Electronic archives information-based automatic management and classification system and method

Abstract

The application discloses an electronic archive informationization automatic management classification system and method, relates to the technical field of electronic archive management and information barrier-free cross, and achieves multi-feature fusion and automatic classification indexing of multi-format electronic archives through an archive automatic classification engine, and constructs a multi-level tree-shaped category system; a user behavior perception module determines the barrier-free demand type of a user based on a multi-feature weighting model without sensing; the application matches a differentiated interaction mode through a multi-modal barrier-free interaction module, links archives of different levels to execute grading desensitization rules, dynamically adjusts category display logic by combining access frequency, user matching degree and archive value through a classification view dynamic scheduling module, collects whole-process operation data through an interaction feedback iteration module, continuously optimizes the model and scheduling rules in an incremental learning mode, and finally realizes safe and compliant, accurately adapted barrier-free electronic archive automatic classification management.

Description

Technical Field

[0001] This invention belongs to the technical field of electronic records management and information accessibility, specifically relating to an automatic management and classification system and method for electronic records. Background Technology

[0002] Currently, electronic record management systems within enterprises and institutions have generally completed digital upgrades. Relying on automatic classification technology, they enable rapid collection and hierarchical management of various types of records, including documents, public records, and business records. These systems simultaneously support both internal business record retrieval and public access to records. With the advancement of age-friendly and information accessibility improvements in government services, record management systems, as public service entry points, need to cover the record retrieval needs of special groups such as the elderly and visually impaired. Furthermore, most organizational archives store internal documents and classified records of varying security levels; the visibility of the classification system must precisely correspond to user permissions and access modes, placing higher demands on the precision and security compliance of accessibility improvements.

[0003] The accessibility modifications to existing electronic record classification systems mostly remain at the surface level, only supporting manual activation of large font mode or screen reader compatibility mode. They fail to automatically identify adaptation needs based on user behavior, resulting in a high learning curve for special groups. The adaptation logic is not specifically optimized to address the multi-level nested structure of record classifications. Visually impaired users may experience semantic confusion when navigating categories using screen readers, and elderly users may struggle to quickly locate target records within multi-level category paths. More critically, existing modifications do not integrate with the record classification management system. In accessibility mode, all categories and record information are displayed indiscriminately, without differentiated anonymization rules for different classification levels, posing a compliance risk of leakage of classified category information. Furthermore, the adaptation rules are fixed configurations, unable to dynamically iterate based on real user feedback, leading to a gradual disconnect between the adaptation effect and actual user needs over time.

[0004] In summary, existing automatic electronic record classification systems cannot simultaneously meet the multiple demands of classification management efficiency, accessibility and inclusiveness of services, and compliance with record confidentiality requirements. A classification management solution is needed that can seamlessly identify users' accessibility needs, link confidentiality level control, and dynamically iterate and optimize, thereby improving the record utilization experience for special groups while ensuring the security of record data. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an automated electronic archive information management and classification system and method. This system can achieve multi-feature fusion and automatic classification indexing of multi-format electronic archives through an automatic archive classification engine, constructing a multi-level tree-shaped category system; it can seamlessly determine user accessibility needs based on a multi-feature weighted model through a user behavior perception module; it can match differentiated interaction modes through a multi-modal accessibility interaction module, linking archive confidentiality levels to execute hierarchical desensitization rules; it can dynamically adjust the category display logic through a classification view dynamic scheduling module combining access frequency, user matching degree, and archive value; and it can collect full-process operation data through an interactive feedback iteration module, continuously optimizing the model and scheduling rules through incremental learning, ultimately achieving secure, compliant, and accurately adaptable automated classification management of electronic archives.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, an electronic archive information-based automatic management and classification system includes the following components: Automatic Archive Classification Engine: Connects to the archive collection interface to complete the content parsing of multi-format electronic archives, deep integration of text semantic features and metadata features, and automatic classification indexing, and constructs a multi-level tree-shaped archive classification category system; User behavior perception module: Collects user operation data in the categorized navigation scenario through front-end embedding, anonymizes and desensitizes the collected behavior data, and determines the user's accessibility needs type based on a multi-feature weighted model and outputs the corresponding user tag; Multimodal accessibility interaction module: Matches differentiated interaction modes based on user tags, and executes hierarchical desensitization broadcast and display rules in conjunction with file confidentiality attributes; Category view dynamic scheduling module: Combines user tags, historical operation habits and file value attributes to dynamically adjust the category display level, sorting priority and interface layout; Interactive feedback iteration module: Collects user operation feedback data throughout the entire process, and uses incremental learning to optimize the user behavior perception model and view scheduling rules respectively.

