An auxiliary secret protection method, system, storage medium and program product

Abstract

The application provides an auxiliary classification method, system, storage medium and program product, and relates to the technical field of information security. The technical scheme decomposes the to-be-classified file into paragraphs and extracts the abstract, then extracts and refines a few summarized cores from the huge and mixed paragraphs through semantic clustering and group abstract generation, avoids the problem that key points are diluted by massive non-classified content, and because the group abstract is the crystallization of paragraph semantics, it retains the context association and logical structure between words while refining the text. The subsequent stage can identify and capture implicit key points that are not defined by fixed keywords, thereby improving the accuracy of key point identification and the accuracy of the auxiliary classification result. At the same time, the business field is determined to ensure that the recall from the classified basis library has relevance and pertinence, and the relevance between the group abstract and the classified basis is used to improve the accuracy of the auxiliary classification result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information security, and particularly relates to an auxiliary classified information determination method and system, a storage medium and a program product. BACKGROUND

[0002] With the rapid development of information technology and the deepening of globalization, the security of information has become a key issue of concern for countries, enterprises and individuals; especially in the field involving sensitive information, accurate classified information division of files is crucial.

[0003] The prior art usually adopts a method based on rules or keyword matching. Specifically, the system will preset a knowledge base containing a large number of classified information keywords and fixed rules. When processing the classified information determination file, the system will perform full-text retrieval on the file content, and count the number of classified information keywords matched or the number of rule items triggered. When the statistical value reaches the preset threshold, the system will recommend a corresponding classified information level.

[0004] However, with the increasing complexity of modern file content, classified information is no longer presented by isolated and explicit keywords, but is more deeply implied in the semantic logic of the context. However, the related art has limited understanding of complex semantics, resulting in low accuracy of classified information point identification, and thus low accuracy of auxiliary classified information determination results. SUMMARY

[0005] The present application provides an auxiliary classified information determination method, system, storage medium and program product for improving the accuracy of classified information point identification and thus improving the accuracy of auxiliary classified information determination results.

[0006] In a first aspect, the present application provides an auxiliary classified information determination method, comprising: receiving a classified information determination file and dividing the classified information determination file into a plurality of text paragraphs; extracting a paragraph abstract of each text paragraph; clustering all paragraph abstracts into a plurality of abstract groups based on semantic similarity between the paragraph abstracts; generating a group abstract for each abstract group; determining a business field of the classified information determination file or the group abstract; recalling corresponding classified information determination basis from a preset classified information determination basis library according to the business field; calculating the relevance of the group abstract and the classified information determination basis respectively to generate a relevance score of the corresponding group abstract; aggregating the group abstracts to form an abstract set based on the relevance score; and generating a classified information determination result based on the abstract set.

[0007] Compared with the full-text processing in the related art, the technical solution divides the to-be-determined secret file into paragraphs and extracts an abstract, and then through semantic clustering and generating a group abstract, the huge and mixed paragraphs are extracted and refined into a few summarized cores (i.e., the group abstract), which reduces the data amount for subsequent analysis and avoids the problem that key points are diluted by a large amount of non-secret content. Since the group abstract is a crystallization of paragraph semantics, it retains the context association and logical structure between words while refining the text. The complete retention of the "context relationship" provides a solid technical foundation for subsequent stages to identify and capture implicit secret points that are not defined by fixed keywords and are contained in the context logical relationship, thereby improving the accuracy of secret point identification and the accuracy of the auxiliary secret determination result. At the same time, the business field to which the secret determination result belongs is determined to ensure that the recall from the secret determination basis library is relevant and targeted, and the relevance between the group abstract and the secret determination basis is used to improve the accuracy of the auxiliary secret determination result.

[0008] In some embodiments of the first aspect, after the step of aggregating the group abstracts to form the abstract set based on the correlation scores, the method further includes, for each group abstract in the abstract set, performing a word segmentation process to obtain a plurality of constituent words; for a single constituent word, retrieving a first synonym and / or a first abbreviation corresponding to the constituent word from a preset synonym library and / or an abbreviation library, and replacing the corresponding constituent word in the group abstract to obtain an abstract variant; and selecting at least one abstract variant as the group abstract and expanding it to the abstract set.

[0009] By using the above technical solution, on the basis of having obtained the abstract set, the corresponding constituent words in the group abstract are retrieved and replaced from the preset synonym library and / or the abbreviation library to generate abstract variants that are semantically the same but have various forms of expression. The abstract variants are expanded to the abstract set to increase the robustness of subsequent recall. Even if the group abstract is not completely consistent with the secret determination basis, the abstract variant can establish effective matching with the secret determination basis, thereby improving the recall rate of the subsequent secret file knowledge base.

[0010] In some embodiments combined with the first aspect, based on the summary set, the step of generating the classification result specifically comprises: extracting keywords for each group summary in the summary set; retrieving second synonyms and / or second abbreviations corresponding to the keywords from a preset synonym library and / or abbreviation library; vectorizing each group summary in the summary set to obtain a group summary vector; performing text matching on file titles and / or classified content in a preset classified file knowledge base using the keywords and / or the second synonyms and / or the second abbreviations to obtain a first candidate file set; performing similarity calculation on the group summary vector and a pre-stored file vector in the classified file knowledge base to obtain a second candidate file set; merging the first candidate file set and the second candidate file set to obtain a classified file list; and generating the classification result based on the classified file list.

[0011] By using the above technical solution, the semantic breadth is expanded by expanding the keywords of the group summary with synonyms and / or abbreviations, and then in the subsequent text matching process, files that are semantically related but have different words can be recalled. By using the group summary vector for similarity calculation, the deep semantic information of the group summary is encoded as a point in a high-dimensional vector space, and by calculating the spatial distance between vectors, files that are related in core concepts and themes but have different expression methods can be recalled, the context and potential context are understood, and deep semantic recall is achieved. Finally, by merging the first candidate file set and the second candidate file set, the combination of precise matching and fuzzy semantics is achieved, which avoids the problem of missing keywords in semantic retrieval and makes up for the shortcoming of keyword retrieval that cannot understand deep semantics.

[0012] In some embodiments combined with the first aspect, based on the classified file list, the step of generating the classification result specifically comprises: calculating the semantic association degree between the classified file list and the group summary, and screening a reference file according to the semantic association degree; splicing a preset classification task instruction, the reference file and a file to be classified to obtain a classification instruction text; inputting the classification instruction text into an inference generation model to obtain a classification result; and the classification result includes a classified level, a classification basis, a classified period and a confidence level.

