A method and system for automatically classifying unstructured data based on business semantic understanding

Abstract

A method and system for automatic classification of unstructured data based on business semantic understanding is disclosed, relating to the domain of computer systems based on a specific computational model. This method constructs a classified domain ontology library containing business concept nodes and logical relationship edges, and builds sensitive semantic template graphs with different security levels based on the ontology library. For the unstructured data to be classified, entity instances and their dependencies are extracted using an entity relationship extraction model and mapped to the ontology library to form a subgraph of facts to be tested. By calculating the matching score between the subgraph of facts to be tested and each sensitive semantic template graph, the security level corresponding to the template graph with the highest matching score exceeding a threshold is selected as the classification result. This application aims to improve the efficiency of unstructured data sharing and analysis across departments.

Description

Technical Field

[0001] This application belongs to the field of computer systems based on specific computational models, and in particular relates to a method and system for automatic classification of unstructured data based on business semantic understanding. Background Technology

[0002] With the deepening of the construction of the energy internet and digital power grid, the power industry has accumulated massive amounts of unstructured data in production operation, equipment maintenance and marketing services. This data contains sensitive information such as power grid topology, equipment operating parameters and user electricity consumption behavior. Accurate automatic classification of unstructured data is a prerequisite for achieving secure data sharing.

[0003] In related technologies, deep learning-based natural language processing techniques are widely used for automated classification. By supervising the training of large-scale labeled power documents, the model can capture the contextual semantic features in the text, thereby automatically identifying text fragments containing hidden sensitive information and predicting their security level. This method based on semantic vector representation improves the automation and generalization ability of unstructured data classification, enabling the system to infer potential sensitive attributes based on lexical and syntactic features when faced with unseen document descriptions.

[0004] However, in the context of the State Grid's equipment defect lifecycle management, there are practical limitations in its application. For example, a defect report, during the processing stage, involves real-time operational risks and is considered highly sensitive data. The model's classification of it as high-risk based on textual characteristics is accurate. However, once the defect enters the eliminated or archived stage, its business attributes transform into historical data for reference, and its sensitivity objectively decreases. Because the relevant technologies primarily classify based on the static semantic features of the text content, the model's output classification results remain consistent across different business stages, even when the report text content remains unchanged. This results in a large amount of desensitized historical data being continuously locked at a high security level, thus limiting the efficiency of data flow during cross-departmental sharing and analysis. Summary of the Invention

[0005] This application provides a method and system for automatic classification of unstructured data based on business semantic understanding, which can improve the efficiency of unstructured data circulation during cross-departmental sharing and analysis.

[0006] Firstly, this application provides an automatic classification method for unstructured data based on business semantic understanding. This method involves constructing a classified domain ontology library, which includes business concept nodes and logical relationship edges connecting these nodes. Based on this library, sensitive semantic template graphs corresponding to different security levels are constructed. These sensitive semantic template graphs are directed graph structures composed of concept nodes representing sensitive information and logical relationship edges representing business logic. A pre-defined entity relationship extraction model is used to process the unstructured data to be classified, outputting a set of entity instances and the dependency relationships between these instances. The set of entity instances is mapped to business concept nodes in the classified domain ontology library, and dependency relationships are mapped to logical relationship edges, constructing a test fact subgraph representing the semantic structure of the unstructured data to be classified. The test fact subgraph is then matched with each sensitive semantic template graph to obtain a matching score. The sensitive semantic template graph with the highest matching score (greater than a preset threshold) is selected as the target template graph. The security level associated with the target template graph is output as the classification result of the unstructured data to be classified.

[0007] By employing the aforementioned technical solution, and constructing a hierarchical domain ontology library and a sensitive semantic template graph, combined with an entity relationship extraction model to process unstructured data, the extracted entity instances and dependencies can be mapped onto the domain ontology to form a subgraph of facts to be tested. The degree of matching with the sensitive semantic template is calculated through graph structure matching, thereby determining the data security level. This semantic understanding-based hierarchical method can deeply understand the business meaning of data content, improving the accuracy of data hierarchical classification. Compared to traditional methods relying solely on keyword matching, this solution can better identify implicit sensitive information associations in the data, reducing hierarchical bias caused by incomplete literal information. By introducing business semantic understanding, the hierarchical method's ability to understand context is enhanced, strengthening its robustness in handling complex unstructured data. The graph structure-based matching mechanism makes the hierarchical process interpretable, facilitating the analysis and optimization of hierarchical results, and improving the credibility of the hierarchical results.

[0008] In conjunction with some implementations of the first aspect, in some implementations, the entity instance set is mapped to business concept nodes in the hierarchical domain ontology, and dependency relations are mapped to logical relation edges to construct a test fact subgraph representing the semantic structure of the unstructured data to be hierarchically classified. Specifically, this includes: using each entity instance in the entity instance set as a graph node and dependency relations as graph edges to construct an initial discrete semantic network; for node pairs with semantic gaps or missing connections in the discrete semantic network, retrieving potential inference chains connecting the node pairs in the hierarchical domain ontology, where the potential inference chain is a path composed of intermediate concept nodes and their logical relations that do not explicitly appear in the unstructured data in the ontology; calculating the relevance evaluation value of the potential inference chain and using the potential inference chain with the highest relevance evaluation value as the completion path; and integrating the intermediate concept nodes and logical relations in the completion path into the discrete semantic network to obtain the test fact subgraph representing the semantic structure of the unstructured data to be hierarchically classified.

[0009] By employing the aforementioned technical solution, during the construction of the fact subgraph to be tested, potential reasoning chains connecting nodes with semantic gaps are retrieved from the hierarchical domain ontology library, and the chain with the highest relevance is selected for semantic completion, thereby expanding the semantic integrity of the fact subgraph to be tested. This semantic completion mechanism compensates for the information gaps in unstructured data representation and improves the comprehensiveness of semantic understanding. By introducing intermediate concept nodes from the ontology library, the ability of the fact subgraph to be tested to express implicit semantic relationships is enhanced. Compared to directly using discrete semantic networks, the completed fact subgraph to be tested can more accurately reflect the semantic structure of the data content and reduce the probability of misjudgment caused by information discontinuity. This ontology-based semantic completion method improves the system's ability to understand ambiguous expressions and implicit relationships.

[0010] In conjunction with some implementation methods of the first aspect, in some implementation methods, graph structure matching calculations are performed between the subgraph of facts to be tested and each sensitive semantic template graph to obtain matching degree scores between the subgraph of facts to be tested and each sensitive semantic template graph. Specifically, this includes: extracting the semantic feature vectors of each node in the subgraph of facts to be tested and the sensitive semantic template graph respectively; calculating the semantic distance between each node based on the semantic feature vectors, and constructing a semantic mapping matrix that characterizes the degree of semantic alignment between each node; establishing a set of node-level correspondences between the subgraph of facts to be tested and the sensitive semantic template graph based on the semantic mapping matrix; identifying the projection edges of each logical relationship edge in the subgraph of facts to be tested onto the sensitive semantic template graph based on the set of node-level correspondences; determining the number of projection edges that have corresponding topological connections in the sensitive semantic template graph; calculating the proportion of the number of projection edges to obtain the structural isomorphism factor; and generating the matching degree scores between the subgraph of facts to be tested and each sensitive semantic template graph by weighted calculation of the semantic similarity probability distribution value and the structural isomorphism factor in the semantic mapping matrix.

[0011] By employing the aforementioned technical solution, graph structure matching calculations are performed by extracting node semantic feature vectors and constructing a semantic mapping matrix. Combined with topological isomorphism analysis, this achieves dual matching at both the semantic and structural levels. Distance calculation based on semantic feature vectors improves the accuracy of node mapping, while the introduction of structural isomorphism factors enhances the ability to recognize overall semantic patterns. This matching method, which comprehensively considers semantic similarity and structural consistency, improves the accuracy of sensitive information pattern recognition. The final matching score is generated through weighted calculation, ensuring that the matching results retain both the detailed features of semantic similarity and the overall features of structural correspondence, thus improving the reliability of hierarchical judgment. This multi-dimensional matching mechanism reduces misjudgments caused by single features and improves the stability of hierarchical results.

[0012] In conjunction with some implementations of the first aspect, in some implementations, before outputting the security level associated with the target template graph as the classification result of the unstructured data to be classified, the method further includes: obtaining entity nodes in the unstructured data to be classified that are not included in the fact subgraph to be tested, as a set of background nodes; calculating the word vector of each background node in the set of background nodes, and generating a global context vector of the unstructured data to be classified based on the word vector of the background nodes; obtaining the business domain vector pre-associated with the target template graph, the business domain vector being used to characterize the business scenario category to which the target template graph belongs; calculating the cosine similarity between the global context vector and the business domain vector as the scenario compatibility; if the scenario compatibility is determined to be less than a preset compatibility threshold, deleting the target template graph and then selecting the sensitive semantic template graph with a matching score greater than the preset threshold and the largest value as the target template graph; if the scenario compatibility is determined to be not less than the preset compatibility threshold, then outputting the security level associated with the target template graph as the classification result of the unstructured data to be classified.