[0007] Furthermore, the automatic archive classification engine uses a pre-trained model based on the BERT-base architecture for classification. The specific steps are as follows: It connects to the archive collection interface to receive electronic archives in multiple formats, and performs format adaptation preprocessing on PDF documents, scanned images, and audio / video files respectively; for scanned documents, it extracts text content using optical character recognition, marking items with a single-field recognition confidence score below 0.85 as items requiring review; for audio / video files, it extracts segmented text through speech-to-text conversion and simultaneously marks timestamps; it simultaneously extracts three types of metadata features: archive document number, generating unit, and date of creation, and maps these metadata after one-hot encoding. The 128-dimensional feature vector is fused with the 768-dimensional text semantic vector by concatenating the first and last parts. The input is used to classify the model and output the corresponding classification label to construct a three-level tree-structured classification system. The classification results are synchronously written to the archive metadata database. The classification model is trained on 120,000 historical labeled archives. When adding a new category, the parameters of the underlying Transformer network are fixed, and only the parameters of the fully connected output layer are fine-tuned. The learning rate is set to 2e-4. The training iteration is 45 rounds with an early stopping strategy. The training is terminated if the accuracy on the validation set does not improve for 5 consecutive rounds. The model can adapt to the dynamic adjustment requirements of the classification system without full retraining.

[0008] Furthermore, the user behavior perception module incorporates a multi-feature weighted recognition model, which calculates a comprehensive score for user accessibility needs using the following formula. : ,in A comprehensive score is given to users' accessibility needs, with a score range of 0-100; - The baseline weights for the five feature dimensions were based on 12,000 user behavior samples from the municipal archives over 12 months. The training and validation sets were divided in a 7:3 ratio. The optimal index for user type identification was selected through five-fold cross-validation and then normalized to three decimal places. The values ​​were 0.237, 0.214, 0.221, 0.188, and 0.140. This is the characteristic value of the page scaling ratio. It is 0 when the scaling ratio is ≤100% and 100 when it is ≥200%. It is calculated as a linear difference within the interval. The single category click interval feature value is the average interval of 5 consecutive category clicks. The value is 0 when the interval is ≤1s and 100 when it is ≥4s, with a linear mapping within the interval. The keyboard navigation percentage feature value is obtained by multiplying the percentage of keyboard navigation operations in the user's last 10 categorized operations by 100. The voice interaction percentage feature value is obtained by multiplying the proportion of voice command operations in the current user session to the total number of operations by 100. The feature value for the percentage of users returning to the previous navigation level in their last 10 navigation actions is calculated by multiplying the percentage by 100. The model has two preset threshold levels. A score ≥71.5 indicates a visually impaired user; a score ≤47.5 indicates a visually impaired user. A score below 71.5 indicates an elderly user. If the value is less than 47.5, the user is identified as a regular user, and the recognition result is synchronized to the two downstream modules in real time.

[0009] Furthermore, the multimodal accessibility interaction module sets up a hierarchical semantic interaction mechanism for visually impaired users; it performs hierarchical semantic translation on the nested structure of the multi-level category tree of archives, converting each level of category into a linear semantic sequence of Y categories at level X, with the current category being the Zth category plus the category name. In multi-level nesting, a prefix identifier is added to the parent category to avoid hierarchical confusion when the screen reader reads the data. When connecting to mainstream screen readers, it broadcasts the commands in a linear sequence order, integrates a voice interaction engine to receive user natural language commands, and uses a built-in archive domain thesaurus to align the commands to entities before employing edit distance. The system calculates the similarity between the command text and the standard category name, with a similarity threshold of 0.67. Categories exceeding the threshold are sorted in descending order of similarity and then presented to the user. The broadcast process is linked to the file's security classification. Top Secret and Confidential files only broadcast the document number and retention period, without broadcasting the content summary and subject keywords. Secret files only broadcast a de-identified subject overview. Non-classified files broadcast the complete core metadata. Users can complete the entire process of category navigation, file filtering, and result browsing via voice commands. The voice broadcast allows users to interrupt at any time and issue new commands, with a system response latency of ≤180ms.

[0010] Furthermore, the multimodal accessibility interaction module is configured with both interface and content tailored to elderly users. It automatically adjusts the basic presentation parameters of the category pages, changing the default font size to 19 points, increasing the line spacing by 32%, and simultaneously enlarging the interactive button size to 1.45 times its original size, with the button click hotspot extending 10px outwards to reduce the probability of click errors by elderly users. It hides non-core function entries such as advanced filtering, batch export, and category customization, retaining only the three core functions: category navigation, keyword search, and file preview. This simplifies the layout of page interactive elements, removing redundant decorative icons and animations to reduce visual and cognitive burden. A one-click "return to homepage" button is fixed on the right side of the page, allowing users to return to the primary category page with a single click. All operation buttons are accompanied by clear text descriptions, avoiding purely icon-based interaction. The color contrast between page text and background is adjusted to 4.7:1, and the display fields for livelihood-related file entries are automatically simplified, showing only the file name and the file retrieval process, while hiding professional management fields, adapting to the visual abilities and usage needs of elderly users.