[0013] By using the above technical solution, by splicing the preset classification task instruction, the reference file and the file to be classified, the reference file has relatively high accuracy and relevance, when the model receives this structured classification instruction text, the inference process is guided and constrained within the range of the provided reference file, rather than relying on the files inside which may have biases or be outdated, thereby suppressing the knowledge hallucination phenomenon of large language models.

[0014] In some embodiments of the first aspect, before the step of extracting the keywords for each group summary in the summary set, the method further comprises: obtaining the text length of each group summary, and determining the target keyword quantity according to the text length; extracting refined keywords in the group summary based on the target keyword quantity; and abbreviating the group summary according to the refined keywords, wherein the refined keywords are used to generate a new group summary for replacing the group summary, and the group summary is used to constrain the new group summary.

[0015] By adopting the above technical solution, although the group summary of the long text is complete in information, it may also produce noise due to too large data quantity. Therefore, by determining the target keyword quantity according to the text length and extracting the refined keywords, and then abbreviating the group summary according to the refined keywords, a more refined and less noisy group summary is generated. Using this new group summary after abbreviation for subsequent text matching and similarity calculation makes the first candidate file set and the second candidate file set more relevant, so as to improve the ability of specific key information and further improve the purpose and accuracy of recall.

[0016] In some embodiments of the first aspect, the determination standard of the target keyword quantity is: In the formula, is the target keyword quantity, is the text length, is a disturbance coefficient, and the value range is [0.5, 1].

[0017] By adopting the above technical solution, the logarithmic function is adopted to adapt to the growth law of information entropy, that is, the marginal benefit of new information decreases with the increase of the text length, so that the quantity of the extracted keywords can be adaptively adjusted with the change of the text length.

[0018] In some embodiments of the first aspect, before the step of generating the classified result based on the summary set, the method further comprises performing at least one of the following operations: performing instruction cleaning on the group summary; identifying sensitive terms in the group summary and performing hierarchical mapping and desensitization replacement; embedding a pre-set structured term conforming to a security regulation in the group summary; matching and rewriting an unauthorized query keyword in the group summary by using a regular expression, and performing summary standardization on the group summary by using a structured query template; obtaining the permission level of a current operation user, comparing the permission level with the classified levels of each file in the files to be recalled, and providing only the file content within the user's permission range.

[0019] By adopting the above technical solutions, the risk of instruction injection is prevented and misoperation is avoided by pre-cleaning the group summary. Secondly, identifying and de-identifying sensitive terms in the group summary and replacing them prevents the leakage or recording of real sensitive information during processing without affecting semantic analysis, achieving a balance between data availability and confidentiality. Thirdly, embedding pre-set structured terms standardizes internal and non-standard expressions, ensuring that all content interacting with external systems or generating reports complies with confidentiality regulations and unit management requirements. Fourthly, matching and rewriting unauthorized query keywords using regular expressions enables real-time monitoring and blocking of query intentions that may lead to illegal information leakage. Then, by embedding pre-set structured terms in the group summary and applying structured query templates for standardization, not only is the professionalism and compliance of the summary content ensured, avoiding ambiguity or legal risks caused by non-standard expressions, but also, by adding a user permission verification step at the initial stage of document retrieval, security is further enhanced when processing classified information.

[0020] Secondly, this application provides an auxiliary confidentiality system, which includes: one or more processors and a memory; the memory is coupled to one or more processors, and the memory is used to store computer program code, which includes computer instructions; one or more processors call the computer instructions to cause the auxiliary confidentiality system to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, this application provides a computer program product containing instructions that, when the computer program product is run on an auxiliary secrecy system, cause the auxiliary secrecy system to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on an auxiliary secrecy system, cause the auxiliary secrecy system to perform the method described in the first aspect and any possible implementation thereof.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0024] 1. Compared to full-text processing in related technologies, this technical solution decomposes the document to be classified into paragraphs and extracts summaries. Then, through semantic clustering and generating group summaries, the large and complex paragraphs are extracted and refined into a few concise core elements (i.e., group summaries). This reduces the amount of data for subsequent analysis and avoids the problem of key classified points being diluted by massive amounts of non-classified content. Furthermore, because group summaries are the crystallization of paragraph semantics, they crucially preserve the contextual relationships and logical structure between words while refining the text. This complete preservation of "contextual relationships" provides a solid technical foundation for subsequent stages to identify and capture implicit classified points that are not defined by fixed keywords but are implied in the contextual logical relationships, thereby improving the accuracy of classified point identification and thus the accuracy of the auxiliary classification results. Simultaneously, determining the relevant business domain ensures that the retrieval from the classification basis database is relevant and targeted, and the correlation between group summaries and classification basis is used to improve the accuracy of the auxiliary classification results.

[0025] 2. By expanding the keywords of the group summary with synonyms and / or abbreviations, semantic breadth is expanded, enabling the recall of documents that are semantically related to the group summary but use different terminology during subsequent text matching. Using group summary vectors for similarity calculation encodes the deep semantic information of the group summary as a point in a high-dimensional vector space. By calculating the spatial distance between vectors, documents that are related in core concepts and topics but not entirely expressed can be recalled, understanding the context and potential language to achieve deep semantic recall. Finally, by merging the first and second candidate document sets, a combination of precise matching and fuzzy semantics is achieved. This avoids the problem of semantic retrieval potentially missing keywords and compensates for the limitation of keyword retrieval in understanding deep semantics.

[0026] 3. By concatenating the pre-set classification task instructions, reference documents, and documents to be classified, where the reference documents are documents with relatively high accuracy and relevance, when the model receives this structured classification instruction text, its reasoning process is guided and constrained within the scope of the provided reference documents, rather than relying on documents that may contain biases or be outdated, thereby suppressing the knowledge illusion phenomenon of large language models. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an auxiliary confidentiality determination method in an embodiment of this application;

[0028] Figure 2 This is another flowchart illustrating the auxiliary confidentiality determination method in the embodiments of this application;

[0029] Figure 3 This is a schematic diagram of the architecture of the auxiliary confidentiality determination method in the embodiments of this application;

[0030] Figure 4 This is a schematic diagram of an exemplary hardware structure of the auxiliary confidentiality system in the embodiments of this application. Detailed Implementation

[0031] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0032] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0033] Please see Figure 1 and Figure 3 , Figure 1 for Figure 1 This is a flowchart illustrating an auxiliary confidentiality determination method in an embodiment of this application. Figure 3 This is a schematic diagram of the architecture of the auxiliary confidentiality determination method in the embodiments of this application;

[0034] An auxiliary method for determining confidentiality includes:

[0035] S101. Receive the file to be classified and divide it into multiple text segments.