[0013] By adopting the above technical solution, a global context vector is generated using a set of background nodes, and its similarity is calculated with the business domain vector of the template graph, thus improving the scenario adaptability of the classification results. This context-based filtering mechanism can identify cases where local features match but the overall scenario does not, reducing the probability of cross-scenario mismatches. A secondary filtering process using a compatibility threshold further improves the accuracy of the classification results. Incorporating background information not involved in graph matching enhances the classification system's ability to perceive the overall data context, improving the rationality of the classification results. This classification method, which considers the global context, improves the system's ability to identify the business scenario to which the data belongs, reducing classification errors caused by scenario misjudgment.

[0014] In conjunction with some implementation methods of the first aspect, in some implementation methods, a global context vector for the unstructured data to be classified is generated based on the word vectors of the background nodes. Specifically, this includes: counting the frequency of occurrence of the words corresponding to each background node in the background node set in the unstructured data to be classified; calculating the inverse document frequency of each background node based on the occurrence frequency and a preset corpus to obtain the weight value of each background node; using the weight value to perform a weighted summation of the word vectors of all background nodes in the background node set to obtain an initial context vector; and normalizing the initial context vector to obtain the global context vector for the unstructured data to be classified.

[0015] By employing the aforementioned technical solution, weight values ​​are obtained by statistically analyzing the frequency of background node words and calculating inverse document frequency using a pre-defined corpus. This ensures a reasonable balance of importance among background nodes when generating the global context vector, avoiding the problem of excessively high weights for ordinary words that might result from relying solely on word frequency. A weighted summation method comprehensively considers the semantic information of all background nodes, and normalization ensures the vector's standardization, enabling the generated global context vector to more accurately represent the overall semantic features of unstructured data. This approach improves the accuracy of the global context vector in expressing the text's theme, thereby enhancing the reliability of subsequent scene-compatible hierarchical results. The introduction of weight values ​​allows the system to better identify and retain background information with significant business characteristics, reducing interference from ordinary words in context understanding and thus improving the system's accuracy in judging document business scenarios.

[0016] In conjunction with some implementations of the first aspect, in some implementations, after performing graph structure matching calculations between the fact subgraph to be tested and each sensitive semantic template graph to obtain the matching degree scores between the fact subgraph to be tested and each sensitive semantic template graph, the method further includes: under the condition that all matching degree scores are less than a preset threshold, constructing a graph editing cost function for each sensitive semantic template graph; calculating the minimum cumulative edit distance to transform the fact subgraph to be tested into a sensitive semantic template graph based on the graph editing cost function; obtaining the total number of nodes contained in the sensitive semantic template graph, and normalizing the minimum cumulative edit distance using the total number of nodes to obtain a structural difference coefficient; selecting the sensitive semantic template graph with a structural difference coefficient less than a preset difference threshold and the smallest value as the final target template graph; and outputting the security level associated with the final target template graph as the classification result of the unstructured data to be classified.

[0017] By adopting the above technical solution, for cases where all matching scores are less than a preset threshold, a secondary matching mechanism based on graph edit distance is introduced. This allows the system to measure the structural transformation cost from the test fact subgraph to each template graph. The minimum cumulative edit distance is used as the evaluation metric, and the structural difference coefficient is obtained through normalization using the total number of nodes, making the degree of difference between template graphs of different sizes comparable. This approach improves the system's ability to handle incomplete matching scenarios. By selecting the template graph with the smallest structural difference coefficient as the final target, the system can find the most structurally similar sensitive information pattern even when direct matching fails. The graph edit distance-based matching mechanism improves the system's accuracy in identifying similar but not completely identical sensitive information patterns, reducing classification failures caused by subtle differences.

[0018] In conjunction with some implementation methods of the first aspect, in some implementation methods, a graph editing cost function is constructed for each sensitive semantic template graph. Specifically, this includes: determining the business concept category to which the node in the fact subgraph to be tested belongs in the hierarchical domain ontology library; for nodes that exist in the fact subgraph to be tested but not in the sensitive semantic template graph, if the business concept category to which the node belongs is a preset non-sensitive auxiliary category, then the cost of the node deletion operation is set to a preset first value; for nodes that exist in the fact subgraph to be tested but not in the sensitive semantic template graph, if the business concept category to which the node belongs is a preset sensitive entity category, then the cost of the node deletion operation is set to a preset second value, where the preset second value is greater than the preset first value; for the first node in the fact subgraph to be tested and the second node in the sensitive semantic template graph, if the business concept category to which the first node belongs is different from the business concept category to which the second node belongs, then the cost of the node replacement operation of replacing the first node with the second node is set to a preset blocking value, where the preset blocking value is greater than the preset second value.

[0019] By adopting the above technical solution, and by differentiating the business concept categories to which nodes belong and setting different editing operation costs, the system can reflect the differences in the importance of different types of nodes when performing graph structure matching. The system sets a lower cost for deleting non-sensitive auxiliary category nodes and a higher cost for deleting sensitive entity category nodes, and sets blocking levels of cost for node replacement operations between different business concept categories. This differentiated cost function design improves the semantic accuracy of graph structure matching. The system can better protect the integrity of sensitive entity information during the matching process, reducing the possibility of sensitive information being incorrectly replaced or deleted. This business semantic-based cost setting method enhances the protection of sensitive information during graph structure transformation and improves the reliability of the hierarchical results.

[0020] In a second aspect, embodiments of this application provide an automatic classification system for unstructured data based on business semantic understanding. This automatic classification system for unstructured data based on business semantic understanding 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, the computer program code including computer instructions, and one or more processors call the computer instructions to cause the system to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a system, cause the system to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer program product that, when run on a system, causes the system to execute the method described in any possible implementation of the first aspect.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. This application provides an automatic classification method for unstructured data based on business semantic understanding. By constructing a classification domain ontology library and a sensitive semantic template graph, and combining it with an entity relationship extraction model to process unstructured data, the extracted entity instances and dependencies can be mapped onto the domain ontology to form a subgraph of facts to be tested. The degree of matching with the sensitive semantic template is calculated through graph structure matching, thereby determining the data security level. This semantic understanding-based classification method can deeply understand the business meaning of the data content, improving the accuracy of data classification. Compared with traditional methods that rely solely on keyword matching, this solution can better identify the implicit sensitive information associations in the data, reducing classification bias caused by incomplete literal information. By introducing business semantic understanding, the classification method's ability to understand context is improved, enhancing its robustness in handling complex unstructured data. The graph structure-based matching mechanism makes the classification process interpretable, facilitating the analysis and optimization of classification results, and improving the credibility of the classification results.

[0024] 2. This application provides an automatic classification method for unstructured data based on business semantic understanding. It utilizes a set of background nodes to generate a global context vector and calculates its similarity with the business domain vector of the template graph, improving the scenario adaptability of the classification results. This context-based filtering mechanism can identify situations where local features match but the overall scenario does not, reducing the probability of cross-scenario mismatches. A secondary filtering process using a compatibility threshold improves the accuracy of the classification results. Including background information not involved in graph matching enhances the classification system's ability to perceive the overall data context, improving the rationality of the classification results. This classification method considering the global context improves the system's ability to identify the business scenario to which the data belongs, reducing classification errors caused by scenario misjudgment.

[0025] 3. This application provides an automatic classification method for unstructured data based on business semantic understanding. For cases where all matching scores are less than a preset threshold, a secondary matching mechanism based on graph edit distance is introduced. This allows the system to measure the structural transformation cost from the test fact subgraph to each template graph. The minimum cumulative edit distance is used as the evaluation metric, and the structural difference coefficient is obtained through normalization using the total number of nodes, making the degree of difference between template graphs of different sizes comparable. This approach improves the system's ability to handle incomplete matching scenarios. By selecting the template graph with the smallest structural difference coefficient as the final target, the system can find the most structurally similar sensitive information pattern even when direct matching fails. The graph edit distance-based matching mechanism improves the system's accuracy in identifying similar but not completely identical sensitive information patterns, reducing classification failures caused by subtle differences. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating an automatic classification method for unstructured data based on business semantic understanding, as described in an embodiment of this application.

[0027] Figure 2 This is another flowchart illustrating an automatic classification method for unstructured data based on business semantic understanding, as described in this application.