[0011] Furthermore, the category view dynamic scheduling module has a built-in category priority sorting model, which calculates the display priority score of a single category using the following formula. : ,in The category is given a priority score; the higher the score, the higher the category is ranked. , , , The weight coefficients for the four evaluation dimensions were constructed by five experts from the fields of record management, accessibility design, and user experience using the analytic hierarchy process (AHP). The matrix was then calibrated after a consistency test (CR = 0.068 < 0.1), and the values ​​were 0.392, 0.315, 0.183, and 0.110. The category access frequency feature value is obtained by taking the cumulative number of visits of the corresponding group of users in the past 27 days for this category, and mapping it to the 0-1 interval after maximum and minimum normalization. The matching feature value for user groups is 1 for categories such as social security and medical insurance, real estate registration, and public services for elderly users; 1 for categories such as official document circulation and business approval for ordinary users; and the corresponding value in the range of 0.2-0.8 for other categories based on business relevance. This is the aging decay coefficient, with a value of e. The power, where The attenuation constant has a value of 0.022. This is the number of days since the last visit to the category, expressed in days (d). The value coefficient for archives is set as follows: 1 for permanent preservation, 0.6 for 30-year preservation, and 0.3 for 10-year preservation. The model sorts the categories at the same level from high to low according to the P-value, and the top 12% of categories are included in the priority display area to shorten the user navigation path.

[0012] Furthermore, the category view dynamic scheduling module executes differentiated hierarchical scheduling rules for different user groups. For visually impaired users, only the first two levels of categories are displayed by default. When the category system exceeds three levels, deeper subcategories are automatically collapsed. Users can trigger an expansion operation via voice or keyboard commands before loading the next level of category data. All categories at the same level are arranged from highest to lowest priority score, with high-frequency and popular categories placed first. The category page is configured with dedicated keyboard shortcuts for five high-frequency public service categories, allowing users to customize the shortcut mappings according to their own usage habits. When a user enters a new category... When the category level is reached, the scheduling module triggers the multimodal accessibility interaction module to execute the level broadcast; for elderly users, a quick category zone is permanently displayed at the top of the category page, automatically extracting the 5 categories with the highest access frequency in the past 27 days and including them in the zone; the quick category set is recalculated and updated every 8 days; for new users without historical operation data, a cold start strategy is adopted, and five general high-frequency livelihood categories such as social security inquiry, marriage records, real estate registration, academic qualification certification, and policy documents are included in the zone by default; the administrator manually adjusts the default configuration and quantity limit of quick categories according to the business type of the organization.

[0013] Furthermore, the interactive feedback iteration module executes a closed-loop optimization process; it continuously collects feedback data during user classification operations, including four core indicators: category jump path, mid-course exit node, number of repeated searches, and final target category location time; it defines mid-course exit as an operation where the user enters a category but does not click on the file and returns to the previous level within 15 seconds; it defines redundant intermediate levels as levels containing only one subcategory, or levels where multiple subcategories have a semantic similarity exceeding 0.9 and have no business differentiation significance; and it performs statistical analysis on all feedback data by user group every 7 days. When the dropout rate of a specific user group under a certain category path exceeds 31.5%, the category path optimization logic is automatically triggered, merging the corresponding redundant intermediate levels and moving the target category one level forward to shorten the user operation path. Every 27 days, the labeled user behavior samples are fed back to the user demand identification model, and the elastic weight consolidation algorithm is used to perform incremental learning to update the feature weights and judgment thresholds, retaining historical classification knowledge to avoid catastrophic forgetting. The classification view scheduling rules are iterated once every quarter based on full feedback data to adapt to the dynamic changes in user group characteristics and usage habits.

[0014] Furthermore, the category view dynamic scheduling module and the multimodal accessibility interaction module work together to execute accessibility control rules for classified categories. For visually impaired users, if a user does not have access to the corresponding classified files, the screen reader will automatically skip classified categories when broadcasting the category list, without indicating the existence of such categories, thus avoiding the leakage of classified category system information. If the user has access, the broadcast will be executed according to the anonymization rules corresponding to the classification level. For elderly users, all classified category entrances will be automatically hidden in the category view of users without access, and only public and internal categories will be displayed. Classified category identifiers for users with access will use a low-saturation style to avoid the page information being too eye-catching and causing misoperation. All classified file operations in accessibility mode will be automatically logged and synchronously written to the file security audit database to meet the compliance requirements of file confidentiality management.

