Generalization processing method and system for economic big data Chinese text data

Abstract

The application relates to the field of generalization encryption technology and discloses a generalization processing method and system for economic big data Chinese text data, which comprises the following steps: firstly, normalizing and segmenting original text data, and counting co-occurrence relations of word pairs by using preset sliding window parameters; then, calculating the correlation strength of the word pairs according to the co-occurrence count and non-negative, and constructing a weighted word graph with the correlation strength as the edge weight; then, calculating a normalized graph matrix and performing feature analysis to obtain a second eigenvalue and a second eigenvector, so as to calculate an edge level sensitivity value and determine a word level sensitivity; screening a high sensitivity word set according to a threshold parameter, generating a window radius in the feature vector coordinate space according to the coordinate difference and a radius parameter of the high sensitivity word and a peak neighbor node, forming a word bucket and merging overlapping intervals; finally, generating a bucket code for each word bucket by using an encryption algorithm, replacing the words in the bucket, and outputting generalization encrypted text data with privacy protection and structural usability.

Description

Technical Field

[0001] This invention relates to the field of generalized encryption technology, and more specifically, to a method and system for generalized processing of text data in economic big data. Background Technology

[0002] With the acceleration of economic digitalization, massive amounts of economic text data have been generated in fields such as finance, trade, and industry monitoring. These data originate from various sources, including corporate announcements, policy documents, news and public opinion, and transaction records. To support applications such as macroeconomic forecasting, risk identification, and credit assessment, graph models are typically used to characterize the potential semantic relationships between texts. Nodes correspond to economic entities or keywords, while edges represent the strength of semantic or statistical associations. However, this type of text data often contains highly sensitive information such as company identities and transaction details, requiring generalization processing to ensure data security when shared across institutions.

[0003] Existing methods often rely on word frequency, category, or hierarchical concepts for term-level replacement and desensitization, failing to adequately consider the weight balance between nodes and edges in a graph structure. This can easily lead to overall information imbalance and make it difficult to achieve an effective balance between privacy protection and data usability. Generalization processing of economic and financial texts often employs semantic clustering, word replacement, and edge weight adjustment. However, the semantics of economic and financial texts exhibit a significant non-uniform distribution on the network: high-frequency economic terms form dense connections, while low-frequency specific entities / technical terms form sparse bridging relationships. When generalization operations uniformly adjust edge weights or merge semantic nodes, the originally balanced semantic network experiences abrupt connectivity changes, with the structural influence of some nodes / connections being abnormally amplified, leading to an imbalance in local information transmission capabilities. This not only reduces the accuracy of subsequent data analysis but may also expose original privacy information through semantic cues remaining at the structural level in the generalized text. Taking financial public opinion monitoring as an example, when multiple texts involve the same financial event (such as a major asset restructuring or default event of an institution), related keywords form a high-density connected subgraph in the co-occurrence graph, showing a local clustering coefficient significantly higher than the global average. If generalization only maps sensitive terms to a vocabulary or generalizes concepts without considering their connectivity features in the graph structure, then the degree centrality, betweenness centrality, and other topological indicators of the generalized nodes will still maintain abnormal numerical distributions. Attackers can use graph topology analysis or spectral clustering methods to identify clusters of nodes with abnormal structures, combine this with publicly available information to perform association reasoning, thereby bypassing word-level generalization protection, inferring the identity of sensitive entities, and posing a risk of privacy leakage.

[0004] From a graph theory perspective, once economic texts are constructed into a graph structure through co-occurrence or dependency relationships, generalization operations that modify edge weights, node groupings, and connection patterns directly alter the graph's connectivity and mathematical properties. At the spectral level, this change manifests as rapid fluctuations in key eigenvalues, leading to a significant increase in the influence of individual connection propagation paths. Because economic and financial corpora naturally possess hierarchical and industry-concentrated characteristics, even minor changes in edge weights can cause significant instability in graph connectivity. As generalization deepens, some bridging nodes evolve into key nodes controlling the overall structure, making encrypted text potentially indirectly identifiable by structural analysis methods, further exacerbating the conflict between privacy protection and information usability. Summary of the Invention

[0005] This invention provides a method and system for generalizing text data in economic big data, which solves the technical problems mentioned in the background.