[0028] Figure 3 This is another flowchart illustrating an automatic classification method for unstructured data based on business semantic understanding, as described in this application.

[0029] Figure 4 This is a schematic diagram of the physical device structure of an automatic classification system for unstructured data based on business semantic understanding, provided in an embodiment of this application. Detailed Implementation

[0030] 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 any or all possible combinations including one or more of the listed items.

[0031] 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.

[0032] The following example is used in conjunction with Figure 1 This application describes an automatic classification method for unstructured data based on business semantic understanding in its embodiments: Please see Figure 1 This is a flowchart illustrating an automatic classification method for unstructured data based on business semantic understanding, as described in an embodiment of this application.

[0033] S101. Construct a hierarchical domain ontology library; The system constructs a hierarchical domain ontology, which includes business concept nodes and logical relationship edges connecting them. A hierarchical domain ontology is a systematically defined shared concept model for the data hierarchical requirements of a specific business domain. It formally describes the concepts, relationships between concepts, and attributes within that domain. Business concept nodes are the basic units in the ontology, representing concrete entities, objects, or abstract concepts in the business domain, such as "customer," "contract," and "transaction amount." Logical relationship edges are the lines connecting these nodes, representing semantic associations between concepts, such as predicate relationships like "signed," "contains," and "belongs to." The term "not limited" means that the construction source of the hierarchical domain ontology is not limited to a single data source; it can include various sources such as industry standard documents, internal enterprise data dictionaries, and historical hierarchical records. When multiple thresholds are involved, such as when selecting high-frequency concepts for ontology construction, the frequency threshold for concept occurrence should be dynamically adjusted according to the size of the corpus, and this frequency threshold is usually lower than the confidence threshold used to determine core business relationships. The process of constructing a hierarchical domain ontology is essentially the process of transforming unstructured business knowledge into a structured knowledge graph that computers can understand. The system first extracts key business terms as concept nodes by analyzing a large amount of business documents and data classification standards. Next, the system analyzes the interaction methods of these terms in actual business scenarios and defines the logical relationship edges between them. The resulting ontology not only contains static concept definitions but also dynamic business logic rules, providing a foundational knowledge framework for subsequent semantic understanding.

[0034] The system can employ a semi-automatic construction approach combining rules and statistics. First, it utilizes part-of-speech tagging and named entity recognition algorithms from natural language processing to extract high-frequency noun phrases as candidate concepts from massive amounts of business documents. Then, it applies TF-IDF (Term Frequency-Inverse Document Frequency) or TextRank algorithms to rank the importance of candidate concepts, filtering out core business concept nodes. For constructing logical relationship edges, the system can use dependency parsing tools to identify the subject-verb-object structure between concepts in sentences, thereby extracting "subject-verb-object" triples and converting them into relationship edges between concepts. Another implementation method is to utilize deep learning models for ontology learning. The system can fine-tune a pre-trained language model on the business corpus to train a specialized entity-relation extraction model. This model can automatically identify potential business concepts and their interrelationships from unstructured text and convert them into ontology files in RDF (Resource Description Framework) or OWL (Web Ontology Language) format. In addition, the system can incorporate expert knowledge for manual verification and correction, and use ontology editing tools to improve the automatically generated ontology library, ensuring the accuracy and completeness of the ontology library.

[0035] S102. Based on a hierarchical domain ontology library, construct sensitive semantic template graphs corresponding to different security levels. The system constructs sensitive semantic template graphs corresponding to different security levels based on a hierarchical domain ontology. Each sensitive semantic template graph is a directed graph structure composed of concept nodes representing sensitive information and logical relationship edges representing business logic. The sensitive semantic template graph is a special type of directed graph structure built upon the hierarchical domain ontology, specifically designed to describe the typical semantic features of data at a specific security level. Concept nodes representing sensitive information are the nodes that play a key role in the template graph; they directly correspond to sensitive elements in the data classification standards, such as "ID number," "bank account number," or "top secret recipe." Logical relationship edges representing business logic describe the existence and association methods of these sensitive elements in specific business scenarios. Different security levels correspond to different template graphs, meaning the system needs to establish corresponding semantic patterns for each level, such as "public," "internal," "secret," and "confidential." It is not explicitly stated that the number and complexity of sensitive semantic template graphs are not limited to a fixed pattern but are flexibly defined according to specific business scenarios and classification rules. When multiple thresholds are involved, for example, when determining the complexity of the template graph, the upper limit threshold for the number of nodes should be greater than the threshold for the number of nodes in the minimum connected subgraph to ensure that the template graph has sufficient expressive power without becoming overly complex. The system constructs sensitive semantic template graphs based on a hierarchical domain ontology library. This process essentially involves subgraph extraction and schema instantiation on top of the ontology library. Following the data security classification guidelines, the system identifies the core sensitive concepts involved in each level, finds these concepts and their surrounding relationships in the ontology library, and abstracts them into template graphs.

[0036] One specific method for constructing sensitive semantic template graphs is subgraph extraction based on expert rules. Security experts manually select specific concept nodes and relation edges from the domain ontology library on a visual interface, based on data classification standards, to form a subgraph representing a specific security level. For example, for the "Confidential" level, experts would select nodes such as "customer name," "ID number," and "account balance," as well as relation edges such as "ownership" and "association" between them, saving them as the template graph for that level. The system serializes and stores these manually defined subgraphs as standard templates. Another implementation method is automatic mining based on historical labeled data. The system collects a large amount of historical data labeled with security levels and uses frequent subgraph mining algorithms to analyze the common substructures of data of the same security level after mapping in the ontology library. The system statistically analyzes the frequency and confidence of these common substructures, automatically converting frequent substructures that meet the conditions into sensitive semantic template graphs for that security level. This method can automatically learn classification rules from the data, reduce manual intervention, and discover hidden semantic patterns that human experts might overlook.

[0037] S103. Use the preset entity relationship extraction model to process the hierarchical unstructured data and output the entity instance set and the dependency relationship between each entity instance in the entity instance set. A pre-trained entity relation extraction model refers to a machine learning or deep learning model that automatically identifies entities and determines relationships between them from unstructured text. Unstructured data to be classified refers to raw data requiring security level determination, typically including text documents, emails, chat logs, etc., whose content lacks a fixed format structure. An entity instance set refers to the set of specific words or phrases extracted from the text, corresponding to concrete objects in the real world, such as "Zhang San," "2023," and "Address A." Dependency relationships between entity instances refer to the interrelationships between these entities in sentence structure or semantic logic, such as "Zhang San" being a "resident" of "Address A." It is not limited to a specific neural network type; the architecture of the entity relation extraction model can be CNN, RNN, Transformer, or their variants. When multiple thresholds are involved, the confidence threshold for the model's output entity category should be independent of the confidence threshold for relation classification. Typically, the threshold for entity recognition is set higher to ensure the accuracy of basic information. The process by which the system uses this model to process data is the process of transforming continuous natural language text into discrete, structured information units. The system first preprocesses the data to be classified, including word segmentation and stop word removal, and then inputs it into the model. The model calculates its internal parameters and outputs all entities contained in the text and their types, and predicts the possible relationship types between each pair of entities.

[0038] To specifically implement entity relation extraction, the system can employ a pipelined approach. First, the system uses a dedicated named entity recognition model to scan the text to be classified, labeling all entity instances of interest and forming an entity instance set. Next, the system inputs the identified entity pairs and the context sentences containing these entities into an independent relation classification model. This model determines whether a predefined relation exists between each pair of entities and outputs a dependency label. Another implementation approach is to use a joint extraction model. The system uses an end-to-end deep learning model to simultaneously complete entity recognition and relation extraction tasks in a single inference process. By sharing parameters of the underlying encoding layer, this model can capture the interdependencies between entity recognition and relation extraction tasks, thus avoiding the error propagation problem in pipelined methods. The model directly outputs a list of triples (entity 1, relation, entity 2), and the system parses these triples to obtain the entity instance set and dependency relations.

[0039] S104. Map the set of entity instances to business concept nodes in the hierarchical domain ontology library, and map the dependency relationships to logical relation edges to construct a test fact subgraph representing the semantic structure of the unstructured data to be hierarchized. The system maps entity instance sets to business concept nodes in a hierarchical domain ontology and maps dependency relationships to logical relationship edges, constructing a test fact subgraph representing the semantic structure of the unstructured data to be hierarchically classified. Specifically, this includes: using each entity instance in the entity instance set as a graph node and dependency relationships as graph edges to construct an initial discrete semantic network; for node pairs with semantic gaps or missing connections in the discrete semantic network, retrieving potential inference chains connecting the node pairs in the hierarchical domain ontology, where potential inference chains are paths composed of intermediate concept nodes and their logical relationships that do not explicitly appear in the unstructured data in the ontology; calculating the relevance evaluation value of the potential inference chains and using the potential inference chain with the highest relevance evaluation value as the completion path; and integrating the intermediate concept nodes and logical relationships in the completion path into the discrete semantic network to obtain the test fact subgraph representing the semantic structure of the unstructured data to be hierarchically classified.

