An electronic archive informatization management method and system

Abstract

This invention discloses an electronic archive information management method and system, relating to the field of data processing technology. The method involves collecting raw electronic archive text data, preprocessing the raw electronic archive text data to obtain a standardized archive dataset, identifying core entities in the standardized archive dataset, constructing an archive text hypergraph network, using a hypergraph attention mechanism to aggregate features of the archive text hypergraph network to obtain a global semantic feature vector, using a cross-attention mechanism to correlate and map the gained input feature vector and the global semantic feature vector to obtain a comprehensive semantic representation of the archive text, performing archive entity association reasoning based on the comprehensive semantic representation of the archive text for the business request to be analyzed to obtain archive association evaluation data, performing risk analysis based on the archive association evaluation data for the business request to be analyzed to obtain analysis results, matching the analysis results to obtain archive management decision instructions, and generating security traceability records.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for electronic archive information management. Background Technology

[0002] In recent years, with the in-depth application of information technology in the field of archival management, the digitization of paper archives and the structured storage of electronic archives have become the foundation of the industry. Existing technologies mainly focus on converting physical archives into digital formats through scanning, OCR recognition, and other means, and statically associating them with relevant business objects (such as engineering projects and equipment assets) to achieve digital archiving and retrieval of archives.

[0003] Currently, Chinese invention patent application CN2021108460569 discloses a method and system for digitizing paper archives based on BIM for highway engineering. This method digitizes paper archives of highway engineering projects; links the digitized structured data with the corresponding BIM components of the highway engineering project; and displays the digitized paper archives based on query requests and the corresponding BIM components of the highway engineering project. However, this related technology lacks a dual correction mechanism combining character topology and pinyin sequence, making it difficult to effectively remove sparse interference signals caused by recognition errors, resulting in low accuracy in constructing standardized archive datasets. Summary of the Invention

[0004] The technical problem addressed by this invention is that existing technologies lack adaptive correction models for visual similarity and homophony deviations generated during OCR recognition. This makes it difficult to maintain semantic fidelity of data under complex background noise, leading to subsequent analysis often being based on noisy data and posing a fundamental accuracy risk. Secondly, at the semantic modeling level, most existing technologies are based on unary mappings or simple binary relations, lacking means to capture multi-dimensional and non-linear logical relationships between archival entities, making it difficult to perform structural rationality analysis from the perspective of a global hypergraph network. This results in the system being diluted by a large amount of normal background information when dealing with hidden semantic conflicts, making it difficult to discover deep-level contextual anomalies that are completely invisible at the single-archive level. Finally, at the management decision-making level, existing platforms lack a quantitative reasoning and evaluation system with logical interpretability. When faced with high-dimensional business requests, the system struggles to automatically extract confidence indicators supported by logical chains, causing the decision-making process to still heavily rely on human experience and lacking a full-process secure decision traceability mechanism, failing to guarantee the authenticity and non-repudiation of information management behaviors.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a method for information management of electronic records, comprising the following steps:

[0006] Step S1: Collect the original electronic archive text data, preprocess the original electronic archive text data to obtain a standardized archive dataset;

[0007] Step S2: Identify the core entities in the standardized archive dataset, construct an archive text hypergraph network, and use the hypergraph attention mechanism to aggregate features of the archive text hypergraph network to obtain a global semantic feature vector.

[0008] Step S3: Use the cross-attention mechanism to perform association mapping on the input feature vector after gain and the global semantic feature vector to obtain the comprehensive semantic representation of the archival text. Based on the comprehensive semantic representation of the archival text, perform archival entity association reasoning on the business request to be analyzed to obtain archival association evaluation data.

[0009] Step S4: Based on the archive association evaluation data, perform risk analysis on the business request to be analyzed, obtain the analysis results, match the analysis results to obtain archive management decision instructions, and generate security traceability records.

[0010] As a preferred embodiment of the electronic archives information management method described in this invention, step S1 specifically includes:

[0011] Step S11: Use the N-gram sliding window algorithm to segment the original electronic archive text data to obtain a candidate word set. Compare the words in the candidate word set with the preset standard archive lexicon. If a word is not in the preset standard archive lexicon and its word frequency is lower than the preset frequency threshold, then mark the word as an interference candidate.

[0012] The original electronic archive text data includes unstructured text, original evidence records from the electronic archive system, and archive metadata;

[0013] Unstructured text includes visually similar characters and homophonic misspellings.

[0014] Step S12: Extract the initials, finals and tones of the interference candidates to form a pinyin sequence. Use the edit distance algorithm to calculate the conversion cost between the pinyin sequence and the candidate original meaning word sequence in the preset standard archive lexicon, and perform normalization processing to obtain the pinyin similarity.

[0015] Step S13: Based on the stroke, radical, and structural features of the interference candidates, extract the topological feature vectors of the characters and calculate the cosine similarity to obtain the character similarity between the interference candidates and the candidate original words.

[0016] Step S14: The similarity scores of pinyin and character shapes are weighted and summed to calculate the comprehensive similarity value. The calculation formula is as follows:

[0017] ;

[0018] in, This represents the overall similarity score. This represents the similarity adjustment coefficient. Indicates the similarity of pinyin. Indicates the similarity of character shapes;

[0019] Step S15: Select the candidate original word with the highest comprehensive similarity value to perform semantic replacement on the interference candidate, and perform noise reduction and format standardization on the text to obtain a standardized archive dataset.

[0020] As a preferred embodiment of the electronic archives information management method described in this invention, step S2 specifically includes:

[0021] Step S21: Identify the core entities in the standardized archive dataset and calculate the contextual diversity scores of the core entities in the preset archive management scenario;

[0022] Step S22: Extract the initial feature vectors of the core entities using the pre-trained language model, and use the core entities as archive entity nodes, with the initial feature vectors as the representation of the archive entity nodes. Construct an archive text hypergraph network using the word order adjacency relationship and grammatical structure association between the core entities.

[0023] Step S23: Based on the archival text hypergraph network, calculate the degree centrality of core entities in the archival text hypergraph network and their frequency of occurrence in the standardized archival dataset. Combine this with the contextual diversity score under the preset archival management scenario to calculate the comprehensive importance weight and relational feature sparseness weight of the core entities. The calculation formula is as follows:

[0024] ;

[0025] ;

[0026] in, This represents the overall importance weight of the core entities. This represents the first preset weighting coefficient. This represents the second preset weighting coefficient. This represents the third preset weighting coefficient. This indicates the degree centrality of the core entity in the archival text hypergraph network. Indicates the frequency of occurrence in the standardized archival dataset. This indicates the contextual diversity score within a predefined file management scenario. This represents the maximum degree centrality of the core entity in the document text hypergraph network. This represents the maximum frequency of occurrence in the standardized archival dataset. This represents the maximum value of the contextual diversity score within the preset file management scenario. The weights represent the rarity of relation features. This represents the total number of logical association paths between all core entities in the standardized archive dataset. Indicates the first The frequency of occurrence of specific relational features in standardized archival datasets. This represents the average degree centrality of core entities in the archival text hypergraph network;

[0027] Step S24: Apply the comprehensive importance weight as a gain coefficient to the initial feature vector of the core entity to obtain the gained input feature vector. The calculation formula is as follows:

[0028] ;

[0029] in, This represents the input feature vector after gain. Represents the initial eigenvector;

[0030] Step S25: Use the hypergraph attention mechanism to aggregate features of the hypergraph network of the archive text, update the node feature vectors corresponding to each archive entity node, and obtain the global semantic feature vector through pooling operation.

[0031] In a preferred embodiment of the electronic archive information management method described in this invention, in step S25, a hypergraph attention mechanism is used to aggregate features of the archive text hypergraph network, update the node feature vectors corresponding to each archive entity node, and obtain the global semantic feature vector through pooling operations. The processing logic includes:

[0032] Step S251: Extract the node feature vectors of each archive entity node in the archive text hypergraph network, and use a preset shared linear transformation matrix to perform spatial mapping on the node feature vectors to obtain the projected feature vectors.

