An archive content intelligent evaluation system based on deep learning

Abstract

The application discloses an intelligent archive content evaluation system based on deep learning, and relates to the technical field of deep learning, which comprises inputting a standardized feature vector into a deep learning model, combining a rule matching benchmark to calculate a time sequence waveform factor and a semantic kurtosis factor, generating an evaluation confidence matrix, containing a cosine similarity and a Spearman correlation coefficient, comparing an index threshold to trigger an abnormal mark and a graded early warning. The application can comprehensively evaluate the archive content from multiple dimensions such as semantic understanding, risk evaluation and value classification by constructing a deep learning evaluation model of multi-model fusion, including a semantic understanding model, a risk evaluation model and a value classification model, and dynamically optimizing a split attribute according to an evaluation rule decision tree, so that an accurate comprehensive evaluation result and a rule matching benchmark are generated, and the comprehensiveness and scientificity of the archive content evaluation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and particularly relates to an archive content intelligent evaluation system based on deep learning. BACKGROUND

[0002] With the acceleration of the informationization process, the archive data generated by various institutions has shown an explosive growth, and the content form is becoming increasingly complex, covering text, tables, images, multimedia and other formats. At the same time, the demand for archive management has gradually shifted from traditional storage and retrieval to dynamic evaluation of content quality, security and utilization value.

[0003] Currently, natural language processing and machine learning techniques have been gradually introduced into the field of archive management, such as classification systems based on keyword matching, risk warning tools based on statistical models, etc. Some research attempts to use deep learning models, such as convolutional neural networks and recurrent neural networks, to perform semantic analysis on archive texts, or to construct a special data set through clustering algorithms. However, existing technologies are mostly limited to single tasks, lack multi-dimensional evaluation capabilities, and model training relies on static rules, making it difficult to dynamically optimize evaluation indicators. In addition, existing systems still have obvious shortcomings in complex semantic understanding, multi-source feature fusion and cross-domain indicator supplementation, resulting in insufficient comprehensiveness and credibility of the evaluation results.

[0004] The main problems of the prior art are reflected in the following aspects: single evaluation dimension, unable to comprehensively measure the integrity, security and utilization value of the archives; poor dynamic adaptability, traditional rule engines are difficult to dynamically adjust the evaluation rules according to the archive type and security level, and lack effective supplement mechanism when the index coverage is insufficient; limited intelligence level, insufficient depth of semantic understanding of unstructured text, risk assessment relies on artificial preset sensitive word library, difficult to capture the potential correlation of time sequence behavior patterns and permission violations; low warning accuracy, abnormal detection is mostly based on static threshold, lack of dynamic confidence analysis combining semantic features and time sequence features

[0005] Therefore, it is necessary to invent an archive content intelligent evaluation system based on deep learning to solve the above problems. SUMMARY

[0006] The purpose of the present application is to provide an archive content intelligent evaluation system based on deep learning to solve the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides the following technical solution: an archive content intelligent evaluation system based on deep learning, comprising the following modules:

[0008] Archive data collection module: obtain archive basic information and content data through API interface, identify archive type and storage format, and construct archive content raw data set containing original content, type label and storage format;

[0009] Archive content preprocessing module: text standardization processing and data cleaning based on density clustering are performed on the archive content raw data set, and word vector, time series feature and semantic distribution feature are extracted to generate standardized archive content feature vector and output thematic data set label;

[0010] Model construction module: used for constructing a deep learning evaluation model of multi-model fusion, training semantic understanding model, risk assessment model and value classification model based on standardized archive content feature vector, wherein the outputs of the semantic understanding model, the risk assessment model and the value classification model are fused by weighting to generate a comprehensive evaluation result, and an evaluation rule decision tree with archive type and secret level as root nodes is generated by dynamically optimizing split attributes through C4.5 algorithm to generate rule matching benchmark;

[0011] Archive index setting module: based on the evaluation rule decision tree, an evaluation index system including integrity, security and utilization value is established, and when the index coverage rate is lower than the preset threshold, adjacent archive type index is called and weighted supplemented through BERT semantic similarity retrieval;

[0012] Archive content intelligent evaluation module: the standardized feature vector is input into the deep learning model, the timing waveform factor and the semantic kurtosis factor are calculated by combining the rule matching benchmark, the evaluation confidence matrix is generated, and the cosine similarity and the Spearman correlation coefficient are included, the index threshold is compared to trigger abnormal marking and grading early warning;

[0013] Evaluation result output module: according to the abnormal detection result, a structured evaluation report is generated through a pre-defined report template, and risk item priority, value grade and repair suggestion are marked.