[0006] The first aspect involves generalized processing methods for text data in economic big data, including: The original text data is normalized and segmented to obtain word sequences. The co-occurrence count of word pairs in the word sequences is counted based on preset sliding window parameters. The association strength of word pairs is calculated based on co-occurrence counts and then nonnegated to construct a weighted word graph with association strength as the edge weight. The normalized graph matrix is ​​calculated based on the weighted word graph. Feature analysis is performed on the normalized graph matrix to obtain the second eigenvalue and the corresponding second eigenvector. The edge-level sensitivity value of each edge is also calculated. For each word, the largest edge-level sensitivity value among its connected edges is taken as the word-level sensitivity of that word. A set of highly sensitive words is obtained by filtering the word-level sensitivity based on a preset threshold parameter. In the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitivity word. The peak neighbor node is the node connected to the high-sensitivity word and has the largest corresponding edge sensitivity value. The window radius is determined based on the coordinate difference between the high-sensitivity word and the peak neighbor node and the preset radius parameter. An initial word bucket is generated and the final bucket set is obtained by merging overlapping intervals. For each final bucket, an encryption algorithm is used to process the preset identifier and bucket number to generate a bucket code. The words in the bucket are replaced with a replacement token in a preset format, and the generalized encrypted text data is output.

[0007] Secondly, a generalized processing system for text data in economic big data, applied to any of the generalized processing methods for text data in economic big data described above, includes: The data processing module performs normalization processing on the raw text data and segments it into word sequences. Based on preset sliding window parameters, it counts the co-occurrence of word pairs in the word sequences. The association strength module calculates the association strength of word pairs based on co-occurrence counts and performs non-negation processing to construct a weighted word graph with association strength as the edge weight; The edge-level sensitivity module calculates a normalized graph matrix based on a weighted word graph, performs feature analysis on the normalized graph matrix to obtain the second eigenvalue and the corresponding second eigenvector, and calculates the edge-level sensitivity value for each edge. The high-sensitivity word module takes the largest edge-level sensitivity value among the connected edges of each word as the word-level sensitivity, and filters the word-level sensitivity based on preset threshold parameters to obtain a set of high-sensitivity words. In the final bucket module, in the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitivity word. The peak neighbor node is the node connected to the high-sensitivity word and has the largest corresponding edge sensitivity value. The window radius is determined based on the coordinate difference between the high-sensitivity word and the peak neighbor node and the preset radius parameter. An initial word bucket is generated and the final bucket set is obtained by merging overlapping intervals. The generalized encryption module uses an encryption algorithm to process the preset identifier and bucket number to generate a bucket code for each final bucket. It then replaces the words in the bucket with a preset replacement token and outputs generalized encrypted text data.

[0008] The beneficial effects of this invention include: by constructing a weighted word graph based on economic text data and combining it with feature analysis of a normalized graph matrix to achieve sensitive identification at the structural level, it can accurately locate key nodes and edges affecting overall connectivity at the spectral level; based on this, it uses the coordinate difference between highly sensitive words and peak-reaching neighbor nodes to generate an adaptive window radius, automatically forming a set of semantically related word buckets, and uses an encryption algorithm to generate bucket-level replacement tokens, thereby achieving semantic generalization at the word level, maintaining connectivity balance at the structural level, and ensuring privacy and security at the encryption level. This invention enables economic texts to eliminate the identifiability of sensitive information after generalization processing while maintaining the statistical regularity of the original graph structure, effectively balancing data security and analytical usability, and is particularly suitable for privacy protection applications involving cross-institutional sharing and large-scale economic corpora. Attached Figure Description

[0009] Figure 1 This is a flowchart of the method for generalizing text data in economic big data according to the present invention; Figure 2 This is a block diagram of the generalized processing system for text data in economic big data according to the present invention. Detailed Implementation