[0033] Step S252: Calculate the attention coefficients between the central node and each neighboring node. The calculation formula is as follows:

[0034] ;

[0035] in, Indicates the attention coefficient. This represents the activation function. This represents a learnable attention vector. This represents the predefined shared linear transformation matrix. Represents the projected feature vector of the center node. This represents the projected feature vector of the neighboring nodes. This represents the vector concatenation operator;

[0036] Step S253: The attention coefficients are normalized using the Softmax function to obtain the node contribution. The archive entity node representations of each neighboring node are then weighted and summed to obtain the aggregate representation of the current central node. The calculation formula is as follows:

[0037] ;

[0038] ;

[0039] in, Indicates the node's contribution. Indicates the central node The set of adjacent domain nodes, Indicates the central node, Indicates the central node Any neighboring node in the set of adjacent domain nodes, Indicates the central node With any neighbor node during the traversal Attention coefficient Represents an exponential function. This represents the aggregate representation of the current central node. This represents the activation function. Indicates neighboring nodes;

[0040] Step S254: Perform pooling operation on the aggregated representations of all current central nodes to obtain the global semantic feature vector.

[0041] As a preferred embodiment of the electronic records information management method of the present invention, step S3 specifically includes:

[0042] Step S31: Use the cross-attention mechanism to perform a correlation mapping between the gained input feature vector and the global semantic feature vector to obtain a comprehensive semantic representation of the archival text.

[0043] Step S32: Based on the comprehensive semantic representation of the archive text, adjust the feature weights and relation feature sparse weights corresponding to each archive entity node in the archive text hypergraph network, and use the path search algorithm to traverse and retrieve the logical association paths connecting different archive entities in the adjusted archive text hypergraph network.

[0044] Step S33: Establish a semantic evaluation model, extract the number of relation features, relation strength, and total path step length in the logical association path, and calculate the inference confidence index. The calculation formula is as follows:

[0045] ;

[0046] in, This represents the confidence index of inference. Indicates the number of relational features. Indicates the first The correlation strength coefficient of the relationship. The total step length of a logically related path. The weights represent the rarity of relation features. A number indicating the relationship;

[0047] Step S34: Obtain the business request to be analyzed and perform vectorization encoding to obtain the request feature vector. Calculate the vector cosine similarity between the representation of each archive entity node and the request feature vector to obtain the semantic similarity score. Then, use a nonlinear fusion operator to process the semantic similarity score and the inference confidence index to obtain the archive association evaluation data.

[0048] As a preferred embodiment of the electronic archive information management method of the present invention, in step S31, a cross-attention mechanism is used to perform correlation mapping between the gained input feature vector and the global semantic feature vector to obtain a comprehensive semantic representation of the archive text. The processing logic includes:

[0049] Step S311: Extract the input feature vector after gain, and map the input feature vector after gain to a query matrix through a linear transformation matrix;

[0050] Extract the global semantic feature vector, and map the global semantic feature vector into a key matrix and a value matrix through the corresponding linear transformation matrix;

[0051] Step S312: Calculate the energy distribution matrix of the query matrix and the key matrix, and scale the energy distribution matrix using the square root of the feature dimension to obtain the scaled point integral distribution map.

[0052] Step S313: The Softmax function is used to normalize the integral distribution of the scaling points to obtain the attention weight matrix.

[0053] Step S314: The value matrix is ​​weighted and mapped using the attention weight matrix to obtain the mapped feature matrix. The mapped feature matrix is ​​then merged with the query matrix using the residual connection operator to obtain the comprehensive semantic representation of the archive text.

[0054] As a preferred embodiment of the electronic records information management method of the present invention, step S314 specifically includes:

[0055] The mapping feature matrix is ​​obtained by multiplying the weight vector of each row in the attention weight matrix with the value matrix. The calculation formula is as follows:

[0056] ;

[0057] in, Represents the mapping feature matrix, Represents the attention weight matrix. Represents a value matrix;

[0058] The mapping feature matrix and the query matrix are added element-wise using the residual addition operator to obtain the fused feature term. The fused feature term is then processed using the layer normalization operator to obtain the comprehensive semantic representation of the archival text.

[0059] As a preferred embodiment of the electronic records information management method of the present invention, step S4 specifically includes:

[0060] Step S41: Invoke the inference confidence index and semantic similarity score corresponding to the business request to be analyzed, and perform weighted fusion processing using a preset nonlinear mapping operator to obtain the comprehensive management risk index. The calculation formula is as follows:

[0061] ;

[0062] in, Indicates the comprehensive management risk index, This represents the activation function. This represents the first preset decision weight coefficient. This represents the second preset decision weight coefficient. Indicates the semantic similarity score. This represents the confidence index of inference. Indicates the bias term;

[0063] Step S42: Extract the comprehensive management risk index and match it with the preset risk discrimination matrix to obtain the analysis results;

[0064] The analysis results include high-risk conflicts, verification concerns, manual risk review, and compliance approval;

[0065] Step S43: Extract the analysis results and match them with the preset decision instruction set to obtain the corresponding file management decision instructions;

[0066] The processing logic for verification and concern determination is as follows:

[0067] When the comprehensive management risk index is within the preset verification range and the reasoning confidence index is lower than the preset confidence threshold, it will receive verification attention and trigger a manual secondary review instruction.

[0068] Step S44: Extract the behavioral metadata when executing the file management decision instruction, combine it with the unique identifier of the analysis result, generate a security traceability record, and encapsulate the security traceability record into an encrypted data block and mount it to the distributed ledger node;

[0069] The security traceability record includes the access subject, risk characteristics, and decision-making basis.

[0070] As a preferred embodiment of the electronic archives information management method of the present invention, in step S42, a comprehensive management risk index is extracted and matched with a preset risk discrimination matrix to obtain the analysis result. The processing logic includes:

[0071] Retrieve the first and second risk thresholds from the preset risk discrimination matrix;

[0072] Among them, the first risk threshold is greater than the second risk threshold;

[0073] The comprehensive management risk index is compared with the first risk threshold. When the comprehensive management risk index is greater than or equal to the first risk threshold, the analysis result is output as a high-risk conflict.

[0074] When the comprehensive management risk index is less than the first risk threshold but greater than or equal to the second risk threshold, the inference confidence index is extracted and compared with a preset confidence threshold, specifically including:

[0075] If the reasoning confidence index is less than the preset confidence threshold, the output analysis result will be "verification attention" and a manual secondary review instruction will be triggered.

[0076] If the reasoning confidence index is greater than or equal to the preset confidence threshold, the output analysis result will be manually reviewed for risk.

[0077] When the comprehensive management risk index is less than the second risk threshold, the output analysis result is "compliant and passed".

[0078] Secondly, an electronic archive information management system includes a text processing module, a semantic network construction module, a correlation analysis module, and a decision tracing module;

[0079] The text processing module is used to collect raw electronic archive text data, preprocess the raw electronic archive text data, and obtain a standardized archive dataset.

[0080] The semantic network construction module is used to identify core entities in the standardized archive dataset, construct an archive text hypergraph network and calculate the comprehensive importance weight and relation feature sparse weight of the core entities, and use the hypergraph attention mechanism to perform feature aggregation on the archive text hypergraph network to obtain a global semantic feature vector.

[0081] The association analysis module is used to perform association mapping on the input feature vector and the global semantic feature vector after gain using the cross attention mechanism to obtain the comprehensive semantic representation of the archival text. Based on the comprehensive semantic representation of the archival text, the module performs archival entity association reasoning on the business request to be analyzed to obtain archival association evaluation data.

[0082] The decision tracing module is used to perform risk analysis on the business request to be analyzed based on the archive association evaluation data, obtain the analysis results, match the analysis results to obtain the archive management decision instructions, and generate a security tracing record.

[0083] The beneficial effects of this invention are as follows: This invention innovatively combines the topological features of candidate words with pinyin similarity for multi-dimensional verification, effectively filtering out visual similarity and homophonic interference caused by OCR recognition deviations during the digitization of archives. This fundamentally solves the semantic fidelity problem of archive datasets, ensuring the accuracy of the underlying data for management decisions. By constructing an archive text hypergraph network and introducing pooling and cross-attention mapping of global semantic feature vectors, deep integration of local entity features and global macro-context is achieved, effectively solving the semantic silo problem in traditional methods and significantly improving the system's accuracy in recognizing hidden and cross-dimensional semantic conflicts between archives. By introducing a reasoning confidence quantification index composed of relation strength, rarity weight, and path step size, this invention constructs a risk decision-making system with logical interpretability and simultaneously generates security traceability records. This elevates electronic archive management from traditional "intuitive experience matching" to "rational logical reasoning," significantly enhancing the authenticity and non-repudiation of management decisions and ensuring the continuous, stable, and efficient operation of electronic archive information management under high-risk, high-frequency business requests. Attached Figure Description

[0084] Figure 1 A flowchart illustrating the steps of an electronic records information management method according to an embodiment of the present invention;

[0085] Figure 2 This is a flowchart of the preprocessing steps for this invention. Detailed Implementation

[0086] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0087] Example, refer to Figures 1-2 This paper provides a method for information management of electronic records, which includes the following steps:

[0088] Step S1: Collect the original electronic archive text data, preprocess the original electronic archive text data to obtain a standardized archive dataset;

[0089] Step S2: Identify the core entities in the standardized archive dataset, construct an archive text hypergraph network, and use the hypergraph attention mechanism to aggregate features of the archive text hypergraph network to obtain a global semantic feature vector.