[0036] Among them, documents pending classification refer to electronic documents or text materials that require classification and identification of their security level.

[0037] It is important to understand that the term "text paragraph" here is a broad concept, referring to a continuous unit of text segmented from a pending confidential document for independent semantic analysis. This unit can be a single paragraph, a text block spanning multiple paragraphs, or any continuous segment within a document. This application does not impose any specific limitations on this term.

[0038] Specific segmentation methods may include, but are not limited to, the following:

[0039] Physical delimiter-based segmentation: Segmenting based on inherent markers in the text that indicate the end of a paragraph (such as two consecutive line breaks).

[0040] Fixed-length segmentation: Set a fixed threshold for the number of words or characters to divide the document into multiple text segments of roughly equal length.

[0041] Semantic aggregation-based segmentation: Using natural language processing (NLP) technology, the document is first segmented into sentences, and then semantically related consecutive sentences are merged into a logical paragraph unit by calculating the semantic similarity or topic coherence between sentences.

[0042] S102. Extract a paragraph summary for each text segment;

[0043] Paragraph summaries refer to concise texts that summarize and represent the content of the corresponding paragraphs.

[0044] In some embodiments, an extractive summarization model can be used to select the most critical sentence or sentences from a text paragraph as a paragraph summary.

[0045] In other embodiments, a generative model (Abstractive Summarization) can be used, in which the model understands the content of a text paragraph and then reorganizes and generates a new, more concise paragraph summary; this is not a limitation.

[0046] S103. Based on the semantic similarity between paragraph summaries, all paragraph summaries are clustered into multiple summary groups;

[0047] Semantic similarity refers to the degree of closeness in meaning between two texts.

[0048] In some embodiments, each paragraph summary is treated as a multi-dimensional vector composed of words. Semantic similarity is determined by comparing the angles or distances between these vectors in space, and vectors that are close in distance are grouped into the same cluster. This results in multiple summary groups, where paragraph summaries within each group are semantically related, while paragraph summaries between different groups are semantically relatively independent.

[0049] In some embodiments, all words in all paragraph summaries are mapped to a vector space, and words with similar semantics are also located close to each other in the space. For each paragraph summary, word vectors corresponding to all words are found. Then, a comprehensive vector representing the entire paragraph is generated by averaging or TF-IDF weighted averaging. The semantic similarity between any two summary sentence vectors is calculated using cosine similarity.

[0050] S104. Generate a group summary for each summary group;

[0051] It should be noted that a group abstract refers to a text that represents the core theme of a group of abstracts, and can be:

[0052] A summary of any paragraph directly selected from a group, for example, according to a preset rule (such as selecting the first one) or randomly selected.

[0053] The most representative paragraph summary is selected from the summary group; for example, in step S103, the paragraph summary that is the most central when clustered in the vector space is selected.

[0054] A completely new summary text is generated by recreating all the summary content within the group.

[0055] In some specific embodiments, all summary content within a group and preset instructions are input into a large language model to obtain a completely new summary text. In other specific embodiments, the summary content is rearranged according to the order in which it appears in the original text; this is not limited here.

[0056] S105. Determine the business area of ​​the document or group summary to be classified;

[0057] It should be noted that the purpose of this step is to assign one or more predefined business domain tags to the document to be classified or its core content (represented as a group summary) so that relevant regulations can be accurately retrieved from the large database of classification criteria.

[0058] It is worth noting that only the business domain of the document to be classified is needed. However, in some other embodiments, considering that the content of the document to be classified may be highly concentrated or may involve multiple overlapping domains, the business domain of the document to be classified may not be able to fully cover the content of the document to be classified. Therefore, the business domain of the group summary is needed. The combination of the business domains of the group summary can cover the content of the document to be classified, making the retrieval in the classification database more comprehensive and accurate.

[0059] In one implementation, a keyword-based matching method can be used. This is particularly suitable for scenarios requiring rapid classification based on explicit features such as filenames and titles. It can quickly assign domain labels by matching domain keywords in the retrieved text with a pre-defined domain feature dictionary.

[0060] In another implementation, a model-based intelligent classification method can be used. The full text or group summaries of the undetermined confidential document are taken as input, and a pre-trained text classifier is used for prediction; this is not limited to this approach.

[0061] S106. Based on the business domain, retrieve the corresponding confidentiality basis from the preset confidentiality basis database;

[0062] The "classification basis database" refers to a pre-built and maintained structured knowledge base that stores authoritative materials such as various documents, standards, and historical classification cases used to determine the classification level. "Recall" here refers to an information retrieval process, that is, searching for and retrieving relevant information from the knowledge base based on specific conditions (in this case, the business domain). The "classification basis" refers to the specific entries recalled from the classification basis database, which will serve as the criteria for subsequent judgments.

[0063] S107. Calculate the correlation between the group summary and the basis for classification, and generate the corresponding correlation score for the group summary.

[0064] It should be noted that the degree of relevance can be understood in a sense as the closeness between the semantic connotation of the group summary and the circumstances or standards stipulated by the basis for classification.

[0065] The criteria for classifying information into classified information essentially describe a series of judgment standards that require classification. The higher the relevance of a summary to a particular criterion for classification, the more the information content carried by the summary matches the conditions or scenarios described by that criterion, both semantically and factually. Therefore, this relevance score can be regarded as a logical confidence or probability indicator; the higher the score, the greater the likelihood that the content will be classified according to this standard.

[0066] In some embodiments, the semantic similarity determination method in step S103 can be used to determine the degree of association.

[0067] In other embodiments, the relevance may depend on the keywords shared between the confidentiality criteria and the group summary. The more and more important the shared keywords (i.e., the rarer they are in the entire knowledge base), the higher the relevance. This is not limited here.