[0015] On the other hand, a method for automatic management and classification of electronic archives is characterized by the following specific steps: S1. The interface for connecting to the archive collection performs format preprocessing on the multi-format electronic archives entering the database, extracts text semantic features and metadata features and completes vector fusion, inputs them into the pre-trained archive domain classification model to complete automatic classification indexing, constructs a multi-level tree-shaped archive classification category system, and writes the classification results into the archive metadata database simultaneously. S2. During the user's access to the category navigation page, the full amount of operation behavior data is collected seamlessly through front-end tracking. After anonymization, the data is input into a multi-feature weighted user demand recognition model to determine the user's accessibility demand type and output the corresponding user tag. For new users with no historical data, the default configuration is the standard interaction mode and data collection is started simultaneously. S3. Based on the determined user tags, automatically match the corresponding multimodal interaction mode, combine the file confidentiality level attribute to execute the corresponding desensitization interaction rules, and output the classified interaction channel adapted to different operation capabilities to the user. S4. Based on users' historical classification operation habits, accessibility needs, and file value attributes, dynamically adjust the display level, sorting priority, and interface presentation parameters of the classification categories to simplify the user operation path. S5. Continuously collect feedback data from the entire user classification process, and regularly use incremental learning to optimize the user identification model and classification view scheduling rules to continuously improve accessibility.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: (1) This invention achieves deep synergy between accessibility adaptation and archive confidentiality management. While providing differentiated classification and interactive services for visually impaired and elderly users, it avoids the risk of leakage of confidential archive information in the accessibility mode from the interaction layer through the desensitization broadcast of confidentiality level linkage and the category hiding mechanism of permission matching. It takes into account both the universality of public archive services and the compliance of confidentiality management, and solves the core pain point of traditional accessibility transformation that only focuses on interface experience and ignores data security.

[0017] (2) This invention constructs a complete optimization closed loop of user-unobtrusive recognition, dynamic view scheduling, and behavior feedback iteration. It can accurately match the adaptation scheme without the user having to manually switch the accessibility mode. At the same time, it continuously corrects the recognition model and classification view rules based on real operation data. The adaptation effect is dynamically optimized as the system runs, effectively reducing the category navigation loss rate and target file location time for special groups. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is an overall flowchart of an electronic archive information-based automatic management and classification system; Figure 2 A flowchart for barrier-free hierarchical control of classified categories in an electronic archive information automatic management and classification system; Figure 3 This is a flowchart of an automated classification method for electronic records information management. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1 : Online document retrieval scenarios for the public at municipal comprehensive archives: In the scenario of online document retrieval by the public at a municipal comprehensive archive, the archive's collection covers multiple categories of archives, including documents, public affairs, and classified information. The online platform receives an average of approximately 320,000 visits annually, with users aged 60 and above and visually impaired users accounting for about 18%. The collection also includes some archives classified as "secret" or higher, requiring a balance between providing accessible services and ensuring confidentiality compliance. This example fully illustrates the system's operation process based on this scenario.

[0022] The system is deployed on the archive's private cloud node, connecting to existing archive collection interfaces. Both existing and new archives are uniformly accessed and processed by the automatic archive classification engine. Upon receiving archives of various formats, the engine first performs format adaptation: PDF native documents have their text layer directly extracted; scanned documents have their text extracted via optical character recognition (OCR), with a single-field recognition confidence threshold set at 0.85. This value is determined according to industry quality control standards for archive digitization; fields below this threshold have an error rate exceeding 12%, and the cost of manual review is lower than the cost of correcting erroneous archives, therefore these are automatically marked as items awaiting review and transferred to manual auditing; audio and video oral archives are converted from speech to text to generate segmented text and marked with timestamps. After preprocessing, the engine extracts three types of metadata: document number, issuing unit, and date of issuance. These are mapped to a 128-dimensional feature vector using one-hot encoding, and then concatenated with the 768-dimensional text semantic vector output by the BERT-base model. This fusion is input into the classification model to output classification labels, constructing a three-level tree-structured category system, and the results are simultaneously written to the metadata database. The classification model was trained on 120,000 historical indexed archives of the museum. When adding new categories, the parameters of the underlying Transformer were fixed, and only the parameters of the fully connected output layer were fine-tuned. The learning rate was set to 2e-4, and the model was iterated for 45 rounds with an early stopping strategy. If the accuracy of the validation set did not improve for 5 consecutive rounds, the training was terminated. Category updates could be completed without full retraining.