[0014] Preferably, the archive content preprocessing module comprises:

[0015] Text processing unit: construct an archive field text library, and perform word segmentation, part-of-speech tagging and named entity recognition on archive text by using natural language processing technology to generate word vector sequence with entity label;

[0016] Data cleaning unit: filter out repeated invalid data through word vector cosine similarity matching, and generate thematic data set according to archive type and theme through DBSCAN density clustering algorithm;

[0017] Feature fusion unit: extract timestamp, secret level identifier and version information metadata to generate time series feature and semantic distribution feature, and map them with word vector to form multi-dimensional feature vector.

[0018] Preferably, the establishment process of the special topic dataset comprises:

[0019] Extracting the agency name, business terminology and event keywords in the archive content, calculating the cosine similarity of the word vector, and applying to the archive field scene, when the similarity is greater than the preset threshold, it is classified into the same special topic dataset;

[0020] According to the archive formation time, storage period and business category, the special topic dataset is time series labeled and weight allocated.

[0021] Preferably, in the model construction module:

[0022] The semantic understanding model adopts the Transformer architecture, and generates a semantic similarity matrix and a topic distribution vector through a multi-head attention mechanism after inputting the standardized archive content feature vector;

[0023] The risk assessment model adopts a long short-term memory network LSTM combined with a conditional random field CRF, inputs the sensitive word frequency statistics, access log sequence analysis and modification trace comparison, and outputs the leakage risk probability value, the tampering risk probability value and the permission violation risk probability value;

[0024] The value classification model adopts a multi-layer perception machine MLP, inputs the utilization frequency, historical evaluation results and industry standards, optimizes the weight matrix through a back propagation algorithm, and generates a value classification label and a scoring rule.

[0025] Preferably, the construction process of the evaluation rule decision tree comprises:

[0026] Taking the archive type and the secret level as the root node, using the C4.5 algorithm combined with the information gain ratio of the time sequence waveform factor and the semantic kurtosis factor, dynamically optimizing the split attribute;

[0027] According to the descending order of the information gain ratio, the intermediate nodes and the leaf nodes are divided, when the information gain of the leaf node is less than the preset threshold, the splitting is stopped, and when the number of samples is insufficient, the splitting is also stopped;

[0028] The time sequence waveform factor and the semantic kurtosis factor output by the leaf node are standardized, and the mean and the standard deviation are calculated as the rule matching reference.

[0029] Preferably, the implementation mode of the archive index setting module comprises:

[0030] The archive content integrity index is extracted from the output result of the semantic understanding model, including the field missing rate and the content consistency score;

[0031] The security index is extracted from the output result of the risk assessment model, including the sensitive information leakage risk value and the access permission compliance score;

[0032] The utilization value index is extracted from the output result of the value classification model, including the academic reference value, the business guidance value and the historical certificate value score;

[0033] When the index coverage is lower than the threshold value, the semantic similarity of adjacent file types is calculated through the BERT model, and the weighted supplementary index is generated according to the similarity weight, wherein the weight distribution rule is that the supplementary weight increases by 10% for every 0.1 increase in similarity.

[0034] Preferably, the file content intelligent evaluation module comprises:

[0035] The decision factor calculation unit calculates the time sequence waveform factor composed of the content update frequency, the access frequency and the associated business frequency based on the time sequence characteristics, and calculates the semantic kurtosis factor composed of the keyword concentration, the semantic sentiment intensity and the theme consistency based on the keyword distribution;

[0036] The intelligent reasoning unit generates the evaluation confidence matrix containing the cosine similarity and the Spearman correlation coefficient by evaluating the decision tree for rule matching;

[0037] The abnormality detection unit compares the index threshold value, and when the difference degree exceeds the threshold value, an early warning signal is triggered according to the risk level, and the risk level includes high, medium and low.

