Electronic document security management system based on dynamic sensitive word detection

Abstract

The application discloses an electronic document security management system based on dynamic sensitive word detection, and belongs to the technical field of computers. According to the technical scheme provided by the embodiment of the application, in response to a sensitive word detection request for a target electronic document, the document type, the document sensitivity, the detection requirement and the dynamic reference information of the target electronic document are determined, and the acquisition of multidimensional data is realized. The document type, the document sensitivity, the detection requirement and the dynamic reference information of the target electronic document are used to determine the target detection mode, that is, the target detection mode is determined by using the acquired multidimensional data, and the target detection mode is more matched with the target electronic document. The target electronic document is detected by using the target detection mode, and a more accurate sensitive word detection result can be obtained, so that more accurate and efficient electronic document security management is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to an electronic document security management system based on dynamic sensitive word detection. BACKGROUND

[0002] In the field of electronic document security management, the related technology mainly detects content through pre-defined static sensitive word library. Such systems usually use unified matching rules to scan documents, such as keyword exact matching or regular expression to identify sensitive content. However, the related technology has obvious limitations: first, the static word library is difficult to adapt to the semantic change of sensitive information, especially when facing the rapid evolution of network new words or industry terms, the system cannot update the detection rules in time, resulting in a significant increase in false negative rate; second, the existing system lacks the ability to understand the context semantics of the document, which is easy to misjudge normal expressions as sensitive content, causing frequent misjudgment; third, the traditional method cannot perform differentiated detection according to document type, sensitivity, and other characteristics, resulting in low detection efficiency. In addition, the existing system performs poorly in handling user personalized detection needs, making it difficult to meet the customization requirements in different scenarios. These defects seriously restrict the actual application effect of the electronic document security management system, and an technical solution that can dynamically adapt to document features and accurately identify sensitive content is urgently needed. SUMMARY

[0003] The embodiment of the present application provides an electronic document security management system based on dynamic sensitive word detection, which can improve the accuracy of sensitive word detection to realize more accurate and efficient electronic document security management. The technical solution is as follows:

[0004] On the one hand, an electronic document security management system based on dynamic sensitive word detection is provided, which includes a processor and a memory, and the processor is used to perform the following steps:

[0005] In response to a sensitive word detection request for a target electronic document, the document type, document sensitivity, detection requirement, and dynamic reference information of the target electronic document are determined, the detection requirement is carried by the sensitive word detection request, and the dynamic reference information is dynamic reference information used to assist in sensitive word detection;

[0006] Based on the document type, the document sensitivity, the detection requirement, and the dynamic reference information, a target detection method for sensitive word detection of the target electronic document is determined;

[0007] The target detection method is used to perform sensitive word detection on the target electronic document, and a sensitive word detection result of the target electronic document is obtained, which includes sensitive words in the target electronic document and the sensitivity of the sensitive words.

[0008] In a possible implementation, the determining, in response to a sensitive word detection request for a target electronic document, a document type, a document sensitivity degree, a detection requirement, and dynamic reference information of the target electronic document, includes:

[0009] In response to a sensitive word detection request for a target electronic document, the target electronic document and the detection requirement are obtained from the sensitive word detection request. The detection requirement is in a text form and is used to represent a detection requirement personalized by a target user, which is a sending user of the sensitive word detection request.

[0010] The target electronic document is subjected to document keyword extraction and abstract extraction, to obtain a plurality of document keywords and a document abstract of the target electronic document.

[0011] The target electronic document is classified based on the document abstract, to determine a document type of the target electronic document. The document type is used to indicate a corresponding field.

[0012] The document sensitivity degree and the dynamic reference information of the target electronic document are determined based on the plurality of document keywords, the document type, and the document abstract.

[0013] In a possible implementation, the classifying, based on the document abstract, the target electronic document to determine a document type of the target electronic document, the document type being used to indicate a corresponding field, includes:

[0014] Document description information of the target electronic document is obtained. The document description information includes a document source, a document purpose, and a document destination. Based on the document description information and the document abstract, a first initial document type and field reference information of the target electronic document are determined. The field reference information is used to assist in determining a field to which the target electronic document belongs. Based on the first initial document type of the target electronic document and the field reference information, the document type of the target electronic document is determined.

[0015] The determining, based on the plurality of document keywords, the document type, and the document abstract, the document sensitivity degree and the dynamic reference information of the target electronic document, includes:

[0016] The document sensitivity degree of the target electronic document is determined based on the plurality of document keywords and the document type. The dynamic reference information is determined based on the document abstract.

[0017] In a possible implementation, the determining, based on the document description information and the document abstract, the first initial document type and the field reference information of the target electronic document, includes:

[0018] determine a document usage scenario of the target electronic document based on the document description information; determine a first initial document type of the target electronic document based on the document usage scenario and the document abstract; and query a first reference information library based on the document abstract to obtain domain reference information of the target electronic document;

[0019] The method further includes determining a document type of the target electronic document based on the first initial document type of the target electronic document and the domain reference information.

[0020] The method further includes determining a plurality of candidate document types based on the domain reference information; determining a second initial document type from the plurality of candidate document types based on the document abstract of the target electronic document, the second initial document type being a candidate document type that has a highest matching degree with the document abstract; and combining the first initial document type and the second initial document type to obtain the document type of the target electronic document.

[0021] In a possible implementation, the method further includes determining a document sensitivity degree of the target electronic document based on the plurality of document keywords and the document type.

[0022] The method further includes determining a document keyword sensitivity score of each of the document keywords; fusing the document keyword sensitivity scores of the plurality of document keywords to obtain a document keyword sensitivity degree of the plurality of document keywords; determining a document type sensitivity degree corresponding to the document type; and fusing the document keyword sensitivity degree and the document type sensitivity degree to obtain the document sensitivity degree of the target electronic document.

[0023] The method further includes determining the dynamic reference information based on the document abstract.

[0024] The method further includes performing keyword extraction on the document abstract to obtain a plurality of abstract keywords; performing semantic feature extraction on the document abstract to obtain an abstract semantic feature of the document abstract; and performing information recall in a second reference information library based on the plurality of abstract keywords and the abstract semantic feature to obtain the dynamic reference information, the second reference information library storing dynamic reference information that is updated in real time.

[0025] In a possible implementation, the method further includes determining a target detection manner for sensitive word detection on the target electronic document based on the document type, the document sensitivity degree, the detection requirement, and the dynamic reference information.

[0026] The method further includes determining a sensitive word reference example based on the document type.

[0027] determine a target detection level and a key detection content based on the document sensitivity degree, the detection requirement, and the document type, the target detection level being used to represent a fine degree of sensitive word detection, and the key detection content being used to represent a focus of sensitive word detection;

[0028] determine a plurality of dynamic sensitive words based on the dynamic reference information;

[0029] determine the target detection manner based on the sensitive word reference example, field information of a target field, the target detection level, the key detection content, and the plurality of dynamic sensitive words, the target field being a field to which the target electronic document belongs.

