An intelligent classification and retrieval system for government files

Abstract

The application discloses a kind of government affair file intelligent classification and retrieval system, belong to file management and information technology field.The system is aimed at the problems, such as low efficiency, high error rate of traditional government affair file management mode, and the problems, such as insufficient safety control in government affair file management of existing general file management system, low search accuracy, etc., intelligent classification is carried out using the deep learning architecture combining CNN and RNN, and accurate search is realized through word vector model and attention mechanism.The system constructs government affair field knowledge graph to provide intelligent decision support, designs multi-level authority management system to guarantee file safety, also has real-time data updating and self-learning mechanism, and can adapt to the rapid update of government affair knowledge.Compared with prior art, the system has significant advantages in classification and retrieval efficiency, accuracy, safety, compatibility and intelligent expandability, etc., can realize the intelligent, efficient and standardized management of government affair file, and meet the complex needs of government affair office.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of file management and information technology, and specifically relates to an intelligent classification and retrieval system for government files. BACKGROUND

[0002] In the development process of government file management, the traditional manual management method has long dominated. Staff members manually classify files according to simple classification catalogs based on experience, and rely on file name keywords to find files in shared folders or file cabinets during retrieval. This method is not only inefficient, but also has a 12%-18% error rate due to the subjectivity of manual judgment. According to research, a 50-person department spends 800-1200 hours per month on file classification, and the average retrieval time is 30-40 minutes per time.

[0003] With the development of information technology, general file management systems such as SharePoint and WPS enterprise cloud documents have emerged. They have basic storage, classification browsing, and keyword retrieval functions, and some have simple permission management. However, for government file management, there are obvious defects. In terms of security control, the permission setting is not fine enough, and more than 70% of government departments using it have feedback that it cannot meet the multi-level permission needs of different classified files. At the same time, there is a lack of understanding of professional terms in government files, and the retrieval accuracy is only 60%-70%. The system involved in the patent with publication number CN119068494A also has similar problems in terms of government specificity.

[0004] Although there are many frontier explorations in the field of academic research on text classification and information retrieval, such as deep learning models based on CNN and RNN and improved vector space retrieval algorithms. However, in practical applications, these technologies face many challenges. On the one hand, in the face of rapid updates in government knowledge, models lack self-learning and adaptive capabilities. The deep learning model introduced by the research team has a classification accuracy rate that drops from 85% to 60% within a month after the release of new policies. The patent technology with publication number CN118861213B also fails to effectively solve this problem. On the other hand, when dealing with complex and diverse government file formats, it is difficult to ensure compatibility and processing accuracy, with an average of 20%-30% of files experiencing parsing errors or content loss. SUMMARY

[0005] The present application solves the problems in the processing of government files by deeply integrating advanced technologies such as deep learning, semantic understanding, and index optimization algorithms. It realizes an organic overall solution to meet the complex needs of file management in government office scenarios, and an intelligent classification and retrieval system for government files.

[0006] The application discloses a government affair file intelligent classification and retrieval system, which comprises a feature extraction module, a knowledge graph module, a retrieval module, an urgency evaluation module and a permission management module.

[0007] The feature extraction module adopts a deep learning architecture combining a convolutional neural network model and a recurrent neural network model, wherein the convolutional neural network model is provided with multiple convolutional layers, and the size of the convolution kernel is dynamically adjusted according to the file features, for extracting local features of the file; the recurrent neural network model adopts an LSTM unit and fuses a self-attention mechanism, for capturing the logical relationship of the text sequence.

[0008] The knowledge graph module constructs a government affair field knowledge graph, fuses cross-field knowledge such as government affairs, economy and society, and has a dynamic updating mechanism, for monitoring government affair data sources in real time and updating entities and relationships.

[0009] The retrieval module converts the file and the retrieval word into vectors by means of a word vector model, calculates the semantic similarity by a cosine similarity formula, and self-adaptively adjusts the attention weight according to the retrieval word heat, the user historical retrieval behavior and the file timeliness by combining the attention mechanism, to realize accurate retrieval.

[0010] The urgency evaluation module determines the timeliness of the government affair file based on a fuzzy rule base and a dynamic rule weight adjustment mechanism, to reflect the urgency of the government affair file.

[0011] The permission management module combines the organization structure of the government department and the file secret level, designs an access control algorithm based on the role and the secret level, simultaneously introduces a risk evaluation mechanism, dynamically adjusts the permission according to the file access risk, and cooperates the permission management with the file encryption, so that different permission users obtain different decryption keys.

[0012] Further, the government affair file intelligent classification and retrieval system comprises a preprocessing module, which pre-processes the uploaded government affair file, including removing repeated and invalid files from the government affair file, classifying and storing the government affair file according to the department, the year and the like, and converting the government affair file into a unified text format and removing the header, the footer and the advertising information in the file.

[0013] Further, the convolutional neural network model is based on a dynamic convolution kernel selection method, and the specific steps are as follows:

[0014] The number of words L, the number of paragraphs P, the title level H and the nesting list depth D of the file are extracted, a multi-dimensional feature vector X=[L, P, H, D] is constructed, a complexity score is calculated, and the file is divided into a high-complexity file, a medium-complexity file and a low-complexity file based on the complexity score.

[0015] The semantic features of high complexity files, medium complexity files and low complexity files are extracted using different scale convolution kernels, and the multi-scale features are fused by weighted splicing;

[0016] The convolutional neural network uses a government field labeled data set in the training stage, and the loss function uses a cross-entropy with a class balance factor:

[0017]

[0018] Wherein, Relieve the class imbalance problem, c is the class number, N c is the number of samples of class c; the optimizer uses AdamW, and the learning rate is η=0.01.

[0019] Further, the complexity score is represented as:

[0020] C=w T ·X norm +b

[0021] Wherein, w=[0.35,0.25,0.20,0.20] T is the weight vector, X norm is the normalized feature of word number L, paragraph number P, title level H and nested list depth D, and b is the bias term.

[0022] Wherein, C≥1.2 is a high complexity file, 0.8≤C<1.2 is a medium complexity file, and C<0.8 is a low complexity file.