[0010] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0011] Example 1 like Figures 1 to 2 As shown, the generalization processing method for text data in economic big data includes: The original text data is normalized and segmented to obtain word sequences. The co-occurrence count of word pairs in the word sequences is counted based on preset sliding window parameters. The association strength of word pairs is calculated based on co-occurrence counts and then nonnegated to construct a weighted word graph with association strength as the edge weight. The normalized graph matrix is ​​calculated based on the weighted word graph. Feature analysis is performed on the normalized graph matrix to obtain the second eigenvalue and the corresponding second eigenvector. The edge-level sensitivity value of each edge is also calculated. For each word, the largest edge-level sensitivity value among its connected edges is taken as the word-level sensitivity of that word. A set of highly sensitive words is obtained by filtering the word-level sensitivity based on a preset threshold parameter. In the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitivity word. The peak neighbor node is the node connected to the high-sensitivity word and has the largest corresponding edge sensitivity value. The window radius is determined based on the coordinate difference between the high-sensitivity word and the peak neighbor node and the preset radius parameter. An initial word bucket is generated and the final bucket set is obtained by merging overlapping intervals. For each final bucket, an encryption algorithm is used to process the preset identifier and bucket number to generate a bucket code. The words in the bucket are replaced with a replacement token in a preset format, and the generalized encrypted text data is output.

[0012] In one embodiment of the present invention, the original text data is normalized and segmented to obtain a word sequence. The co-occurrence count of word pairs in the word sequence is then calculated based on preset sliding window parameters, including: After performing normalization on the original text data, word segmentation is performed to obtain a word sequence: ;in, Indicates the total length of the word sequence. Indicates the first word in the word sequence Words in a given position, ; Set preset sliding window parameters For each position in the word sequence Define the forward window set : ;in, This is the offset within the window; Any two words in the word sequence Harmony Calculate the ordered co-occurrence count : ;in, Indicates an indicator function, when and The value is 1 if the condition is met, and 0 otherwise. Any two words in the word sequence Harmony Symmetricization of ordered co-occurrence counts yields disordered co-occurrence counts. : .

[0013] The sources of raw text data include, but are not limited to, the following types: Corporate-related documents: such as corporate announcements issued by listed companies (including financial report disclosures and announcements of major investment / cooperation matters), and internal corporate operation reports (such as supply chain management documents and text descriptions corresponding to customer cooperation agreements); Policy and regulatory documents: such as economic regulation policy documents issued by the government and regulatory agencies (such as detailed rules of monetary policy and industry support policies), and financial regulatory rules (such as textual explanations of bank credit approval regulations); Market and public opinion texts: such as news reports published by financial media (including industry dynamics and market trend analysis), and discussion texts about economic events on social media or professional forums (such as comments on supply and demand changes in a certain industry). Transaction-related documents: such as transaction records from financial institutions (e.g., terms in loan agreements, textual information on cross-border trade documents), and explanatory texts corresponding to inter-company transaction vouchers (e.g., order notes, economic-related descriptions in logistics information).

[0014] Raw text data has two properties: Strong economic relevance: The text content closely revolves around economic activities, containing a large number of economic terms, such as GDP growth rate, monetary policy, supply chain finance, and specific economic entity identifiers, such as the full / abbreviated name of the enterprise, the code of the financial institution, and transaction details, such as the textual description of the transaction amount and the name of the partner.

[0015] High privacy sensitivity: The text contains sensitive information that requires strict protection, such as internal corporate transaction data, operational strategy texts of specific institutions, and identity / transaction information related to customers. If shared directly across institutions, it is easy to cause privacy leakage risks. This is also the motivation for achieving privacy protection through generalized encryption.

[0016] Standardization and word segmentation include: Standardization processing: Standardize the original economic texts (such as corporate announcements, financial public opinion, etc.), such as unifying the character encoding to NFKC format and converting all characters to lowercase to eliminate the problem of different forms of the same word caused by encoding differences and inconsistencies in capitalization. For example, avoid misinterpreting GDP and gdp as two different words.

[0017] Word segmentation: Use word segmentation tools adapted to Chinese economic texts, such as word segmenters optimized for technical terms, to split the normalized text into a continuous sequence of words, that is, a list of individual words in the original text order.

[0018] The word sequence obtained after splitting is a set of words arranged in the original text order. The total length of the word sequence refers to the total number of words contained in the split text, and the word at position t refers to the specific word corresponding to position t in the list.