[0030] In a possible implementation, the determining of the sensitive word reference example based on the document type comprises:

[0031] querying a sensitive word example library by using the document type to obtain the sensitive word reference example corresponding to the document type, the sensitive word reference example including a plurality of cases of being identified as a sensitive word;

[0032] The determining of the target detection level and the key detection content based on the document sensitivity degree, the detection requirement, and the document type comprises:

[0033] determining a target detection dimension and a target detection granularity based on the document sensitivity degree, determining a special detection rule based on the detection requirement, determining the target detection level based on the target detection dimension and the target detection granularity, and determining the key detection content based on the special detection rule and the document type;

[0034] The determining of the plurality of dynamic sensitive words based on the dynamic reference information comprises:

[0035] performing recall in a sensitive word library based on the dynamic reference information to obtain the plurality of dynamic sensitive words.

[0036] In a possible implementation, the determining of the target detection manner based on the sensitive word reference example and field information of a target field, the target detection level, the key detection content, and the plurality of dynamic sensitive words comprises:

[0037] generating a target sensitive word detection example based on the sensitive word reference example and the plurality of dynamic sensitive words;

[0038] determining an initial detection manner based on field information of the target field, the target detection level, and the key detection content;

[0039] Determine the target detection manner based on the target sensitive word detection example and the initial detection manner.

[0040] In a possible implementation, the generating of the target sensitive word detection example based on the sensitive word reference example and the plurality of dynamic sensitive words comprises:

[0041] filling the sensitive word reference example and the plurality of dynamic sensitive words into a first prompt text template to obtain a first prompt text, the first prompt text template being used to indicate that the sensitive word reference example is expanded by using the plurality of dynamic sensitive words; inputting the first prompt text into a target language model, encoding and decoding the first prompt text by using the target language model, and outputting the target sensitive word detection example;

[0042] The determining of the initial detection manner based on the field information of the target field, the target detection level, and the key detection content comprises:

[0043] determining a field detection parameter corresponding to the target field based on the field information of the target field, the field detection parameter being used to represent a field detection manner of the target field; and matching the field detection parameter, the target detection level, and the key detection content in a plurality of candidate detection manners to obtain the initial detection manner, the initial detection manner being a candidate detection manner that is successfully matched;

[0044] The determining of the target detection manner based on the target sensitive word detection example and the initial detection manner comprises:

[0045] filling the target sensitive word detection example and the initial detection manner into a second prompt text template to obtain a second prompt text, the second prompt text template being used to indicate that the initial detection manner is corrected by using the target sensitive word detection example; inputting the second prompt text into a target language model, encoding and decoding the second prompt text by using the target language model, and outputting the target detection manner.

[0046] In a possible implementation, the sensitive word detection of the target electronic document by using the target detection manner to obtain a sensitive word detection result of the target electronic document comprises:

[0047] detecting sensitive words in the target electronic document by using the target detection manner to obtain a plurality of candidate sensitive words in the target electronic document and a candidate sensitive word degree of each of the candidate sensitive words;

[0048] The domain information based on a target domain to which the target electronic document belongs is used to correct the plurality of candidate sensitive words and the candidate sensitive word degree of each candidate sensitive word, so as to obtain a sensitive word detection result of the target electronic document.

[0049] In one aspect, an electronic document security management device based on dynamic sensitive word detection is provided, and the device comprises:

[0050] A first determination module determines a document type, a document sensitive degree, a detection requirement, and dynamic reference information of a target electronic document in response to a sensitive word detection request of the target electronic document, the detection requirement is carried by the sensitive word detection request, and the dynamic reference information is dynamic reference information used to assist in sensitive word detection;

[0051] A second determination module determines a target detection mode for sensitive word detection of the target electronic document based on the document type, the document sensitive degree, the detection requirement, and the dynamic reference information;

[0052] A detection module performs sensitive word detection on the target electronic document by using the target detection mode, and obtains a sensitive word detection result of the target electronic document, the sensitive word detection result comprising sensitive words in the target electronic document and a sensitive word degree of the sensitive words.

[0053] Through the technical solutions provided by the embodiments of the present application, the document type, the document sensitive degree, the detection requirement, and the dynamic reference information of the target electronic document are determined in response to the sensitive word detection request of the target electronic document, and the acquisition of multi-dimensional data is realized. The target detection mode is determined by using the document type, the document sensitive degree, the detection requirement, and the dynamic reference information of the target electronic document, that is, the target detection mode is determined by using the acquired multi-dimensional data, and the target detection mode is more matched with the target electronic document. The sensitive word detection is performed on the target electronic document by using the target detection mode, and a more accurate sensitive word detection result can be obtained, so as to realize more accurate and efficient electronic document security management. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0055] Figure 1 is a schematic diagram of an implementation environment provided by the embodiments of the present application;

[0056] Figure 2is a flowchart for determining a sensitive word detection result provided by an embodiment of the present application.

[0057] Figure 3 is a flowchart for determining a document sensitivity and dynamic reference information provided by an embodiment of the present application.

[0058] Figure 4 is a flowchart for determining a target detection method provided by an embodiment of the present application.

[0059] Figure 5 is another flowchart for determining a sensitive word detection result provided by an embodiment of the present application.

[0060] Figure 6 is a structural schematic diagram of an electronic document security management device based on dynamic sensitive word detection provided by an embodiment of the present application.

[0061] Figure 7 is a structural schematic diagram of an electronic document security management system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0063] In the present application, the terms "first", "second", and the like are used to distinguish the same or similar items with basically the same function and purpose, and it should be understood that there is no logical or time sequence relationship between "first", "second", and "nth", and the number and execution order are not limited.

[0064] Artificial intelligence (AI) is the use of digital computers or machine controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain better results.

[0065] Machine learning (ML) is a multi-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge sub-models to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence.

[0066] Electronic document: An electronic document refers to a file or collection of information created, stored, transmitted and processed in digital form. It replaces traditional paper documents and is managed using computers and electronic devices.

[0067] Sensitive word detection: Sensitive word detection refers to the process of using computer technology (usually text analysis, natural language processing, pattern matching, etc.) to automatically identify and locate specific predefined words, phrases, or patterns in a given text content (such as a paragraph, an article, user comments, chat records, electronic documents, etc.). These defined words or patterns are usually considered sensitive or inappropriate.

[0068] In related technologies, electronic document security management mainly relies on pre-defined static sensitive word libraries for content detection, using uniform matching rules to scan documents. For example, sensitive content is identified through keyword exact matching or regular expression. However, static word libraries are difficult to adapt to semantically variable sensitive information, such as network neologisms or industry terminology evolution, leading to missed reports or misjudgments. For example, in medical industry documents, the update of professional terms may make the original word library unable to identify new sensitive expressions.

[0069] To solve the above problems, the inventors found that there is a contradiction between the fixedness of static word libraries and the dynamic characteristics of documents. Through analysis, it is found that document type differences, sensitivity fluctuations, and user personalized needs are not effectively integrated. Further thinking found that if the multi-dimensional attributes of documents and dynamic information are combined, a more adaptable detection model can be constructed. For example, different detection strategies are used for legal documents and social texts, forming a dynamic detection mechanism based on multi-dimensional parameter coordination.

[0070] Figure 1 is an implementation environment schematic diagram of the technical scheme provided by the embodiments of the present application, referring to Figure 1 , the implementation environment can include a terminal 110 and an electronic document security management system (hereinafter referred to as system) 140.

[0071] The terminal 110 is connected to the system 140 through a wireless network or a wired network. Optionally, the terminal 110 is a smartphone, a tablet computer, a notebook computer, a desktop computer, etc., but is not limited thereto. The terminal 110 installs and runs an application program supporting electronic document security management based on dynamic sensitive word detection.