[0090] Step S3: Use the cross-attention mechanism to perform association mapping on the input feature vector after gain and the global semantic feature vector to obtain the comprehensive semantic representation of the archival text. Based on the comprehensive semantic representation of the archival text, perform archival entity association reasoning on the business request to be analyzed to obtain archival association evaluation data.

[0091] Step S4: Based on the archive association evaluation data, perform risk analysis on the business request to be analyzed, obtain the analysis results, match the analysis results to obtain archive management decision instructions, and generate security traceability records.

[0092] In specific implementation, step S1 includes:

[0093] Step S11: Use the N-gram sliding window algorithm to segment the original electronic archive text data to obtain a candidate word set. Compare the words in the candidate word set with the preset standard archive lexicon. If a word is not in the preset standard archive lexicon and its word frequency is lower than the preset frequency threshold, then mark the word as an interference candidate.

[0094] The original electronic archival text data includes unstructured text, original evidence records from the electronic archival system, and archival metadata;

[0095] Unstructured text includes visually similar characters and homophonic misspellings.

[0096] Step S12: Extract the initials, finals and tones of the interference candidates to form a pinyin sequence. Use the edit distance algorithm to calculate the conversion cost between the pinyin sequence and the candidate original meaning word sequence in the preset standard archive lexicon, and perform normalization processing to obtain the pinyin similarity.

[0097] Step S13: Based on the stroke, radical, and structural features of the interference candidates, extract the topological feature vectors of the characters and calculate the cosine similarity to obtain the character similarity between the interference candidates and the candidate original words.

[0098] Step S14: The similarity scores of pinyin and character shapes are weighted and summed to calculate the comprehensive similarity value. The calculation formula is as follows:

[0099] ;

[0100] in, This represents the overall similarity score. This represents the similarity adjustment coefficient. Indicates the similarity of pinyin. Indicates the similarity of character shapes;

[0101] Step S15: Select the candidate original word with the highest comprehensive similarity value to perform semantic replacement on the interference candidate, and perform noise reduction and format standardization on the text to obtain a standardized archive dataset.

[0102] Specifically, the N-gram sliding window algorithm is used to segment the original electronic archive text data. The process involves sliding a window of length N across the text sequence character by character. Each slide extracts N consecutive characters within the window, forming a candidate word. For example, for the text "electronic archives," Bi-gram will yield three candidate words: [electronic, sub-file, archive]. Compared to traditional dictionary-based segmentation, this method captures all possible character combinations without omission, making it particularly effective in identifying new "pseudo-words" caused by spelling errors, thus laying the foundation for subsequent noise recognition.

[0103] A pre-built standard archival thesaurus is an authoritative, independently constructed database containing all standard terms, normative names, and codes in the field of archival management. Building such a thesaurus typically requires integrating industry standards, historical compliance archives, and expert knowledge, and is continuously maintained and updated. Its fundamental purpose is to provide a single, correct target reference for the entire corrective process. Each candidate term is compared to this thesaurus; if a term is not in the thesaurus, it is initially marked as suspicious.

[0104] However, not all words not in the dictionary are errors; some may be reasonable low-frequency new words. Therefore, word frequency is introduced as a statistic for filtering. Word frequency refers to the number of times a candidate word appears in the entire original text dataset. The system sets a preset frequency threshold. This threshold is usually not a fixed value but is dynamically determined by statistical analysis of a low quantile of the word frequencies of all unregistered words. The advantage of this is that it achieves adaptive filtering: words that appear only once or twice and are very likely to be recognition noise, such as errors of similar-looking characters, are judged as interference candidates and enter the subsequent complex correction process; while words that appear more frequently and may be real but not included proper nouns are temporarily ignored, avoiding over-correction and balancing processing accuracy and computational cost.

[0105] For words identified as interference candidates, a dual correction mechanism is activated. The first step is speech correction. The system extracts the pinyin sequence of the interference, including the initial consonant, final vowel, and tone. Then, it calculates the edit distance between the pinyin sequence and the pinyin sequences of each candidate original meaning word in the dictionary. The edit distance algorithm quantifies the difference between the two sequences by calculating the minimum number of single-character edit operations required to transform from one sequence to another; this number of operations is the transformation cost. The lower the cost, the more similar the pinyin is. Edit operations include insertion, deletion, and replacement.

[0106] Secondly, character shape correction requires extracting topological feature vectors. This is achieved by analyzing the stroke types, stroke order, radical composition, and spatial structure of Chinese characters. Stroke types include horizontal, vertical, left-falling, and right-falling strokes, while spatial structures include left-right, top-bottom, and enclosing. These discrete features are digitally encoded using embedding representation and combined into a high-dimensional vector, which represents the visual morphology of the character. Then, the cosine similarity between the interference terms and the character shape vectors of the candidate original meaning words is calculated to measure visual similarity.

[0107] After obtaining the pinyin similarity and character shape similarity separately, the two are weighted and summed to obtain a comprehensive similarity value. The similarity adjustment coefficient can be empirically adjusted according to the prevalence of speech errors and visual errors in specific scenarios. The candidate original word with the highest comprehensive similarity value is selected as the final correction result. Among all possible standard words, the correct option that is most similar to the current interference item in terms of both "sound" and "shape" is semantically replaced. This selected standard word replaces the interference item in the original text, thereby completing denoising and standardization, and obtaining a high-quality "standardized archival dataset".

[0108] Compared to traditional archival processing techniques based on dictionary-based precise matching or simple statistical cleaning, the system constructed in this invention has a significant advantage in achieving deep repair and non-destructive restoration of underlying recognition biases in archives. Traditional techniques, lacking an understanding of text generation mechanisms, often fail to distinguish between recognition biases and meaningless noise, easily leading to the loss of crucial information. This solution, by establishing a dual-path detection mechanism of sound and form, can simulate and capture visual similarity interference and homophonic miswriting patterns caused by harsh recognition environments. While filtering random noise, it accurately recovers the semantic truth of the original archives, fundamentally solving the pain points of low data fidelity and critical semantic gaps in traditional methods when processing unstructured archives.

[0109] In specific implementation, step S2 includes:

[0110] Step S21: Identify the core entities in the standardized archive dataset and calculate the contextual diversity scores of the core entities in the preset archive management scenario;

[0111] Step S22: Extract the initial feature vectors of the core entities using the pre-trained language model, and use the core entities as archive entity nodes, with the initial feature vectors as the representation of the archive entity nodes. Construct an archive text hypergraph network using the word order adjacency relationship and grammatical structure association between the core entities.

[0112] Step S23: Based on the archival text hypergraph network, calculate the degree centrality of core entities in the archival text hypergraph network and their frequency of occurrence in the standardized archival dataset. Combine this with the contextual diversity score under the preset archival management scenario to calculate the comprehensive importance weight and relational feature sparseness weight of the core entities. The calculation formula is as follows:

[0113] ;

[0114] ;

[0115] in, This represents the overall importance weight of the core entities. This represents the first preset weighting coefficient. This represents the second preset weighting coefficient. This represents the third preset weighting coefficient. This indicates the degree centrality of the core entity in the archival text hypergraph network. Indicates the frequency of occurrence in the standardized archival dataset. This indicates the contextual diversity score within a predefined file management scenario. This represents the maximum degree centrality of the core entity in the document text hypergraph network. This represents the maximum frequency of occurrence in the standardized archival dataset. This represents the maximum value of the contextual diversity score within the preset file management scenario. The weights represent the rarity of relation features. This represents the total number of logical association paths between all core entities in the standardized archive dataset. Indicates the first The frequency of occurrence of specific relational features in standardized archival datasets. This represents the average degree centrality of core entities in the archival text hypergraph network;

[0116] Step S24: Apply the comprehensive importance weight as a gain coefficient to the initial feature vector of the core entity to obtain the gained input feature vector. The calculation formula is as follows:

[0117] ;

[0118] in, This represents the input feature vector after gain. Represents the initial eigenvector;

[0119] Step S25: Use the hypergraph attention mechanism to aggregate features of the hypergraph network of the archive text, update the node feature vectors corresponding to each archive entity node, and obtain the global semantic feature vector through pooling operation.