[0068] In some preferred embodiments, this can be achieved using a vector space model. Specifically, the text content of the group summary and the basis for classification can first be converted into corresponding file vectors using models such as TF-IDF, Word2Vec, or BERT. Then, the correlation between any two file vectors is quantified by calculating the cosine similarity, Euclidean distance, or other vector space distance. The closer the cosine similarity value is to 1, the more relevant the two file contents are, and the higher the correlation.

[0069] S108. Based on the relevance score, the group summaries are aggregated to form a summary set;

[0070] It should be noted that the aggregation here is more of a screening and collection process, that is, selecting those parts that are highly relevant to the classification criteria from all group summaries based on their relevance scores. The scenario is that it is necessary to select the parts that truly pose a risk of classification from numerous group summaries and filter out other irrelevant noise information (group summaries).

[0071] S109. Based on the summary set, generate a confidentiality result.

[0072] It should be noted that there are already many mature implementation methods for generating the final confidentiality result based on the digest set, which will not be elaborated here, or you can refer to the detailed description in the subsequent preferred embodiments.

[0073] It is important to emphasize that the core value proposition of this embodiment lies in the fact that, through a series of processes from S101 to S108, a large and complex file of documents awaiting classification has been successfully transformed into a highly condensed and semantically clear set of key evidence (summary set). This transformation process filters out a large amount of redundant information unrelated to classification and does not overlook any potential key information.

[0074] As can be seen, compared to full-text processing in related technologies, this technical solution decomposes the document to be classified into paragraphs and extracts summaries. Then, through semantic clustering and generating group summaries, it extracts and refines the large and complex paragraphs into a few concise core elements (i.e., group summaries). This reduces the amount of data for subsequent analysis and avoids the problem of key classified points being diluted by massive amounts of non-classified content. Furthermore, because group summaries are the crystallization of paragraph semantics, they crucially preserve the contextual relationships and logical structure between words while refining the text. It is this complete preservation of "contextual relationships" that provides a solid technical foundation for subsequent stages to identify and capture implicit classified points that are not defined by fixed keywords but are implied in the contextual logical relationships, thereby improving the accuracy of classified point identification and thus the accuracy of the auxiliary classification results. Simultaneously, determining the relevant business domain ensures that the retrieval from the classification basis database is relevant and targeted, and the correlation between group summaries and classification basis is used to improve the accuracy of the auxiliary classification results.

[0075] In a preferred embodiment, step S108 can be further divided into: S1081, setting a correlation degree screening criterion.

[0076] It should be noted that the screening criteria can be a preset relevance threshold, such as a value between 0 and 1 (e.g., 0.7); or, the screening criteria can be a preset quantity or proportion, such as selecting the top-N group summaries with the highest relevance scores, or selecting the group summaries with the top 20% relevance scores. This application does not impose specific limitations on this.

[0077] S1082. Based on the aforementioned screening criteria, all group abstracts are screened.

[0078] S1083. The group summaries retained after filtering constitute the summary set.

[0079] In actual use, there may be a problem in step S108, namely that the expression in the document to be classified may be inconsistent with the expression of the basis for classification (actual semantics are related), and the degree of inconsistency in group summary may be more serious, resulting in poor recall rate of subsequent classified document knowledge base.

[0080] In some preferred embodiments, therefore after step S108, the method further includes:

[0081] S201. For each group of abstracts in the abstract set, perform word segmentation to obtain multiple constituent words;

[0082] Among them, constituent words refer to the word sequences obtained after word segmentation, which are the smallest units that constitute the semantics of the original group summary.

[0083] S202. For a single constituent word, retrieve the first synonym and / or first abbreviation corresponding to the constituent word from the preset thesaurus and / or abbreviation corpus, and replace the corresponding constituent word in the group summary to obtain a summary variant;

[0084] Specifically, each constituent word in the original group summary is traversed and used as a query key to search a pre-defined knowledge base. This knowledge base preferably includes a thesaurus and an abbreviation database. When the first synonym or abbreviation corresponding to a constituent word is found, a replacement operation is performed, that is, the corresponding constituent word in the original group summary is replaced with the retrieved new word, thereby generating one or more summary variants.

[0085] It's easy to understand that the essence of this step is to find variants for the key constituent words and use these variants to create multiple summary variants that are equivalent to the original group summary but have different expressions. An original group summary may generate a set containing multiple summary variants because it contains multiple replaceable words. This set (containing the original summary and all its summary variants) will be used together as input for subsequent matching steps.

[0086] It is worth noting that the core of this step is to enrich the expression of the query by utilizing external knowledge bases, bridging the differences in usage habits between different text sources, thereby overcoming the recall omission problem caused by the lexical gap. By actively generating summary variants containing multiple equivalent expressions, we strive to ensure that no matter what version of the classified documents in the subsequent classified document knowledge base is used, at least one version of the summary in the expanded set can be successfully matched, thereby improving the recall rate and robustness.

[0087] S203. Select at least one summary variant as a group summary and expand it into the summary set.

[0088] It should be noted that not all generated abstract variants are adopted; there may be a selection criterion.

[0089] In some embodiments, the selection strategy here can be very simple, such as selecting all generated variants;

[0090] In other embodiments, for example, only those variants whose replacement words come from a high-confidence professional thesaurus may be selected, or each original summary may be limited to a maximum of N variants to prevent the summary set from becoming excessively bloated. No further limitations are specified here.

[0091] As can be seen, based on the existing summary set, by retrieving and replacing corresponding constituent words in the group summary from a pre-defined thesaurus and / or abbreviation database, semantically identical but expressively diverse summary variants can be generated. Expanding the summary variants into the summary set increases the robustness of subsequent recall; even if the group summary does not completely match the classification criteria, an effective match can be established between the summary variants and the classification criteria, thereby improving the recall rate of the subsequent classified document knowledge base.

[0092] In some preferred embodiments, there are security and compliance risks when the summary set is used for subsequent retrieval or generation tasks. Therefore, in some preferred embodiments...

[0093] Before step S109, at least one of the following operations is performed:

[0094] It should be noted that the following are several exemplary implementations of the aforementioned at least one operation, and there is no strict execution order among them; they can be executed individually or in combination:

[0095] S301. Perform instruction cleaning on the group summary to prevent instruction injection risks.