[0038] Preferably, the evaluation result output module comprises a repair suggestion generation unit, which is used for triggering the following operations according to the risk level:

[0039] The first level risk: a forced encryption storage instruction and a permission limitation instruction are generated;

[0040] The second level risk: a content check instruction and a log audit instruction are generated;

[0041] The third level risk: a regular review marking instruction is generated.

[0042] The technical effects and advantages of the present application are as follows:

[0043] 1、The present application can comprehensively collect data, standardize processing and feature extraction of file content through the file data collection module, file content preprocessing module and the like, generate multi-dimensional feature vectors, and provide high-quality data basis for subsequent evaluation model training and intelligent evaluation, thereby effectively improving the efficiency and accuracy of file content processing;

[0044] 2、The application can comprehensively evaluate the archive content from multiple dimensions such as semantic understanding, risk assessment and value classification, generate accurate comprehensive evaluation results and rule matching benchmarks, and improve the comprehensiveness and scientificity of the archive content evaluation, by constructing a multi-model fusion deep learning evaluation model including a semantic understanding model, a risk assessment model and a value classification model, and dynamically optimizing split attributes in combination with evaluation rule decision trees.

[0045] 3、The application can realize dynamic and intelligent evaluation of archive content, timely discover abnormalities and provide corresponding repair suggestions, and enhance the practicability and reliability of the system, by establishing an evaluation index system including integrity, security and utilization value through the archive index setting module, supplementing by BERT semantic similarity retrieval when the index coverage is insufficient, and calculating timing waveform factors and semantic kurtosis factors, generating an evaluation confidence matrix and triggering abnormal marking and hierarchical early warning through the archive content intelligent evaluation module. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a schematic diagram of the overall structure of the application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the application.

[0048] The application provides a deep learning-based archive content intelligent evaluation system as shown in Figure 1 The system comprises the following modules:

[0049] The archive data acquisition module: acquires archive basic information and content data through an API interface, identifies archive types and storage formats, and constructs an archive content raw data set including raw content, type labels and storage formats;

[0050] It should be noted that the API interface supports multiple data source access, is compatible with different system data format standards, and ensures the universality and compatibility of data acquisition; the basic information includes archive number, formation time, security level, storage period, affiliated organization and responsible person; the content data includes text content, tables, attachments, pictures and multimedia files; and the raw content is the actual content data of the archive;

[0051] An archive content preprocessing module: text standardization processing and data cleaning based on density clustering are performed on the archive content original data set, and a vocabulary vector, a time series feature and a semantic distribution feature are extracted to generate a standardized archive content feature vector, and an output special topic data set label;

[0052] Further, in the above technical solution, the archive content preprocessing module comprises:

[0053] A text processing unit: an archive field text library is constructed, natural language processing technology is used to perform word segmentation, part-of-speech tagging and named entity recognition on the archive text, and a word vector sequence with entity labels is generated;

[0054] A data cleaning unit: repeated invalid data is filtered through cosine similarity matching of the word vector, and a special topic data set is generated according to the archive type and theme through the DBSCAN density clustering algorithm;

[0055] A feature fusion unit: timestamp, classification mark and version information metadata are extracted to generate time series features and semantic distribution features, and a multi-dimensional feature vector is formed by mapping with the word vector.

[0056] Further, in the above technical solution, the establishment process of the special topic data set comprises:

[0057] Extracting the names of institutions, business terms and event keywords in the archive content, calculating the cosine similarity of the word vector, and for the archive field application scenario, when the similarity is greater than a preset threshold, it is classified into the same special topic data set;

[0058] According to the archive formation time, the storage period and the business category, the special topic data set is time series labeled and weight allocated.