[0023] When a user visits the category navigation page, the user behavior perception module seamlessly collects five operational data points—page zoom ratio, click interval, and keyboard navigation percentage—through front-end embedding. An internal multi-feature weighted model calculates the comprehensive accessibility requirement score S. The baseline weights are calculated based on 12,000 anonymous behavioral samples from the library over 12 months, divided into training and validation sets at a 7:3 ratio. Five-fold cross-validation was used to select the optimal indicator, with F1 score values ​​of 0.237, 0.214, 0.221, 0.186, and 0.139. Feature values ​​are mapped to the 0-100 range according to a unified rule. The model presets two threshold levels: S≥71.5 indicates a visually impaired user, and 47.5≤S<71.5 indicates an elderly user. These thresholds are determined based on user behavior distribution, balancing accuracy and false positive control. User tags are synchronized to the two downstream modules in real time.

[0024] The multimodal accessibility module matches the corresponding interaction mode. In the visually impaired user mode, the module performs hierarchical semantic translation of the category tree, converting it into a linear semantic sequence with hierarchical numbers. When there are multiple levels of nesting, the parent category target identification is added to avoid hierarchical confusion when reading the screen. It integrates a voice interaction engine, uses a built-in thesaurus of document domains for entity alignment, and calculates the similarity between the instructions and standard categories using edit distance. The threshold is set to 0.67 to balance recall and precision. Categories exceeding the threshold are sorted and fed back according to similarity. The broadcast is linked to the document classification: top secret and confidential documents only broadcast the document number and retention period, secret documents broadcast an anonymized topic overview, and non-confidential documents broadcast the complete metadata. In the elderly user mode, the category font is adjusted to 19 points, the line spacing is increased by 32%, the buttons are enlarged by 1.45 times and the hot zone extends by 10px, non-core functions are hidden, the color contrast is adjusted to 4.7:1, and the display fields of livelihood-related items are simplified.

[0025] The category view dynamic scheduling module incorporates a priority ranking model, calculating a priority score P based on four dimensions: access frequency, user matching degree, time-related decay, and file value. Weights are determined using the analytic hierarchy process (AHP), with a consistency test CR of 0.068 < 0.1, and values ​​of 0.392, 0.314, 0.183, and 0.109. The model sorts categories at the same level by score, with the top 12% included in the priority display area. Visually impaired users are shown only the first two levels of categories by default, with deeper levels automatically collapsed. Five high-frequency categories are configured with keyboard shortcuts. Elderly users have a dedicated quick category area at the top, extracting the five most accessed categories from the past 27 days and updating every 8 days. New users are loaded with five public service categories by default.

[0026] The interactive feedback iteration module continuously collects four types of feedback data. Mid-course exit is defined as returning without clicking on a file within 15 seconds of entering a category. Redundant levels are defined as levels containing only one subcategory or subcategories with semantic similarity exceeding 0.9 and no business distinction. Statistical analysis is performed every 7 days. Optimization is triggered when the single-path mid-course exit rate exceeds 31.2%, merging redundant levels. Every 27 days, an elastic weight consolidation algorithm is used to recycle incremental training, imposing constraints on historically high-importance weights to reduce the disturbance of historical knowledge by new sample updates, avoid catastrophic forgetting, and synchronously update weights and thresholds. The scheduling rules are iterated quarterly.

[0027] In summary, this embodiment fully covers the entire system module operation process, achieving a deep integration of automatic electronic archive classification and barrier-free adaptation. While ensuring the security of classified archives, it lowers the operational threshold for special groups and adapts to the dual service and management needs of public archives.

[0028] Example 2 : Scenario of internal archives management platform for provincial state-owned enterprises: In the context of internal archive management within provincial state-owned enterprises, corporate archives cover multiple categories including production, operations, personnel, and classified technologies. Access is restricted to internal employees, including current employees, retired employees, and employees with disabilities. Classified technical archives account for approximately 23% of the total, requiring strict confidentiality management. Simultaneously, the access rights of retired and disabled employees must be guaranteed. This example focuses on illustrating the system's classified information linkage control mechanism.

[0029] The system is deployed on the enterprise's intranet server and interfaces with the enterprise's document management system. After all documents are automatically categorized and indexed by the document classification engine, they are synchronously linked to the employee permission system and document security level labels. The user behavior awareness module collects employee operation data in the intranet environment, identifies accessibility requirements, and simultaneously links with the enterprise's unified identity authentication system to obtain employee permission levels.