[0019] The preset sliding window parameter is a pre-defined positive integer representing the maximum allowed distance between two words when counting co-occurrences. This parameter is set to avoid counting distant word pairs that have no actual semantic connection. For example, in a 1000-word text, if the distance between the beginning "bank" and the end "exchange rate" is far greater than the sliding window, their co-occurrence is meaningless.

[0020] The forward window set is for each word at position t in the word sequence. The forward window set refers to all positions shifted forward by 1 from position t to the range of the sliding window parameter. For example, if the sliding window parameter is 5, the forward window set for position t=3 would be positions 4, 5, 6, 7, and 8. If the total length is insufficient, the set is taken up to the maximum length. The forward window set only counts the co-occurrence relationship between the word at position t and the words in its forward window, ensuring that co-occurrence statistics focus on semantically closely related word pairs.

[0021] Ordered co-occurrence refers to the occurrence of two words in a clear order, that is, only the case where word a appears at a certain position t and word b appears in the forward window of that position t is counted (the case where word b comes before word a is not counted).

[0022] Iterate through all possible positions t and all possible forward window offsets, and sum up all the results (1 or 0) of the indicator function that meet the conditions. The final sum is the ordered co-occurrence count of word a and word b, which represents the total number of times word a appears before word b and the distance is within the sliding window.

[0023] The purpose of symmetry processing is that in economic texts, the semantic relationship between many word pairs does not depend on their order. For example, monetary policy and interest rates represent related concepts regardless of which comes first. Ordered co-occurrence counting only counts a single order and cannot fully reflect the actual strength of the relationship between word pairs. Therefore, symmetry processing is needed to eliminate the influence of order.

[0024] The unordered co-occurrence count represents the total number of times word a and word b co-occur within the sliding window (regardless of their order), which better meets the actual statistical needs of semantic association between word pairs in economic texts.

[0025] In one embodiment of the present invention, the association strength of word pairs is calculated based on co-occurrence counts and nonnegated to construct a weighted word graph with association strength as edge weights, including: Each word in the word sequence is used as a set of nodes in a weighted word graph. ; Calculate the total co-occurrence of all unordered word pairs based on unordered co-occurrence counting. : ; Any two words in the word sequence Harmony Calculate the joint probability of word pairs and the marginal probability of words respectively: , ;in, For words Harmony The joint co-occurrence probability, For words The probability of edge co-occurrence; Calculation words Harmony The original value of the correlation strength is then nonnegated: ; ; in, It is the natural logarithm function. The final correlation strength after nonnegation; Based on node set and final correlation strength Constructing a weighted undirected word graph : Representing edge sets, including: ; The weight matrix includes: when hour, ;when Sometimes, or hour, .

[0026] Using each word in a word sequence as the node set of a weighted word graph means taking each independent word in the word sequence after splitting the original text as a node (i.e., the basic building block of the graph) in the weighted word graph. For example, after splitting an economic text, we get words such as monetary policy, interest rate, GDP, and inflation, each word corresponding to a node; the node set is the total set of all these independent words.

[0027] To calculate the total co-occurrence of all unordered word pairs based on unordered co-occurrence counts, it is necessary to combine this with unordered co-occurrence counts: All unordered word pairs refer to all pairs of different words in the node set (such as monetary policy-interest rate, GDP-inflation, etc., and word A-word B and word B-word A are considered the same pair and are not counted repeatedly). The total co-occurrence count is the sum of all the unordered co-occurrence counts of all these different unordered word pairs. For example, the unordered co-occurrence count of monetary policy-interest rate is 8, GDP-inflation is 5, and the total count of all other word pairs is 200. Then the total co-occurrence count is 8+5+200=213.

[0028] Joint probability of word pairs: For any two words, such as monetary policy and interest rate, the joint probability is the result of dividing the disordered co-occurrence count of the two words by the total co-occurrence count. For example, if the disordered co-occurrence count of monetary policy and interest rate is 8 and the total co-occurrence count is 213, then the joint probability is 8 / 213≈0.037. The joint probability reflects the frequency of the two words appearing together within the sliding window; the larger the value, the more frequently the two words are associated.