[0040] The subgraph of facts to be tested is a graph structure generated based on the data to be graded, reflecting the actual semantic state within the data. The initial discrete semantic network refers to a preliminary graph directly composed of extracted entities and relationships; at this stage, the graph may be disconnected. Semantic breaks or missing connections refer to situations in the discrete semantic network where, although some entity nodes are logically related, there is no direct linguistic description connecting them in the text. A potential inference chain refers to a path existing in the grading domain ontology that connects two seemingly isolated nodes; this path consists of a series of intermediate concept nodes and logical relationships. The relevance assessment value is a quantitative score of the rationality of the potential inference chain, while the completed path is the inference chain with the highest score. The process of constructing the subgraph of facts to be tested is not merely a simple mapping, but a process of semantic completion and inference. The system first aligns the extracted entities to the concepts in the ontology, forming a preliminary network. Then, the system detects breakpoints in the network, uses rich background knowledge from the ontology for inference, finds missing intermediate links, and connects the broken semantic fragments to form a complete and connected graph structure.

[0041] In implementing this step, the initial network is first constructed through mapping. The system uses a string fuzzy matching algorithm or vector-based semantic similarity calculation to compare each word in the entity instance set with the concept names in the hierarchical domain ontology. If the matching degree exceeds a threshold, a mapping relationship is established, with the entity as a graph node and its instance value as a node attribute. Simultaneously, the extracted dependency relations are matched with logical relation edges in the ontology to construct the initial discrete semantic network. For path completion, the system uses a graph search-based algorithm. For any two nodes in the discrete network that are not directly connected but are close to each other, the system performs a bidirectional breadth-first search or A* search algorithm in the entire hierarchical domain ontology to find all possible paths connecting these two nodes. The system calculates the length of each path, the weight of the included nodes, and the type weight of the edges, and comprehensively derives a relevance evaluation value. For example, the shorter the path and the more core business concepts it contains, the higher the evaluation value. The system selects the path with the highest evaluation value as the completion path, inserting the intermediate nodes and relation edges on the path into the discrete semantic network to complete the construction of the fact subgraph to be tested. Another approach is to use graph embedding techniques for link prediction. The system pre-maps all nodes and edges in the ontology to a low-dimensional vector space using algorithms such as TransE or GraphSAGE. When constructing the subgraph of the facts to be tested, for pairs of nodes with missing connections, the system calculates the probability that they have some kind of relationship in the vector space. If the probability is higher than a set threshold, the system infers the most likely intermediate nodes and relationships based on the vector operation results, thereby completing the subgraph.

[0042] S105. Perform graph structure matching calculations between the subgraph of facts to be tested and each sensitive semantic template graph to obtain the matching degree score between the subgraph of facts to be tested and each sensitive semantic template graph. The system performs graph structure matching calculations between the fact subgraph to be tested and each sensitive semantic template graph to obtain matching scores. Specifically, this includes: extracting semantic feature vectors of each node in both the fact subgraph to be tested and the sensitive semantic template graph; calculating the semantic distance between nodes based on the semantic feature vectors and constructing a semantic mapping matrix representing the degree of semantic alignment between nodes; establishing a set of node-level correspondences between the fact subgraph to be tested and the sensitive semantic template graph based on the semantic mapping matrix; identifying the projection edges of each logical relationship edge in the fact subgraph to be tested onto the sensitive semantic template graph based on the set of node-level correspondences; determining the number of projection edges with corresponding topological connections in the sensitive semantic template graph; calculating the proportion of the number of projection edges to obtain the structural isomorphism factor; and generating matching scores between the fact subgraph to be tested and each sensitive semantic template graph by weighted calculation of the semantic similarity probability distribution value in the semantic mapping matrix and the structural isomorphism factor.

[0043] Graph structure matching computation refers to the process of comparing the similarity between two graphs in terms of topological structure and node semantics. Semantic feature vectors are numerical representations of node semantic information, typically composed of high-dimensional real vectors. Semantic distance is a measure of the difference between two vectors in space; the smaller the distance, the closer the semantics. The semantic mapping matrix is ​​a two-dimensional matrix that records the pairwise similarity between nodes in the test graph and nodes in the template graph. The set of node-level correspondences is the optimal node matching pair determined based on the matrix. Projected edges refer to whether there are corresponding edges between edges in the test graph and their corresponding nodes in the template graph. The structural isomorphism factor is a value between 0 and 1, reflecting the similarity ratio of the two graphs in their edge connection patterns. The matching score is the final quantitative result used to determine the security level of the test data. It is not limited to a specific embedding model; the method for calculating semantic feature vectors is not limited to that of a specific embedding model. When multiple thresholds are involved, the similarity threshold for determining node correspondences should be higher than the total score threshold for determining a successful match to ensure the accuracy of local matching. When the system performs this step, it aims to quantify the fit between the test data and various security level standards. The system does not only compare whether the node names are the same, but also delves into the meaning that the nodes represent and the organizational structure between them.

[0044] In its implementation, the system first extracts features using a graph neural network. The system inputs the subgraph of facts to be tested and the sensitive semantic template graph into a graph convolutional network or a graph attention network, respectively. The model aggregates information about each node and its neighboring nodes, generating node feature vectors containing both structural and semantic information. Next, the system calculates the Euclidean distance or cosine similarity between each node vector in the subgraph of the test graph and each node vector in the template graph, constructing a semantic mapping matrix. Based on this matrix, the system uses the Hungarian algorithm or a greedy strategy to find the node-to-node correspondences that maximize the total similarity, establishing a set of node-level correspondences. For structural matching, the system traverses each edge in the subgraph of facts to be tested, checking whether there are edges of the same or compatible type between the two nodes connected to it and their corresponding nodes in the template graph. The system counts the number of successfully projected edges, divides it by the total number of edges in the subgraph of the test graph or the template graph, and obtains the structural isomorphism factor. Finally, the system weighted sums the mean similarity of corresponding nodes in the semantic mapping matrix with the structural isomorphism factor to obtain the final matching score. Another implementation method is to use the maximum common subgraph algorithm. The system transforms the graph matching problem into finding the longest common subgraph between two graphs. A variant of the Bron-Kerbosch algorithm is used to search for common substructures between the test graph and the template graph. The system calculates the proportion of nodes and edges in the longest common subgraph to the total number of nodes and edges in the template graph. Simultaneously, it combines text similarity based on node attributes to calculate a comprehensive similarity score as the matching score.

[0045] S106. Select the sensitive semantic template image with the largest matching score that is greater than the preset threshold as the target template image; The preset threshold is a numerical standard set by the system to filter out invalid results with excessively low matching scores, ensuring the reliability of the classification judgment. The target template image refers to the sensitive semantic template image that is ultimately identified as the one that best represents the characteristics of the data to be tested. The maximum value means that among all template images that meet the threshold condition, the one with the highest similarity is selected, following the best matching principle. It is not limited to this; the specific value of the preset threshold can be adjusted according to the error tolerance of the actual application scenario. For example, in security scenarios, the threshold setting will be relatively low to avoid missed judgments. When multiple thresholds are involved, the preset threshold should be significantly higher than the average score generated by random matching. The system will sort and filter all matching scores calculated in the previous step. If no template image has a matching score exceeding the threshold, the system may determine that the data is unclassified or at a normal level, or mark it as awaiting manual review. If multiple template images exceed the threshold, the system compares their specific scores and selects the one with the highest score.

[0046] S107. Output the security level associated with the target template graph as the classification result of the unstructured data to be classified.

[0047] The security level associated with the target template graph refers to the predefined classification labels, such as L1, L2, L3, L4, or specific security classification names, defined when constructing the template graph. The classification result output refers to the system assigning the final determined level label to the data under test, possibly accompanied by related metadata updates or log recordings. It is not limited to simple text labels; the output format can include structured JSON objects, database record updates, or triggering downstream security policy execution. When multiple thresholds are involved, this step does not directly involve threshold comparison but directly references the decision results from previous steps. The system's execution of this step is the final output of the entire method. The system searches for the attribute information of the target template graph to obtain its corresponding security level identifier. Then, the system binds this identifier to the original unstructured data to be classified. This may involve modifying file attributes, updating fields in the database, or generating a classification report. This result will directly guide subsequent data management measures, such as access control, encrypted storage, or audit monitoring.