[0030] The multimodal accessibility module and the dynamic scheduling module for the classification view work together to execute confidentiality control rules. For visually impaired employees, if an employee does not have access to the corresponding security level, the screen reader automatically skips confidential categories when displaying the category list, without indicating their existence, to avoid leaking the confidentiality category system. If the employee has access, the display follows the anonymization rules corresponding to the security level; top-secret files only display the file number, without showing any content information. For elderly retired employees, the classification view for accounts without access automatically hides all confidential category entries, displaying only public and internal administrative files. For accounts with access, confidential category icons use a low-saturation style to reduce visual attention and prevent accidental operation. All confidential file operations in accessibility mode are automatically logged, including information such as operator, operation time, accessed file number, and operation type, and are synchronously written to the enterprise file security audit database to meet confidentiality compliance audit requirements.

[0031] The interactive feedback iteration module optimizes the iteration cycle based on the characteristics of internal user groups. Employee user behavior samples are recycled every 20 days for incremental training to adapt to the operating habits of internal users. The mid-way exit rate threshold for confidential category paths has been adjusted to 28% to strictly control the user experience of confidential category navigation, reduce the invalid dwell time of users in confidential category areas, and reduce the risk of leakage.

[0032] In summary, this embodiment addresses the high confidentiality requirements of internal corporate archives by strengthening the confidentiality control mechanism in an accessible mode. It deeply integrates access verification, de-identified interaction, and audit trail recording, ensuring the rights of special employees to access archives while strictly implementing corporate archive confidentiality management regulations.

[0033] Example 3 : Scenario for document retrieval terminals in district and county government service centers: In the scenario of self-service inquiry terminals in district and county government service centers, the terminals are deployed next to the service windows in the government service hall, mainly providing file inquiry services to people seeking services offline. Users are primarily middle-aged and elderly people, most of whom are unfamiliar with operating smart devices. The terminals simultaneously handle the inquiry functions of publicly available government files and internal service files, requiring a balance between ease of use and data security. This embodiment focuses on explaining the system's age-friendly terminal adaptation logic.

[0034] The system is deployed locally on government service terminals and connects to the government archives collection interface. After the automatic archives classification engine completes the classification, the category system is restructured according to the theme of government services to adapt to the search habits of the public. The user behavior perception module adjusts the feature weights of the recognition model based on the characteristics of terminal touch operation, increasing the weight ratio of three features: click interval, page zoom, and operation backtracking, to adapt to the behavioral characteristics of terminal touch operation and improve the recognition accuracy for elderly users.

[0035] The multimodal accessibility module optimizes interaction logic for various terminal scenarios. In elderly user mode, large fonts and high-contrast display are enabled by default, and operation steps are simplified to three or fewer, with voice guidance prompts for each step. Query results automatically display the service process and required materials, while hiding specialized management fields. In visually impaired user mode, full-process voice interaction is supported, allowing for category navigation and file searches without touching the screen. The terminal has a built-in speaker and noise-canceling microphone to ensure accurate voice recognition even in noisy environments like lobby halls.

[0036] The category view dynamic scheduling module optimizes the quick category zone for terminal scenarios, defaulting to high-frequency service categories such as real estate inquiry, social security inquiry, and marriage registration. Administrators can adjust the zone content based on quarterly service hotspots. The category hierarchy is compressed to two levels by default, reducing the number of clicks for users and simplifying operations. The interaction feedback iteration module analyzes terminal operation data every 7 days, optimizing guidance prompts for category paths with high dropout rates and supplementing operation guidance voice to improve the success rate of operations for elderly users.

[0037] In summary, this embodiment optimizes the age-friendly interaction and view scheduling logic for offline scenarios of government self-service terminals, simplifies the operation path while retaining core functions, and effectively lowers the threshold for middle-aged and elderly people to use government archive query terminals.

[0038] Example 4 : System adaptation effect verification To verify the actual operational effectiveness of the system, a comparative test was conducted using a sample of real users from a municipal archives. The traditional manual accessibility mode switching archive classification system was used as a control group. The technical effectiveness of the system was verified from three dimensions: user identification, operational efficiency, and security and compliance.

[0039] The test selected 60 users, including 20 ordinary users, 20 elderly users, and 20 visually impaired users. Each group of users completed 10 designated file query tasks using both the control group system and this system, and relevant indicators were statistically analyzed. The test results showed that, in terms of user type identification, the overall accuracy rate of this system reached 89.2%, with no need for users to manually switch modes throughout the process; in the control group, the user-initiated activation rate of manually switching modes was only 47%, indicating that a large number of special users did not enjoy accessibility services. In terms of operational efficiency, the average time for visually impaired users to locate target categories was reduced from 42 seconds to 24 seconds, and the task operation success rate increased from 62% to 88%; the average operation time for elderly users was reduced from 37 seconds to 21 seconds, and the task operation success rate increased from 68% to 91%. In terms of security and compliance, the control group's accessibility mode posed a risk of full broadcasting of confidential information. This system, through a security-level linkage desensitization mechanism, did not exhibit any violations of confidential information display rules during the test, and all operations were auditable and traceable.