[0072] The system 140 is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The system 140 provides background services for the application program running on the terminal 110.

[0073] The application provides an electronic document security management system based on dynamic sensitive word detection, comprising a processor and a memory, referring to Figure 2 The processor is used to execute the following steps:

[0074] 201. In response to a sensitive word detection request for a target electronic document, determine the document type, document sensitivity, detection requirement and dynamic reference information of the target electronic document;

[0075] 202. Based on the document type, document sensitivity, detection requirement and dynamic reference information of the target electronic document, determine the target detection method;

[0076] 203. Perform sensitive word detection using the target detection method, and output the sensitive word detection result containing sensitive words and their sensitivity.

[0077] The document type refers to the field classification to which the target electronic document belongs, which can be realized by document abstract classification or keyword clustering, and is used to adapt to detection standards in different fields. The document sensitivity refers to the probability evaluation of sensitive information involved in the document content, which can be specifically calculated by keyword sensitivity score weighting, and is used to dynamically adjust the detection intensity. The detection requirement refers to the detection rule customized by the user, which is extracted by analyzing the text description in the request, and is used to support personalized configuration. The dynamic reference information refers to the auxiliary detection data updated in real time, which is recalled from the external database through semantic feature matching, and is used to supplement the timeliness deficiency of the static word library.

[0078] Specifically, when receiving a sensitive word detection request, the system first extracts the abstract and keywords of the target electronic document, determines its type in combination with the document source, purpose and other information. For example, a medical report is classified as a health field, triggering the corresponding medical sensitive word library. Subsequently, the overall sensitivity of the document is calculated by analyzing the keyword sensitivity and the document type weight. For example, a document containing "gene editing" may be evaluated as a high sensitivity level. After analyzing the detection requirement, for example, the user requires to detect the privacy clause, the system fuses the latest regulation data in the dynamic reference information with the static word library to generate the target detection method containing the newly added sensitive words such as "biometric data". Finally, through multi-dimensional matching and semantic analysis, the sensitive words and their quantitative sensitive values (sensitivity) are output.

[0079] Compared with the related art, the traditional method only relies on fixed word library for mechanical matching, and cannot adjust the detection strategy according to the document characteristics and external changes. The present scheme dynamically fuses the document type, sensitivity, user requirement and real-time information, so that the detection method is automatically optimized according to the document properties and environmental changes.

[0080] By the technical solution, the application realizes fine sensitive word detection on electronic documents, and effectively solves the problem that a static word library cannot dynamically adapt to semantic changes. For example, when detecting social media documents, the system combines a network popular language dynamic library to identify emerging sensitive words; when detecting legal contracts, a professional term library is matched through domain type, and the sensitive degree of clauses is quantified. In this way, both missed detection caused by lagging word libraries and cross-domain misjudgment are avoided, and individualized detection requirements in different scenarios are met.

[0081] Referring to Figure 3 The processor can further perform the following steps.

[0082] 301. In response to a sensitive word detection request for a target electronic document, obtaining the target electronic document and detection requirements from the sensitive word detection request, the detection requirements being in text form and used to represent the detection requirements of a target user, the target user being a sending user of the sensitive word detection request;

[0083] 302. Performing document keyword extraction and abstract extraction on the target electronic document to obtain a plurality of document keywords and a document abstract of the electronic document;

[0084] 303. Classifying the target electronic document based on the document abstract to determine a document type of the target electronic document, the document type being used to indicate a domain corresponding to the target electronic document;

[0085] 304. Determining a document sensitive degree and dynamic reference information of the target electronic document based on the plurality of document keywords, the document type, and the document abstract.

[0086] Among them, the detection requirement in text form refers to the user's personalized detection requirement described in natural language. Specifically, it can be implemented by extracting keywords and semantic rules through text analysis technology, such as identifying the detection range, sensitive level and other elements in the user's requirement through a natural language processing model. Document keyword extraction refers to extracting representative words from electronic documents. Specifically, it can be implemented by using TF-IDF algorithm or sequence labeling model based on deep learning to capture the core content features of the document. Document summary extraction refers to generating a concise text reflecting the main idea of the document. Specifically, it can be implemented by using a text summary generation model, such as an abstract summary model based on Transformer, to preserve the semantic integrity of the document. Document type classification refers to classifying the document according to its content features. Specifically, it can be implemented by using a text classification model combined with a domain knowledge graph, such as identifying financial, medical and other vertical fields through a pre-trained classifier. Document sensitivity refers to the potential risk level of sensitive information involved in the document content. Specifically, it can be calculated by matching keywords and using a semantic analysis model to calculate a sensitive probability value, such as building a sensitivity scoring model based on logistic regression or neural network. Dynamic reference information refers to real-time updated auxiliary detection information. Specifically, it can be obtained by accessing external database interfaces to get the latest sensitive events, industry standards and other dynamic data, such as synchronizing hot information from public opinion monitoring systems through API interfaces.

[0087] Specifically, when receiving a sensitive word detection request, first parse the target electronic document and the text form detection requirement submitted by the user from the request, and extract the key elements in the requirement through natural language processing technology. Then, the target document is extracted for keywords and abstracts, where keyword extraction focuses on high-frequency core words, and abstract generation preserves the semantic context of the document. Based on the generated document abstract, use a text classification model combined with a domain knowledge base to identify the document type, such as determining whether it belongs to the legal or scientific field by matching professional terms in the abstract. After determining the document type, combine the extracted keyword set and domain features to calculate the overall sensitivity of the document through a sensitivity scoring model, and obtain real-time reference data related to the current document from external dynamic information sources. For example, when detecting medical field documents, dynamic reference information can include the latest drug safety warnings or medical regulation changes.

[0088] Compared with the related art, the traditional method relies on fixed sensitive word library for mechanical matching and cannot adjust the detection strategy according to the actual content characteristics of the document. The scheme dynamically extracts the keywords and abstracts of the target electronic document, combines real-time external information, and constructs a detection parameter system based on content semantics. For example, the related art may miss detection when processing emerging network language, while the scheme can incorporate newly emerging sensitive words in time through the real-time updating mechanism of the dynamic reference information library. At the same time, the detection strategy adjustment based on document type classification enables the financial field document to automatically enable the detection rules of sensitive words related to fund transactions, and the medical field document focuses on the identification of patient privacy related terms.

[0089] Through the above technical scheme, the present application realizes the dynamic adaptation of the detection parameters and the content characteristics of the document, and solves the problem that the static word library cannot adapt to individual needs and content changes. For example, when the user proposes the demand of focusing on detecting commercial secrets, the system can strengthen the detection weight of technical terms based on the keywords in the detection demand text, and at the same time, improve the sensitivity judgment threshold by combining the research and development information in the document abstract. For documents containing new industry terms, the latest sensitive words are supplemented through the dynamic reference information library to avoid missing detection due to lag in word library updating. In addition, the accurate classification of document types enables the detection process to call domain-specific detection models, such as automatically loading legal term feature libraries when detecting legal documents, to improve the accuracy of sensitive word recognition in specific fields.