[0023] Further, the feature extraction process of high complexity files, medium complexity files and low complexity files is respectively:

[0024] For high complexity files, a 7×7 convolution kernel is used to extract long-range semantic features, and a dilated convolution is used to expand the receptive field, with a dilated rate r=2, and an output high complexity feature map F large ;

[0025] For medium complexity files, a 5×5 convolution kernel is used to extract semantic features, combined with a gating mechanism to filter noise features, and the gating weight g i,j is:

[0026] g i,j =σ(W g ·F i,j +b g )

[0027] Wherein, σ is the Sigmoid function, W g is the weight matrix in the gating mechanism, used for weighting the feature F i,j , F i,jis the feature vector at the (i, j) position in the feature map obtained after the convolution operation, containing the local feature information of the file, b g is the bias term in the gating mechanism, used to adjust the overall level of the gating output; the output is the complexity feature map:

[0028] F mid is the feature vector at the (i, j) position in the feature map obtained after the convolution operation, containing the local feature information of the file, b i,j ⊙F raw

[0029] F raw is the feature vector at the (i, j) position in the feature map obtained after the convolution operation, containing the local feature information of the file, b spatial is:

[0030] A spatial = Sigmoid(W s · AvgPool(F sm )+W c · MaxPool(F sm ))

[0031] F small = A spatial ⊙F raw

[0032] Sigmoid is a normalization function, W s and W c are weight matrices, AvgPool is an average pooling operation, F sm is the original feature map; F small is the low complexity feature map after attention weighting, is, MaxPool is a maximum pooling operation;

[0033] Finally, the multi-scale features are fused by weighted splicing:

[0034] F fusion = α·F large + β·F mid + γ·F small

[0035] where F fusion is the fused feature map;

[0036] The weight coefficients α, β, γ are dynamically calculated by the complexity score C, and the calculation formula is:

[0037]

[0038] k1k2k3 are parameters used to dynamically calculate the weight coefficients, which are constants determined in the model training or design process.

[0039] Further, the training stage adopts a government affairs field annotation dataset, and the government affairs field annotation dataset includes 10 categories and 500,000 files;

[0040] Specifically, the government affairs field annotation dataset includes ten core annotation categories, covering official document type classification, policy field division, agency entity recognition, policy term extraction, business process step annotation, livelihood appeal classification, legal regulation clause positioning, time node annotation, responsibility subject association and emergency level classification.

[0041] Further, the emergency level evaluation module constructs a fuzzy logic model through a mixed membership function;

[0042] The mixed membership function includes three variables of time sensitivity, influence range and associated event level;

[0043] The right half-triangle function is used to define the membership degree of time sensitivity:

[0044]

[0045] μ High (T s ) represents the membership degree when the time sensitivity is "high", which is used to measure the matching degree of the emergency level of the government document in the time dimension with the "high" emergency level, and the value range is between 0 and 1, and the larger the value represents that the document is closer to the "high" emergency level in time; T s is a time sensitivity variable used to describe the deadline or time-related attributes of the government document, which is an input variable for calculating the membership degree;

[0046] This function increases linearly when T s ≥ 2, and the maximum time sensitivity is located at T s = 3, reflecting that the closer the deadline, the higher the emergency level;

[0047] The trapezoidal function is used to define the membership degree of the influence range:

[0048]

[0049] μ Regional (S i ) represents the membership degree when the influence range is "regional", which is used to measure the matching degree of the influence range of the government document with the "regional" range, and the value range is between 0 and 1, and the value is 1 when the influence range of the document completely meets the definition of "regional", and the closer to 0 represents the lower matching degree with the "regional" range, S iThe variable representing the scope of influence describes the size of the influence of government documents and is the input variable for calculating the membership degree.

[0050] This function is in S i Within the interval [0.5, 2.5], the membership degree remains at 1, and it decays linearly on both sides, making it suitable for describing the smooth process of policy influence transitioning from within a department to the entire city.

[0051] For the association of time levels, a Gaussian function is used to define the membership degree of the associated event levels:

[0052]

[0053] μ Important (E l E represents the degree of membership when the associated event level is "important". It measures the degree of matching between the level of the associated event and the "important" level in a government document. A higher value indicates that the associated event better meets the "important" standard. The value ranges from 0 to 1. l The association level variable is used to describe the importance level of association events in government documents and is the input variable for calculating the membership degree.

[0054] Furthermore, a fuzzy rule base containing 27 rules was constructed based on time sensitivity, scope of impact, and level of related events;

[0055] The generation process of the rule base begins with in-depth analysis of the input variables and the definition of fuzzy sets. Time sensitivity, scope of influence, and level of associated events are divided into three fuzzy sets according to their value ranges: time sensitivity is divided into "low", "medium" and "high", scope of influence is divided into "local", "regional" and "global", and level of associated events is divided into "normal", "important" and "urgent".

[0056] Furthermore, based on the definition of fuzzy sets, the rule base is constructed following an IF-THEN logical structure, covering all possible combinations of input variables:

[0057] Among them, low, local, and normal correspond to the level of urgency as flat items;

[0058] Low, local, and important correspond to the level of urgency as flat items;

[0059] Low, Local, and Emergency correspond to the urgency level of "urgent";

[0060] Low, regional, and regular levels of urgency are classified as standard items;

[0061] Low, Regional, and Important levels correspond to the urgency level as "Extreme".

[0062] Low, Regional, and Emergency levels correspond to the urgency level of "Extreme".

[0063] Low, global, routine corresponds to normal urgency level;

[0064] Low, global, important corresponds to high urgency level;

[0065] Low, global, urgent corresponds to critical urgency level;

[0066] Medium, local, routine corresponds to normal urgency level;

[0067] Medium, local, important corresponds to normal urgency level;

[0068] Medium, local, urgent corresponds to high urgency level;

[0069] Medium, regional, routine corresponds to normal urgency level;

[0070] Medium, regional, important corresponds to high urgency level;

[0071] Medium, regional, urgent corresponds to critical urgency level;

[0072] Medium, global, routine corresponds to high urgency level;

[0073] Medium, global, important corresponds to critical urgency level;

[0074] Medium, global, urgent corresponds to critical urgency level;

[0075] High, local, routine corresponds to high urgency level;

[0076] High, local, important corresponds to high urgency level;

[0077] High, local, urgent corresponds to critical urgency level;

[0078] High, regional, routine corresponds to high urgency level;

[0079] High, regional, important corresponds to critical urgency level;

[0080] High, regional, urgent corresponds to critical urgency level;

[0081] High, global, routine corresponds to critical urgency level;

[0082] High, global, important corresponds to critical urgency level;

[0083] High, global, urgent corresponds to critical urgency level.