[0059] It should be noted that the core task of the text processing unit is to convert unstructured archive text into a machine-understandable semantic vector sequence. The processing steps are as follows: first, collect archive field specific vocabulary such as "classification", "storage period", "volume establishment unit", etc. to form an archive field text library to ensure the accuracy of word segmentation and semantic analysis. Combined with natural language processing technology, continuous text is first divided into independent words, such as "XXXX year financial report" into "XXXX year", "finance" and "report". Then, each word is labeled with a part of speech, such as noun, verb, time word, etc. For example, "archive" is labeled as a verb, and "XXXX year" is labeled as a time word. Then, proper nouns are recognized, such as the name of the institution "XX Archives Bureau", the name "AA", the time "XXXX year X month", etc. A word vector sequence with entity labels is generated, such as [(“XX Archives Bureau”, institution), (“AA”, person name)], and finally the word vector sequence with entity labels is output;

[0060] The core task of the data cleaning unit is to filter invalid data, cluster by file type and topic, form a special data set, and the processing steps are to calculate the cosine similarity of the word vector, judge the degree of repetition of the file content, if the similarity ≥ preset threshold 0.8, the preset threshold is set according to the actual situation, it is considered as repeated data and eliminated, such as two highly similar meeting records only one is kept; Based on the word vector sequence, combined with the file type, such as "personnel files", "financial files" and the topic keywords, such as "employment", "reimbursement", "audit", as clustering features, and then according to the DBSCAN density clustering algorithm, similar file contents are clustered into the same topic, such as all financial files containing "reimbursement process" and "travel expenses" keywords are classified into "travel reimbursement special topic", and finally the special data set divided by type and topic is attached with labels, such as "personnel-employment files", "financial-reimbursement vouchers";

[0061] The core task of the feature fusion unit is to integrate text semantic features, time attributes and metadata to generate standardized feature vectors, and the processing steps are to extract timestamps from the file, such as forming the time "XXXX-XX-XX", secret level identification, such as "secret", "confidential", "top secret", version information, such as "V1.0", "revised edition" and other metadata, based on the timestamp to calculate the file formation time, update frequency, etc., such as "storage period 10 years", "updated 3 times in the past year", and analyze the distribution density of keywords in the text, such as "sensitive word appearance frequency", "theme word concentration", then map the word vector sequence with time sequence features and semantic distribution features to form a multi-dimensional feature vector containing text semantics, time attributes, secret levels and other information, such as a 100-dimensional vector, the first 50 dimensions are word semantic features, the middle 30 dimensions are time and secret level features, and the last 20 dimensions are semantic distribution features, and finally output the standardized file content feature vector, such as [0.8, 0.3, 0.9,...] and the special data set;

[0062] The specific construction process of the special topic data set is: first, semantic clustering based on keywords and cosine similarity, first keyword extraction, using named entity recognition technology to extract agency names such as "XX City Archives Bureau" and "Finance Department", business terms such as "secret level identification", "archiving", "storage period", event keywords such as "audit rectification", "contract renewal", and "personnel transfer", then perform word vector conversion, input the keywords into a pre-trained word embedding model such as Word2Vec or BERT, generate high-dimensional semantic vectors such as 100-dimensional vectors, the vector values reflect the semantic features of the words, and finally perform cosine similarity calculation and clustering. The cosine similarity of different archive keyword vectors is calculated, and the threshold is set to 0.8. When the average similarity of the keyword vectors of two archives is greater than or equal to 0.8, they are classified into the same special topic data set; the second step is time series labeling and weight distribution. Time series labeling is to extract the generation time of the archives, divide the time interval by year and quarter, such as "XXXXQ1" and "XXXXQ4", where XXXX is the specific year, Q1 is the spring quarter, and Q4 is the winter quarter. According to the type of the archives, the storage period is labeled, such as "permanent", "10 years" and "short term", which are mapped to numerical values, such as permanent mapped to value 5, 10 years mapped to value 3, and short term mapped to value 2. According to the classification of the archives, such as "finance", "personnel", "administration" and "infrastructure", the category labels are generated, such as "C01-finance" and "C02-personnel"; the rules for weight distribution are: recent archives have higher weights; permanent archives have higher weights than short-term archives; core business archives have higher weights, which are set according to the business priority of the agency; the weights of the three are summed, such as time weight x 0.5 + storage period x 0.3 + business category x 0.2, to generate the weight of each archive in the special topic data set. The final special topic data set structure is shown in the table:

[0063] File ID Key word set Forming time Storage period Business category Synthetic weight D001 ["XXXX year financial audit",...] XXXX-XX-XX 10 years Financial-audit 0.3