[0120] Specifically, a "preset file management scenario" is not a static dictionary, but a framework that defines specific analytical objectives and business rules. Examples include compliance audits of project payment, cross-checking of personnel files, and patent infringement risk assessment. The construction of preset file management scenarios is typically based on domain knowledge and exists in the form of configuration files or rule bases, guiding the system to understand the business meaning of entities and relationships. The core reason for its introduction is that the importance of the same entity can vary drastically across different business scenarios. For example, a project manager might be a critical core node in a safety incident accountability scenario, but might be just an ordinary entity in a file borrowing frequency statistics scenario. Through preset scenarios, the system can dynamically assign different contextual diversity scores to entities—this score measures the richness of a core entity's business context when it appears in different sentences, paragraphs, or combinations with other entities. Identifying core entities typically employs named entity recognition technology combined with a domain dictionary, while calculating its contextual diversity score can be done by analyzing the quantity and distribution of different business-related words or entity types within the surrounding window across all occurrences of the entity.

[0121] The initial feature vector of the core entity is extracted using a pre-trained language model. By inputting the context sentence or paragraph containing the entity into the pre-trained language model, the vector representation of the corresponding position of the entity is obtained. This vector initially contains rich general semantic information.

[0122] The core innovation of this step is the construction of a hypergraph network for archival texts. Unlike traditional binary relation graphs, which can only represent the relationship between two entities, hypergraphs allow an edge to connect any number of nodes. The construction method is as follows: each core entity is treated as a node, and then, based on the word order adjacency and syntactic structure associations in the text, groups of entities that participate in the same syntactic structure or local word order are connected by a hyperedge. For example, multiple entities appearing consecutively in word order adjacency constitute an event sequence; in syntactic structure associations, multiple entities with syntactic relationships such as subject-verb-object are combined through dependency parsing. In this way, the interactions between multiple entities in a complex contract clause or a descriptive text can be fully captured by a single hyperedge, thus more naturally representing the inherent multi-dimensional, high-order semantic relationships in archival texts.

[0123] Based on the constructed hypergraph network of archival text, the system performs multi-dimensional statistical measurements. The degree centrality of core entities within the hypergraph is statistically analyzed, typically referring to the number of hyperedges to which the entity node belongs. The more hyperedges an entity is associated with, the more pivotal its position is in the text's semantic structure. Simultaneously, the system also calculates the frequency of the entity's occurrence in the standardized archival dataset, a fundamental statistical importance indicator. Finally, the system combines these measures with the aforementioned contextual diversity score modulated by business scenarios, calculating a comprehensive importance weight using the formula in step S23. This weight is no longer a reflection of a single frequency or network degree, but a composite indicator integrating the frequency of occurrence in the standardized archival dataset, the degree centrality of the core entity in the archival text hypergraph network, and contextual diversity. Ultimately, this weight is applied as a gain coefficient to the initial feature vector of the core entities, and then information aggregation is performed through the hypergraph attention mechanism. This results in a global semantic feature vector that not only contains the semantics of the text itself but also highlights the core entity information with high value and complex relationships in specific business scenarios. Compared to traditional methods that rely solely on word frequency or simple co-occurrence for weight calculation, this method represents a leap from "planar statistics" to "contextualized and structured deep perception." The first preset weight coefficient is set to 0.5. In the hypergraph network of archival text, degree centrality directly reflects the pivotal nature and connectivity breadth of an entity node within the semantic structure, serving as a core indicator for measuring the structural importance of an entity. This invention captures high-order, multi-dimensional relationships between archival entities by constructing a hypergraph network; therefore, the structural information represented by degree centrality plays a decisive role in subsequent feature aggregation and global semantic representation. Setting the first preset weight coefficient to 0.5 highlights the dominant role of degree centrality in the comprehensive importance weight calculation, ensuring that key pivotal entities in the network receive higher weight gains. The second preset weight coefficient is set to 0.3. This coefficient is used to weight the frequency of occurrence in the standardized archival dataset, and frequency of occurrence is a fundamental indicator for measuring the statistical importance of an entity. Assigning a weight of 0.3 indicates that while emphasizing the structural pivotality, the system also fully respects the universality and importance of entities objectively presented in historical data, avoiding the neglect of certain low-frequency but critical entities that might result from relying solely on network structure. The third preset weight coefficient is set to 0.2. This coefficient is used to weight the contextual diversity score in the preset file management scenario, reflecting the entity's contextualized business value. Assigning a weight of 0.2 means that the system leaves an adjustable window for external business rules and domain knowledge in the core weight calculation. This allows the final importance of the same entity in different business scenarios such as auditing, verification, and analysis to be dynamically fine-tuned according to scenario requirements, achieving a leap from "general semantics" to "business semantics."

[0124] In specific implementation, in step S25, the hypergraph attention mechanism is used to aggregate features of the archive text hypergraph network, update the node feature vectors corresponding to each archive entity node, and obtain the global semantic feature vector through pooling operation. The processing logic includes:

[0125] Step S251: Extract the node feature vectors of each archive entity node in the archive text hypergraph network, and use a preset shared linear transformation matrix to perform spatial mapping on the node feature vectors to obtain the projected feature vectors.

[0126] Step S252: Calculate the attention coefficients between the central node and each neighboring node. The calculation formula is as follows:

[0127] ;

[0128] in, Indicates the attention coefficient. This represents the activation function. This represents a learnable attention vector. This represents the predefined shared linear transformation matrix. Represents the projected feature vector of the center node. This represents the projected feature vector of the neighboring nodes. This represents the vector concatenation operator;

[0129] Step S253: The attention coefficients are normalized using the Softmax function to obtain the node contribution. The archive entity node representations of each neighboring node are then weighted and summed to obtain the aggregate representation of the current central node. The calculation formula is as follows:

[0130] ;

[0131] ;

[0132] in, Indicates the node's contribution. Indicates the central node The set of adjacent domain nodes, Indicates the central node, Indicates the central node Any neighboring node in the set of adjacent domain nodes, Indicates the central node With any neighbor node during the traversal Attention coefficient Represents an exponential function. This represents the aggregate representation of the current central node. This represents the activation function. Indicates neighboring nodes;

[0133] Step S254: Perform pooling operation on the aggregated representations of all current central nodes to obtain the global semantic feature vector.

[0134] Specifically, a pre-defined shared linear transformation matrix is ​​used to spatially map the node feature vectors. The core approach involves performing matrix multiplication on this matrix with the original feature vector of each archival entity node in the archival text hypergraph network. This transforms all nodes from their original feature space to a common, new feature space, resulting in projected feature vectors. The fundamental reason for this is the establishment of a unified feature transformation standard shared by all nodes. This ensures the comparability of the transformed features across all nodes, providing a fair basis for subsequent calculations of attention relationships between nodes. Furthermore, this learnable matrix can automatically optimize to uncover and enhance the feature dimensions most useful for the current archival semantic aggregation task, improving the efficiency of feature representation.

[0135] Pooling is performed on the aggregated representations of all current central nodes, specifically global average pooling. The aggregated representations of all nodes are treated as a set, and the arithmetic mean of the values ​​of all nodes in each feature dimension is calculated. This operation yields a vector with dimensions identical to those of individual nodes, but each dimension contains global statistical information about the network—the global semantic feature vector. This step is crucial, elevating information from individual to holistic, from local structure to global semantics. This global vector no longer reflects the characteristics of any single node but captures the overall distribution and core patterns of the entire archival text hypergraph network across all feature dimensions, providing a highly generalized and information-rich global context for subsequent association reasoning. Compared to traditional methods that simply concatenate or select partial node features, this pooling approach more evenly and comprehensively compresses and preserves the semantic information of the entire network, avoiding information loss or bias.