[0084] More specifically, the dynamic rule weight adjustment mechanism comprises:

[0085] initial weight of each rule All are 1, indicating that all rules contribute to the output in the initial state; by collecting real-time feedback of users on the results of permission allocation, combined with the model predicted emergency degree matching degree r predicted , dynamically adjust the rule weight:

[0086]

[0087] represents the weight of the i-th rule at t time; represents the weight of the i-th rule at t+1 time; η represents the learning rate; r actual represents the actual emergency degree matching degree;

[0088] Dynamic weight adjustment also works with the risk assessment model, and the cooperation is realized through the weight scaling factor λ R , the calculation formula of λ R is:

[0089]

[0090] R represents the real-time risk value, R crit represents the risk critical value;

[0091] In order to prevent long-term weight deviation from causing the model to deviate from the initial design, the system sets a weight decay mechanism:

[0092]

[0093] Among them, the attenuation coefficient γ=0.001, which ensures that the weight gradually returns to the initial value when there is no continuous feedback, and maintains the stability of the system.

[0094] The beneficial effects achieved by the present application are:

[0095] In terms of efficiency and accuracy, the system shows excellent advantages. The intelligent classification architecture based on CNN and RNN revolutionizes the low-efficiency mode of traditional manual classification, and can complete the classification of a large number of government files in an instant, greatly shortening the file classification period. The retrieval link is also excellent, and the retrieval driven by semantic understanding combined with the attention mechanism discards the low-efficiency mode of traditional file name keyword retrieval, shortens the original retrieval time of 30-40 minutes per time to several seconds or several minutes, and improves the classification accuracy rate to more than 95%. The retrieval accuracy rate is improved from 60%-70% of the general system to more than 90%, which accurately meets the demand of government office for fast and accurate acquisition of files.

[0096] From the perspective of security and compatibility, the system effectively solves the problems of the prior art. The multi-level permission management system makes fine-grained permission allocation for different secret-level files and user roles, greatly strengthening file security management and preventing government file information leakage compared with the extensive permission setting of general file management systems. At the same time, by using OCR technology and format conversion processing flow, the system can accurately identify and process government files of various complex formats such as PDF, Word, Excel, pictures and scanned copies, solving the format compatibility problem and ensuring the integrity and accuracy of file processing.

[0097] In terms of intelligence and scalability, the system brings a new experience to government office work. The intelligent decision-making supported by the knowledge graph enables the system to associate relevant information when classifying and searching files, such as searching for "environmental protection policies in a certain region", which can be associated with law enforcement cases and enterprise environmental protection measures, providing comprehensive knowledge services. Real-time data updating and self-learning mechanism enable the system to keep pace with the knowledge update pace in the government field. In addition, the extensible design of the micro-service architecture facilitates flexible expansion of functions according to business needs, and the rich interfaces reserved for integration with other government systems break the isolation of existing systems, improve the overall collaborative efficiency and intelligent level of government office work, and provide staff with the convenience of working anytime and anywhere, and optimize the user experience. BRIEF DESCRIPTION OF DRAWINGS

[0098] Figure 1 A classification flowchart of a government file intelligent classification and retrieval system;

[0099] Figure 2 A retrieval flowchart of a government file intelligent classification and retrieval system. DETAILED DESCRIPTION

[0100] The advantages and characteristics of the present application will become more apparent with the description. However, these examples are only exemplary and do not constitute any limitation on the scope of the present application. Those skilled in the art should understand that the details and forms of the technical solutions of the present application can be modified or replaced without departing from the spirit and scope of the present application, and such modifications and replacements fall within the protection scope of the present application.

[0101] The present application aims to build a government file intelligent classification and retrieval system to solve the problems of the prior art in government file management. The system is deployed in the government office local area network, and the server uses high-performance multi-core processors, large-capacity memories and high-speed storage devices to meet the storage and rapid processing needs of a large number of government files. The client includes desktop computers and notebook computers used by staff, which interact with the server through the local area network.

[0102] Specifically, the government document intelligent classification and retrieval system includes a preprocessing module, a feature extraction module, a knowledge graph module, a retrieval module, an urgency assessment module, and a permission management module.

[0103] The preprocessing module collects and organizes data: Collect government documents produced by various departments of a city government in the past five years, including policies and regulations, administrative approval documents, meeting minutes, etc., totaling about 500,000. File formats include PDF, Word, Excel, images, and scanned copies. These files are preliminarily sorted, duplicates and invalid files are removed, and they are stored by department, year, etc.

[0104] The feature extraction module uses the sorted file data to train an intelligent classification model based on a convolutional neural network model (CNN) and a recurrent neural network model (RNN). In the CNN part, multiple convolution layers are set, with convolution kernel sizes of 3x3 and 5x5, to extract local features such as policy terms and key clauses from the file. The RNN part uses LSTM units to capture long-distance dependencies in text and understand the logical order between policy clauses. After multiple rounds of training, the model's classification accuracy reaches 96%.

[0105] In traditional convolutional neural network models (CNN), the size of the convolution kernel is usually fixed, making it difficult for the model to adapt flexibly to the feature extraction needs of different types of government documents. The invention innovatively designs a dynamic convolution kernel selection method. During the file preprocessing stage, the system automatically analyzes the file's word count, structural complexity, and other features. For policy and regulation files with a large number of words and complex structures, the system will dynamically select larger convolution kernels, such as 7x7 or 9x9, which can more comprehensively capture long sequence information and complex semantic structures in the file. For short content and simple structure notification files, the system will use smaller convolution kernels, such as 2x2 or 3x3, to accurately locate key information and reduce unnecessary computational load. Through this dynamic adjustment mechanism, the model can more efficiently extract features of different files, significantly improving the accuracy and efficiency of classification.