[0064] Model construction module: for constructing a deep learning evaluation model for multi-model fusion, based on standardized archive content feature vectors to train semantic understanding models, risk assessment models and value classification models, wherein the outputs of the semantic understanding models, risk assessment models and value classification models are combined by weighted fusion to generate a comprehensive evaluation result, and an evaluation rule decision tree with archive type and secret level as root nodes is generated by dynamically optimizing the split attribute through C4.5 algorithm to generate rule matching benchmarks;

[0065] Further, in the above technical solution, the model construction module comprises:

[0066] The semantic understanding model adopts a Transformer architecture, which generates a semantic similarity matrix and a topic distribution vector through a multi-head attention mechanism after inputting the standardized archive content feature vector;

[0067] The risk assessment model adopts a long short-term memory network (LSTM) combined with a conditional random field (CRF), inputs sensitive word frequency statistics, access log sequence analysis and modification trace comparison, and outputs a leakage risk probability value, a tampering risk probability value and a permission violation risk probability value;

[0068] The value classification model adopts a multi-layer perceptron (MLP), inputs utilization frequency, historical evaluation results and industry standards, optimizes a weight matrix through a back propagation algorithm, and generates a value classification label and a scoring rule;

[0069] It should be known that an input layer of the semantic understanding model is to convert a word vector into a Token embedding, generate an input tensor in combination with position encoding, for example, encode “XXXX-XX-XX” into a position vector to reserve timing information; a multi-head attention layer is to independently calculate a Query, Key and Value matrix for each head; a feedforward neural network layer is to perform nonlinear transformation on multi-head attention output, use a GeLU activation function to enhance semantic representation capability; and an output layer is a semantic similarity matrix and a theme distribution vector, the semantic similarity matrix is to calculate a semantic cosine similarity of a current file and a historical file, and the theme distribution vector is to generate a theme probability distribution, for example, a probability of procurement is 0.7, a probability of finance is 0.2, and a probability of others is 0.1, which is used for subsequent input of a value classification model;

[0070] The risk assessment model inputs sensitive word frequency statistics, such as the frequency of sensitive words in the file, such as "ID card number" and "bank account number", access log sequence, such as the format "[timestamp] [user ID] [operation type, such as read, modify and delete] [permission level]", and modification trace comparison, such as the text difference segment before and after recording modification, modification user and time; the input access log is converted into a time sequence vector in the LSTM layer, and the timestamp is converted into a time interval feature, such as the number of hours from the last access; the network configuration is a two-layer bidirectional LSTM with a hidden layer dimension of 256, capturing the dependency relationship between time steps, such as the time sequence pattern of continuous multiple abnormal permission access; the output of each time step is a hidden state vector, representing the risk characteristics at that moment, such as the sensitive word change amplitude of a certain modification operation; the sequence labeling in the CRF layer is based on the hidden state output by the LSTM, calculating the transition probability of adjacent time step operations, such as whether the read to delete operation conforms to the permission rules, the loss function uses the CRF log likelihood loss, combined with the preset risk rules, such as confidential level files are prohibited from being modified by unauthorized users, to optimize the risk label sequence prediction; the output of the output layer is the leakage risk probability value, the tampering risk probability value and the permission violation risk probability value, which are used to trigger graded early warning; the leakage risk probability value is calculated according to the sensitive word propagation path, such as the probability of exporting through email attachments, ranging from 0 to 1; the tampering risk probability value is based on the semantic consistency of the modification trace, such as the probability of logical conflict between the modified content and the original text, and the permission violation risk probability value is compared with the matching degree of the operation permission and the file secret level, such as the probability of a normal user accessing a confidential level file;

