according to claim 1 is characterized in that: The step of testing the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and adjusting and optimizing the generative large language model for sensitive information review according to the test result, specifically includes: Testing the performance of the generative large language model for sensitive information review based on the verification document data and outputting a corresponding confusion matrix, and evaluating the sensitive information recognition effect of the generative large language model for sensitive information review according to the confusion matrix; The elements of the confusion matrix include true positive examples, false positive examples, true negative examples and false negative examples, wherein the true positive examples are correct identification of sensitive information; the false positive examples are incorrect identification of sensitive information; the true negative examples are correct identification of non-sensitive information; and the false negative examples are incorrect identification of non-sensitive information; Based on the confusion matrix, the sensitive information identification index corresponding to the generative large language model for sensitive information review is calculated, wherein the sensitive information identification index includes accuracy, precision, recall, F1 score and false positive rate; The calculation formula of the accuracy is: In the formula, ZQL represents the accuracy of the generative large language model used for sensitive information review; TP represents true positive examples, that is, correctly identifying sensitive information; TN represents true negative examples, that is, correctly identifying non-sensitive information; FP represents false positive examples, that is, incorrectly identifying sensitive information; FN represents false negative examples, that is, incorrectly identifying non-sensitive information; The calculation formula of the accuracy is: Where JQL represents the precision of the generative large language model used for sensitive information review; TP represents true positives, i.e., correctly identifying sensitive information; FP represents false positives, i.e., incorrectly identifying sensitive information; The calculation formula of the recall rate is: Where ZHL represents the recall rate corresponding to the generative large language model for sensitive information review; TP represents true positive examples, i.e., correctly identifying sensitive information; FN represents false negative examples, i.e., incorrectly identifying non-sensitive information; The calculation formula of the F1 score is: In the formula, F1 represents the F1 score corresponding to the generative large language model for sensitive information review; JQL represents the precision corresponding to the generative large language model for sensitive information review; ZHL represents the recall corresponding to the generative large language model for sensitive information review; The calculation formula of the false positive rate is: Where JZL represents the false positive rate corresponding to the generative large language model for sensitive information review; FP represents false positive examples, i.e., incorrectly identifying sensitive information; TN represents true negative examples, i.e., correctly identifying non-sensitive information; The sensitive information recognition indicators are analyzed, and the precision and recall are weighed, the sensitive information recognition effect of the generative large language model for sensitive information review is evaluated, and the generative large language model for sensitive information review is adjusted and optimized.

8. The sensitive information review system for official documents based on a generative large language model according to claim 1 is characterized in that: The position information of the current document sensitive information in the current document data includes paragraph number, line number, character start position and character end position.

9. A method for reviewing sensitive information in official documents based on a generative large language model, applied to a system for reviewing sensitive information in official documents based on a generative large language model as claimed in any one of claims 1 to 8, wherein the review method comprises: Obtain historical official document data based on the official document database; Performing preprocessing operations on the historical official document data to generate corresponding target official document data; Classifying the target official document data to generate corresponding training official document data and verification official document data; Using a recurrent neural network to train a generative large language model for sensitive information review based on the training document data, and testing the sensitive information recognition effect of the generative large language model for sensitive information review based on the verification document data, and adjusting and optimizing the generative large language model for sensitive information review according to the test results; The current official document data uploaded by the user end is received, and after preprocessing, is input into the generative large language model for sensitive information review, the current official document sensitive information corresponding to the current official document data is identified and marked, the location information of the current official document sensitive information in the current official document data is recorded, the corresponding official document sensitive information review result is output and displayed to the user end through the user interaction interface.

Images