[0106] Specifically, the dynamic convolution kernel selection method (Dynamic Kernel Selection Algorithm, DKSA) solves the semantic granularity difference problem in government document classification through a multi-scale feature fusion mechanism. Traditional CNN models use fixed-size convolution kernels (such as 5x5), which are difficult to capture both the long passage dependencies of policies and regulations and the local keywords of notifications and announcements, resulting in a significant drop in classification accuracy in complex file scenarios. The core innovation of the improved dynamic convolution kernel selection method is to introduce a file structure complexity quantification indicator and an adaptive kernel size mapping function. The specific implementation steps are as follows:

[0107] 1. Structure complexity quantification model: Extract the number of words L, the number of sections P, the title level H and the nested list depth D in the preprocessing stage, and construct a multi-dimensional feature vector X = [L, P, H, D]. After dimension reduction by principal component analysis (PCA), calculate the complexity score C as follows:

[0108] C = w T · X norm + b

[0109] where w = [0.35, 0.25, 0.20, 0.20] T is the weight vector, X norm is the normalized feature of the number of words L, the number of sections P, the title level H and the nested list depth D (such as ), and b = 0.1 is the bias term. According to the value of C, the file is divided into three categories: high complexity (C ≥ 1.2), medium complexity (0.8 ≤ C < 1.2), and low complexity (C < 0.8).

[0110] 2. Dynamic convolution kernel mapping and multi-scale feature fusion: DKSA dynamically selects the convolution kernel size and fuses multi-scale features for different complexity files. In this embodiment, the file is divided into three categories, and the processing process is as follows:

[0111] High complexity file: Use 7x7 convolution kernel to extract long-range semantic features, and superimpose dilated convolution (DilatedConvolution) to expand the receptive field, with dilated rate r = 2, output feature map F large .

[0112] Medium complexity file: Use 5x5 convolution kernel, combined with gating mechanism (Gating Mechanism) to filter noise features, gating weight g i,j = σ(W g · F i,j + b g ), where σ is the Sigmoid function, and the output F mid = g raw ⊙ F .

[0113] Low complexity file: Use 3x3 convolution kernel, and insert attention module (CBAM) to strengthen key areas, the spatial attention weight calculation method is:

[0114] A spatial = Sigmoid(W s · AvgPool(F small ) + W c · MaxPool(F small ))

[0115] Finally, the multi-scale features are fused by weighted concatenation, and the concatenation method is:

[0116] F fusion = a F large + b F mid + g F small

[0117] Where the weight coefficients a, b, g are dynamically calculated by the complexity score C, and the calculation formula is:

[0118]

[0119]

[0120] 3. Model training and deployment, the training stage adopts the government field annotation dataset (10 categories, 500,000 files), the government field annotation dataset includes 10 categories, 500,000 files, specifically, the government field annotation dataset contains ten core annotation categories, covering official document type classification, policy field division, agency entity recognition, policy term extraction, business process step annotation, livelihood appeal classification, legal regulation clause positioning, time node annotation, responsibility subject association and emergency degree classification.

[0121] The loss function is cross-entropy with a class balance factor, and the definition formula is:

[0122]

[0123] Where, Relieve the class imbalance problem (N c The number of samples of class c). The optimizer uses AdamW, and the learning rate is η = 0.01.

[0124] The retrieval module uses a word vector model (such as Word2Vec) to convert the text in the file into a vector representation and build a file vector library. At the same time, the attention mechanism model is trained to accurately focus on key information during the retrieval process. Through the learning of a large number of user retrieval records, the accuracy of the retrieval model reaches 92%.

[0125] Specifically, the retrieval model is personalized and optimized: traditional retrieval models usually adopt a unified retrieval strategy, which cannot meet the personalized needs of different users. The present application constructs a personalized retrieval model based on user roles and historical retrieval behaviors. The system will conduct in-depth analysis on the user's historical retrieval records, extract the user's retrieval preferences, such as the fields of interest, commonly used search terms, etc., and convert these preferences into a user preference vector. When the user performs retrieval, the system will not only convert the search terms into vectors, but also combine the user preference vector with them to jointly calculate the similarity with the file vector. For example, for a user responsible for environmental protection, the system will preferentially display files related to environmental protection in the search results, and according to the user's historical preferences, the related files will be more accurately sorted. This personalized retrieval method can better meet the diverse needs of different users, significantly improving the retrieval experience and efficiency.

[0126] The knowledge graph module constructs a knowledge graph: entities and relationships are extracted from collected government files and government open data to construct a government knowledge graph. For example, policy names, issuing departments, applicable scope, and other entities are extracted from policy and regulation files, as well as publishing, applicable, and other relationships. The knowledge graph covers about 100,000 entities and 500,000 relationships.

[0127] Fusion of multi-modal information: existing technologies often rely only on text information when classifying files, ignoring other important information dimensions in the files. The present application proposes a method of fusing multi-modal information, in addition to processing the file text, it also makes full use of the file metadata and image information (if the file contains images). For the metadata of the file, such as the issuing department, the date of issuance, etc., the system will process it and convert it into a vector form that can be understood by the model. For the image information in the file, the system first uses image processing techniques such as edge detection, feature point extraction, etc. to convert the image into a feature vector. Then, these metadata vectors, image feature vectors, and text feature vectors are fused, so that the model can integrate multiple information when making classification decisions, so as to more accurately judge the file category, especially suitable for processing complex government files containing multiple information forms.

[0128] Semantic expansion combined with knowledge graph: Existing retrieval techniques have limitations in semantic understanding, making it difficult to fully exploit the potential semantics of search terms. The invention uses a knowledge graph in the government affairs field to expand the semantics of search terms. When the user inputs a search term such as "new energy subsidy policy", the system will automatically associate related concepts such as "new energy vehicle types" and "subsidy standard formulation basis" based on the knowledge graph, and use these related concepts as expanded search terms. At the same time, a new semantic fusion algorithm is introduced, which can consider the similarity of the original search term, the expanded search term and the file vector, as well as their association strength in the knowledge graph, to reorder the search results. In this way, the search results are more comprehensive and accurate, providing more valuable information to users.