[0136] This application introduces a pre-defined shared linear transformation matrix to spatially map node feature vectors and combines it with a hypergraph attention mechanism. Compared to traditional feature aggregation methods based on fixed rules or simple averaging, this achieves a dynamic and accurate measurement of the strength of associations between archival entities. Traditional methods struggle to distinguish the differences in semantic contributions of different neighboring nodes to the central node when fusing node information. This solution projects node features onto a unified semantic space using the aforementioned matrix and utilizes learnable attention vectors to calculate attention coefficients based on the dynamic relationship between the projected feature vectors of the central node and those of neighboring nodes. This automatically identifies and focuses on the most valuable highly associated neighbors for the current semantic aggregation. This mechanism ensures that the aggregation representation of each node's current central node is deeply integrated with its key contextual information, significantly improving the accuracy of characterizing complex archival semantic structures.

[0137] The pooling operation employed in this application offers significant advantages in information integrity compression and global semantic representation performance compared to traditional methods that directly concatenate features or select only representative nodes. Traditional methods are susceptible to information redundancy or sampling bias, making it difficult to form a compact and comprehensive global representation. This approach, however, pools the aggregated representations of all current central nodes, condensing all locally semantically enhanced node features in the network into a fixed-dimensional global semantic feature vector. This operation filters out individual node identity details while fully preserving the global semantic distribution and core patterns encoded by the overall structure of the hypergraph network, thereby generating a high-purity, highly generalizable archival semantic summary.

[0138] In specific implementation, step S3 includes:

[0139] Step S31: Use the cross-attention mechanism to perform a correlation mapping between the gained input feature vector and the global semantic feature vector to obtain a comprehensive semantic representation of the archival text.

[0140] Step S32: Based on the comprehensive semantic representation of the archive text, adjust the feature weights and relation feature sparse weights corresponding to each archive entity node in the archive text hypergraph network, and use the path search algorithm to traverse and retrieve the logical association paths connecting different archive entities in the adjusted archive text hypergraph network.

[0141] Step S33: Establish a semantic evaluation model, extract the number of relation features, relation strength, and total path step length in the logical association path, and calculate the inference confidence index. The calculation formula is as follows:

[0142] ;

[0143] in, This represents the confidence index of inference. Indicates the number of relational features. Indicates the first The correlation strength coefficient of the relationship. The total step length of a logically related path. The weights represent the rarity of relation features. A number indicating the relationship;

[0144] Step S34: Obtain the business request to be analyzed and perform vectorization encoding to obtain the request feature vector. Calculate the vector cosine similarity between the representation of each archive entity node and the request feature vector to obtain the semantic similarity score. Then, use a nonlinear fusion operator to process the semantic similarity score and the inference confidence index to obtain the archive association evaluation data.

[0145] Specifically, adjusting the feature weights and relation feature sparse weights corresponding to each archival entity node in the archival text hypergraph network involves using the comprehensive semantic representation of the archival text as a reference benchmark or contextual condition. Through a lightweight attention layer, the relative importance of each node feature and each relation feature within the current global semantic context is reassessed. For the feature weights corresponding to each archival entity node, the system calculates the relevance of each node to the global semantic core and dynamically scales the original weights accordingly, enhancing the influence of entities closely related to the current archival theme. For relation feature sparse weights, contextual corrections are made based on the importance of the nodes connected to the relation in the adjusted network and the semantic consistency of the relation pattern within the global context. For example, a relation that is originally rare may have its weight further increased if it connects two core entities that are both very important in the current comprehensive semantic context, highlighting the potential value of this unique association. This adjustment process transforms the hypergraph network from a static structural representation into a dynamic semantic graph deeply integrated with the overall content of the current archives.

[0146] In the dynamic semantic graph formed after weight adjustment, a path search algorithm is used to traverse and retrieve logically related paths connecting different archival entities. The aim is to discover potential, deep connections between entities that conform to business logic. This process is not a simple full graph traversal, but a semantically guided intelligent search. The system transforms the adjusted node weights and relation sparseness weights into cost or heuristic information in the graph search. Subsequently, an improved Dijkstra path search algorithm is used to explore the hypergraph network starting from a given initial archival entity node. This algorithm prioritizes traversing paths that pass through high-weight nodes and high-importance relations, while being constrained by path length, ultimately outputting one or more semantically and structurally economical and reliable connection paths. These paths clearly reveal the logical chains formed between archival entities through a series of intermediate nodes and relational features.

[0147] To match the natural language expression of the business request to be analyzed with the archival content, it needs to be vectorized. This step is achieved by using a pre-trained language model that is from the same source as the archival entity feature extraction. The text of the business request is input into the model to obtain its sentence-level or semantic representation vector, i.e., the request feature vector. This ensures that the semantic representation of the business request and the feature vectors of the entity nodes in the archive are in the same semantic space.

[0148] To integrate semantic similarity scores based on content matching and inference confidence metrics based on logical reasoning, the system utilizes a nonlinear fusion operator to process both. This nonlinear fusion operator is typically a multilayer perceptron containing a nonlinear activation function. Taking the semantic similarity score and inference confidence metric as inputs, the system captures the complex, nonlinear combination patterns between them through its internal learnable parameters and nonlinear transformations. After processing by the nonlinear fusion operator, a comprehensive and more discriminative archive association evaluation data is output. This data quantitatively reflects the strength and reliability of the association between the analyzed business request and the archive content, providing accurate and interpretable quantitative evidence for the final decision. The entire process constitutes a complete inference loop, from global semantic understanding to specific path discovery, and then to accurate quantitative matching with external requests.

[0149] The core function of the inference confidence index is to provide a quantifiable and interpretable reliability measure for the complex relationships between archival entities, thereby transforming semantic analysis results into key evidence supporting automated risk decision-making. By comprehensively considering the strength and rarity of each relationship in the association path, as well as the overall path length, a single and intuitive confidence value is calculated. This design effectively addresses the shortcomings of traditional methods, such as the ambiguity of the association inference process and the reliance on human experience in decision-making. It significantly improves the interpretability of system decisions. When the system makes a risk judgment, it can trace which factors led to the increase or decrease in confidence, such as whether the association chain is too long or contains some common but fragile connections. Secondly, it achieves accurate risk classification and identification, keenly discovering those hidden associations that appear semantically related but have weak or abnormal underlying logic, thus marking such cases as objects of concern requiring key verification. Finally, as one of the core factors of comprehensive risk assessment, this indicator works in conjunction with semantic similarity, enabling the system to dynamically allocate management resources, automatically handle requests with high confidence and high risk, and prioritize manual review for requests with low confidence and medium risk. This significantly improves the automation level and response efficiency of archival information management while ensuring security.

[0150] This application provides an intelligent association analysis method that deeply integrates semantic understanding and logical reasoning. Traditional methods fail to reveal the complex logical chains between archival entities and lack quantitative assessment of the reliability of the reasoning process, resulting in superficial and unreliable analysis results. This application constructs a complete technical chain including cross-attention association mapping, dynamic hypergraph weight adjustment, logical path search, and quantitative confidence evaluation. It upgrades the matching of business requests and archival content from traditional "content similarity calculation" to "structured logical association reasoning and quantitative evaluation." This enables the system to produce association evaluation data that is not only semantically relevant but also possesses inherent logical support and reliability metrics. Thus, it provides interpretable, quantifiable, and deeply reliable decision-making basis for archival analysis in high-risk business scenarios, achieving a fundamental leap from information retrieval to knowledge reasoning.

[0151] In specific implementation, in step S31, a cross-attention mechanism is used to perform an association mapping between the gained input feature vector and the global semantic feature vector to obtain a comprehensive semantic representation of the archival text. The processing logic includes:

[0152] Step S311: Extract the input feature vector after gain, and map the input feature vector after gain to a query matrix through a linear transformation matrix;

[0153] Extract the global semantic feature vector, and map the global semantic feature vector into a key matrix and a value matrix through the corresponding linear transformation matrix;

[0154] Step S312: Calculate the energy distribution matrix of the query matrix and the key matrix, and scale the energy distribution matrix using the square root of the feature dimension to obtain the scaled point integral distribution map.

[0155] Step S313: The Softmax function is used to normalize the integral distribution of the scaling points to obtain the attention weight matrix.

[0156] Step S314: The value matrix is ​​weighted and mapped using the attention weight matrix to obtain the mapped feature matrix. The mapped feature matrix is ​​then merged with the query matrix using the residual connection operator to obtain the comprehensive semantic representation of the archive text.