[0090] The present application further proposes a technical scheme based on document description information and document abstract to determine the first initial document type and domain reference information of the target electronic document, based on the first initial document type and domain reference information to determine the document type, based on the document keywords and document type to determine the document sensitivity, and based on the document abstract to determine the dynamic reference information.

[0091] Among them, the document description information refers to the structured metadata containing the source of the document, the purpose of the document and the place of the document. Specifically, it can be implemented by extracting the metadata field of the electronic document using the document attribute parsing module, which is used to provide the context information of the document use scenario. The first initial document type refers to the classification result preliminarily determined by the document use scenario and the abstract content. Specifically, it can be implemented by using a rule-based classifier to establish the preliminary field association of the document. The domain reference information refers to the dynamic knowledge data that assists in determining the domain of the document. Specifically, it can be implemented by using the association query of the domain knowledge graph to correct the domain deviation of the initial classification result. The document sensitivity refers to the quantitative evaluation value of the sensitivity of the keywords and the sensitivity of the document type. Specifically, it can be implemented by using a weighted fusion algorithm to dynamically adjust the strictness of the detection strategy. The dynamic reference information refers to a real-time updated semantic sensitive feature set. Specifically, it can be implemented by using a natural language processing model to extract semantic features of the abstract to capture the semantic association of emerging sensitive words.

[0092] Specifically, when processing the target electronic document, first, the metadata thereof is parsed to obtain the document source, purpose and destination, which constitute the document description information. By analyzing the relevance of the document purpose and the summary content, a scene matching algorithm is used to generate a first initial document type. At the same time, the document summary is input into the domain knowledge base for semantic retrieval to recall the domain reference information related to the summary theme. In combination with the initial classification result and the domain knowledge, a multi-dimensional fusion algorithm is used to determine the final document type. In the sensitivity evaluation stage, the extracted document keywords are scored for sensitivity, and in combination with the preset sensitivity weight corresponding to the document type, a comprehensive sensitivity value is generated through weighted calculation. In the dynamic reference information acquisition process, the semantic vector features of the summary are extracted using a deep learning model, and in combination with the real-time updated sensitive word library, the semantic similarity is matched to recall the relevant dynamic sensitive feature set in the current environment.

[0093] Compared with related technologies, the traditional method only relies on a static keyword library for document classification and sensitivity judgment, and cannot adapt to network neologisms and semantic evolution. The present scheme solves the problem of missing context information in the document classification process by fusing structured metadata and semantic analysis technology to build a dynamically updated domain knowledge reference system. At the same time, a dual evaluation mechanism of keyword sensitivity and document type sensitivity is adopted to overcome the misjudgment problem caused by single-dimensional evaluation. The real-time recall mechanism of dynamic reference information breaks through the technical bottleneck of static word library update lag, and realizes the rapid response of the sensitive word detection system to semantic changes.

[0094] Through the above technical solutions, the present application can accurately identify the domain and sensitivity level of the document and dynamically adjust the detection strategy. In the financial field document processing scene, it can effectively identify the sensitive variants of industry terms; in the medical data management scene, it can timely capture newly emerging privacy protection keywords. By establishing the association mechanism of document attributes, semantic features and dynamic knowledge base, the adaptability of the sensitive word detection system to complex semantic environment is significantly improved.

[0095] The present application further proposes determining a document use scene of the target electronic document based on the document description information, determining a first initial document type based on the document use scene and the document summary, querying domain reference information based on the document summary in a first reference information library, determining a plurality of candidate document types based on the domain reference information, determining a second initial document type from the candidate document types based on the document summary, and combining the first initial document type and the second initial document type to obtain the document type of the target electronic document.

[0096] The document description information refers to metadata including a document source, a use, and a destination, and can be extracted from a document attribute field by using a natural language processing technology, and is used to capture a scene feature of the document. The document use scene refers to an actual application environment of the document in a business process, and can be determined by analyzing a document use field in combination with a semantic reasoning model, and is used to establish an association between the scene feature and the document type. The first initial document type refers to a preliminary classification result generated based on the scene feature, and can be output after the scene classification model is used to perform pattern matching on the document use scene, and is used to implement type determination in the scene dimension. The domain reference information refers to a domain classification standard related to the document summary, and can be obtained by vectorization retrieval in an industry classification knowledge base containing real-time updated industry classification knowledge, and is used to introduce external dynamic classification basis. The candidate document type refers to a set of potential document categories conforming to the domain feature, and can be generated by performing semantic grouping on the domain reference information by using a clustering algorithm, and is used to narrow the classification range. The second initial document type refers to a candidate type with the highest matching degree with the document summary, and can be determined by calculating a cosine similarity between the document summary and a description text of the candidate type, and is used to implement deep alignment of the content feature and the domain feature.

[0097] Specifically, the source and use fields in the document description information are parsed and input into the scene classification model to generate vector data representing the application scene, and joint reasoning is performed in combination with the semantic feature of the document summary to output the first initial document type. Meanwhile, the document summary is converted into a vector form and subjected to a nearest neighbor search in a dynamically updated industry classification knowledge base to recall a domain classification standard most relevant to the current semantics as the domain reference information. After the candidate document type set is selected based on the domain reference information, a semantic similarity algorithm is used to calculate the matching degree of the document summary with each candidate type, and the type with the highest matching degree is selected as the second initial document type. Finally, the first initial type generated based on the scene feature and the second initial type generated based on the content feature are logically combined to form a document type determination result integrating the three-party information of the scene, the content, and the domain.

[0098] Compared with related technologies, the traditional method only relies on the document content or static classification rules for type determination, and is prone to misjudgment due to semantic ambiguity or lagging classification standards. The scheme fuses the scene feature, dynamic domain knowledge, and content semantic matching to construct a multi-dimensional classification decision mechanism, effectively solving the limitations of single-dimensional classification.

[0099] Through the above technical solutions, the present application realizes accurate classification of the document type, dynamically integrates the scene feature and real-time data of the domain knowledge base, and significantly improves the degree of fit between the classification result and the current industry standard. At the same time, the dual initial type combination mechanism effectively balances the weight distribution of the scene feature and the content feature, avoids classification errors caused by single feature deviation, and thus enhances the adaptability of subsequent sensitive word detection.

[0100] The application further proposes determining a document keyword sensitive score of each document keyword, weighting and fusing the document keyword sensitive scores of multiple document keywords to obtain a document keyword sensitive degree; determining a document type sensitive degree corresponding to a document type; weighting and fusing the document keyword sensitive degree and the document type sensitive degree to obtain a document sensitive degree of the target electronic document; extracting keywords from the document abstract to obtain multiple abstract keywords; extracting semantic features from the document abstract to obtain abstract semantic features; performing information recall in a second reference information library by using the multiple abstract keywords and the abstract semantic features to obtain dynamic reference information, and the second reference information library stores real-time updated dynamic reference information.