[0129] Dynamic entity and relationship update: Due to the rapid update of knowledge in the government affairs field, traditional knowledge graph update methods often cannot keep up with changes. The invention establishes a dynamic entity and relationship update mechanism. The system monitors government data sources such as government official websites and policy publishing platforms in real time, and starts the entity and relationship extraction program as soon as new policies, regulations, department adjustments, etc. are found. Using natural language processing technology and information extraction algorithms, new entities and relationships are extracted from newly published information, and the structure and content of the knowledge graph are updated in a timely manner. For example, when a new environmental protection policy is released, the system can quickly identify new entities such as new environmental protection supervision measures, and new relationships such as the responsibility relationship between the policy and related enterprises, ensuring that the knowledge graph always remains up-to-date and provides accurate knowledge support for the system's classification and retrieval.

[0130] Cross-domain knowledge fusion: To broaden the application scope of the government document management system, the invention proposes a cross-domain knowledge fusion method. In addition to integrating knowledge within the government affairs field, it also fuses government knowledge with economic, social and other fields. For example, in the environmental protection policy field, the system not only focuses on the content of the policy itself, but also correlates economic data such as the market size of the environmental protection industry, the investment of enterprises in environmental protection projects, and social livelihood impacts such as the degree of improvement of the living environment of residents and the impact on employment. By building a cross-domain knowledge fusion model, the system can mine the deep relationships between knowledge in different fields, providing more comprehensive and in-depth knowledge support for government decision-making, meeting the complex decision-making needs of government work.

[0131] The urgency assessment module is used to determine the timeliness of government documents: In the design and implementation of the fuzzy logic model, through the fine processing of multi-dimensional input variables, the innovative application of mixed membership functions, the construction of a comprehensive fuzzy rule base, and the introduction of a dynamic rule weight adjustment mechanism, the accuracy and adaptability of the urgency assessment of government documents are significantly improved.

[0132] The fuzzy processing of the multi-dimensional input variable and the design of the mixed membership function include: for three input variables of time sensitivity (T s ), influence range (S i ) and associated event level (E l ), a mixed membership function is designed;

[0133] The mixed membership function specifically includes that the right half triangular function is used to define the membership degree of “high” emergency degree for time sensitivity:

[0134]

[0135] This function linearly increases when T s ≥ 2, and the vertex is located at T s = 3 (the maximum time sensitivity), which intuitively reflects the actual scene that “the closer the deadline, the higher the emergency degree”.

[0136] The mixed membership function specifically includes that the trapezoidal function is used to define the membership degree of “regional” range for the influence range:

[0137]

[0138] The function keeps the membership degree as 1 in the interval of S l ∈ [0.5, 2.5], linearly decays on both sides, and is suitable for describing the smooth process of the transition of the influence range from the department to the city policy.

[0139] The mixed membership function specifically includes that the Gaussian function is used to define the membership degree of “important” event for the associated time level:

[0140]

[0141] The symmetry and smoothness of the Gaussian function can effectively handle the uncertainty of the associated event level (such as the subjectivity of the assessment of the event level), and the setting of the standard deviation σ = 0.5 makes the membership degree higher in the interval of E l = 2 ± 0.5, which conforms to the fuzzy boundary characteristics of “important” events.

[0142] For parameter setting, the purposes of the settings of the present embodiment include but are not limited to: the vertex position (such as T s = 2) is determined based on the statistical analysis of the historical government file processing data, ensuring that more than 80% of the time windows of the extremely urgent files are covered; the standard deviation (such as σ = 0.5) is optimized through cross-validation, balancing the sensitivity and noise resistance of the model; the function type selection (triangle, trapezoidal, Gaussian) is selected according to the variable characteristics: the time sensitivity needs to be clear threshold, the influence range needs to be widely covered, and the associated event needs to be smoothly transitioned.

[0143] Further, the traditional permission management system only relies on simple rules and cannot cope with complex scenarios with multiple factors coupled. The patent constructs a fuzzy rule base containing 27 rules, covering all combinations of input variables, which is implemented as follows:

[0144] The comprehensive fuzzy rule base aims to build a comprehensive and detailed rule system by covering all possible input variable combinations to accurately map the urgency determination logic in complex business scenarios. The rule base generation process starts with a deep analysis of input variables and fuzzy set definition. Time sensitivity (T s ), impact range (S i ) and associated event level (E l ) are divided into three fuzzy sets respectively: time sensitivity is divided into "low", "medium" and "high", impact range is divided into "local", "regional" and "global", and associated event level is divided into "routine", "important" and "urgent". Each fuzzy set is mathematically expressed by a specific type of membership function (triangle, trapezoidal and Gaussian function) to reflect the differences in different variable characteristics.

[0145] Based on the definition of fuzzy sets, the construction of the rule base follows the "IF-THEN" logic structure, covering all possible combinations of input variables, a total of 3x3x3=27 rules. Each rule maps the fuzzy set combination of input variables to the fuzzy set of output variables (urgency E); where the urgency E includes three types: urgent, urgent and normal, and the specific rules corresponding to the urgency are shown in Table 1.

[0146] Table 1: Urgency determination logic

[0147]

[0148]

[0149] The dynamic rule weight adjustment mechanism, the patent proposes a feedback-driven rule weight optimization mechanism to solve the problem that the traditional static rule base is difficult to adapt to changes in business scenarios. The initial weight of each rule is set to 1, indicating that all rules contribute equally to the output in the initial state. During system operation, by collecting real-time feedback from users on the results of permission allocation (such as a satisfaction score of 1-5), combined with the model's predicted urgency matching degree r predicted (calculated from the defuzzification score E and the actual labeled urgency level), the rule weight is dynamically adjusted.

[0150]

[0151] Let represent the weight of the i-th rule at time t. It reflects the degree of influence of the rule on the final output at the current time. Initially, the weights of all rules are usually set to 1, and are continuously adjusted based on feedback as the system runs. Let represent the weight of the i-th rule at time t+1. η is the new weight value of the i-th rule after one weight adjustment, and η represents the learning rate, used to control the magnitude of the rule weight adjustment. It is a preset constant, with a reference value of 0.01. actual This represents the degree of matching between the actual urgency level and the urgency level predicted by the model. It is a value determined based on the match between the actual urgency level and the urgency level predicted by the model. It is used to measure the degree of conformity between the model prediction and the actual situation, and thus guides the adjustment of rule weights.