[0048] In practice, the system maintains a mapping table or database table that records the correspondence between template image IDs and security levels. After the target template image ID is determined in step S106, the system queries this table to obtain the corresponding level label. The system can return this label to the caller in JSON format via an API interface. Simultaneously, the system can write the classification results into the metadata of unstructured data. For example, when processing a Word document, the system uses the Office Open XML SDK to modify the document's custom attributes and add a SecurityClassification field; when processing a text field in a database, the system executes an SQL update statement to fill the level into the corresponding level column. Another implementation method is to trigger a linked response. While outputting the classification results, the system publishes a classification completion event via a message queue. The downstream data leakage prevention system subscribes to this event, and upon receiving the confidentiality level result, immediately and automatically encrypts the file and moves it to a dedicated secure storage area.

[0049] In the above embodiments, by constructing a hierarchical domain ontology library and a sensitive semantic template graph, and combining it with an entity relationship extraction model to process unstructured data, the extracted entity instances and dependency relationships can be mapped onto the domain ontology to form a subgraph of facts to be tested. The degree of matching with the sensitive semantic template is calculated through graph structure matching, thereby determining the data security level. This semantic understanding-based hierarchical method can deeply understand the business meaning of data content, improving the accuracy of data hierarchical classification. Compared with traditional methods that rely solely on keyword matching, this scheme can better identify the implicit sensitive information associations in the data, reducing hierarchical bias caused by incomplete literal information. By introducing business semantic understanding, the hierarchical method's ability to understand context is improved, enhancing its robustness in processing complex unstructured data. The graph structure-based matching mechanism makes the hierarchical process interpretable, facilitating the analysis and optimization of hierarchical results, and improving the credibility of the hierarchical results.

[0050] Building upon the above embodiments, to further improve the accuracy and reliability of the classification results, this application also provides another automatic classification method for unstructured data based on business semantic understanding. This method analyzes background information in the unstructured data to be classified that does not directly participate in graph structure matching, and establishes a scenario-compatible verification mechanism to evaluate the degree of matching between the test data and the target template graph at the business scenario level. The following section combines... Figure 2 Another method for automatic classification of unstructured data based on business semantic understanding is described in the embodiments of this application: Please refer to Figure 2 This is another flowchart illustrating an automatic classification method for unstructured data based on business semantic understanding in an embodiment of this application.

[0051] S201. Obtain entity nodes in the unstructured data to be graded that are not included in the fact subgraph to be tested, and use them as the background node set. Entity nodes refer to nodes that represent specific objects, concepts, or attributes in a knowledge graph or semantic network. The background node set refers to the remaining entity nodes that exist in the original data but were not selected into the fact subgraph to be tested; these constitute the context in which the core event occurs. The acquisition operation refers to the process by which the system traverses the graph structure or index list to filter node data that meets specific conditions. In this step, the system first identifies all node IDs that constitute the fact subgraph to be tested, then traverses all nodes in the complete graph structure corresponding to the unstructured data to be graded, extracting nodes whose IDs are not in the node list of the fact subgraph to be tested, and aggregating them to form the background node set. This process aims to capture environmental information that seems peripheral but is crucial to determining the overall business scenario, such as time, location, non-core participants, or descriptive phrases.

[0052] This step can be implemented using a graph database query language. The first approach utilizes the difference query function of the graph database. The system first queries the full graph structure corresponding to the unstructured data to be classified based on its unique identifier, obtaining a set A of all nodes. Simultaneously, based on the subgraph of facts generated in the previous steps, the system obtains a set B of all nodes in the subgraph. Next, the system performs set operations, calculating the difference (AB) between set A and set B. The elements in this difference set are the entity nodes not included in the subgraph of facts to be classified. The system instantiates these node objects and stores them in a list structure as the background node set. The second approach is based on a label filtering algorithm. During the construction of the subgraph of facts to be classified, the system assigns a Boolean label (like a core node) to each selected node. When background nodes are needed, the system iterates through all entity nodes corresponding to the unstructured data to be classified, checking their label attributes. Nodes with a label of False or without that label are identified as background nodes and added to the background node set.

[0053] S202. Calculate the word vector of each background node in the background node set, and generate a global context vector of the unstructured data to be classified based on the word vector of the background node. The process involves calculating the word vector of each background node in the background node set and generating a global context vector for the unstructured data to be classified based on the word vectors of the background nodes. Specifically, this includes: calculating the frequency of occurrence of the words corresponding to each background node in the background node set in the unstructured data to be classified; calculating the inverse document frequency of each background node based on the occurrence frequency and a pre-defined corpus to obtain the weight value of each background node; using the weight value to perform a weighted summation of the word vectors of all background nodes in the background node set to obtain an initial context vector; and normalizing the initial context vector to obtain the global context vector of the unstructured data to be classified.

[0054] The word vector of a background node refers to the numerical representation of the text words corresponding to the background node in a high-dimensional vector space, which captures the semantic features of the words. The global context vector of the unstructured data to be graded is a comprehensive vector formed by aggregating the semantic information of all background nodes, representing the overall business atmosphere or contextual tendency of the entire data. Occurrence frequency refers to the number of times a specific word appears in the unstructured data text to be processed. The pre-collected corpus refers to a pre-collected and organized text dataset covering relevant business areas, used to provide a statistical reference benchmark. Inverse document frequency is a statistical indicator used to measure the importance of words; it reflects the prevalence of a word in the entire corpus, with more prevalent words having lower weights. The weight value is a value calculated based on occurrence frequency and inverse document frequency, used to quantify the contribution of each background node to the overall context. The initial context vector is an intermediate result after weighted summation, without scaling. Normalization refers to the process of converting vectors into unit vectors or scaling their numerical range to a specific interval to eliminate the influence of vector length on similarity calculation. In this step, the system not only converts the text into a vector, but also highlights key background information through a weighting mechanism, thereby synthesizing a feature vector that can accurately describe the overall business environment of the data.

[0055] Specifically, this refinement technique can be implemented using a TF-IDF weighted word embedding model. The first approach combines this with a pre-trained static word vector model. The system first loads a pre-trained word vector lookup table. For each node in the background node set, the system extracts its text label and counts the frequency of that label in the original text of the unstructured data to be graded. Simultaneously, the system queries the index of a pre-defined corpus, calculates the inverse document frequency (IVF) of the text label, and multiplies the TF and IDF to obtain the weight value of the node. Next, the system retrieves the word vector of the node from the lookup table and multiplies it by the corresponding weight value. Finally, the system accumulates all the weighted word vectors dimension-wise to obtain an initial context vector, and normalizes the initial context vector using the L2 norm to output a global context vector. The second approach utilizes a dynamic context embedding model. The system inputs the complete text of the unstructured data to be graded into a pre-trained language model and obtains the context-related embedding vector for each token. For a node in the background node set, the system finds its corresponding token position in the text and extracts the corresponding output layer vector as the word vector for that node. The weights are still calculated based on TF-IDF logic, but during weighted summation, the system dynamically adjusts the weights using an attention mechanism. Specifically, the attention scores between background nodes and core fact subgraph nodes are calculated as auxiliary weights and fused with the TF-IDF weights. Finally, a global context vector is generated through weighted averaging and a LayerNorm normalization layer.

[0056] S203. Obtain the business domain vector pre-associated with the target template image. The business domain vector is used to represent the business scenario category to which the target template image belongs. The target template graph refers to the candidate sensitive semantic template selected in the previous steps by the graph structure matching algorithm, which has a high matching score with the subgraph of facts to be tested in terms of structure and local semantics. The business domain vector is a pre-calculated and stored high-dimensional numerical vector; it is not just a simple label, but a mathematical abstraction of the business scenario to which the template graph applies. Representation refers to displaying or representing the essential characteristics of things in a specific way. The business scenario category refers to the specific macro-environmental classification of data generation and use. In this step, the system needs to access a knowledge base or database storing the template graph metadata. Each sensitive semantic template graph, when created or added to the database, is labeled with its business domain by experts or algorithms, and a centralized feature vector, i.e., the business domain vector, is generated based on a large amount of representative text data in that domain. The system indexes the associated metadata record based on the unique identifier ID of the currently selected target template graph and reads the business domain vector from it. This step is a crucial bridge connecting micro-graph structure matching and macro-scenario verification, providing a benchmark for subsequent compatibility calculations.