[0071] The input in the value classification model utilizes frequency, historical evaluation results and industry standards. The frequency is the number of times the archives are searched and downloaded, which is a continuous numerical value. The historical evaluation results are past value classification labels such as high value, medium value, etc. and scores such as 1-5 points. The input layer is to perform one-hot encoding on discrete features such as historical evaluation labels, and to normalize continuous features such as utilization frequency to the [0, 1] interval. The hidden layer is a 2-layer fully connected layer. The first layer has 128 neurons, and the activation function is ReLU. The second layer has 64 neurons, and the activation function is Sigmoid. For example, the input "utilization frequency 0.8 + historical score 4 points + industry standard business guidance weight 0.4" is calculated through a weight matrix to obtain the initial score of each value dimension. The output layer outputs the value classification label and the scoring rule. The value classification label generates a multi-classification probability through a Softmax function, such as high value 0.6, medium value 0.3, and low value 0.1. The scoring rule is a linear combination of the scores of each dimension, such as academic reference (0.3 x 4) + business guidance (0.4 x 5) + historical evidence (0.3 x 3) = 4.1 points, with 1 decimal place. The final output format is a binary tuple, such as value classification label: high value, comprehensive score: 4.1, which is used for utilization value evaluation in the archive index setting module.

[0072] Further, in the above technical solution, the construction process of the evaluation rule decision tree includes:

[0073] Taking the archive type and the secret level as the root node, the C4.5 algorithm is used to dynamically optimize the split attribute by combining the information gain ratio of the time sequence waveform factor and the semantic kurtosis factor.

[0074] The intermediate nodes and the leaf nodes are divided in descending order of information gain ratio. When the information gain of the leaf node is less than a preset threshold, the splitting is stopped. When the number of samples is insufficient, the splitting is also stopped.

[0075] The time sequence waveform factor and the semantic kurtosis factor output by the leaf node are standardized, and the mean and the standard deviation are calculated as the rule matching reference.

[0076] It should be noted that the preset threshold is set based on industry conventional practice.

[0077] Archive index setting module: based on the evaluation rule decision tree, an evaluation index system including integrity, security and utilization value is established. When the index coverage rate is lower than a preset threshold, adjacent archive type indicators are retrieved and weighted to supplement.

[0078] Further, in the above technical solution, the implementation of the archive index setting module includes:

[0079] extracting an archive content integrity index from the semantic understanding model output result, including a field missing rate and a content consistency score;

[0080] extracting a security index from the risk assessment model output result, including a sensitive information leakage risk value and an access permission compliance score;

[0081] extracting a utilization value index from the value classification model output result, including an academic reference value, a business guidance value, and a historical evidence value score;

[0082] When the index coverage rate is lower than the threshold, the semantic similarity of adjacent archive types is calculated by a BERT model, and a weighted supplementary index is generated according to the similarity weight, wherein the weight allocation rule is that the supplementary weight increases by 10% for every 0.1 increase in similarity.

[0083] It should be noted that the calculation formula of the field missing rate is field missing rate = number of missing required fields ÷ total number of required fields; the content consistency score is calculated based on the multi-head attention mechanism of the Transformer model, a semantic similarity matrix of the archive text is generated, the cosine similarity average between paragraphs is calculated, and the theme consistency is checked in combination with the theme distribution vector to determine whether it is lower than the preset threshold, if it is lower, then it is deducted, and the output range is 0-100 points, the higher the score, the stronger the content consistency; the academic reference value score is calculated based on the number of times the archive is cited by an academic search system and the frequency of being referenced by research projects, and the output range is 0-5 points; the business guidance value score is calculated based on the access frequency of the archive by business departments, the number of downloads, and the number of times the associated business processes are triggered, and the output range is 0-5 points; the historical evidence value score is calculated based on the accuracy of the archive formation time, i.e. the integrity of the time stamp, and the non-tamperability of the content modification trace, and the output range is 0-5 points.

[0084] Archive content intelligent evaluation module: input the standardized feature vector into the deep learning model, calculate the time series waveform factor and the semantic kurtosis factor in combination with the rule matching benchmark, generate the evaluation confidence matrix, including the cosine similarity and the Spearman correlation coefficient, compare the index threshold to trigger the abnormal marker and the graded early warning;

[0085] Further, in the above technical solution, the archive content intelligent evaluation module comprises:

[0086] Decision factor calculation unit: calculate the time series waveform factor composed of content update frequency, access frequency, and associated business frequency based on time series characteristics, and calculate the semantic kurtosis factor composed of keyword concentration, semantic sentiment intensity, and theme consistency based on keyword distribution;

[0087] Intelligent inference unit: rule matching by evaluating decision tree, generating evaluation confidence matrix containing cosine similarity and Spearman correlation coefficient;