[0157] Specifically, in step S311, the gained input feature vector and global semantic feature vector are subjected to linear projection transformations, respectively. The gained input feature vector is then multiplied with the query transformation matrix to map it to the query matrix. Simultaneously, the same global semantic feature vector is multiplied with both the key transformation matrix and the value transformation matrix to map it to the key matrix and the value matrix, respectively. The ingenuity of this design lies in assigning different roles and functions to features originating from different levels within the attention mechanism through different projection matrices, enabling the global semantics to respond to the query of each specific entity in a structured manner.

[0158] By multiplying the query matrix by the transpose of the key matrix, an energy distribution matrix is ​​obtained. Each element in the energy distribution matrix quantifies the original association strength between a specific entity query and a certain "key" feature in the global semantics. To prevent the dot product result from being too large due to the high dimension of the feature vectors in the energy distribution matrix, which could affect the stability and gradient of the subsequent Softmax function, the system introduces a crucial scaling step: dividing each element in the calculated energy distribution matrix by the square root of the feature dimension of the key vector. This standardization operation yields a scaled dot integral distribution map, ensuring that the calculation of attention weights is within a numerical range where the gradient is more easily optimized. This is the mathematical foundation for the stable and efficient operation of the entire mechanism.

[0159] Compared to traditional methods that directly calculate similarity, this scheme first uses independent linear transformation matrices to map local entity features and global semantic features into function-separated query matrices, key matrices, and value matrices, respectively. This achieves accurate "retrieval-response" associations within a semantically aligned interaction space, effectively overcoming the biases caused by direct matching of heterogeneous features. Furthermore, by scaling the energy distribution matrix based on the square root of the feature dimension, the system avoids the numerical instability of high-dimensional dot products, ensuring a reasonable distribution of attention weights and gradient optimization, enabling the model to stably capture all key semantic associations from salient to weak. Finally, residual connections are used to merge the mapped global features with the original local query features, generating a comprehensive semantic representation of the archival text. This design deeply integrates the global context while fully preserving the core details of local entities, achieving enhanced rather than submerged information fusion. Overall, this series of operations ensures that the final semantic representation possesses both accurate local semantics and a rich global perspective.

[0160] In specific implementation, step S314 includes:

[0161] The mapping feature matrix is ​​obtained by multiplying the weight vector of each row in the attention weight matrix with the value matrix. The calculation formula is as follows:

[0162] ;

[0163] in, Represents the mapping feature matrix, Represents the attention weight matrix. Represents a value matrix;

[0164] The mapping feature matrix and the query matrix are added element-wise using the residual addition operator to obtain the fused feature term. The fused feature term is then processed using the layer normalization operator to obtain the comprehensive semantic representation of the archival text.

[0165] Specifically, the layer normalization operator calculates the mean and variance of each row in the fused feature term matrix across the entire feature dimension. The layer normalization operator then uses this mean and variance to standardize the sample vector, transforming it into a distribution with a mean of 0 and a variance of 1. This process involves not only simple scaling but also the introduction of two learnable parameter vectors—a scaling factor and a shift factor—allowing the standardized features to be flexibly rescaled and translated according to task requirements. The beneficial effects of applying the layer normalization operator to the fused feature terms are direct and crucial: it effectively alleviates the internal covariate shift problem that may be caused by attention weighting and matrix multiplication, ensuring that different sample features are within a stable and comparable numerical range, thereby significantly accelerating model training convergence and enhancing its generalization ability. The features output after this normalization step become the final comprehensive semantic representation of the archival text, which integrates global contextual information, preserves local query details, and has a robust numerical distribution.

[0166] This application utilizes the attention weight matrix to perform matrix multiplication on the value matrix. Compared to the traditional method of global undifferentiated information aggregation, this achieves a high degree of selectivity and precise targeting in the extraction of archival information. It ensures that only the global semantic components that best match the current business intent are extracted into the mapping feature matrix, greatly improving the system's ability to capture key clues in complex archival contexts. At the same time, it introduces a residual addition operator to add the mapping feature matrix and the query matrix element-wise, followed by layer normalization operator processing. Compared to the traditional direct feature replacement method, this not only fully preserves the original semantic gain of local entities while introducing the global context, effectively preventing information dilution and numerical instability problems in deep learning, but also ensures that the final generated comprehensive semantic representation of archival text has extremely high fidelity and robustness.

[0167] In specific implementation, step S4 includes:

[0168] Step S41: Invoke the inference confidence index and semantic similarity score corresponding to the business request to be analyzed, and perform weighted fusion processing using a preset nonlinear mapping operator to obtain the comprehensive management risk index. The calculation formula is as follows:

[0169] ;

[0170] in, Indicates the comprehensive management risk index, This represents the activation function. This represents the first preset decision weight coefficient. This represents the second preset decision weight coefficient. Indicates the semantic similarity score. This represents the confidence index of inference. Indicates the bias term;

[0171] Step S42: Extract the comprehensive management risk index and match it with the preset risk discrimination matrix to obtain the analysis results;

[0172] The analysis results include high-risk conflicts, verification concerns, manual risk review, and compliance approval;

[0173] Step S43: Extract the analysis results and match them with the preset decision instruction set to obtain the corresponding file management decision instructions;

[0174] The processing logic for verification and concern determination is as follows:

[0175] When the comprehensive management risk index is within the preset verification range and the reasoning confidence index is lower than the preset confidence threshold, it will receive verification attention and trigger a manual secondary review instruction.

[0176] Step S44: Extract the behavioral metadata when executing the file management decision instruction, combine it with the unique identifier of the analysis result, generate a security traceability record, and encapsulate the security traceability record into an encrypted data block and mount it to the distributed ledger node;

[0177] Security traceability records include the access subject, risk characteristics, and decision-making basis.

[0178] Specifically, the preset nonlinear mapping operator consists of an activation function, a first preset decision weight coefficient, a second preset decision weight coefficient, and a bias term. After linearly weighting the semantic similarity score and the inference confidence index, it is input into a Sigmoid function for transformation, thus obtaining a comprehensive management risk index. The risk of file association is not simply determined by the sum of semantic similarity and inference confidence; there is a complex nonlinear interaction between the two. For example, high similarity accompanied by low confidence may indicate potential semantic fraud or logical conflict, and its risk nature is quite different from the case where both are high. The nonlinear mapping operator can automatically learn this complex risk combination pattern through model training, thereby generating a more business-discriminatory and interpretable comprehensive risk quantification index. The first preset decision weight coefficient is set to 0.4, and the second preset decision weight coefficient is correspondingly set to 0.6. This setting is based on the core principle of "inference first, semantic verification," meaning that in the risk assessment framework constructed in this invention, more emphasis should be placed on the logical reasoning evidence chain generated by the system through hypergraph networks, path search, and confidence calculation, rather than relying solely on surface text similarity. The reasoning confidence index quantifies the reliability and logical strength of the association path between archival entities, and is a key innovative output of this solution to achieve the leap from information retrieval to knowledge reasoning; while the semantic similarity score provides a basic content matching reference. Giving reasoning confidence a higher weight (0.6) enables the comprehensive management risk index to more sensitively reflect the credibility of deep reasoning, thereby directly addressing the deficiency of traditional technologies in lacking an interpretable quantitative reasoning system. The advantage of this setting is that it not only highlights the advanced nature and distinguishability of the logic-driven decision-making of this invention, but also makes the operation of the subsequent preset risk discrimination matrix more accurate—especially when the risk index is in the middle gray zone, the system can intelligently distinguish between "verification concern" and "manual risk review" based on whether the reasoning confidence is below the threshold, realizing dynamic diversion and optimization of risk disposal. At the same time, retaining an appropriate weight (0.4) for semantic similarity avoids the limitations of a single index, ensuring that the system can still maintain a robust response in scenarios where the logical path is not obvious but the semantics are highly related, reflecting the design idea of ​​evidence balance. In addition, this weight setting provides an initial starting point that conforms to business logic for model training, which helps guide the algorithm to converge quickly and ultimately comprehensively improves the risk identification accuracy, decision interpretability, and automated response efficiency of the archive information management system.