[0029] Marginal probability of a word: For a single word, such as "monetary policy," its marginal probability is the sum of the disordered co-occurrence counts of the word with all other unordered word pairs, divided by the total co-occurrence count. For example, the disordered co-occurrence counts of "monetary policy" with "interest rate," "inflation," and "exchange rate" are 8, 6, and 4 respectively, totaling 18, with a total co-occurrence count of 213. Therefore, the marginal probability is 18 / 213≈0.085. The marginal probability reflects the activity level of the word in co-occurring with other words in the entire text; a larger value indicates that the word is more central in semantic association.

[0030] Calculate the original value of the association strength: Use the natural logarithm function, taking the joint probability of the word pair divided by the product of the marginal probabilities of the two words as the input logarithm. If the frequency of the two words appearing together is much higher than the product of their independent frequencies (i.e., joint probability > marginal probability product), the original value is positive, indicating that there is a non-accidental semantic association between the two words, such as monetary policy and interest rate; if the joint probability ≈ marginal probability product, the original value is close to 0, indicating that the association between the two words is weak; if the joint probability < marginal probability product, the original value is negative, indicating that the two words are almost unrelated, such as monetary policy and ice cream.

[0031] Nonnegation: All negative numbers and 0s in the original association strength values ​​are adjusted to 0, retaining only positive numbers to obtain the final association strength. This is because in the semantic graph of economic texts, negatively related or unrelated word pairs are meaningless for subsequent structural analysis. For example, the negative association between monetary policy and ice cream does not need to be reflected in the graph. Nonnegation simplifies the graph structure, retaining only the strength values ​​of actual semantic associations.

[0032] An edge set is the collective term for connected word pairs in a graph, containing only unordered word pairs with a final association strength greater than 0 (i.e., word pairs with actual semantic association). For example, the final association strength of monetary policy and interest rate is 2.1 (>0), so there is an edge between these two nodes; the strength of monetary policy and ice cream is 0, so there is no connection. This ensures that only meaningful semantic connections are preserved in the graph.

[0033] The weight matrix is ​​a matrix that quantifies the association strength of each edge: for word pairs with edges, such as monetary policy-interest rate, the value at the corresponding position in the matrix is ​​equal to the final association strength; for word pairs without edges, such as monetary policy-ice cream or a word-to-itself pair, the value at the corresponding position in the matrix is ​​0.

[0034] Since all word pairs are constructed based on disordered co-occurrence counts and disordered association strengths, the edges are undirected (the edge from word A to word B is the same as the edge from word B to word A, and has the same strength), which conforms to the symmetry of semantic associations between word pairs in economic texts. For example, the association between monetary policy affecting interest rates and interest rates reflecting monetary policy is bidirectional.

[0035] In one embodiment of the present invention, a normalized graph matrix is ​​calculated based on a weighted word graph, a second eigenvalue and a corresponding second eigenvector are obtained by feature analysis of the normalized graph matrix, and an edge-level sensitivity value for each edge is calculated, including: Based on weighted undirected word graph Construct the angle matrix For angle matrix The dimension is equal to the set of nodes. The number of nodes, and satisfying: ; ; ; Define the set of nodes such as dimensions The identity matrix of the number of nodes Calculate the normalized graph matrix : ;in, For the angle matrix The inverse square root matrix; For normalized graph matrix Perform eigenvalue decomposition to solve for the second eigenvalue of the following constraint. and the corresponding second eigenvector ,include: First constraint ; Second constraint ; Third constraint ; in, For a dimension equal to the set of nodes A vector of all 1s representing the number of nodes; Weighted undirected word graph Each edge Calculate the edge sensitivity value : ;in, Represents the second eigenvector At the node The component at that location.

[0036] The diagonal matrix is ​​a diagonal matrix used to characterize the sum of connection strengths of each node in the word graph (only the diagonal positions of the matrix have values, and the off-diagonal positions are all 0), as shown below: Matrix dimension: The dimension is equal to the number of nodes in the weighted word graph, that is, each node corresponds to a diagonal element of the matrix, ensuring that the connection strength of each node can be quantified individually.