[0096] Instruction cleaning refers to scanning the text content of the group summary to identify and process any errors that may occur during the subsequent execution of the recalled file set and the identification and execution of operation instructions in the summary.

[0097] In one specific embodiment, a pre-defined "instruction feature library" is constructed. This library contains various potentially dangerous instruction patterns, such as SQL command keywords, system command lines, and script language syntax. The text content of each group summary is matched against the instruction feature library for pattern matching. Once an operation instruction string matching a pre-defined pattern is found, a pre-defined cleaning strategy is executed. This cleaning strategy includes, but is not limited to:

[0098] Remove the matched instruction string completely from the group summary.

[0099] Replace the instruction string with an empty string.

[0100] Special characters in the instruction string are escaped so that they lose their ability to be executed as instructions and exist only as ordinary text; no restrictions are imposed here.

[0101] S302. Identify sensitive terms in the group summary and perform hierarchical mapping and desensitization replacement;

[0102] Sensitive terms refer to specific information contained in the text that could cause harm if leaked. Hierarchical mapping involves assigning identified sensitive terms to a pre-defined security classification system based on their sensitivity level (e.g., top secret, confidential, secret), providing a basis for subsequent differentiated processing. De-identification replacement involves replacing the original sensitive terms with a placeholder that does not contain specific sensitive information but retains its type.

[0103] It should be noted that before using group summaries as query statements to retrieve information or generate conclusions, they should be checked to ensure that they do not contain any sensitive information, thus preventing the leakage of sensitive information in subsequent processing.

[0104] In one specific embodiment, this step can be implemented in the following way:

[0105] Each group summary is processed using a pre-built dictionary of sensitive terms. This dictionary stores a list of sensitive terms and labels each term with its corresponding sensitivity level and a placeholder for the type of replacement. The processing flow is as follows:

[0106] The text content of the group summary is matched against a dictionary of sensitive terms to identify one or more sensitive terms contained therein.

[0107] For identified sensitive terms, look up their preset sensitivity level and corresponding type placeholder in the dictionary.

[0108] The sensitive terms in the original group summary are replaced with the type placeholders found in the query, generating a new group summary that has been desensitized.

[0109] The final anonymized group summary retains the original sentence structure and semantics of non-sensitive parts while masking core sensitive entity information. This anonymized group summary will be used securely in all subsequent steps.

[0110] For example:

[0111] Group summary: Technical parameters of a certain xx.

[0112] Rewrite rules:

[0113] Security Classification Label: Add a security classification label (such as `[Confidential]`) and hide the specific project name.

[0114] Blurring: Replace "xx" with items from a specific domain.

[0115] Final Summary: [Confidential] Technical Specifications for Domain-Specific Projects.

[0116] Effect: It maintains the query intent while avoiding the direct exposure of sensitive information.

[0117] S303. Embed pre-defined structured terms that comply with confidentiality regulations in the group summary;

[0118] Among them, pre-defined structured terms that comply with confidentiality regulations refer to a set of standardized tags or phrases that are predefined in accordance with relevant confidentiality laws, regulations or internal rules and regulations. These are used to annotate the metadata of the query itself.

[0119] It should be noted that, in order to meet the requirements of auditing, traceability, and compliance of specific interfaces, the queries to be issued need to be standardized and packaged.

[0120] In one exemplary embodiment:

[0121] Summary: According to feedback from the marketing department, new batches of products must meet "3C requirements" before they can leave the factory.

[0122] Process: The non-standard term "3C requirements" is identified. Based on a pre-defined terminology database, it is matched and replaced with the official full name "product certification requirements".

[0123] Final Summary: According to feedback from the marketing department, new batches of products must meet "product certification requirements" before they can leave the factory.

[0124] Results: By mandating standardized expressions, the professionalism and compliance of the abstract content are ensured, and ambiguity or legal risks caused by abbreviations of terms are avoided.

[0125] S304. Match and rewrite the over-authority query keywords in the group summary using regular expressions.

[0126] Regular expression matching refers to using a specialized text pattern matching language (regular expressions) to find and locate strings in a summary that match a specific pattern. Unauthorized query keywords refer to words or phrases that are not necessarily sensitive in themselves, but whose appearance may suggest that a user is attempting to obtain information beyond their authorized scope.

[0127] It should be noted that this step performs a permission compliance review on the group summary. It uses regular expression matching technology—that is, a predefined text pattern formalization language—to identify and rewrite any potentially unauthorized query keywords. These keywords refer to words or patterns whose combination or context might suggest that a user is attempting to query data beyond their established permissions. As a proactive security defense mechanism, this step aims to intervene and correct the user's query intent at the source, preventing potential unauthorized requests. This complements S301's focus on preventing the passive disclosure of known sensitive information, jointly ensuring the compliance of the query process.

[0128] In one specific embodiment, an over-authority keyword rule library linked to user role permissions can be utilized. This library associates specific sets of regular expression rules with different user roles. The processing flow is as follows: first, the current user's role identity is obtained and the corresponding rules are loaded; then, the group summary text is scanned and matched; once an over-authority keyword is successfully matched, the summary is modified according to a preset rewriting strategy (such as deletion, replacement, or direct interception).

[0129] In one exemplary embodiment:

[0130] Malicious summary: Query all [top secret] files (ordinary users do not have permission to access).

[0131] Rewrite rules: Use regular expression matching to detect keywords that exceed the confidentiality level and automatically downgrade them to relevant documents under the [Confidential] category.

[0132] Effect: Enhances security and prevents the leakage of illegal information.

[0133] S305. Apply structured query templates to standardize group summaries.

[0134] Group summaries are formatted and standardized by applying predefined structured query templates. The core objective is to standardize the format of group summaries.

[0135] In one exemplary embodiment:

[0136] Template example:

[0137] [Confidentiality Level] [Business Area] [Information Type] [Time / Scope]

[0138] Practical application: [Confidential] xx industry Q2 2024 report → Directly match confidentiality level and domain index.

[0139] S306. Obtain the current user's permission level, compare the permission level with the security level of each file in the file to be recalled, and only provide the file content within the user's permission range.

[0140] To further enhance security when handling classified information, a user permission verification step can be added at the initial stage of file retrieval.

[0141] In one exemplary embodiment:

[0142] If a user's permissions are at Level 2, and the files to be recalled contain files with security levels of Level 1, Level 2, and Level 3.