[0152] The learning rate η = 0.01 controls the adjustment range to prevent drastic fluctuations in weights. For example, if a rule frequently leads to excessively high permission allocation (low user ratings) in the environmental protection bureau scenario, its weight will gradually decrease, thereby reducing its impact on the final output. Conversely, in rules for handling emergencies, if user feedback indicates that increased permissions help with rapid response, the weight will increase accordingly. This mechanism significantly improves the system's scenario adaptability by adaptively learning the business characteristics of different departments (e.g., the environmental protection bureau focuses more on the level of related events, while the finance bureau emphasizes time sensitivity).

[0153] Furthermore, the dynamic weight adjustment also works in conjunction with the risk assessment model. When the real-time risk value RR (calculated from access frequency, IP anomaly index, etc.) is high, the system will temporarily increase the weight of security-related rules, such as "IFE". l =The rule of "urgent THEN deweighting" increases the weight to prioritize blocking potential risks. This synergy is achieved through a weight scaling factor λ. R Realize, λ R The calculation formula is:

[0154]

[0155] R represents the real-time risk value, R crit Indicates the risk threshold;

[0156] Where R crit =1.5 is the risk threshold. For example, when R = 1.8, λ R =2.2, the weight of the relevant security rules is amplified to 2.2 times the original value, thereby strengthening access control in high-risk scenarios. Meanwhile, to prevent long-term weight shifts from causing the model to deviate from its initial design, the system has a weight decay mechanism, with the formula:

[0157]

[0158] Where the attenuation coefficient γ = 0.001, ensure the weight in no continuous feedback gradually back to the initial value, maintain the stability of the system.

[0159] Through the design of rule base and the coordination optimization of dynamic weight adjustment, the fuzzy logic model of the patent has achieved multiple technical breakthroughs in the assessment of the emergency level of government files. Compared with traditional methods, the combination of 27 fully covered rules and dynamic weight mechanism expands the granularity of authority allocation from 3 levels to 6 levels, significantly improving the classification accuracy. At the same time, the learning rate η and the decay coefficient γ of weight adjustment are optimized through grid search and cross-validation, ensuring a balance between rapid adaptability and long-term stability. For example, in the simulation test, after receiving 100 user feedbacks, the weight adjustment improves the efficiency of emergency handling in the environmental protection bureau scenario by 50%, and the time-sensitive file processing delay in the finance department is shortened to within 1 second. This technical advantage not only reflects the theoretical innovation of the mathematical model, but also proves its potential for wide application in the intelligent management of government affairs through actual business scenarios.

[0160] Dynamic permission adjustment based on risk assessment: According to the organizational structure and file classification of government departments, set up a permission management system. Divide users into roles such as ordinary clerks, department leaders, and department heads, and assign them corresponding file access permissions. For example, ordinary clerks can only access public and some internal files, department leaders can access files below the confidential level, and department heads have the highest permissions.

[0161] Traditional permission management usually adopts a static permission allocation method, which cannot respond to potential risks in the file access process in a timely manner. This invention introduces a dynamic permission adjustment mechanism based on risk assessment. The system will monitor the access frequency of files, user access behavior patterns, and other information in real time, and use risk assessment algorithms to quantitatively evaluate file access risks. When abnormal access behavior is detected, such as frequent access to a large number of high-class files in a short period of time, access from abnormal IP addresses, etc., the system will immediately start the dynamic permission adjustment program to temporarily restrict or freeze the access permissions of related users until the risk is resolved. This dynamic permission management method can more effectively protect the security of government files and prevent data leakage and illegal access.

[0162] Encryption and permission collaborative management: To further enhance the security of government files, the invention collaboratively designs permission management and file encryption. The system uses advanced encryption algorithms to encrypt and store government files. Different encryption keys of different strength are used for files of different secret levels. In terms of permission management, only users with corresponding permissions can obtain specific decryption keys. For example, ordinary clerks can only obtain decryption keys for public files and part of internal files, and the validity period and number of uses of the keys are also limited according to the file secret level and access risk. Through this encryption and permission collaborative management method, even if the file data is illegally obtained, the file content cannot be viewed without the correct decryption key, greatly improving the security of the file.

[0163] III. System operation and operation

[0164] As shown in Figure 1 , the file classification process is as follows:

[0165] File upload: staff upload newly generated government files to the system through the client. For example, a staff member of the Municipal Environmental Protection Bureau writes a policy document on the pollution control of a certain river and uploads it in Word format.

[0166] Preprocessing: The system automatically preprocesses the uploaded file, including format conversion (converting Word files to a unified text format), noise removal (removing irrelevant information such as headers, footers, and advertisements from the file).

[0167] Feature extraction: CNN is used to extract features from the preprocessed file. The convolution kernel slides over the file text, extracting local features such as "river pollution control" and "environmental protection measures". These features are passed to the RNN.

[0168] Classification decision: RNN determines the file category based on the features extracted by CNN and its understanding of the text sequence. In this example, the model determines that the file belongs to the "environmental policy and regulation" category and stores it in the corresponding database folder.

[0169] As shown in Figure 2 , the file retrieval process is as follows:

[0170] Retrieval term input: staff enter the retrieval term on the client, such as "related policies for the pollution control of a certain river".

[0171] Semantic understanding: The system converts the retrieval term into a vector form through the word vector model and calculates the semantic similarity with the vectors in the file vector library. At the same time, the attention mechanism assigns weights to different parts of the file according to the focus of the retrieval term, highlighting the key information related to the retrieval term.

[0172] Result filtering and sorting: The system filters out relevant files based on semantic similarity and attention weights, and sorts them from high to low relevance. For example, in addition to direct pollution control policy files, environmental monitoring reports related to the river, and relevant law enforcement cases are also associated with the files.

[0173] Result display: The search results are displayed in a list on the client side, with each file showing the file name, file summary, and relevance score. Staff can click on the file to view it.

[0174] Permission management and security: When staff try to access a file, the system verifies it according to the permission management system. For example, a regular clerk tries to access a confidential financial budget file, and the system will prompt that the permission is insufficient and refuse access. Only the department head with the corresponding permission can successfully open the file.