[0037] Values ​​of diagonal elements: The value of a diagonal element in the angle matrix is ​​equal to the sum of the correlation strengths of all edges connecting that node. For example, if the edge correlation strengths between the node monetary policy and interest rates, inflation, and exchange rates are 2.1, 1.8, and 1.5 respectively, then the value of the corresponding diagonal element is 2.1 + 1.8 + 1.5 = 5.4. If the node has no connected edges (an extreme case), then the value of the diagonal element is 0.

[0038] Values ​​of off-diagonal elements: All elements at off-diagonal positions have a value of 0. This is an inherent property of diagonal matrices, focusing only on the total number of connections within a node and not on the direct relationships between nodes.

[0039] A normalized graph matrix is ​​a matrix obtained by standardizing the weight matrix of a weighted word graph. Its core function is to eliminate the bias caused by the number of connections between nodes (i.e., the diagonal values ​​of the angle matrix), allowing edge weights to more objectively reflect the true semantic relationships between nodes. A normalized graph matrix can characterize the structural characteristics of a word graph, avoiding the distortion of edge weights caused by excessive connections between high-frequency nodes.

[0040] By decomposing the normalized graph matrix, a series of eigenvalues ​​and corresponding eigenvectors are obtained. The first eigenvalue and first eigenvector usually correspond to the case where all nodes are indistinguishable (e.g., a vector of all 1s), and have no structural distinguishing meaning. However, the second eigenvalue and second eigenvector can reflect the cluster structure differences of the word graph. That is, which nodes belong to the same semantic cluster and which nodes are key nodes for cross-cluster connections.

[0041] The first constraint (characteristic equation relationship): Multiplying the normalized graph matrix by the second eigenvector results in the product of the second eigenvalue and the second eigenvector. This is the fundamental definition of eigenvalues ​​and eigenvectors, ensuring that the vector is an eigenvector that reflects the matrix structure, and the value is the corresponding eigenvalue.

[0042] The second constraint (unit norm): The norm of the second eigenvector is equal to 1. This eliminates the influence of differences in vector scale. Without a norm constraint, the same eigenvalue may correspond to countless eigenvectors with different scales. By unifying it to the unit norm, the range of values ​​for the eigenvectors is fixed.

[0043] The third constraint (orthogonality to the all-1 vector): The inner product of the second eigenvector and the all-1 vector (a vector where all elements are 1 and the dimension equals the number of nodes) is 0 (i.e., orthogonal). The all-1 vector represents a state where all nodes are indifferent, and orthogonality means that the second eigenvector can break through the indifferent state and accurately reflect the differences between nodes. That is, the values ​​of nodes within a cluster on the second eigenvector are similar, while the values ​​of nodes across clusters are significantly different.

[0044] The magnitude of the second feature value reflects the overall connectivity of the word graph. The larger the value, the stronger the connectivity of the word graph (tight connections between clusters); the smaller the value, the weaker the connectivity (sparse connections between clusters, with bottlenecks). The distribution of the second feature vector values ​​reflects the cluster affiliation of nodes; nodes with similar values ​​belong to the same semantic cluster, while nodes with large differences in values ​​belong to different clusters.

[0045] Edge sensitivity value is an indicator that measures the sensitivity of each edge to the connectivity of the word graph (i.e., the second feature value). The larger the value, the stronger the impact on the overall connectivity of the word graph, which is called a structure-sensitive edge.

[0046] The edge sensitivity value is equal to the association strength of the edge multiplied by the square of the difference between the components of the two nodes on the second eigenvector. Where: Squared component difference: If two nodes belong to the same cluster, the component difference is small, and the squared value is even smaller; if they belong to different clusters (cross-cluster edge), the component difference is large, and the squared value is even larger. That is, this part directly quantifies the cross-cluster attribute of the edge. The stronger the cross-cluster attribute, the greater the impact on connectivity.

[0047] Multiply by the edge's association strength: Edges with strong association strength contribute more to the connection between nodes, and when combined with cross-cluster attributes, their sensitivity to overall connectivity is further amplified.