[0057] This step can be implemented through database join queries or cache retrieval. The first method is attribute querying based on a relational or graph database. The system maintains a template metadata table containing fields such as "Template ID," "Template Name," "Business Domain ID," and "Business Domain Vector." Once the target template graph is determined, the system uses SQL or graph query statements, with the template ID as the key, to directly retrieve the corresponding binary large object or array string stored in the "Business Domain Vector" field, and deserializes it into vector format. The second method utilizes a memory caching system. Considering that the business domain vectors of the template graph are static and frequently accessed data, the system pre-loads the mapping relationship between all template graph IDs and business domain vectors into the memory cache at startup. When step S203 is executed, the system directly retrieves the vector data from memory through key-value pair lookup. This method greatly reduces disk I / O operations and significantly improves data retrieval speed in high-concurrency scenarios.

[0058] S204. Calculate the cosine similarity between the global context vector and the business domain vector as the scene compatibility. The global context vector is the vector generated in step S202, representing the overall environmental characteristics of the data to be classified. The business domain vector is the vector obtained in step S203, representing the standard scene characteristics of the target template image. Cosine similarity is a mathematical method that evaluates the similarity between two vectors by calculating the cosine of the angle between them. Its value range is usually between -1 and 1. The closer the value is to 1, the more consistent the directions of the two vectors are, i.e., the more semantically similar they are. Scene compatibility is a specific indicator defined in this application, used to quantify the degree of fit between the data to be classified and the target template image in the macro business background. It directly reuses the calculation result of cosine similarity. In this step, the system takes two high-dimensional vectors as input, performs dot product and modulus product operations, and finally obtains the cosine value. This calculation process is essentially judging whether "the environment in which this data is located" and "the scene in which this sensitive template usually appears" overlap in the high-dimensional semantic space. If the two overlap significantly, it means that the template image is not only structurally matched, but also reasonable in business logic; otherwise, there may be a mismatch across scenes.

[0059] S205. If the scene compatibility is determined to be less than the preset compatibility threshold, delete the target template image and then select the sensitive semantic template image with the largest matching score that is greater than the preset threshold as the target template image. Qualified. A scenario compatibility score below this threshold means that while the data to be graded matches the template graph in its local structure, there are significant differences in its overall business context, resulting in a false positive due to "similar in form but not in essence." Deleting the target template graph means removing it from the current candidate template list or marking it as eliminated in this grading process, preventing it from participating in the final judgment. Selecting the sensitive semantic template graph with the highest matching score above the preset threshold means the system reverts to the template matching ranking list to find the next best candidate. The matching score is the graph structure matching score calculated in the previous step. In this step, the system performs a logical judgment: if the scenario compatibility calculated in S204 is lower than the standard, the system determines that the current best match is invalid. To avoid grading failure, the system must initiate a backtracking mechanism to find the template with the second-highest structural matching score among the remaining candidate templates that meets the basic matching threshold, promote it to a new target template graph, and prepare to re-perform the S203 and S204 verification processes, or directly adopt the suboptimal solution.

[0060] The first approach is iterative processing based on a priority queue. The system maintains a priority queue of candidate templates sorted in descending order of structural matching score. When the first element of the queue is retrieved as the target template and its scene compatibility is found to be insufficient, the system directly performs a dequeue operation, permanently removing it. Subsequently, the system checks if the queue is empty. If not empty, the system reads the new first element and checks if its matching score is still greater than the preset basic structure matching threshold. If so, it is marked as the new target template and a recursive call or loop jump instruction is triggered, allowing it to enter the next round of scene verification. The second approach is list traversal based on status markers. The system has a list containing all candidate template objects, each with a status field. When verification fails, the system updates the status of the current target template to "eliminated." Next, the system traverses the list, filters out all templates with a status of "pending verification," sorts them by matching score, selects the one with the highest score exceeding the threshold, updates its status to "currently processing," and uses it as the new target template.

[0061] S206. If the scene compatibility is determined to be no less than the preset compatibility threshold, execute the step of outputting the security level associated with the target template graph as the classification result of the unstructured data to be classified.

[0062] A scenario compatibility score not less than a preset compatibility threshold indicates that the data to be graded not only closely matches the template graph in its microscopic graph structure but also aligns with the domain of the template graph in its macroscopic business context. The security level associated with the target template graph refers to the confidentiality level corresponding to this sensitive semantic template within a predefined security policy. The grading result output refers to the process by which the system finally determines the data security label and persistently stores or sends it to downstream security management components. In this step, the system confirms successful verification and no longer searches for other candidate templates. The system reads the metadata attributes of the current target template graph and extracts its security level field. Subsequently, the system binds this security level to the ID of the unstructured data to be graded, generating a complete grading record. This step marks the successful completion of the automatic grading process; the system ensures high credibility of the output results through double verification.

[0063] In the above embodiments, a global context vector is generated using a set of background nodes, and its similarity is calculated with the business domain vector of the template graph, improving the scenario adaptability of the classification results. This context-based filtering mechanism can identify situations where local features match but the overall scenario does not, reducing the probability of cross-scenario mismatches. Secondary filtering by setting a compatibility threshold improves the accuracy of the classification results. Including background information not involved in graph matching enhances the classification system's ability to perceive the overall data context, improving the rationality of the classification results. This classification method that considers global context improves the system's ability to identify the business scenario to which the data belongs, reducing classification errors caused by scenario misjudgment.

[0064] In addition to the aforementioned optimization scheme based on scenario compatibility, this application also provides another method for automatic classification of unstructured data based on business semantic understanding, addressing situations where direct matching is ineffective. This method evaluates the similarity between the test data and the template graph by calculating the transformation cost between graph structures, providing an effective solution for handling scenarios with incomplete matching. The following section will combine... Figure 3 This application describes another method for automatic classification of unstructured data based on business semantic understanding: Please refer to [link to relevant documentation]. Figure 3 This is another flowchart illustrating an automatic classification method for unstructured data based on business semantic understanding, as described in this application.

[0065] S301. If all matching scores are less than the preset threshold, construct a graph editing cost function for each sensitive semantic template graph.

[0066] When all matching scores are determined to be less than a preset threshold, the system constructs a graph editing cost function for each sensitive semantic template graph. Specifically, this includes: determining the business concept category of the node in the test fact subgraph within the hierarchical domain ontology; for nodes present in the test fact subgraph but not in the sensitive semantic template graph, if the node's business concept category is a preset non-sensitive auxiliary category, setting the cost of the node deletion operation to a preset first value; for nodes present in the test fact subgraph but not in the sensitive semantic template graph, if the node's business concept category is a preset sensitive entity category, setting the cost of the node deletion operation to a preset second value, which is greater than the preset first value; and for the first node in the test fact subgraph and the second node in the sensitive semantic template graph, if the business concept categories of the first and second nodes are different, setting the cost of replacing the first node with the second node to a preset blocking value, which is greater than the preset second value.

[0067] A sensitive semantic template graph is a pre-constructed knowledge graph structure representing a specific sensitive business scenario or data pattern, containing sensitive entities and their relationships. A graph editing cost function is a mathematical model or set of rules used to define the operational costs required to transform one graph structure into another; these operations typically include node insertion, deletion, and replacement, as well as edge insertion, deletion, and replacement. A hierarchical domain ontology is a knowledge base containing all concepts, attributes, relationships, and their hierarchical structures within a specific business domain, used to provide semantic-level classification criteria. A test fact subgraph is a graph structure extracted from the unstructured data to be classified, reflecting the current data content. Business concept categories refer to the categories to which nodes belong in the ontology, such as "person's name," "amount," and "address." Pre-defined non-sensitive auxiliary categories refer to categories that do not play a decisive role in determining the data sensitivity level, such as the node categories corresponding to general conjunctions or meaningless auxiliary words. Pre-defined sensitive entity categories refer to categories directly associated with sensitive information, such as "ID number" and "bank card number." The preset first value, preset second value, and preset blocking value represent the cost weights of different operations, with the preset second value being greater than the preset first value, and the preset blocking value being greater than the preset second value. This relationship reflects the system's tolerance for different types of differences: deleting non-sensitive information has the lowest cost, deleting sensitive information has a higher cost, and replacing different types of business concepts has the highest cost. When executing this step, the system first checks all matching results from the previous stage to confirm that no template graph's matching score meets the requirements for direct matching. Subsequently, the system enters a secondary matching process, traversing each candidate sensitive semantic template graph and tailoring a set of editing cost rules for each template graph. The core of this set of rules lies in combining semantic understanding and identifying the specific business attributes of each node in the subgraph under test by querying a hierarchical domain ontology library. Based on these attributes, the system differentiates the cost of deletion or replacement operations. For example, if an unimportant auxiliary word node appears in the graph under test, the system assigns it a low deletion cost; if a critical sensitive entity node appears, the system assigns it a high deletion cost; and if the node type in the graph under test is completely different from the node type in the template graph, the system assigns it an extremely high blocking cost to prevent incorrect matching.