[0175] Format compatibility processing: When staff upload a scanned file, the system first converts the image text into editable text using OCR technology. After character segmentation, feature extraction, and classification recognition, the recognition accuracy reaches 98%. Then, the system converts it into a unified format for subsequent processing, ensuring the integrity and accuracy of the file content.

[0176] Four, system maintenance and optimization

[0177] Data update: Regularly collect new government files and incrementally train the classification and retrieval models to adapt to new policies, regulations, and business needs. For example, update the data once a month, and the model accuracy remains stable in the long run.

[0178] User feedback processing: Collect user feedback on search results, such as accuracy and whether it meets their needs. Based on feedback, adjust the parameters of the retrieval algorithm to optimize the search results. For example, users feedback that some search results are less relevant, and by analyzing feedback data, adjust the weight distribution of the attention mechanism to improve search accuracy.

[0179] System performance monitoring: Real-time monitoring of server CPU, memory, disk I / O, and other performance indicators to ensure stable operation of the system. When performance bottlenecks are found, optimize them, such as increasing server memory, optimizing database query statements, etc.

[0180] Improve the accuracy of classification and retrieval, use intelligent algorithms to accurately understand the semantics of government files, reduce classification errors caused by subjective judgment, and reduce classification error rate to a very low level, and improve the search accuracy to more than 90%.

[0181] In view of the problem of insufficient safety management of the general file management system, the application will build a fine permission management system to meet the multi-level security requirements of different secret levels of government files. When facing the rapid update of government knowledge and complex file formats, the system has strong self-learning, self-adaptive ability and excellent format compatibility, ensuring that it can adapt to new policies and regulations in a timely manner, accurately process various formats of government files, and realize the intelligent, efficient and standardized management of government files.

[0182] The above is only the specific steps of the application, and does not constitute any limitation on the protection scope of the application; any technical solution formed by equivalent transformation or equivalent replacement falls within the protection scope of the application; the parts not elaborated in the application belong to the known technology of those skilled in the art.

Claims

1. A government document intelligent classification and retrieval system, characterized in that, The government affair file intelligent classification and retrieval system comprises a feature extraction module, a knowledge graph module, a retrieval module, an urgency evaluation module and a permission management module; The feature extraction module is used for intelligent classification of government affair files, adopts a deep learning architecture combining a convolutional neural network model and a recurrent neural network model, wherein the convolutional neural network model is provided with multiple convolutional layers, the size of the convolution kernel is dynamically adjusted according to the file features, and is used for extracting local features of the file; the recurrent neural network model adopts an LSTM unit and fuses a self-attention mechanism, and is used for capturing the logical relationship of a text sequence; The knowledge graph module constructs a government affair field knowledge graph, fuses cross-field knowledge of government affairs, economy and society, and has a dynamic updating mechanism, and monitors government affair data sources in real time, and updates entities and relationships; The retrieval module converts files and retrieval words into vectors by means of a word vector model, calculates semantic similarity by a cosine similarity formula, and self-adaptively adjusts attention weights according to retrieval word heat, user historical retrieval behaviors and file timeliness, so as to realize accurate retrieval; The urgency evaluation module determines the timeliness of government affair files based on a fuzzy rule base and a dynamic rule weight adjustment mechanism, so as to reflect the urgency of government affair files; The permission management module combines a government department organization structure and a file secret level, designs an access control algorithm based on roles and secret levels, simultaneously introduces a risk evaluation mechanism, dynamically adjusts permissions according to file access risks, and cooperates permission management with file encryption, so that different permission users obtain different decryption keys.

2. The government document intelligent classification and retrieval system according to claim 1, wherein, The government affair file intelligent classification and retrieval system comprises a preprocessing module, which pre-processes uploaded government affair files, including removing repeated and invalid files, classifying and storing the government affair files according to departments and years, and converting the government affair files into a unified text format and removing headers, footers and advertising information in the files. The convolutional neural network model is based on a dynamic convolution kernel selection method, and the specific steps are as follows:

3. The government document intelligent classification and retrieval system of claim 1, wherein Extract the number of words L, the number of paragraphs P, the title level H and the nesting list depth D of the file, construct a multi-dimensional feature vector X = [L, P, H, D], calculate the complexity score, and divide the file into a high complexity file, a medium complexity file and a low complexity file based on the complexity score; Different size convolution kernels are used to extract semantic features of the high complexity file, the medium complexity file and the low complexity file, and the multi-scale features are fused by weighted splicing; The convolutional neural network adopts a government affair field labeled data set in the training stage, and the loss function adopts a cross-entropy with a class balance factor: The complexity score is expressed as: where, to alleviate the class imbalance problem, c is the class index, N c is the number of samples in class c; the optimizer is AdamW with a learning rate of η = 0.

01.

4. The government document intelligent classification and retrieval system according to claim 3, wherein, When C is greater than or equal to 1.2, it is a high complexity file, when 0.8 is less than or equal to C and less than 1.2, it is a medium complexity file, and when C is less than 0.8, it is a low complexity file. C = w T • X norm + b where w = [0.35, 0.25, 0.20, 0.20] T is a weight vector, X norm is the normalized features of the word number L, the section number P, the title level H and the nested list depth D, and b is the bias term. The feature extraction processes of the high complexity file, the medium complexity file and the low complexity file are as follows:

5. The government document intelligent classification and retrieval system according to claim 3, wherein, Finally, the multi-scale features are fused by weighted splicing: For high complexity files, a 7x7 convolution kernel is used to extract long-range semantic features, and a hollow convolution is used to expand the receptive field with a hollow rate r = 2, output high complexity feature map F large ; For medium complexity files, a 5x5 convolution kernel is used to extract semantic features, combined with a gating mechanism to filter noise features, and the gating weight g i,j is: g i,j = σ(W g · F i,j + b g ) where σ is a Sigmoid function, W g is a weight matrix in the gating mechanism, used to weight the features F i,j , F i,j is a feature vector at position (i, j) in the feature map obtained after convolution operation, containing local feature information of the file, b g is a bias term in the gating mechanism, used to adjust the overall level of the gating output; the output complexity feature map: F mid = g i,j O F raw F raw is the original feature map without the gating mechanism when extracting medium complexity file semantic features using a 5x5 convolution kernel; for low complexity files, a 3x3 convolution kernel is used to extract semantic features, and an attention module is inserted to strengthen key areas. The spatial attention weight A spatial is: A spatial = Sigmoid(W s · AvgPool(F sm ) + W c · MaxPool(F sm )) F small = A spatial O F raw sigmoid is a normalization function, W s and W c are weight matrices, AvgPool is an average pooling operation, F sm is the original feature map; F small is the low complexity feature map weighted by attention, is, MaxPool is a max pooling operation; The weight coefficients a, b and g are dynamically calculated according to the complexity score C, and the calculation formula is: F fusion = a · F large + β · F mid + γ · F small wherein F fusion is a fused feature map; k1, k2 and k3 are parameters for dynamically calculating the weight coefficients, and are constants determined in the model training or design process. ​ 6. The government document intelligent classification and retrieval system according to claim 3, wherein, The training phase adopts a government field annotation dataset, and the government field annotation dataset includes 10 categories and 500,000 files; Specifically, the government field annotation dataset contains ten core annotation categories, covering official document type classification, policy field division, agency entity recognition, policy term extraction, business process step annotation, livelihood appeal classification, legal regulation clause positioning, time node annotation, responsibility subject association, and emergency level classification.

7. The government document intelligent classification and retrieval system of claim 1, wherein The emergency level evaluation module constructs a fuzzy logic model through a hybrid membership function; The hybrid membership function includes three variables of time sensitivity, influence range, and associated event level; The right half-triangle function is used to define the membership degree of time sensitivity: μ High (T s ) is the membership degree representing the time sensitivity as "high", which is used to measure the matching degree of the emergency level of the government file in the time dimension with the "high" emergency level, and its value ranges from 0 to 1, and the larger the value, the closer the file is to the standard of "high" emergency level in time; T s is a time sensitivity variable representing the deadline or time-related attributes of the government file, which is an input variable for calculating the membership degree; This function increases linearly for T s ≥ 2, and the maximum time sensitivity is located at T s = 3, reflecting that the closer the deadline, the higher the urgency. The trapezoidal function is used to define the membership degree of influence range: μ Regional (S i ) is a membership degree indicating the influence range as "region", used to measure the matching degree of the influence range of the government document with the "region" range, the value range is between 0 and 1, and the value is 1, indicating that the influence range of the document completely conforms to the definition of "region", and the closer to 0 indicates the lower matching degree with the "region" range, S i is an influence range variable, used to describe the size of the influence range of the government document, which is an input variable for calculating the membership degree. This function is in S i The maximum value of ∈[0.5,2.5] is 1, and the linear decay on both sides is suitable for describing the smooth process of the transition from the department to the city policy. The Gaussian function is used to define the membership degree of associated event level: μ Important (E l ) is the membership degree of the event level being "important", which is used to measure the matching degree of the event level associated with the government document and the "important" level. The greater the value, the more consistent the associated event with the "important" standard, and the value range is between 0 and 1, E l is the associated event level variable, which is used to describe the importance level of the associated event of the government document, and is the input variable for calculating the membership degree.

8. The government document intelligent classification and retrieval system of claim 7, wherein, Based on time sensitivity, influence range, and associated event level, a fuzzy rule base containing 27 rules is constructed; Among them, the generation process of the rule base starts with the deep analysis of the input variables and the definition of fuzzy sets, in which time sensitivity, influence range, and associated event level are divided into three fuzzy sets according to the value range: time sensitivity is divided into "low", "medium", and "high", influence range is divided into "local", "regional", and "global", and associated event level is divided into "routine", "important", and "urgent".

9. The government document intelligent classification and retrieval system of claim 8, wherein, Based on the definition of fuzzy sets, the construction of the rule base follows the IF-THEN logic structure, covering all possible combinations of input variables: Among them, low, local, and routine correspond to the emergency level of flat; Low, local, and important correspond to the emergency level of flat; Low, local, and urgent correspond to the emergency level of urgent; Low, regional, and routine correspond to the emergency level of flat; Low, regional, and important correspond to the emergency level of urgent; Low, regional, and urgent correspond to the emergency level of urgent; Low, global, and routine correspond to the emergency level of flat; Low, global, and important correspond to the emergency level of urgent; Low, global, and urgent correspond to the emergency level of special urgent; Medium, local, and routine correspond to the emergency level of flat; Medium, local, and important correspond to the emergency level of flat; Medium, local, and urgent correspond to the emergency level of urgent; Medium, regional, and routine correspond to the emergency level of flat; Medium, regional, and important correspond to the emergency level of urgent; Medium, regional, and urgent correspond to the emergency level of special urgent; Medium, global, and routine correspond to the emergency level of urgent; Medium, global, and important correspond to the emergency level of special urgent; Medium, global, and urgent correspond to the emergency level of special urgent; High, local, and routine correspond to the emergency level of urgent; High, local, and important correspond to the emergency level of urgent; High, local, and urgent correspond to the emergency level of special urgent; High, regional, and routine correspond to the emergency level of urgent; High, regional, and important correspond to the emergency level of special urgent; High, regional, and urgent correspond to the emergency level of special urgent; High, global, and routine correspond to the emergency level of special urgent; High, global, and important correspond to the emergency level of special urgent; High, global, and urgent correspond to the emergency level of special urgent.

10. The intelligent classification and retrieval system for government documents according to claim 1, wherein, The dynamic rule weight adjustment mechanism includes: The initial weight of each rule is set to 1, indicating that all rules contribute equally to the output in the initial state; by collecting real-time feedback from users on the results of permission allocation, combined with the model-predicted emergency degree matching degree r predicted , dynamically adjust the rule weight: represents the weight of the i-th rule at time t; represents the weight of the i-th rule at time t+1; η represents a learning rate; r actual represents the actual urgency degree matching degree; The dynamic weight adjustment also works in coordination with the risk assessment model, which is coordinated through a weight scaling factor λ R The implementation, λ R The calculation formula is: R represents a real-time risk value, R crit represents a risk threshold value; To prevent long-term weight deviation from causing the model to deviate from the initial design, the system sets a weight decay mechanism: Wherein, the attenuation coefficient γ = 0.001, ensure the weight in no continuous feedback gradually regression initial value, maintain the stability of the system.

Images