[0179] The pre-configured decision instruction set is a rule base that maps different analysis results to specific executable operations. Specifically, after obtaining analysis results such as high-risk conflicts or verification concerns, these are used as keys to query the pre-configured decision instruction set, retrieving and triggering the corresponding file management decision instructions. For example, a high-risk conflict might map to automatically freezing file access and notifying the security officer, while a verification concern would trigger a manual secondary review instruction. Its fundamental purpose is to automate, standardize, and make risk response real-time, ensuring that every risk assessment conclusion can be translated into predetermined management actions unambiguously and without delay. This eliminates response delays or operational inconsistencies that may result from human intervention, thereby significantly improving the efficiency and standardization of risk management.

[0180] When a record management decision is executed, the system automatically captures relevant metadata, such as timestamps, user IDs, record identifiers, and instruction content. Simultaneously, it associates this with a unique identifier triggered by the analysis results. This information is then structured and assembled according to a predetermined format to generate a complete security traceability record. This record comprehensively documents the entire chain of information: "when, who, based on what risk analysis, on which record, and what operation was performed." To ensure the record's permanent trustworthiness and immutability, the system encrypts the record, encapsulates it into an encrypted data block, and then submits and uploads it to a distributed ledger node. The fundamental reason for introducing distributed ledger technology is to leverage its decentralized, immutable, and traceable characteristics to provide the highest level of data integrity assurance and audit credibility for the entire record management decision-making process. Once a record is on the chain, no single party can modify or delete it, thus providing a solid and reliable technological foundation for post-event auditing, accountability, and compliance verification.

[0181] The preset risk assessment matrix is ​​not a simple single-value threshold, but a logical judgment framework that integrates two dimensions: a comprehensive management risk index and an inference confidence index. Its construction is based on domain knowledge, historical risk cases, and compliance requirements. It divides the risk index into major, medium, and low-level ranges by setting a first and second risk threshold, and simultaneously sets a preset confidence threshold to assess the reliability of the inference process. The entire matrix operates as follows: after the system calculates the two quantitative indicators, it automatically substitutes them into the matrix for matching—for example, when the risk index is higher than the first threshold, it is directly judged as a high-risk conflict; when the risk index is in the middle range, the confidence index is further cross-validated; if it is lower than the preset confidence threshold, it is judged as requiring intervention and verification attention; otherwise, it enters manual risk review; if the risk index is lower than the second threshold, it is judged as compliant. The fundamental reason for adopting this matrix-based judgment method is that it overcomes the inherent defect of traditional single-dimensional risk scoring in being unable to distinguish the "nature of risk." By cross-checking the reliability of evidence with the severity of risk, this matrix can accurately identify "gray area" business requests that appear to have moderate risk levels but for which the system itself lacks sufficient control. This triggers more targeted management actions, such as secondary manual review, achieving a leap from "rough alerts" to "intelligent handling." It significantly improves the accuracy of high-risk identification and effectively avoids misjudgments or omissions caused by system uncertainties. This makes the decision-making logic of the entire archive risk management system closer to expert experience, and the response more accurate and efficient.

[0182] In specific implementation, in step S42, the comprehensive management risk index is extracted and matched with the preset risk discrimination matrix to obtain the analysis results. The processing logic includes:

[0183] Retrieve the first and second risk thresholds from the preset risk discrimination matrix;

[0184] Among them, the first risk threshold is greater than the second risk threshold;

[0185] The comprehensive management risk index is compared with the first risk threshold. When the comprehensive management risk index is greater than or equal to the first risk threshold, the analysis result is output as a high-risk conflict.

[0186] When the comprehensive management risk index is less than the first risk threshold but greater than or equal to the second risk threshold, the inference confidence index is extracted and compared with a preset confidence threshold, specifically including:

[0187] If the reasoning confidence index is less than the preset confidence threshold, the output analysis result will be "verification attention" and a manual secondary review instruction will be triggered.

[0188] If the reasoning confidence index is greater than or equal to the preset confidence threshold, the output analysis result will be manually reviewed for risk.

[0189] When the comprehensive management risk index is less than the second risk threshold, the output analysis result is "compliant and passed".

[0190] Specifically, this application constructs a risk discrimination mechanism based on the synergy of dual thresholds and inference confidence. Compared to traditional single-threshold alarms or linear grading models, its significant advantage lies in achieving accurate classification and dynamic handling of risks in document management. Traditional technologies often rely solely on fixed numerical limits for risk assessment, easily generating numerous false alarms or missed alarms in critical areas, leading to a surge in management costs. This solution, through a tiered design of the first and second risk thresholds, can quickly separate extremely high-risk conflict items from absolutely compliant pass items, greatly improving the system's initial screening efficiency. More importantly, for potential risks in the gray area, this solution introduces an inference confidence index as a secondary judgment dimension. By distinguishing between "verification concern" and "manual review," it accurately triggers different levels of management instructions, effectively alleviating the redundancy and pressure of manual review common in traditional methods. This ensures that management decisions possess both the agility of automated processing and the rigor of logical evidence chains, fundamentally optimizing the resource allocation efficiency and risk control accuracy of the electronic document information management system.

[0191] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0192] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for information management of electronic records, characterized in that, Includes the following steps: Step S1: Collect the original electronic archive text data, preprocess the original electronic archive text data to obtain a standardized archive dataset. The processing logic is as follows: Step S11: Use the N-gram sliding window algorithm to segment the original electronic archive text data to obtain a candidate word set. Compare the words in the candidate word set with the preset standard archive lexicon. If a word is not in the preset standard archive lexicon and its word frequency is lower than the preset frequency threshold, then mark the word as an interference candidate. The original electronic archive text data includes unstructured text, original evidence records from the electronic archive system, and archive metadata; Unstructured text includes visually similar characters and homophonic misspellings. Step S12: Extract the initials, finals and tones of the interference candidates to form a pinyin sequence. Use the edit distance algorithm to calculate the conversion cost between the pinyin sequence and the candidate original meaning word sequence in the preset standard archive lexicon, and perform normalization processing to obtain the pinyin similarity. Step S13: Based on the stroke, radical, and structural features of the interference candidates, extract the topological feature vectors of the characters and calculate the cosine similarity to obtain the character similarity between the interference candidates and the candidate original words. Step S14: The pinyin similarity and character shape similarity are weighted and summed to calculate the comprehensive similarity value. The calculation formula is as follows: ； Where S represents the overall similarity value, This represents the similarity adjustment coefficient. Indicates the similarity of pinyin. Indicates the similarity of character shapes; Step S15: Select the candidate original word with the highest comprehensive similarity value to perform semantic replacement on the interference candidate, and perform noise reduction and format standardization on the text to obtain a standardized archive dataset. Step S2: Identify the core entities in the standardized archive dataset, construct an archive text hypergraph network, and use the hypergraph attention mechanism to aggregate features of the archive text hypergraph network to obtain a global semantic feature vector. Step S3: Use the cross-attention mechanism to perform association mapping on the input feature vector after gain and the global semantic feature vector to obtain the comprehensive semantic representation of the archival text. Based on the comprehensive semantic representation of the archival text, perform archival entity association reasoning on the business request to be analyzed to obtain archival association evaluation data. Step S4: Based on the archive association evaluation data, perform risk analysis on the business request to be analyzed, obtain the analysis results, match the analysis results to obtain archive management decision instructions, and generate security traceability records.

2. The electronic records information management method as described in claim 1, characterized in that, Step S2 specifically includes: Step S21: Identify the core entities in the standardized archive dataset and calculate the contextual diversity scores of the core entities in the preset archive management scenario; Step S22: Extract the initial feature vectors of the core entities using the pre-trained language model, and use the core entities as archive entity nodes, with the initial feature vectors as the representation of the archive entity nodes. Construct an archive text hypergraph network using the word order adjacency relationship and grammatical structure association between the core entities. Step S23: Based on the archival text hypergraph network, calculate the degree centrality of core entities in the archival text hypergraph network and their frequency of occurrence in the standardized archival dataset. Combine this with the contextual diversity score under the preset archival management scenario to calculate the comprehensive importance weight and relational feature sparseness weight of the core entities. The calculation formula is as follows: ； ； Where W represents the overall importance weight of the core entity, This represents the first preset weighting coefficient. This represents the second preset weighting coefficient. denoted by the third preset weight coefficient, d represents the degree centrality of the core entity in the archival text hypergraph network, f represents the frequency of occurrence in the standardized archival dataset, and c represents the contextual diversity score under the preset archival management scenario. This represents the maximum degree centrality of the core entity in the document text hypergraph network. This represents the maximum frequency of occurrence in the standardized archival dataset. This represents the maximum value of the contextual diversity score within the preset file management scenario. This represents the sparseness weight of relational features, where N represents the total number of logical association paths between all core entities in the standardized archival dataset. Indicates the first The frequency of occurrence of specific relational features in standardized archival datasets. This represents the average degree centrality of core entities in the archival text hypergraph network; Step S24: Apply the comprehensive importance weight as a gain coefficient to the initial feature vector of the core entity to obtain the gained input feature vector. The calculation formula is as follows: ； Where V represents the input feature vector after gain. Represents the initial eigenvector; Step S25: Use the hypergraph attention mechanism to aggregate features of the hypergraph network of the archive text, update the node feature vectors corresponding to each archive entity node, and obtain the global semantic feature vector through pooling operation.