[0040] In summary, this system significantly outperforms traditional solutions in terms of recognition accuracy, operational efficiency, and security and compliance. It can effectively improve the record retrieval experience for special groups while ensuring the security of record data, and has clear practical application value.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An electronic archive information-based automatic management and classification system, characterized in that, The system includes the following components: Automatic Archive Classification Engine: Connects to the archive collection interface to complete the content parsing of multi-format electronic archives, deep integration of text semantic features and metadata features, and automatic classification indexing, and constructs a multi-level tree-shaped archive classification category system; User behavior perception module: Collects user operation data in the categorized navigation scenario through front-end embedding, anonymizes and desensitizes the collected behavior data, and determines the user's accessibility needs type based on a multi-feature weighted model and outputs the corresponding user tag; Multimodal accessibility interaction module: Matches differentiated interaction modes based on user tags, and executes hierarchical desensitization broadcast and display rules in conjunction with file confidentiality attributes; Category view dynamic scheduling module: Combines user tags, historical operation habits and file value attributes to dynamically adjust the category display level, sorting priority and interface layout; Interactive feedback iteration module: Collects user operation feedback data throughout the entire process, and uses incremental learning to optimize the user behavior perception model and view scheduling rules respectively.

2. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The automatic archive classification engine uses a pre-trained model based on the BERT-base architecture to achieve classification. The specific steps are as follows: it connects to the archive collection interface to receive electronic archives in multiple formats, and performs format adaptation preprocessing on PDF documents, scanned images, and audio and video files respectively; it extracts text content from scanned documents through optical character recognition, and marks items with a single field recognition confidence score of less than 0.85 as items to be reviewed. Audio and video files are processed by speech-to-text conversion to extract segmented text and simultaneously marked with timestamps. Three types of metadata features—archive number, issuing unit, and date of creation—are extracted synchronously. These metadata features are then one-hot encoded and mapped to a 128-dimensional feature vector, which is then fused with a 768-dimensional text semantic vector using a first-to-last concatenation method. This fusion is input into a classification model, which outputs corresponding classification labels, constructing a three-level tree-structured classification system. The classification results are simultaneously written to the archival metadata database. The classification model is trained on 120,000 historically labeled archives. When adding new categories, the underlying Transformer network parameters are fixed, with only the fully connected parameters of the output layer fine-tuned. The learning rate is set to 2e-4, and training is iterated for 45 epochs with an early stopping strategy. Training is terminated if the accuracy on the validation set shows no improvement for five consecutive epochs.

3. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The user behavior perception module incorporates a multi-feature weighted recognition model, which calculates a comprehensive score for user accessibility needs using the following formula. : ,in A comprehensive score is given to users' accessibility needs; - The baseline weights for the five feature dimensions; This is a characteristic value representing the page scaling ratio; This refers to the single category click interval feature value; The feature value representing the proportion of keyboard navigation; The characteristic value representing the proportion of voice interaction; This represents the characteristic value of the proportion of backtracking along the classification path.

4. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The multimodal accessibility interaction module is designed for visually impaired users, employing a semantic interaction mechanism linked to different security levels. It performs hierarchical semantic translation on the nested structure of multi-level category trees, converting each level of category into a linear semantic sequence of Y categories at level X, with the current category being the Zth category, plus the category name. In multi-level nesting, a prefix is ​​added to the parent category. When connecting to mainstream screen readers, the module plays the commands sequentially in a linear sequence. It integrates a voice interaction engine to receive user natural language commands, uses a built-in thesaurus for entity alignment of commands, and calculates the similarity between the command text and the standard category name using edit distance. A similarity threshold of 0.67 is set, and categories exceeding this threshold are sorted in descending order of similarity before being presented to the user. The playback process is linked to the security level of the archives. For top-secret and confidential archives, only the document number and retention period are played, omitting the content summary and keywords. For secret archives, only a de-identified topic overview is played. For non-confidential archives, the core metadata is played in full. Users can complete the entire process of category navigation, archive filtering, and result browsing via voice commands. The voice playback allows users to interrupt at any time and issue new commands.

5. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The multimodal accessibility module is configured to be age-friendly in both interface and content for elderly users; it automatically adjusts the basic presentation parameters of the category pages; it hides non-core function entrances such as advanced filtering, batch export, and category customization, retaining only the three core functions: category navigation, keyword search, and file preview; a one-click return-to-homepage button is fixed on the right side of the page, allowing users to return to the primary category page with a single click; all operation buttons are accompanied by clear text descriptions; the color contrast between page text and background is adjusted to 4.7:1; and the display fields for livelihood-related file entries are automatically simplified, showing only the file name and the file retrieval process.

6. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The category view dynamic scheduling module has a built-in category priority ranking model, which calculates the display priority score of a single category using the following formula. : ,in Display priority scores for each category; , , , These are the weighting coefficients for the four evaluation dimensions; This refers to the category access frequency characteristic value; For user group matching feature values; This is the time-related decay coefficient; This is the archival value coefficient.

7. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The category view dynamic scheduling module executes differentiated hierarchical scheduling rules for different user groups. For visually impaired users, only the first two levels of categories of the classification system are displayed by default. When the classification system exceeds three levels, the deep subcategories are automatically collapsed. After the user triggers the expansion operation through voice or keyboard commands, the next level of category data is loaded. All categories at the same level are arranged from high to low priority scores, and high-frequency popular categories are placed at the front. The category page provides dedicated keyboard shortcuts for five frequently used public service categories, allowing users to customize and modify the shortcut mappings according to their own usage habits. When a user enters a new category level, the scheduling module triggers the multimodal accessibility interaction module to execute the level broadcast; for elderly users, a quick category zone is displayed at the top of the category page, and the 5 categories with the highest access frequency in the past 27 days are automatically extracted and included in the zone. The quick category collection is recalculated and updated every 8 days; a cold start strategy is adopted for new users with no historical operation data, and five general high-frequency livelihood categories such as social security inquiry, marriage record, real estate registration, academic qualification certification and policy documents are included in the special zone by default; Administrators can manually adjust the default configuration and quantity limit of quick categories based on the business type of their organization.

8. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The interactive feedback iteration module executes a closed-loop optimization process; it continuously collects feedback data during user classification operations, including four core indicators: category jump path, mid-way exit node, number of repeated searches, and final target category location time; it defines mid-way exit as an operation where the user enters a category but does not click on the file and returns to the previous level within 15 seconds; it defines redundant intermediate levels as levels containing only one sub-category, or levels where the semantic similarity of multiple sub-categories exceeds 0.9 and has no business differentiation significance; every 7 days, it performs statistical analysis on all feedback data by user group; when the mid-way exit rate of a specific user group under a certain category path exceeds 31.5%, it automatically triggers the category path optimization logic, merges the corresponding redundant intermediate levels, moves the target category one level forward, and shortens the user operation path; every 27 days, it feeds back the labeled user behavior samples to the user demand identification model, and uses an elastic weight consolidation algorithm to perform incremental learning to update feature weights and judgment thresholds; The categorized view scheduling rules are iterated once per quarter based on full feedback data.

9. The electronic archives information-based automatic management and classification system according to claim 1, characterized in that, The dynamic scheduling module of the classification view works in conjunction with the multimodal accessibility interaction module to execute accessibility control rules for classified categories. For visually impaired users, if a user does not have access to the corresponding classified files, the screen reader will automatically skip classified categories when reading the category list and will not indicate the existence of such categories. If the user has access, the reading will be performed according to the anonymization rules corresponding to the classification level. For elderly users, all classified category entries will be automatically hidden in the classification view of users without access, and only public and internal categories will be displayed. Classified category identifiers for users with access will use a low-saturation style. All classified file operations in accessibility mode will be automatically logged and synchronously written to the file security audit database.

10. A method for automatic management and classification of electronic archives, applicable to the automatic management and classification system for electronic archives as described in any one of claims 1-9, characterized in that, The specific steps of this method are as follows: S1. The interface for connecting to the archive collection performs format preprocessing on the multi-format electronic archives entering the database, extracts text semantic features and metadata features and completes vector fusion, inputs them into the pre-trained archive domain classification model to complete automatic classification indexing, constructs a multi-level tree-shaped archive classification category system, and writes the classification results into the archive metadata database simultaneously. S2. During the user's access to the category navigation page, the full amount of operation behavior data is collected seamlessly through front-end tracking. After anonymization, the data is input into a multi-feature weighted user demand recognition model to determine the user's accessibility demand type and output the corresponding user tag. For new users with no historical data, the default configuration is the standard interaction mode and data collection is started simultaneously. S3. Based on the determined user tags, automatically match the corresponding multimodal interaction mode, combine the file confidentiality level attribute to execute the corresponding desensitization interaction rules, and output the classified interaction channel adapted to different operation capabilities to the user. S4. Based on users' historical classification operation habits, accessibility needs, and file value attributes, dynamically adjust the display level, sorting priority, and interface presentation parameters of the classification categories to simplify the user operation path. S5. Continuously collect feedback data from the entire user classification process and periodically use incremental learning to optimize the user identification model and classification view scheduling rules.

Images