[0101] The document keyword sensitive score is a numerical value quantitatively evaluated by an algorithm for the sensitivity of each keyword in the document, which can be specifically implemented by using a statistical model based on the frequency of historical sensitive words, and is used to reflect the sensitivity of a single keyword in a specific context. Weighted fusion refers to weight distribution and calculation of the comprehensive value of the sensitive scores of different keywords according to a preset rule, which can be specifically implemented by using a linear weighting or a nonlinear function model, and is used to solve the one-sidedness problem of single keyword sensitivity evaluation. The document type sensitive degree is a sensitivity benchmark value preset according to different document types, which can be specifically implemented by using a risk level mapping table of different document types in a domain knowledge base, and is used to reflect the inherent sensitivity characteristics of the document belonging to a domain. Abstract keyword extraction refers to screening representative core words from the document abstract, which can be specifically implemented by using a TF-IDF algorithm or a word frequency-based sorting method, and is used to capture the core sensitive information of the document. Semantic feature extraction refers to converting the document abstract into a vector or feature sequence that can represent semantics, which can be specifically implemented by using a BERT model or an LSTM neural network, and is used to identify context-related implicit sensitive content. The second reference information library refers to a database storing real-time updated dynamic sensitive words and associated semantic patterns, which can be specifically implemented by using a network crawler to capture the latest sensitive words and establish an indexing mechanism, and is used to solve the problem of static word library update lag.

[0102] Specifically, the document keyword sensitivity score is calculated by a statistical model, for example, the frequency of a specific keyword appearing in historical sensitive documents is counted, and different weights are given according to its context position in the current document. In the weighted fusion process, high-frequency sensitive words are given higher weights, and low-frequency sensitive words are given lower weights, thereby forming a comprehensive document keyword sensitivity. The document type sensitivity is obtained by querying a preset domain risk level table, for example, the sensitivity benchmark value of a financial document is higher than that of a general text document. By twice weighting the keyword sensitivity and the document type sensitivity, for example, using a dynamically adjusted fusion coefficient, the final document sensitivity reflects both the specific content characteristics and the domain sensitivity requirements. In the dynamic reference information acquisition link, after the abstract keywords and semantic features are extracted by algorithms, the real-time data in the second reference information library is combined, for example, the keywords are matched with the dynamically updated sensitive words in the library, and at the same time, the associated sensitive patterns within the similarity threshold range are matched through semantic feature matching, thereby recalling the most relevant dynamic reference information.

[0103] Compared with related technologies, the traditional method only relies on static word library for keyword matching and cannot adapt to semantic changes and domain differences. For example, the related technology uses a fixed sensitive word list for scanning, and when new network words or industry terminology evolve, the static word library cannot be covered in time, resulting in an increased missed detection rate. The present scheme solves the deviation problem of single-dimensional evaluation through the dual evaluation mechanism of document type sensitivity and keyword sensitivity, and through the synergistic effect of semantic features and dynamic database, the sensitive word detection can adapt to language evolution in real time. For example, when detecting a document containing emerging network language, semantic feature matching can identify semantically sensitive words that are not included, and the dynamic database provides real-time updated sensitive word data support.

[0104] Through the above technical solutions, the present application realizes dynamic evaluation of document sensitivity, and solves the evaluation deviation problem caused by the inability of static word library to adapt to content changes. For example, when detecting a technical document containing industry terminology updates, the document type sensitivity can automatically match the latest risk level in this field, and the dynamic reference information provides real-time data of newly added sensitive words in this field, so that the sensitive word detection result is more accurate. At the same time, through the synergistic effect of semantic features and dynamic database, semantically sensitive words that are not explicitly included can be identified, for example, sensitive content expressed through word variants or metaphors can be detected, effectively reducing the missed detection rate.

[0105] Referring to Figure 4 The processor can further perform the following steps:

[0106] 401, determine sensitive word reference examples based on document type;

[0107] 402、determine the target detection level and the key detection content based on the document sensitivity, the detection requirement and the document type;

[0108] 403、determine a plurality of dynamic sensitive words based on dynamic reference information;

[0109] 404、generate a target detection method in combination with sensitive word reference examples, target field information, target detection level, key detection content and dynamic sensitive words.

[0110] Among them, the sensitive word reference example refers to the sensitive word recognition example matched with the document type, which can be realized by sensitive word example library query method, and the typical sensitive word form of the field is obtained by matching the document type to provide a basic reference for the detection method. The target detection level refers to the detection accuracy level divided according to the document sensitivity, which can be realized by weight allocation algorithm, and the detection depth parameter is calculated by the document security level and the user requirement to dynamically adjust the detection resource allocation. The key detection content refers to the sensitive information category that needs to be paid attention to first, which can be realized by rule engine, and the detection focus is determined by analyzing the key word group in the detection requirement and the document type characteristics. The dynamic sensitive word refers to the real-time updated sensitive word set, which can be realized by information recall mechanism, and the latest sensitive word data is obtained by dynamic reference information library to break through the update lag limitation of static word library.

[0111] Specifically, the sensitive word reference example is obtained by matching the document type to establish a basic detection framework consistent with the field characteristics. The target detection level and the key detection content are generated by using a multi-dimensional parameter fusion algorithm in combination with the document sensitivity quantitative index and the user-defined detection requirement to realize the dynamic adaptation of the detection granularity and the detection direction. The dynamic sensitive word is recalled in real time based on the dynamic reference information library to form a sensitive word set containing the latest semantic features. Finally, the field information, the detection level parameter, the key content identifier and the dynamic sensitive word are input into the detection model to generate a composite detection method that combines static rules and dynamic strategies, so that the detection logic not only retains the field characteristics but also has the semantic evolution adaptation ability.

[0112] Compared with related technologies, the traditional method relies on fixed sensitive word library for mechanical matching and cannot identify emerging network language or industry term variants. The present scheme generates a detection method in combination with field characteristics by real-time acquisition of external update data based on dynamic reference information, which effectively responds to semantic dynamic changes. Related technologies use unified detection rules, resulting in the lack of field characteristics. The present scheme references examples based on document type matching to make the detection method deeply consistent with the sensitive word features of the field to which the document belongs. Related technologies have fixed detection granularity and cannot adapt to different security level documents. The present scheme dynamically adjusts the detection level based on the document sensitivity to realize the balanced optimization of detection accuracy and processing efficiency.

[0113] By the technical solution, the application solves the problem of detection accuracy reduction caused by the fact that the static word library cannot adapt to semantic dynamic changes, and effectively captures emerging sensitive words through a dynamic sensitive word updating mechanism. The problem of misjudgment caused by the traditional detection method ignoring the field characteristics is overcome, and the matching degree of the detection strategy and the document attribute is improved through field information fusion. The defect that the fixed detection rule is difficult to adapt to different security level documents is improved, and intelligent allocation of detection resources is realized through multi-level detection parameter configuration.

[0114] The application further proposes to query the sensitive word reference example in the sensitive word example library based on the document type, determine the target detection dimension and the target detection granularity based on the document sensitivity, determine the special detection rule based on the detection requirement, determine the target detection level based on the target detection dimension and the target detection granularity, determine the key detection content based on the special detection rule and the document type, and recall the multiple dynamic sensitive words in the sensitive word library based on the dynamic reference information.

[0115] The sensitive word reference example refers to the typical sensitive word application scene retrieved in the preset example library through the document type, which can be implemented by vectorizing and matching the document type using a natural language processing model, and is used to provide field-related semantic reference for the detection model. The target detection dimension refers to the sensitive information category that needs to be covered in the detection process, which can be dynamically adjusted by the weight parameter corresponding to the document sensitivity, and is used to match the detection requirements of different security levels. The target detection granularity refers to the strictness of sensitive word matching in the detection process, which can be implemented by setting a fuzzy matching threshold or a semantic similarity threshold, and is used to balance the false positive rate and the missed detection rate. The special detection rule refers to the detection condition defined by the user, which can be implemented by extracting keywords and logical relationships from the text description in the detection requirement, and is used to realize personalized detection strategy. The dynamic reference information refers to the real-time updated auxiliary detection data, which can be implemented by crawling network public opinion data or industry term database to supplement the emerging sensitive words not covered by the static word library.