[0048] Edge sensitivity values ​​can be used to directly filter out the edges that are most critical to the connectivity of the word graph. These edges are usually core bridging edges across clusters, such as the edge between monetary policy and GDP in economic texts.

[0049] In one embodiment of the present invention, for each word, the largest edge-level sensitivity value among its connected edges is taken as the word-level sensitivity of that word. A set of highly sensitive words is obtained by filtering word-level sensitivities based on a preset threshold parameter, including: Determine the node The set of adjacent nodes ; Select node The set of adjacent nodes The largest edge sensitivity value in the middle is used as the node. Word-level sensitivity; If node If the word-level sensitivity is greater than or equal to the preset word-level sensitivity, then the node will be... Mark words as highly sensitive words and merge all highly sensitive words to obtain a set of highly sensitive words.

[0050] The adjacency set is the set of all other nodes that are connected to a single node in a weighted undirected word graph by an edge, as follows: Source of the set: Based on a weighted undirected word graph, where edges in the word graph represent semantic associations (association strength > 0) between nodes (words). Therefore, the adjacent nodes of a certain node are all the nodes connected to that node through edges with association strength > 0.

[0051] For example, if the node "Monetary Policy" is connected to interest rate, inflation, and exchange rate in the word graph (with association strengths of 2.1, 1.8, and 1.5 respectively), then the set of adjacent nodes for "Monetary Policy" is {interest rate, inflation, and exchange rate}; if the node has no connected edges (an extreme case), then the set of adjacent nodes is empty.

[0052] Word-level sensitivity is an indicator that measures the degree to which a single node affects the connectivity of a word graph. Its calculation logic revolves around the maximum edge-level sensitivity value of the edges associated with the node, as follows: From the edge-level sensitivity values ​​of all adjacent edges of the current node, select the one with the largest value as the word-level sensitivity of that node. For example, the edge-level sensitivity values ​​of the node monetary policy with interest rates, inflation, and exchange rates are 5.2, 3.8, and 4.1, respectively. The largest value, 5.2, is the word-level sensitivity of monetary policy.

[0053] The influence of a node on the word graph structure is determined by its most sensitive edge, which is the key support connecting the node to other parts of the word graph. If the weight of this edge changes, it will have the greatest impact on the connectivity of the word graph. Therefore, using the sensitivity value of this edge to characterize the node's sensitivity can reflect the structural importance of the node.

[0054] The high-sensitivity word set is a set of nodes selected from all nodes whose word-level sensitivity reaches a preset threshold, as detailed below: The meaning of the preset threshold: The preset word-level sensitivity is a value set in advance based on the privacy protection requirements of economic text and the characteristics of word graph structure (the threshold is set to 4.0 based on historical generalization experience). Its function is to distinguish between nodes with a large impact on the structure and nodes with a small impact. That is, the higher the threshold, the fewer highly sensitive words are selected, and they are all core nodes with a strong structural impact; the lower the threshold, the more highly sensitive words are selected, covering more nodes with a certain structural impact.

[0055] Screening rules: Each node's word-level sensitivity is compared to a preset threshold. If a node's word-level sensitivity is greater than or equal to the threshold, it is marked as a high-sensitivity word; otherwise, it is not marked. All nodes marked as high-sensitivity words are aggregated to obtain a high-sensitivity word set. For example, if the preset threshold is 4.0, nodes "monetary policy" (sensitivity 5.2) and "GDP" (sensitivity 4.5) both meet the threshold and are included in the set; node "inflation" (sensitivity 3.8) does not meet the threshold and is not included.

[0056] Selection objective: Highly sensitive words are key structural nodes in the word graph. Changes in the semantic associations of these nodes can significantly affect the connectivity of the word graph. If generalization is not performed, they can easily become a breakthrough point for privacy leakage (for example, by inferring sensitive information in the original text through the association patterns of these nodes).