3. The electronic records information management method as described in claim 2, characterized in that, In step S25, the hypergraph attention mechanism is used to aggregate features of the archive text hypergraph network, update the node feature vectors corresponding to each archive entity node, and obtain the global semantic feature vector through pooling operations. The processing logic includes: Step S251: Extract the node feature vectors of each archive entity node in the archive text hypergraph network, and use a preset shared linear transformation matrix to perform spatial mapping on the node feature vectors to obtain the projected feature vectors. Step S252: Calculate the attention coefficients between the central node and each neighboring node. The calculation formula is as follows: ； in, Indicates the attention coefficient. This represents the activation function. This represents a learnable attention vector. This represents the predefined shared linear transformation matrix. Represents the projected feature vector of the center node. This represents the projected feature vector of the neighboring nodes. This represents the vector concatenation operator; Step S253: The attention coefficients are normalized using the Softmax function to obtain the node contribution. The archive entity node representations of each neighboring node are then weighted and summed to obtain the aggregate representation of the current central node. The calculation formula is as follows: ； ； in, Indicates the node's contribution. Let k represent the set of adjacency nodes of the central node i, where i represents the central node and k represents any neighboring node in the set of adjacency nodes of the central node i. Let represent the attention coefficient between the center node i and any neighbor node k during the traversal. Represents an exponential function. This represents the aggregate representation of the current central node. Let j represent the activation function and j represent the neighboring nodes; Step S254: Perform pooling operation on the aggregated representations of all current central nodes to obtain the global semantic feature vector.

4. The electronic records information management method as described in claim 3, characterized in that, Step S3 specifically includes: Step S31: Use the cross-attention mechanism to perform a correlation mapping between the gained input feature vector and the global semantic feature vector to obtain a comprehensive semantic representation of the archival text. Step S32: Based on the comprehensive semantic representation of the archive text, adjust the feature weights and relation feature sparse weights corresponding to each archive entity node in the archive text hypergraph network, and use the path search algorithm to traverse and retrieve the logical association paths connecting different archive entities in the adjusted archive text hypergraph network. Step S33: Establish a semantic evaluation model, extract the number of relation features, relation strength, and total path step length in the logical association path, and calculate the inference confidence index. The calculation formula is as follows: ； in, This represents the confidence index of the inference, where n represents the number of relation features. Let L represent the association strength coefficient of the m-th relation, and L represent the total step length of the logical association path. This represents the rarity weight of relation features, where m represents the relation number. Step S34: Obtain the business request to be analyzed and perform vectorization encoding to obtain the request feature vector. Calculate the vector cosine similarity between the representation of each archive entity node and the request feature vector to obtain the semantic similarity score. Then, use a nonlinear fusion operator to process the semantic similarity score and the inference confidence index to obtain the archive association evaluation data.

5. The electronic records information management method as described in claim 4, characterized in that, In step S31, a cross-attention mechanism is used to correlate and map the gained input feature vector and the global semantic feature vector to obtain a comprehensive semantic representation of the archival text. The processing logic includes: Step S311: Extract the input feature vector after gain, and map the input feature vector after gain to a query matrix through a linear transformation matrix; Extract the global semantic feature vector, and map the global semantic feature vector into a key matrix and a value matrix through the corresponding linear transformation matrix; Step S312: Calculate the energy distribution matrix of the query matrix and the key matrix, and scale the energy distribution matrix using the square root of the feature dimension to obtain the scaled point integral distribution map. Step S313: The Softmax function is used to normalize the integral distribution of the scaling points to obtain the attention weight matrix. Step S314: The value matrix is ​​weighted and mapped using the attention weight matrix to obtain the mapped feature matrix. The mapped feature matrix is ​​then merged with the query matrix using the residual connection operator to obtain the comprehensive semantic representation of the archive text.

6. The electronic records information management method as described in claim 5, characterized in that, Step S314 specifically includes: The mapping feature matrix is ​​obtained by multiplying the weight vector of each row in the attention weight matrix with the value matrix. The calculation formula is as follows: ； in, Let A represent the mapping feature matrix, B represent the attention weight matrix, and B represent the value matrix. The mapping feature matrix and the query matrix are added element-wise using the residual addition operator to obtain the fused feature term. The fused feature term is then processed using the layer normalization operator to obtain the comprehensive semantic representation of the archival text.

7. The electronic records information management method as described in claim 6, characterized in that, Step S4 specifically includes: Step S41: Invoke the inference confidence index and semantic similarity score corresponding to the business request to be analyzed, and perform weighted fusion processing using a preset nonlinear mapping operator to obtain the comprehensive management risk index. The calculation formula is as follows: ； Where E represents the comprehensive management risk index, This represents the activation function. This represents the first preset decision weight coefficient. This represents the second preset decision weight coefficient, and K represents the semantic similarity score. represents the confidence index of inference, and b represents the bias term; Step S42: Extract the comprehensive management risk index and match it with the preset risk discrimination matrix to obtain the analysis results; The analysis results include high-risk conflicts, verification concerns, manual risk review, and compliance approval; Step S43: Extract the analysis results and match them with the preset decision instruction set to obtain the corresponding file management decision instructions; The processing logic for verification and concern determination is as follows: When the comprehensive management risk index is within the preset verification range and the reasoning confidence index is lower than the preset confidence threshold, it will receive verification attention and trigger a manual secondary review instruction. Step S44: Extract the behavioral metadata when executing the file management decision instruction, combine it with the unique identifier of the analysis result, generate a security traceability record, and encapsulate the security traceability record into an encrypted data block and mount it to the distributed ledger node; The security traceability record includes the access subject, risk characteristics, and decision-making basis.

8. The electronic records information management method as described in claim 7, characterized in that, In step S42, the comprehensive management risk index is extracted and matched with a preset risk discrimination matrix to obtain the analysis results. The processing logic includes: Retrieve the first and second risk thresholds from the preset risk discrimination matrix; Among them, the first risk threshold is greater than the second risk threshold; The comprehensive management risk index is compared with the first risk threshold. When the comprehensive management risk index is greater than or equal to the first risk threshold, the analysis result is output as a high-risk conflict. When the comprehensive management risk index is less than the first risk threshold but greater than or equal to the second risk threshold, the inference confidence index is extracted and compared with a preset confidence threshold, specifically including: If the reasoning confidence index is less than the preset confidence threshold, the output analysis result will be "verification attention" and a manual secondary review instruction will be triggered. If the reasoning confidence index is greater than or equal to the preset confidence threshold, the output analysis result will be manually reviewed for risk. When the comprehensive management risk index is less than the second risk threshold, the output analysis result is "compliant and passed".

9. An electronic records information management system, applied in the electronic records information management method as described in any one of claims 1-8, characterized in that, It includes a text processing module, a semantic network construction module, an association analysis module, and a decision tracing module; The text processing module is used to collect raw electronic archive text data, preprocess the raw electronic archive text data, and obtain a standardized archive dataset. The semantic network construction module is used to identify core entities in the standardized archive dataset, construct an archive text hypergraph network and calculate the comprehensive importance weight and relation feature sparse weight of the core entities, and use the hypergraph attention mechanism to perform feature aggregation on the archive text hypergraph network to obtain a global semantic feature vector. The association analysis module is used to perform association mapping on the input feature vector and the global semantic feature vector after gain using the cross attention mechanism to obtain the comprehensive semantic representation of the archival text. Based on the comprehensive semantic representation of the archival text, the module performs archival entity association reasoning on the business request to be analyzed to obtain archival association evaluation data. The decision tracing module is used to perform risk analysis on the business request to be analyzed based on the archive association evaluation data, obtain the analysis results, match the analysis results to obtain the archive management decision instructions, and generate a security tracing record.

Images