[0116] Specifically, when the sensitive word detection request is triggered, the system first extracts the common sensitive word use cases in the field from the example library according to the document type, such as sensitive expressions related to fund flow in the financial field. Then, the detection dimension is automatically selected according to the document sensitivity, such as high-sensitive documents covering multi-dimensional detection of privacy leakage, trade secrets, etc. The detection granularity is dynamically adjusted according to the sensitivity, such as loose fuzzy matching for ordinary documents and strict semantic analysis for confidential documents. The text form detection requirements submitted by the user are parsed into special rules, such as requiring to focus on detecting paragraphs related to specific customer names. Dynamic reference information is obtained by real-time retrieval of external databases to obtain the latest sensitive words, such as recently emerging industry slang or network popular language. These elements jointly constitute the detection method, so that the system can not only inherit the domain knowledge of the example library, but also dynamically adapt to semantic changes and personalized needs.

[0117] Compared with related technologies, the traditional method relies on fixed word library for mechanical matching, which cannot identify word shape changes or emerging words, such as being unable to detect pinyin abbreviations or homophonic words. The present scheme continuously updates the word library through a dynamic reference information recall mechanism, such as including newly emerging sensitive words on social media in the detection range. The traditional method uses a unified detection standard, which is difficult to adapt to documents of different security levels, such as using the same detection intensity for ordinary announcements and confidential contracts. The present scheme automatically adjusts the detection dimension and granularity according to the document sensitivity, such as increasing semantic correlation analysis and context verification for confidential contracts. The traditional method lacks personalized detection capability, such as being unable to adjust the detection focus according to user needs. The present scheme generates special rules by parsing detection requirements, such as focusing on detecting patient privacy information in medical documents.

[0118] Through the above technical solutions, the present application realizes the dynamic adaptation of sensitive word detection range and semantic environment, solves the problem of missed detection caused by lagging update of static word library. The dynamic adjustment of detection dimension and granularity optimizes the detection efficiency and avoids the risk of false judgment of high-sensitive documents. The personalized rule generation mechanism enables the detection process to accurately match user needs, such as strengthening the compliance verification of clauses in legal documents. The real-time recall of dynamic reference information ensures that the detection model continuously covers emerging sensitive words.

[0119] The present application further proposes to generate target sensitive word detection examples based on sensitive word reference examples and multiple dynamic sensitive words, determine an initial detection method based on field information of a target field, a target detection level, and key detection content, and determine a target detection method based on the target sensitive word detection examples and the initial detection method.

[0120] The sensitive word reference example refers to a typical sensitive word recognition case associated with the document type and is pre-stored. Specifically, the example set stored in the sensitive word example library according to the industry classification can be used to implement the sensitive word reference example, which is used to provide a basic detection rule. The dynamic sensitive word refers to the current sensitive word obtained through the real-time updating of the reference information library. Specifically, the network crawler can be used to capture social media new words or industry term update data to implement the dynamic sensitive word, which is used to cover new sensitive content generated by semantic evolution. The target detection example refers to the detection example generated by combining the basic rule and the dynamic word. Specifically, the semantic fusion of the reference example and the dynamic word can be implemented by using a natural language processing model, which is used to construct an extensible detection benchmark. The initial detection mode refers to the preliminary detection strategy generated according to the domain characteristics of the document. Specifically, the detection parameters and detection levels can be matched by using the domain knowledge graph to implement the initial detection mode, which is used to adapt to the detection requirements of different application scenarios.

[0121] Specifically, when the target detection mode needs to be determined, the pre-defined sensitive word reference example and the real-time obtained dynamic sensitive word are first input into the language model for semantic fusion to generate a detection example set containing the basic rule and the dynamic word. Then, the detection level parameters are matched according to the knowledge graph of the domain to which the document belongs, and the initial detection rule is generated in combination with the key detection content. Finally, the detection example set is input into the language model to modify the initial detection rule semantically, forming a detection mode that can cover both static rules and dynamic changes. For example, in the financial domain document detection, the initial detection mode is set to focus on the recognition of digital sensitive information, but after fusing the newly added financial fraud term examples from social media, the detection mode is modified to increase the scanning logic for new fraud keywords.

[0122] Compared with related technologies, the traditional method only relies on a fixed word library for keyword matching and cannot recognize new sensitive words generated by semantic evolution and domain characteristic differences. The present scheme expands the detection example set by dynamic sensitive words, so that the detection mode can cover network new words and industry term changes; generates an initial detection strategy by matching the domain information, so as to ensure that the detection rule is consistent with the document application scenario; and breaks through the rigid limitation of the traditional rule engine by using the example modification mechanism, forming a dynamically adjustable detection logic.

[0123] Through the above technical solutions, the present application can effectively recognize new sensitive words generated by semantic evolution, such as variant expressions in network slang; can automatically adjust the detection focus for different domain documents, such as medical documents focusing on privacy information detection; and the detection rule has a dynamic optimization capability, which can automatically adjust the detection logic through example update when new sensitive information patterns appear.

[0124] The application further proposes filling the sensitive word reference example and the plurality of dynamic sensitive words into the first prompt text template to obtain a first prompt text, the first prompt text template being used to indicate that the plurality of dynamic sensitive words are used to expand the sensitive word reference example; inputting the first prompt text into a target language model, encoding and decoding the first prompt text by the target language model to output a target sensitive word detection example; determining a domain detection parameter corresponding to a target domain based on domain information of the target domain, the domain detection parameter being used to represent a domain detection manner of the target domain; matching the domain detection parameter, the target detection level and the key detection content in a plurality of candidate detection manners to obtain an initial detection manner; filling the target sensitive word detection example and the initial detection manner into a second prompt text template to obtain a second prompt text, the second prompt text template being used to indicate that the target sensitive word detection example is used to correct the initial detection manner; inputting the second prompt text into the target language model to output a target detection manner.

[0125] The first prompt text template is a preset text framework containing dynamic sensitive word insertion instructions, which can be implemented by a string template with placeholders, and is used to guide the language model to perform semantic fusion of dynamic sensitive words and static examples. The target language model is a pre-trained model with natural language processing capability, which can be implemented by a large-scale neural network model based on the Transformer architecture, and is used to generate detection examples covering dynamic semantic changes. The second prompt text template is a preset text framework of detection rule correction instructions, which can be implemented by a template string containing detection manner optimization instructions, and is used to establish a feedback mechanism between detection examples and detection rules. The domain detection parameter is a configuration parameter reflecting the detection requirements of sensitive words in a specific domain, which can be implemented by numerical parameters such as domain keyword weight and semantic similarity threshold, and is used to ensure that the detection manner matches the domain characteristics.