[0068] Two specific methods can be used to implement this. The first method is a static mapping method based on a rule engine. The system maintains a predefined cost mapping table, which details the correspondence between various concepts in the hierarchical domain ontology and the cost of editing operations. When the system identifies a node in the subgraph to be tested as belonging to a preset non-sensitive auxiliary category, it directly looks up the corresponding preset first value in the table; if the node is identified as belonging to a preset sensitive entity category, it looks up the table to obtain a preset second value; for node pairs with mismatched types, a preset blocking value is directly assigned. This method is highly efficient and the rules are clear. The second method is a dynamic calculation method based on semantic similarity. The system does not rely entirely on fixed values, but uses a word vector model to calculate the distance between the concept of the node to be tested and the concept of the template node in the semantic space. For deletion operations, the system dynamically generates a cost value based on the information entropy or importance of the node concept; the greater the information content of the node, the higher its deletion cost; the smaller the information content of the node, the lower its deletion cost. For replacement operations, the system calculates the cosine distance between the two concept vectors; the greater the distance, the higher the replacement cost; if it exceeds a certain limit, it is considered a blocking operation.

[0069] S302. Based on the graph editing cost function, calculate the minimum cumulative edit distance to transform the subgraph of facts to be tested into a sensitive semantic template graph; The minimum cumulative edit distance refers to the minimum total cost of transforming the test fact subgraph into a sensitive semantic template graph through a series of allowed edit operations. This total cost is the sum of the single-step costs of all executed operations, and the single-step cost is explicitly defined by the graph edit cost function in the previous step. This distance reflects the degree of difference between the two graph structures after considering business semantic weights. When the system executes this step, it takes the test fact subgraph as the starting state and the sensitive semantic template graph as the target state. The system needs to find an optimal path in a huge search space, which represents a series of operations that minimizes the total cost of transforming the starting state into the target state. Since the calculation of graph edit distance is an NP-hard problem in computational complexity theory, it is impractical to directly exhaustively enumerate all possible combinations of operations. Therefore, the system usually uses heuristic or approximate algorithms to solve it. During the calculation, the system considers not only node-level transformations but also edge transformations. If two nodes are preserved or replaced, the connection between them also needs to be preserved or modified accordingly, which will also generate corresponding edit costs. The final calculated value is not only a geometric distance but also a semantic distance that incorporates business sensitivity understanding.

[0070] Regarding the specific implementation of this step, namely how to efficiently calculate the minimum cumulative edit distance, the following two specific methods can be adopted. The first method is to use the A* search algorithm. The system treats the calculation of graph edit distance as a shortest path search problem in the state space. The algorithm maintains a priority queue to store the currently explored intermediate states and their estimated total costs. The estimated total cost consists of two parts: the actual edit cost incurred and the heuristic estimated cost from the current state to the target state. The heuristic function can be designed as a lower bound estimate based on the difference in node label sets. The A* algorithm, by prioritizing the expansion of states with the minimum estimated cost, can quickly converge to the optimal or suboptimal solution, especially performing well when the graph size is moderate. The second method is to use an approximate algorithm for bipartite graph matching, such as the Hungarian algorithm or a variant of the Kuhn-Monkres algorithm. The system treats the node set of the graph to be tested and the node set of the template graph as two parts of a bipartite graph, with the edge weights set as the replacement cost between nodes. By solving for the maximum or minimum weight matching, the system can quickly find the optimal correspondence between nodes, thus approximating the graph edit distance. Although this method focuses primarily on node matching and ignores some topological information, its computation speed is much faster than the exact search algorithm, making it very suitable for handling large-scale graph structures or scenarios with high real-time requirements.

[0071] S303. Obtain the total number of nodes contained in the sensitive semantic template graph, and use the total number of nodes to normalize the minimum cumulative edit distance to obtain the structural difference coefficient. In this step, the total number of nodes refers to the sum of all nodes in the currently compared sensitive semantic template graph. Normalization is a data processing technique designed to eliminate the incomparability between data of different dimensions or scales, mapping the values ​​to a standard range. The structural difference coefficient is the final index after normalization, used to objectively measure the degree of difference between the fact subgraph to be tested and a specific template graph, excluding the influence of the complexity of the template graph itself. When performing this step, the system first reads the metadata of the current sensitive semantic template graph and extracts the total number of nodes it contains. This step is crucial because the minimum cumulative edit distance is an absolute value. For example, for a complex template containing 100 nodes, an edit distance of 5 may mean a very high similarity; while for a simple template containing only 5 nodes, the same edit distance of 5 may mean a complete mismatch. Therefore, the system must eliminate this scale bias through mathematical operations. The system typically uses the minimum cumulative edit distance calculated in step S302 as the numerator and the total number of nodes as part of the denominator, and performs a division operation. The quotient is the structural difference coefficient. The smaller the coefficient, the closer the structure of the data to be tested is to that of the template; the larger the coefficient, the greater the difference.

[0072] Regarding the specific implementation of this step, namely how to perform normalization to obtain a scientific structural difference coefficient, the following two specific methods can be adopted. The first method is linear normalization based on the largest scale. The formula logic is: Structural Difference Coefficient = Minimum Cumulative Edit Distance / max(Number of nodes in the subgraph of the fact to be tested, Number of nodes in the sensitive semantic template graph). This method considers the larger of the two graphs as the benchmark. If the graph to be tested is much larger than the template graph, the denominator becomes larger and the coefficient becomes smaller, but this may mask the core structural mismatch problem; conversely, if the template graph is very large, this formula can reflect the degree of matching well. This method is simple to calculate and intuitive. The second method is weighted normalization based on the size of the template graph. The formula logic is: Structural Difference Coefficient = Minimum Cumulative Edit Distance / (Number of nodes in the sensitive semantic template graph * Adjustment Factor). The adjustment factor is a constant set based on business experience. This method strictly uses the standard template as the benchmark, emphasizing the degree of deviation of the data to be tested from the standard template. If the data to be tested is missing key nodes in the template, the edit distance will increase, and the coefficient will increase significantly, thus accurately reflecting the non-compliance or mismatch status. This approach focuses more on assessing whether the data under test meets specific compliance patterns.

[0073] S304. Select the sensitive semantic template image with the smallest structural difference coefficient that is less than the preset difference threshold as the final target template image. In this step, the preset difference threshold serves as a decision boundary, filtering out template graphs that, although their difference coefficients are calculated, are actually too large and lack reference value. The final target template graph refers to the standard graph that the system deems best representative of the true attributes of the data under test among all candidate templates. When the system performs this step, it is actually conducting a sorting and filtering process. The system collects the structural difference coefficients calculated for all sensitive semantic template graphs, forming a result set. First, the system filters this result set, eliminating all options with structural difference coefficients greater than or equal to the preset difference threshold. This step is to prevent forced matching; that is, when the data under test is dissimilar to all templates, the system should refuse to classify or mark it as unknown, rather than forcibly matching it to a template with the smallest difference but still completely irrelevant. Among the remaining filtered options, the system compares their structural difference coefficient values ​​and finds the template graph corresponding to the one with the smallest value. This smallest coefficient means that the subgraph of the fact under test can be transformed into this template graph with the fewest and lowest cost transformations, therefore they are the most isomorphic in terms of business semantics. Once a unique minimum value is determined, that template graph is locked as the final target template graph. If multiple template graphs have the same minimum difference coefficient, the system can make a decision based on preset priority rules.

[0074] S305. Output the security level associated with the final target template diagram as the classification result of the unstructured data to be classified.

[0075] In this step, the final target template graph is the standard graph most similar to the data structure under test, as determined in the previous step. The associated security level refers to the predefined security classification label on each sensitive semantic template graph, such as "Public," "Internal," "Secret," "Confidential," or "L1," "L2," "L3," "L4," etc. The unstructured data to be classified is the raw input object for the entire process. The classification result output refers to the final judgment conclusion given by the system, usually presented in the form of metadata tags, log records, or API responses. When performing this step, the system directly reads the security level attributes bound to the final target template graph by querying the attribute fields of the metadata store or knowledge graph. This process is a simple mapping operation because each template has already been assigned a clear security level by business experts when building the sensitive semantic template library. After reading the level information, the system marks it on the original unstructured data. This may involve updating file attributes, writing classification tags to the database, or triggering corresponding security control policies.