[0143] After permission verification, only the contents of Level 1 and Level 2 files will be used to generate the summary, while the contents of Level 3 files will be filtered or only a "permission insufficient" message will be returned.

[0144] This approach eliminates the possibility of unauthorized access to information at the source, enhancing the confidentiality and security of the entire process.

[0145] It is evident that by pre-cleaning the group summary of instructions, the risk of instruction injection is prevented, and misoperation is avoided. Secondly, identifying and de-identifying sensitive terms in the group summary can prevent the leakage or recording of real sensitive information during processing without affecting semantic analysis, achieving a balance between data availability and confidentiality. Thirdly, embedding pre-defined structured terms standardizes internal and non-standard expressions, ensuring that all content interacting with external systems or generating final reports complies with confidentiality regulations and unit management requirements. Fourthly, by matching and rewriting unauthorized query keywords using regular expressions, query intentions that may lead to illegal information leakage can be monitored and blocked in real time. Then, by embedding pre-defined structured terms in the group summary and applying structured query templates for standardization, not only is the professionalism and compliance of the summary content ensured, avoiding ambiguity or legal risks caused by non-standard expressions, but finally, adding a user permission verification step at the initial stage of document retrieval further enhances security when processing classified information.

[0146] In practical use, if the documents whose security level is subsequently determined are limited to the summary set, it may lead to poor matching ability for specific key terms, resulting in poor recall and low precision.

[0147] Please see Figure 2 and Figure 3 , Figure 2 This is another flowchart illustrating the auxiliary confidentiality determination method in the embodiments of this application. Figure 3 This is a schematic diagram of the architecture of the auxiliary confidentiality determination method in the embodiments of this application;

[0148] To further improve this application, in a preferred embodiment, step S109 specifically includes:

[0149] S401. Extract keywords for each group of abstracts in the abstract set;

[0150] In some embodiments, a word's importance is directly proportional to its frequency (TF) in the current summary and inversely proportional to its frequency (IDF) across the entire summary set (corpus). Simply put, a word that appears frequently in the current summary group but rarely in other summary groups is likely a keyword for that summary. Specifically, the TF of each word in the current summary group is first calculated to measure its local importance; then, the IDF of that word across the summary set is calculated to measure its global scarcity. The TF-IDF score, obtained by multiplying the two, quantifies the uniqueness and representativeness of a word to the current summary group. Therefore, a word that appears frequently in the current summary group but is relatively rare in other summary groups will receive a high TF-IDF score and is thus identified as a keyword.

[0151] In other embodiments, the importance of a word depends on the importance and number of other words linked to it. If a word co-occurs with many other important words in the text, it is also likely to be a keyword, without limitation here. Specifically, words in the group summary are treated as nodes in a graph, and edges are established between words co-occurring within a certain window to construct a word graph. Words corresponding to nodes in the graph with a number of connected edges greater than a preset threshold are identified as keywords, without limitation here.

[0152] In some preferred embodiments, to avoid excessive subsequent query tasks, the number of keywords is limited, for example, it cannot exceed a fixed threshold.

[0153] S402. Retrieve the second synonym and / or second abbreviation corresponding to the keyword from the preset thesaurus and / or abbreviation database;

[0154] It should be noted that the preset thesaurus and / or preset abbreviation database in this step, and the preset thesaurus and / or preset abbreviation database in step S202, can refer to the same unified thesaurus and / or abbreviation database entity, or they can refer to two or more physically or logically independent databases. The core of this invention lies in the function of using these databases for retrieval, rather than the number of databases or their implementation. Therefore, all scenarios that can achieve the above function, whether using a single shared database or multiple independent databases, fall within the protection scope of this invention.

[0155] The principle and process of this step are similar to those of step S202. The relevant principles and processes can be found in step S202 and will not be repeated here.

[0156] In some preferred embodiments, in order to avoid excessive subsequent query tasks, the number of extended terms (second synonyms and / or second abbreviations) is limited, for example, it cannot exceed a fixed threshold, or it can be combined with the corresponding keywords to not exceed a fixed threshold, which is not limited here.

[0157] S403. Vectorize each group summary in the summary set to obtain a group summary vector;

[0158] It is worth noting that this step is started in parallel with S401 / S402.

[0159] It should be noted that there are already many mature methods for vectorizing text, which will not be elaborated here.

[0160] S404. Use keywords and / or second synonyms and / or second abbreviations to perform text matching on the file titles and / or key information in the preset classified document knowledge base to obtain the first candidate file set;

[0161] The classified document knowledge base refers to a database or file system that stores a large number of historical classified documents and their related information (such as titles, classification levels, and key points). Key points refer to specific text paragraphs in a document that are explicitly identified as classified. The first candidate document set refers to a list of classified documents directly related to the query keywords, retrieved through a text matching channel.

[0162] It should be noted that the keywords and / or second synonyms and / or second abbreviations are searched in the classified document knowledge base. All files containing these keywords (or their synonyms / abbreviations) are obtained and sorted according to their matching degree (such as word frequency, document frequency, etc.). This sorted list of files constitutes the first candidate file set.

[0163] S405. Calculate the similarity between the group summary vector and the pre-stored file vectors in the classified document knowledge base to obtain the second candidate file set;

[0164] The principle and process of similarity calculation in this step are similar to those in steps S103 and S107. The relevant principles and processes can be referred to in steps S103 and S107, and will not be repeated here.

[0165] Continuing the previous example: After similarity calculation, a series of similarity scores will be obtained between the group summary vector and each pre-stored file vector in the knowledge base. Next, a second candidate file set will be constructed from these classified files using at least one of the following preset filtering strategies:

[0166] Set a similarity threshold (e.g., a similarity score greater than 0.85). Iterate through all similarity scores and automatically select classified documents with scores higher than this preset threshold to form a second candidate file set.

[0167] All classified documents are sorted in descending order of their similarity scores to the group summary vector, and then the top N documents (e.g., the Top 5) are selected as the second candidate document set. No specific restrictions are imposed here.

[0168] S406. Merge the first candidate document set and the second candidate document set to obtain a list of classified documents;

[0169] It should be noted that by leveraging the complementary strengths of different retrieval strategies, a better final result is obtained than any single strategy. This approach retains the precision of text matching (files appearing in the first candidate file set) while incorporating the breadth of vector retrieval (files appearing in the second candidate file set), thus achieving both high recall and high precision in the final list of classified documents.