[0126] Specifically, when it is detected that the sensitive word detection manner needs to be updated, first, the reference example in the static example library is combined with the real-time acquired dynamic sensitive word to generate a first prompt text. For example, when a new word on the network is detected, it is input into the language model together with the original example to generate a detection example containing a combination of new and old sensitive words. Then, the corresponding detection parameter is matched according to the characteristics of the target domain, for example, a higher professional term recognition weight is set in the medical field, and the detection rule of the digital combination is strengthened in the financial field. Finally, the generated detection example and the initial detection manner are input into the language model through the second prompt text, so that the detection rule can be automatically adjusted to adapt to the new example characteristics, such as increasing the fuzzy matching rule or adjusting the semantic similarity threshold.

[0127] Compared with the related art, the related art relies on a static example library maintained manually for rule matching, and cannot respond to the evolution of network new words or industry terms in a timely manner. However, the present scheme can capture semantic changes in real time through dynamic generation of detection examples and adaptive rule optimization. For example, when a new type of sensitive word variant appears, the related art needs to manually update the word library and rules, while the present scheme automatically generates detection examples and adjusts the matching rules through a language model, avoiding the lag of manual intervention.

[0128] Through the above technical solutions, the present application can automatically generate detection examples covering dynamic semantic changes, effectively identifying network new words and term variants that are difficult to detect by traditional methods. At the same time, through the matching of detection parameters and the optimization mechanism of rules adapted to the field, the detection method is accurately matched with the sensitive features of the target field, reducing false positives and missed detections caused by rigid rules. For example, in social media document detection, metaphorical expressions of network popular language can be automatically identified; in patent document detection, sensitive combinations of technical terms can be accurately captured.

[0129] Referring to Figure 5 , the processor can further perform the following steps:

[0130] 501. Perform sensitive word detection on the target electronic document using a target detection method to obtain a plurality of candidate sensitive words and their candidate sensitive word degrees;

[0131] 502. Based on the field information of the field to which the target electronic document belongs, correct the candidate sensitive words and their candidate sensitive word degrees to obtain the sensitive word detection result of the target electronic document.

[0132] Wherein, the field to which the target electronic document belongs is the target field, the target detection method refers to the detection strategy generated according to the document type and dynamic reference information, which can be specifically implemented by using a natural language processing model combined with a dynamic sensitive word library, and is used to realize preliminary sensitive word recognition and sensitive degree evaluation. The candidate sensitive word degree refers to the sensitive word confidence value calculated by a semantic analysis model, which can be specifically implemented by using a probability model based on context correlation degree, and is used to quantify the accuracy of sensitive word recognition. The field information refers to the professional term library and detection rules of the field to which the target electronic document belongs, which can be specifically implemented by using a field knowledge graph or an industry sensitive word standard library, and is used to eliminate false positives caused by semantic differences across fields.

[0133] Specifically, the technical solution improves detection accuracy through a two-stage processing mechanism. In the first stage, a target detection method that fuses dynamic reference information is used to scan the document, identify candidate sensitive words, and generate a preliminary sensitive degree score. In the second stage, a domain-specific domain information library is called to perform semantic verification and weight correction on the candidate sensitive words. For example, in the medical field scenario, a medical terminology library is loaded to perform semantic disambiguation on professional terms such as "positive", reducing the probability of false judgment in non-medical scenarios. At the same time, the sensitive degree score is adjusted according to the domain detection rules, for example, setting a higher sensitive threshold for "yield" and other terms in the financial field. This double-checking mechanism not only retains the flexibility of general detection, but also achieves precise adaptation of detection results through domain knowledge injection.

[0134] Compared with related technologies, the traditional static sensitive word library cannot distinguish the semantic differences between domains, for example, the sensitive degree determination standards for "virus" in the computer field and the medical field are significantly different. The present solution effectively solves the misjudgment problem of homonyms across domains by introducing a domain information correction link, and supplements the recognition ability of the detection system for industry slang and domain variant expressions. The sensitive degree evaluation deviation caused by the uniform detection standard of related technologies is corrected through the domain weight adjustment mechanism in the present solution.

[0135] Through the above technical solution, the present application can accurately identify domain-specific sensitive words and eliminate cross-domain misjudgments, such as avoiding misjudging "sue" in legal documents as a sensitive word. The sensitive degree evaluation deviation problem is effectively controlled through domain weight adjustment. The technical solution significantly improves the accuracy and adaptability of sensitive word detection in cross-domain application scenarios.

[0136] All the above optional technical solutions can be combined to form optional embodiments of the present application, which will not be repeated here.

[0137] Figure 6 is a structural schematic diagram of an electronic document security management device based on dynamic sensitive word detection provided by an embodiment of the present application, referring to Figure 6 The device comprises:

[0138] A first determination module 601 is configured to determine the document type, document sensitive degree, detection requirement, and dynamic reference information of a target electronic document in response to a sensitive word detection request of the target electronic document, wherein the detection requirement is carried by the sensitive word detection request, and the dynamic reference information is dynamic reference information used to assist in sensitive word detection.

[0139] A second determination module 602 is configured to determine a target detection method for sensitive word detection of the target electronic document based on the document type, the document sensitive degree, the detection requirement, and the dynamic reference information.

[0140] The detection module 603 is configured to perform sensitive word detection on the target electronic document by using the target detection manner, and obtain a sensitive word detection result of the target electronic document, where the sensitive word detection result includes sensitive words in the target electronic document and sensitive word degrees of the sensitive words.

[0141] It should be noted that the electronic document security management apparatus based on dynamic sensitive word detection provided in the above embodiments is only used as an example for the division of the above functional modules in the electronic document security management, and in actual applications, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the computer device is divided into different functional modules to complete all or part of the above described functions. In addition, the electronic document security management apparatus based on dynamic sensitive word detection provided in the above embodiments and the method embodiment executed by the processor belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0142] Figure 7 FIG. 7 is a structural schematic diagram of an electronic document security management system provided by an embodiment of the present application. The system 700 can have great differences due to different configurations or performances, and can include one or more processors (Central Processing Units, CPU) 701 and one or more memories 702. The one or more memories 702 store at least one computer program, which is loaded and executed by the one or more processors 701 to implement the method provided in each of the above method embodiments. Of course, the system 700 can also have a wired or wireless network interface, a keyboard, an input and output interface, and other components for realizing the functions of the device, so as to perform input and output. The system 700 can also include other components for realizing the functions of the device, which will not be described here.

[0143] In some embodiments, the computer program related to the embodiments of the present application can be deployed on one computer device for execution, or on multiple computer devices located in one place for execution, or on multiple computer devices distributed in multiple places and interconnected through a communication network for execution. The multiple computer devices distributed in multiple places and interconnected through a communication network can constitute a blockchain system.

[0144] Those skilled in the art can understand that all or part of the above described embodiments can be completed by hardware, or by a program instructing relevant hardware, and the program can be stored in a computer readable storage medium, such as a read-only memory, a magnetic disk or an optical disk.

[0145] The above merely is the optional embodiment of the present application, and does not use to limit the present application, any modification, equivalent replacement, improvement and so on, which are made in the spirit and principle of the present application, should be included in the protection scope of the present application.