[0076] In the above embodiments, for cases where all matching scores are less than a preset threshold, a secondary matching mechanism based on graph edit distance is introduced. This allows the system to measure the structural transformation cost from the test fact subgraph to each template graph. The minimum cumulative edit distance is used as the evaluation metric, and the structural difference coefficient is obtained through normalization using the total number of nodes. This makes the degree of difference between template graphs of different sizes comparable, improving the system's ability to handle incomplete matching scenarios. By selecting the template graph with the smallest structural difference coefficient as the final target, the system can find the most structurally similar sensitive information pattern even when direct matching fails. The graph edit distance-based matching mechanism improves the system's accuracy in identifying similar but not completely identical sensitive information patterns, reducing classification failures caused by subtle differences.

[0077] The system in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 4 This is a schematic diagram of the physical device structure of an automatic classification system for unstructured data based on business semantic understanding, provided in an embodiment of this application.

[0078] It should be noted that, Figure 4 The structure of the system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0079] like Figure 4As shown, the system includes a Central Processing Unit (CPU) 401, which can perform various appropriate actions and processes based on a program stored in Read-Only Memory (ROM) 402 or a program loaded from storage portion 408 into Random Access Memory (RAM) 403, such as executing the methods described in the above embodiments. The RAM 403 also stores various programs and data required for system operation. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An Input / Output (I / O) interface 405 is also connected to the bus 404.

[0080] The following components are connected to I / O interface 405: input section 406 including a camera, infrared sensor, etc.; output section 407 including a liquid crystal display (LCD) and speakers, etc.; storage section 408 including a hard disk, etc.; and communication section 409 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. Communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from it can be installed into storage section 408 as needed.

[0081] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit (CPU) 401, it performs the various functions defined in the present invention.

[0082] It should be noted that the computer-readable medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein a computer-readable computer program is carried. The transmitted data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0083] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0084] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the system described in the above embodiments; or it may exist independently and not assembled into the system. The storage medium carries one or more computer programs that, when executed by a processor of a system, cause the system to implement the methods provided in the above embodiments.

[0085] 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.

[0086] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as "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 "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0087] 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.

[0088] 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. A method for automatic classification of unstructured data based on business semantic understanding, characterized in that, include: Construct a hierarchical domain ontology library, which includes business concept nodes and logical relationship edges connecting the business concept nodes; Based on the hierarchical domain ontology library, a sensitive semantic template graph corresponding to different security levels is constructed. The sensitive semantic template graph is a directed graph structure composed of concept nodes representing sensitive information and logical relationship edges representing business logic. The unstructured data to be processed is processed using a preset entity relationship extraction model, and the output is a set of entity instances and the dependency relationships between the entity instances in the set. The entity instance set is mapped to the business concept nodes in the hierarchical domain ontology library, and the dependency relationship is mapped to logical relationship edges to construct a test fact subgraph representing the semantic structure of the unstructured data to be hierarchized. The graph structure matching calculation is performed between the subgraph of facts to be tested and each of the sensitive semantic template graphs to obtain the matching degree score between the subgraph of facts to be tested and each of the sensitive semantic template graphs; The sensitive semantic template image with the largest matching score greater than a preset threshold is selected as the target template image; The security level associated with the target template graph is output as the classification result of the unstructured data to be classified.

2. The method according to claim 1, characterized in that, The step of mapping the entity instance set to business concept nodes in the hierarchical domain ontology library, and mapping the dependency relationships to logical relation edges, to construct a test fact subgraph representing the semantic structure of the unstructured data to be hierarchized, specifically includes: The entity instances in the entity instance set are used as graph nodes, and the dependency relationships are used as graph edges to construct an initial discrete semantic network. For node pairs in the discrete semantic network that have semantic gaps or missing connections, potential inference chains connecting the node pairs are retrieved in the hierarchical domain ontology library. The potential inference chain is a path composed of intermediate concept nodes and their logical relationships that do not appear explicitly in the unstructured data in the ontology library. Calculate the relevance evaluation value of the potential inference chain, and take the potential inference chain with the highest relevance evaluation value as the completion path; The intermediate concept nodes and logical relationships in the completion path are integrated into the discrete semantic network to obtain the test fact subgraph representing the semantic structure of the unstructured data to be graded.

3. The method according to claim 1, characterized in that, The step of performing graph structure matching calculations between the subgraph of facts to be tested and each of the sensitive semantic template graphs to obtain the matching score between the subgraph of facts to be tested and each of the sensitive semantic template graphs specifically includes: Semantic feature vectors of each node in the subgraph of facts to be tested and the sensitive semantic template graph are extracted respectively; Based on the semantic feature vectors, the semantic distance between each node is calculated, and a semantic mapping matrix representing the degree of semantic alignment between each node is constructed; Based on the semantic mapping matrix, a set of node-level correspondences is established between the subgraph of facts to be tested and the sensitive semantic template graph; Based on the node-level correspondence set, identify the projection edges of each logical relationship edge in the subgraph of the fact to be tested onto the sensitive semantic template graph; Determine the number of projection edges that have corresponding topological connections in the sensitive semantic template graph; Calculate the proportion of the number of projected edges to obtain the structural isomorphism factor; By combining the semantic similarity probability distribution values ​​in the semantic mapping matrix with the structural isomorphism factor, a matching score between the subgraph of the fact to be tested and each of the sensitive semantic template graphs is generated through weighted calculation.

4. The method according to claim 1, characterized in that, Before outputting the security level associated with the target template graph as the classification result of the unstructured data to be classified, the method further includes: Obtain entity nodes in the unstructured data to be graded that are not included in the fact subgraph to be tested, and use them as a set of background nodes; Calculate the word vector of each background node in the background node set, and generate a global context vector of the unstructured data to be classified based on the word vector of the background nodes; Obtain the business domain vector pre-associated with the target template image; the business domain vector is used to characterize the business scenario category to which the target template image belongs. Calculate the cosine similarity between the global context vector and the business domain vector, and use it as the scene compatibility. If the scene compatibility is determined to be less than a preset compatibility threshold, the target template image is deleted and the step of selecting the sensitive semantic template image with the largest matching score that is greater than the preset threshold is executed as the target template image. If the scene compatibility is determined to be no less than the preset compatibility threshold, the step of outputting the security level associated with the target template graph as the classification result of the unstructured data to be classified is performed.

5. The method according to claim 4, characterized in that, The process of generating a global context vector for the unstructured data to be classified based on the word vectors of the background nodes specifically includes: The frequency of occurrence of the words corresponding to each background node in the background node set in the unstructured data to be classified is statistically analyzed. Based on the occurrence frequency and a preset corpus, the inverse document frequency of each background node is calculated to obtain the weight value of each background node; The initial context vector is obtained by weighting and summing the word vectors of all background nodes in the background node set using the weight values. The initial context vector is normalized to obtain the global context vector of the unstructured data to be classified.

6. The method according to claim 1, characterized in that, After performing graph structure matching calculations on the subgraph of facts to be tested and each of the sensitive semantic template graphs to obtain the matching score between the subgraph of facts to be tested and each of the sensitive semantic template graphs, the method further includes: If all the matching scores are determined to be less than the preset threshold, a graph editing cost function is constructed for each of the sensitive semantic template graphs. Based on the graph editing cost function, calculate the minimum cumulative edit distance to transform the subgraph of facts to be tested into the sensitive semantic template graph; Obtain the total number of nodes contained in the sensitive semantic template graph, and use the total number of nodes to normalize the minimum cumulative edit distance to obtain the structural difference coefficient; The sensitive semantic template image with the smallest structural difference coefficient that is less than a preset difference threshold is selected as the final target template image. The security level associated with the final target template graph is output as the classification result of the unstructured data to be classified.

7. The method according to claim 6, characterized in that, The construction of a graph editing cost function for each of the sensitive semantic template graphs specifically includes: In the hierarchical domain ontology library, determine the business concept category to which the node in the fact subgraph to be tested belongs in the hierarchical domain ontology library; For nodes that exist in the subgraph of facts to be tested but not in the sensitive semantic template graph, if the business concept category to which the node belongs is a preset non-sensitive auxiliary category, then the cost of the node deletion operation is set to a preset first value. For nodes that exist in the subgraph of facts to be tested but not in the sensitive semantic template graph, if the business concept category to which the node belongs is a preset sensitive entity category, the cost of deleting the node is set to a preset second value, which is greater than the preset first value. For the first node in the subgraph of facts to be tested and the second node in the sensitive semantic template graph, if the business concept category to which the first node belongs is different from the business concept category to which the second node belongs, the cost of replacing the first node with the second node is set to a preset blocking value, which is greater than the preset second value.

8. An automatic classification system for unstructured data based on business semantic understanding, characterized in that, The system includes: One or more processors and a memory; the memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the system, the system performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the system, the system performs the method as described in any one of claims 1-7.

Images