[0170] In some other preferred embodiments, in order to improve the processing quality, a pre-screening is performed on the list of classified documents to remove low-quality candidate documents. The screening is based on the similarity calculated in step S405 and the text matching degree in step S404. Candidate documents with similarity or matching degree lower than a preset threshold will be filtered out, which will not be elaborated here.

[0171] S407. Based on the list of classified documents, generate the classification result.

[0172] It is evident that by expanding the keywords of the group summary with synonyms and / or abbreviations, semantic breadth is broadened, enabling the recall of documents that are semantically related to the group summary but use different terminology during subsequent text matching. Using group summary vectors for similarity calculation encodes the deep semantic information of the group summary as a point in a high-dimensional vector space. By calculating the spatial distance between vectors, documents that are related in core concepts and topics but not entirely expressed can be recalled, understanding the context and potential language to achieve deep semantic recall. Finally, by merging the first and second candidate document sets, a combination of precise matching and fuzzy semantics is achieved, avoiding the potential omission of keywords in semantic retrieval and compensating for the limitation of keyword retrieval in understanding deep semantics.

[0173] In practical use, merging the first and second candidate document sets, combining the advantages of different candidate document sets, can uncover a large number of documents conceptually related to the query content. However, in scenarios with complex group summaries, the high information density may lead to the risk of losing specific key information.

[0174] Therefore, in some preferred embodiments, before step S401, the following steps are also included:

[0175] S501. Obtain the text length of each group summary and determine the number of target keywords based on the text length;

[0176] Text length refers to a quantitative indicator used to measure the amount of text content, which can be expressed as the number of characters, the number of words, or the number of tokens under a specific model.

[0177] In some embodiments, for each group summary, its text length is calculated, and a preset function or rule is applied to determine the number of target keywords based on this length value; the longer the length value, the more target keywords are included.

[0178] In some preferred embodiments, the criteria for determining the number of target keywords are as follows: In the formula, For the target number of keywords, For text length, is the disturbance coefficient, with a value range of [0.5, 1].

[0179] It can be seen that using a logarithmic function To adapt to the growth law of information entropy, that is, as the text length increases, the marginal benefit of new information decreases, the number of extracted keywords can be reasonably adjusted adaptively with the change of text length.

[0180] S502. Based on the number of target keywords, extract refined keywords from the group abstract;

[0181] The principle and process of this step are similar to those of step S401. The relevant principles and processes can be found in step S401 and will not be repeated here.

[0182] S503. Abbreviate the group summary based on the refined keywords, where the refined keywords are used to generate a new group summary to replace the original group summary, and the original group summary is used to constrain the new group summary.

[0183] Here, the abbreviation refers to an action of text reconstruction, namely, using refined keywords to construct a completely new and shorter version of the summary. Constraints represent a limitation and guidance, namely the complete semantics and context of the original group summary, providing background on how to organize and understand the new group summary, ensuring that the meaning of the new group summary does not deviate from the main idea of ​​the original text.

[0184] In some embodiments, the extracted refined keywords are concatenated in the order in which they appear in the original text. The constraint here is that the order of the keywords preserves a certain logic of the original text.

[0185] In other embodiments, the system can use the original group summary as context to guide a generative model (such as GPT) to write a grammatically correct short sentence containing only these refined keywords. The original group summary ensures that the model does not fabricate relationships out of thin air during generation, but rather organizes them around the theme of the original text.

[0186] It is evident that while group summaries of long texts are complete in information, they may also introduce noise due to their large data volume. Therefore, by determining the number of target keywords based on the text length and extracting refined keywords, the group summaries are then abbreviated based on these refined keywords to generate more concise and less noisy group summaries. This abbreviated group summaries are then used for subsequent text matching and similarity calculations, making the first and second candidate document sets more relevant and improving the ability to retrieve specific key information, thereby enhancing the purposefulness and accuracy of the recall.

[0187] The above embodiments S101 to S109 and embodiments S401 to S407 are based on the summary set and the list of classified documents, respectively. In order to more clearly illustrate how to generate the final classification result based on the information in the summary set and the list of classified documents, a specific and exemplary implementation method for step S407 will be provided below.

[0188] In some embodiments, step S407 specifically includes:

[0189] S601. Calculate the semantic correlation between the list of classified documents and the group summary, and select reference documents based on the semantic correlation.

[0190] The principle and process of semantic relevance in this step are similar to those in step S107. The relevant principles and processes can be found in step S107 and will not be repeated here.

[0191] It should be noted that although the list of classified documents has been filtered, it is still quite large. If all this evidence of varying quality is provided to the subsequent inference model, it may not only exceed the model's context length limit, but also introduce noise and interfere with the model's judgment. Therefore, a final filtering is performed in this step. The specific process can refer to any of the following implementation methods:

[0192] Set a predefined semantic relevance threshold, iterate through all classified documents and their relevance scores, and only retain classified documents with scores higher than or equal to the threshold to obtain reference documents.

[0193] The classified documents are sorted from highest to lowest according to their semantic relevance scores. Then, the top N documents are selected from the sorted list as reference documents. In some preferred embodiments, if the number of classified documents that meet the requirements is less than or equal to N, all of them are retained; this is not a limitation here.

[0194] S602. Concatenate a preset classification task instruction, a reference file, and a file to be classified to obtain the classification instruction text.

[0195] The pre-defined, confidential task instruction refers to a carefully designed natural language text, or prompt, used to guide the inference-generative model on how to perform its task. It defines the model's role, task objective, and output format. Concatenation refers to joining multiple independent text blocks into a single, continuous text string according to a specific logical order. The confidential instruction text is the final, complete input text that can be directly fed into the inference-generative model after concatenation; it contains all information about the task requirements, reference cases, and the problem to be addressed.

[0196] S603. Input the classification instruction text into the inference generation model to obtain the classification result; wherein, the classification result includes the classification level, the basis for classification, the confidentiality period and the confidence level.

[0197] It is evident that by concatenating the pre-set classification task instructions, reference documents, and documents to be classified, where the reference documents are documents with relatively high accuracy and relevance, when the model receives this structured classification instruction text, its reasoning process is guided and constrained within the scope of the provided reference documents, rather than relying on documents that may contain biases or be outdated, thereby suppressing the knowledge illusion phenomenon of large language models.