Claims

1. An electronic document security management system based on dynamic sensitive word detection, characterized in that, The system includes a processor and a memory, the processor being configured to perform the following steps: In response to a sensitive word detection request for a target electronic document, the document type, document sensitivity level, detection requirements, and dynamic reference information of the target electronic document are determined. The detection requirements are carried by the sensitive word detection request, and the dynamic reference information is dynamic reference information used to assist in sensitive word detection. Based on the document type, the document sensitivity level, the detection requirements, and the dynamic reference information, determine the target detection method when performing sensitive word detection on the target electronic document; The method for determining the target detection method when performing sensitive word detection on the target electronic document based on the document type, the document sensitivity, the detection requirements, and the dynamic reference information includes: Based on the document type, sensitive word reference examples are determined; based on the document sensitivity, the detection requirements, and the document type, the target detection level and key detection content are determined, whereby the target detection level indicates the level of detail in sensitive word detection, and the key detection content indicates the focus of sensitive word detection; based on the dynamic reference information, multiple dynamic sensitive words are determined; based on the sensitive word reference examples, the domain information of the target domain, the target detection level, the key detection content, and the multiple dynamic sensitive words, the target detection method is determined, whereby the target domain is the domain to which the target electronic document belongs; The target detection method is determined based on the sensitive word reference examples, the target domain information, the target detection level, the key detection content, and the multiple dynamic sensitive words. The target domain is the domain to which the target electronic document belongs, including: Based on the aforementioned sensitive word reference examples and the multiple dynamic sensitive words, a target sensitive word detection example is generated; based on the domain information of the target domain, the target detection level, and the key detection content, an initial detection method is determined; based on the target sensitive word detection example and the initial detection method, the target detection method is determined. The target electronic document is subjected to sensitive word detection using the aforementioned target detection method to obtain the sensitive word detection result of the target electronic document. The sensitive word detection result includes the sensitive words in the target electronic document and the degree of sensitivity of the sensitive words.

2. The system according to claim 1, characterized in that, In response to a sensitive word detection request for a target electronic document, the document type, document sensitivity level, detection requirements, and dynamic reference information of the target electronic document are determined, including: In response to a sensitive word detection request for a target electronic document, the target electronic document and the detection requirement are obtained from the sensitive word detection request. The detection requirement is in text form and is used to represent the personalized detection requirements of the target user. The target user is the user who sent the sensitive word detection request. The target electronic document is subjected to document keyword extraction and summary extraction to obtain multiple document keywords and document summary of the electronic document; The target electronic document is classified based on the document summary to determine the document type of the target electronic document, and the document type is used to indicate the corresponding field; Based on the multiple document keywords, the document type, and the document summary, the document sensitivity and dynamic reference information of the target electronic document are determined.

3. The system according to claim 2, characterized in that, The step of classifying the target electronic document based on the document summary to determine the document type of the target electronic document, wherein the document type is used to indicate the corresponding field, includes: Obtain the document description information of the target electronic document, the document description information including the document source, document purpose, and document destination; based on the document description information and the document summary, determine the first initial document type and domain reference information of the target electronic document, the domain reference information being used to assist in determining the domain to which the target electronic document belongs; based on the first initial document type and the domain reference information of the target electronic document, determine the document type of the target electronic document; The process of determining the document sensitivity and dynamic reference information of the target electronic document based on the multiple document keywords, the document type, and the document summary includes: Based on the multiple document keywords and the document type, the document sensitivity of the target electronic document is determined; based on the document summary, the dynamic reference information is determined.

4. The system according to claim 3, characterized in that, The step of determining the first initial document type and domain reference information of the target electronic document based on the document description information and the document summary includes: Based on the document description information, the document usage scenario of the target electronic document is determined; based on the document usage scenario and the document summary, the first initial document type of the target electronic document is determined; based on the document summary, a query is performed in the first reference information database to obtain the domain reference information of the target electronic document; The step of determining the document type of the target electronic document based on the first initial document type and the domain reference information includes: Based on the domain reference information, multiple candidate document types are determined; based on the document digest of the target electronic document, a second initial document type is determined from the multiple candidate document types, wherein the second initial document type is the candidate document type with the highest matching degree with the document digest; the first initial document type and the second initial document type are combined to obtain the document type of the target electronic document.

5. The system according to claim 3, characterized in that, The process of determining the document sensitivity of the target electronic document based on the multiple document keywords and the document type includes: Determine the document keyword sensitivity score for each of the document keywords, and then perform a weighted fusion of the document keyword sensitivity scores of the multiple document keywords to obtain the document keyword sensitivity degree of the multiple document keywords; determine the document type sensitivity degree corresponding to the document type; and then perform a weighted fusion of the document keyword sensitivity degree and the document type sensitivity degree to obtain the document sensitivity degree of the target electronic document. The determination of the dynamic reference information based on the document digest includes: Keyword extraction is performed on the document summary to obtain multiple summary keywords; semantic feature extraction is performed on the document summary to obtain summary semantic features; information retrieval is performed in a second reference information database using the multiple summary keywords and the summary semantic features to obtain the dynamic reference information, wherein the second reference information database stores dynamically updated information in real time.

6. The system according to claim 1, characterized in that, The reference examples for determining sensitive words based on the document type include: The document type is used to query the sensitive word example library to obtain the sensitive word reference example corresponding to the document type. The sensitive word reference example includes various situations that are identified as sensitive words. The process of determining the target detection level and key detection content based on the document sensitivity, detection requirements, and document type includes: Based on the document sensitivity, the target detection dimension and target detection granularity are determined; based on the detection requirements, special detection rules are determined; based on the target detection dimension and target detection granularity, the target detection level is determined; based on the special detection rules and the document type, the key detection content is determined. Based on the dynamic reference information, multiple dynamic sensitive words are determined, including: Based on the dynamic reference information, a recall is performed in the sensitive word database to obtain the multiple dynamic sensitive words.

7. The system according to claim 1, characterized in that, The generation of target sensitive word detection examples based on the sensitive word reference examples and the plurality of dynamic sensitive words includes: The sensitive word reference example and the plurality of dynamic sensitive words are filled into the first prompt text template to obtain the first prompt text. The first prompt text template is used to indicate that the sensitive word reference example is expanded with the plurality of dynamic sensitive words. The first prompt text is input into the target language model, and the target language model encodes and decodes the first prompt text to output the target sensitive word detection example. The determination of the initial detection method based on the domain information of the target domain, the target detection level, and the key detection content includes: Based on the domain information of the target domain, domain detection parameters corresponding to the target domain are determined. The domain detection parameters are used to represent the domain detection method of the target domain. The domain detection parameters, the target detection level, and the key detection content are used to match multiple candidate detection methods to obtain the initial detection method. The initial detection method is a successfully matched candidate detection method. The determination of the target detection method based on the target sensitive word detection example and the initial detection method includes: The target sensitive word detection example and the initial detection method are filled into the second prompt text template to obtain the second prompt text. The second prompt text template is used to indicate that the target sensitive word detection example is used to correct the initial detection method. The second prompt text is input into the target language model, and the target language model encodes and decodes the second prompt text to output the target detection method.

8. The system according to claim 1, characterized in that, The step of performing sensitive word detection on the target electronic document using the aforementioned target detection method to obtain the sensitive word detection result of the target electronic document includes: The target electronic document is subjected to sensitive word detection using the target detection method described above, thereby obtaining multiple candidate sensitive words in the target electronic document and the degree of sensitivity of each candidate sensitive word; Based on the domain information of the target field to which the target electronic document belongs, the multiple candidate sensitive words and the candidate sensitive word degree of each candidate sensitive word are corrected to obtain the sensitive word detection result of the target electronic document.

Images