[0088] Abnormality detection unit: comparing index threshold, triggering early warning signal according to risk level when difference exceeds threshold, said risk level including high, medium and low;

[0089] It is known that the content update frequency is the number of modifications of the statistical archives within a set time period, such as monthly or quarterly, and the mean and standard deviation are calculated by sliding the time window; the access frequency is to analyze the access log and count the number of accesses within a unit time, such as daily or weekly, and the access frequency is weighted according to the access level; the associated business frequency is to analyze the association between the archive content and the business database through keyword matching, and the association is classified and counted according to the business category; the keyword concentration is to calculate the keyword weight based on the TF-IDF algorithm, and the cumulative proportion of the top N high-weight keywords is counted; the semantic sentiment intensity uses a pre-trained sentiment analysis model, such as BERT-Emotion, to perform sentiment scoring on the archive text, and outputs the probability value of positive, negative or neutral sentiment; the theme consistency is to extract the theme distribution of the archive content through the LDA theme model, and calculate the KL divergence between themes, if the divergence is lower than the threshold, such as 0.1, it is determined that the theme consistency is high;

[0090] The comparison method of the index threshold is to obtain the dynamic threshold of integrity, security and utilization value from the archive index setting module, such as integrity threshold = 0.9, security threshold = 0.85, and utilization value threshold = 0.8, and then calculate the difference degree with the threshold according to the comprehensive score in the confidence matrix, said difference degree is the standardized Z value;

[0091] Evaluation result output module: according to the abnormality detection result, generating a structured evaluation report through a pre-defined report template, marking the risk item priority, value level and repair suggestion;

[0092] Further, in the above technical solution, the evaluation result output module includes a repair suggestion generation unit for triggering the following operations according to the risk level:

[0093] First level risk: generating forced encryption storage instruction and permission restriction instruction;

[0094] Second level risk: generating content verification instruction and log audit instruction;

[0095] Third level risk: generating regular review mark instruction;

[0096] It is to be understood that the condition for the first level of risk is that the absolute value of the standardized Z value is greater than 1.5, the condition for the second level of risk is that the absolute value of the standardized Z value is between 0.5 and 1.5, and the condition for the third level of risk is that the absolute value of the standardized Z value is less than 0.5.

[0097] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will still be able to modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some of the technical features, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based intelligent evaluation system for archival content, characterized in that, Includes the following modules: Archive data acquisition module: Acquires basic information and content data of archives through API interface, identifies archive types and storage formats, and constructs original dataset of archive content including original content, type tags, and storage formats; The archival content preprocessing module performs text standardization and density-based clustering data cleaning on the original archival content dataset, extracts lexical vectors, time series features, and semantic distribution features, generates standardized archival content feature vectors, and outputs thematic dataset labels. Model building module: Used to build a deep learning evaluation model that integrates multiple models. It trains a semantic understanding model, a risk assessment model, and a value classification model based on standardized archival content feature vectors. The outputs of the semantic understanding model, the risk assessment model, and the value classification model are weighted and fused to generate a comprehensive evaluation result. The evaluation rule decision tree with archival type and security level as the root node is dynamically optimized for splitting attributes using the C4.5 algorithm to generate a rule matching benchmark. Archive indicator setting module: Based on the evaluation rule decision tree, an evaluation indicator system including integrity, security and utilization value is established. When the indicator coverage is lower than the preset threshold, the indicators of adjacent archive types are retrieved through BERT semantic similarity retrieval for weighted supplementation. The intelligent assessment module for archive content inputs standardized feature vectors into a deep learning model, combines rule matching benchmarks to calculate time-series waveform factors and semantic kurtosis factors, generates an assessment confidence matrix, including cosine similarity and Spearman correlation coefficient, and compares the threshold indicators to trigger anomaly marking and graded warnings. Assessment Result Output Module: Based on the anomaly detection results, a structured assessment report is generated using a predefined report template, which marks the priority, value level, and remediation recommendations of risk items.