[0198] The foregoing embodiments have described each part of this application separately. To facilitate a comprehensive understanding of the overall concept of the application, the following will be combined with... Figure 3 To aid understanding (it should be noted that the following embodiments mainly illustrate a part of the framework of the method of this application, rather than presenting all possible implementation details). Figure 3 The diagram illustrates the system architecture and data flow of an auxiliary confidentiality determination method provided in one embodiment of this application. The following will describe part of the method's flow with reference to this diagram.

[0199] The entire solution is first divided into thinking (task breakdown), searching, reranking, reading, and generating.

[0200] Thinking (Task Decomposition): Undetermined Classified Documents → Task Decomposition (Summary Extraction) (Please refer to steps S101 to S104) → Set of Classified Document Summaries → Summary Expansion (Please refer to steps S201 to S203) → Set of Rewritten Variations of Summary 1, Summary 2, or Summary 3 → Iterate through all summaries and their corresponding variations → Generate text abbreviations;

[0201] Search: Path 1: Keyword Recall Keyword Extraction (see step S401) → Synonym Expansion (see step S402) → Construct Search Expression → (Expression Search) Confidential Document Knowledge Base (title + key points) → Keyword Recall Confidential Document Set (see step S404);

[0202] Path 2: Abstract vectorization (see step S403) → (Vector query) Classified document knowledge base (title + key points) embedding → Retrieve classified documents with high similarity (see step S405);

[0203] Path 3: Keyword recall of classified document set and recall of classified documents with high similarity → candidate classified document set (see step S406);

[0204] Rerank: Path 4: Deduplication and truncation → Select TopN (see step S405);

[0205] Path 5: Summary, Top N selection and relevance assessment reading task instructions → Prompt concatenation;

[0206] Reading: Prompt assembly → Large model → Calculate correlation (see step S601) → Select Top 3 according to correlation (see step S601);

[0207] Generate: Top 3 similar classified documents, documents to be classified, and summary and classification instructions) → prompt concatenation (see step S602) → classification model (see step S603) → return classification result (see step S603).

[0208] The following describes an exemplary auxiliary confidentiality system 700 provided in an embodiment of this application. Figure 4 This is an exemplary hardware structure diagram of the auxiliary security system 700 provided in the embodiments of this application.

[0209] In some embodiments, the auxiliary confidentiality system 700 is a computer device or includes a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0210] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0211] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0212] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0213] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0214] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. An auxiliary method for determining confidentiality, characterized in that, include: Receive a file to be encrypted and divide it into multiple text segments; Extract a paragraph summary for each of the text paragraphs; Based on the semantic similarity between the paragraph summaries, all the paragraph summaries are clustered into multiple summary groups; a group summary is generated for each summary group. Determine the business domain of the pending confidential document or the group summary; Based on the business domain, retrieve the corresponding confidentiality basis from the preset confidentiality basis database; Calculate the correlation degree between the group summary and the confidentiality basis, and generate the corresponding correlation degree score for the group summary; Based on the relevance score, the group summaries are aggregated to form a summary set; Based on the summary set, a keying result is generated; keywords are extracted for each group of summaries in the summary set; Retrieve the second synonym and / or second abbreviation corresponding to the keyword from the preset thesaurus and / or abbreviation corpus; Each group summary in the summary set is vectorized to obtain a group summary vector. By using keywords and / or the second synonyms and / or the second abbreviations, text matching is performed on the file titles and / or key information in the preset classified document knowledge base to obtain the first candidate file set; The similarity between the group summary vector and the file vectors pre-stored in the classified document knowledge base is calculated to obtain a second candidate file set; The first candidate file set and the second candidate file set are merged to obtain a list of classified documents; Based on the list of classified documents, the classification result is generated.

2. The method according to claim 1, characterized in that, After the step of aggregating the group summaries into a summary set based on the relevance score, the method further includes: For each group of summaries in the summary set, word segmentation is performed to obtain multiple constituent words; For a single constituent word, retrieve the first synonym and / or first abbreviation corresponding to the constituent word from a preset thesaurus and / or abbreviation corpus, and replace the corresponding constituent word in the group summary to obtain a summary variant; At least one of the said summary variants is selected as the group summary and expanded into the summary set.

3. The method according to claim 1, characterized in that, The step of generating the classification result based on the list of classified documents specifically includes: Calculate the semantic correlation between the list of classified documents and the group summary, and filter out reference documents based on the semantic correlation. A pre-defined classification task instruction, the reference file, and the file to be classified are concatenated to obtain the classification instruction text; The classification instruction text is input into the inference generation model to obtain the classification result; wherein, the classification result includes the classification level, the basis for classification, the confidentiality period and the confidence level.

4. The method according to claim 1, characterized in that, Before the step of extracting keywords for each group of summaries in the summary set, the method further includes: Obtain the text length of each group summary, and determine the number of target keywords based on the text length; based on the number of target keywords, extract refined keywords from the group summaries; The group summary is abbreviated based on the refined keywords, wherein the refined keywords are used to generate a new group summary to replace the group summary, and the group summary is used to constrain the new group summary.

5. The method according to claim 4, characterized in that, The criteria for determining the number of target keywords are as follows: In the formula, x is the number of target keywords, len is the text length, and r is the perturbation coefficient, with a value range of [0.5, 1].

6. The method according to claim 1 or 2, characterized in that, Before the step of generating a confidentiality result based on the digest set, the method further includes performing at least one of the following operations: Perform instruction cleaning on the group summary; Identify sensitive terms in the group summary and perform hierarchical mapping and desensitization replacement; The pre-defined structured terminology that complies with confidentiality regulations is embedded in the group summary; The unauthorized query keywords in the group summary are matched and rewritten using regular expressions; The group summaries are standardized using a structured query template. Obtain the current user's permission level, compare the permission level with the security level of each file in the file to be recalled, and only provide the file content within the user's permission range.

7. An auxiliary security system, characterized in that, The auxiliary secrecy system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the auxiliary secrecy system to perform the method as described in any one of claims 1-6.

8. A computer program product containing instructions, characterized in that, When the computer program product is run on the auxiliary secrecy system, it causes the auxiliary secrecy system to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the auxiliary secrecy system, it causes the auxiliary secrecy system to perform the method as described in any one of claims 1-6.

Images