2. The intelligent evaluation system for archival content based on deep learning according to claim 1, characterized in that, The archive content preprocessing module includes: Text processing unit: Constructs a text library for the archival domain, uses natural language processing technology to perform word segmentation, part-of-speech tagging, and named entity recognition on archival texts, and generates a sequence of lexical vectors with entity tags; Data cleaning unit: Filters duplicate and invalid data by cosine similarity matching of word vectors, and generates thematic datasets by archive type and topic using DBSCAN density clustering algorithm; Feature fusion unit: Extracts metadata such as timestamp, security level identifier and version information, generates time series features and semantic distribution features, and maps them with lexical vectors to form multi-dimensional feature vectors.

3. The intelligent evaluation system for archival content based on deep learning according to claim 2, characterized in that, The process of establishing the thematic dataset includes: Extract institution names, business terms, and event keywords from the archive content, calculate the cosine similarity of the word vectors, and classify them into the same thematic dataset according to the application scenarios in the archive field when the similarity is greater than a preset threshold. The thematic datasets are labeled with time series and weights according to the time of their creation, retention period, and business category.

4. The intelligent evaluation system for archival content based on deep learning according to claim 1, characterized in that, In the model building module: The semantic understanding model adopts the Transformer architecture. After inputting standardized archive content feature vectors, it generates a semantic similarity matrix and topic distribution vectors through a multi-head attention mechanism. The risk assessment model uses a Long Short-Term Memory (LSTM) network combined with a Conditional Random Field (CRF). It takes sensitive word frequency statistics, access log sequence analysis, and modification trace comparison as inputs and outputs the probability values ​​of leakage risk, tampering risk, and permission violation risk. The value classification model uses a multilayer perceptron (MLP) and takes utilization frequency, historical evaluation results, and industry standards as input. It optimizes the weight matrix through backpropagation algorithm to generate value classification labels and scoring rules.

5. The intelligent evaluation system for archival content based on deep learning according to claim 1, characterized in that, The process of constructing the evaluation rule decision tree includes: Using file type and security classification as the root node, the C4.5 algorithm is used to dynamically optimize the splitting attributes by combining the information gain ratio of temporal waveform factor and semantic kurtosis factor. The intermediate nodes and leaf nodes are divided in descending order of information gain ratio. Splitting stops when the information gain of a leaf node is less than a preset threshold, and also stops when the number of samples is insufficient. The temporal waveform factor and semantic kurtosis factor output by the leaf nodes are standardized, and the mean and standard deviation are calculated as the benchmark for rule matching.

6. The intelligent evaluation system for archival content based on deep learning according to claim 1, characterized in that, The implementation methods of the archive indicator setting module include: Extract archival content integrity metrics from the output of the semantic understanding model, including field missing rate and content consistency score; Security indicators are extracted from the output of the risk assessment model, including the risk value of sensitive information leakage and the compliance score of access permissions; Extract utilization value indicators from the output of the value classification model, including academic reference value, business guidance value, and historical voucher value scores; When the indicator coverage is below the threshold, the semantic similarity between adjacent file types is calculated using the BERT model, and a weighted supplementary indicator is generated according to the similarity weight. The weight allocation rule is: for every 0.1 increase in similarity, the supplementary weight increases by 10%.

7. The intelligent evaluation system for archival content based on deep learning according to claim 1, characterized in that, The intelligent assessment module for archive content includes: Decision factor calculation unit: Calculates time-series waveform factors composed of content update frequency, access frequency, and related business frequency based on time-series characteristics; calculates semantic kurtosis factors composed of keyword concentration, semantic sentiment intensity, and topic consistency based on keyword distribution. Intelligent reasoning unit: performs rule matching by evaluating decision trees and generates an evaluation confidence matrix that includes cosine similarity and Spearman correlation coefficient; Anomaly detection unit: When the difference exceeds the threshold, an early warning signal is triggered according to the risk level, which includes high, medium and low.

8. The intelligent evaluation system for archival content based on deep learning according to claim 1, characterized in that, The assessment result output module includes a repair suggestion generation unit, which is used to trigger the following operations based on the risk level: High-priority risk: Generating mandatory encrypted storage instructions and permission restriction instructions; Medium-priority risks: Generating content verification instructions and log auditing instructions; Low-priority risk: Generate periodic review flag instructions.

Images