[0057] In one embodiment of the present invention, in the coordinate space of the second feature vector, a peak neighbor node is selected for each highly sensitive word. The peak neighbor node is the node connected to the highly sensitive word and has the largest corresponding edge sensitivity value. Based on the coordinate difference between the highly sensitive word and the peak neighbor node and a preset radius parameter, a window radius is determined, an initial word bucket is generated, and a final bucket set is obtained by merging overlapping intervals, including: Based on the second feature vector For any node Define the corresponding one-dimensional coordinates ; Determine the first The set of adjacent nodes corresponding to the nth highly sensitive words and the edge sensitivity values ​​are calculated, and the nth high-sensitivity word is calculated. The peak neighbor node of a high-sensitivity word Among them, the peak-reaching neighbor node is the _th The node adjacent to the most sensitive word and the node with the largest edge sensitivity value; Regarding the first Calculate the window radius corresponding to each highly sensitive word. ;in, For the preset radius parameter, Indicates the first A one-dimensional coordinate of a highly sensitive word Represents the second eigenvector At the node The component at the location; For the Initial bucket of highly sensitive words ; Map the coordinate range corresponding to each initial word bucket to a feature interval. Calculate the union of all feature intervals and extract all maximally connected subintervals from the union. ; Based on maximally connected subintervals Generate the final bucket set : ;in, For the final number of buckets, for index, for index, Indicates the first One-dimensional coordinates of highly sensitive words.

[0058] Defining a node's one-dimensional coordinates based on the second feature vector transforms the abstract nodes (words) in the word graph into specific positions on a one-dimensional number axis, allowing structural relationships between nodes (within-cluster / across-cluster) to be reflected through coordinate distance, as detailed below: The role of the second feature vector is to characterize the cluster structure of the word graph (the vector components of nodes within a cluster are similar, while the components of nodes across clusters are significantly different). Therefore, the component value corresponding to each node in the second feature vector is directly used as the one-dimensional coordinate of that node.

[0059] For example, if in the second feature vector, the component values ​​of the node monetary policy are 0.32, the interest rate is 0.35, and GDP is -0.28, then their one-dimensional coordinates are 0.32, 0.35, and -0.28, respectively. That is, the coordinates of monetary policy and interest rate are close (belonging to the same semantic cluster), while the coordinates of GDP differ greatly from both (belonging to different semantic clusters). Reducing the high-dimensional word graph structure (the relationships between nodes) to one-dimensional coordinates allows for the division of word buckets by coordinate intervals. This transforms the generation logic of word buckets from semantic association to geometric proximity, simplifying the grouping operation and accurately matching the cluster structure of the word graph.

[0060] The peak neighbor node is the adjacent node that has the greatest impact on the connectivity of the word graph among the high-sensitivity word association edges, as detailed below: For the i-th highly sensitive word, first determine its set of neighboring nodes (all nodes connected to the highly sensitive word by an edge), and then select the neighboring node with the largest edge sensitivity value to the highly sensitive word from the set, which is the peak neighbor node.

[0061] For example, the adjacent nodes of the highly sensitive term "monetary policy" are interest rate, inflation, and exchange rate, with corresponding edge sensitivity values ​​of 5.2, 3.8, and 4.1, respectively. The highest value of 5.2 corresponds to the adjacent node interest rate, so interest rate is the peak neighbor node of monetary policy.

[0062] The edge between the peak neighbor node and the high-sensitivity word is the most critical structurally sensitive edge of the high-sensitivity word.

[0063] The window radius is a parameter that controls the size of the initial word bucket corresponding to a single high-sensitivity word, as detailed below: Preset radius parameter: A positive number less than or equal to 1 (e.g., 0.5) set in advance to adjust the size of the window radius and balance the generalization range and semantic precision.

[0064] Window radius = preset radius parameter × coordinate difference between the high-sensitivity word and the peak neighbor node. For example, the coordinate of the high-sensitivity word "monetary policy" is 0.32, the coordinate of the peak neighbor "interest rate" is 0.35, the coordinate difference is 0.03, the preset radius parameter is 0.5, then the window radius is 0.5 × 0.03 = 0.015.

[0065] The window radius is dynamically determined by using coordinate differences and preset parameters. For highly sensitive words and peak neighbors with close structural associations, the window radius is small (generalization range is concentrated); for words with slightly looser associations, the radius can be increased by adjusting the parameters to ensure that the size of each initial word bucket is adapted to its structural association strength, avoiding overgeneralization or undergeneralization caused by a uniform radius.

[0066] The initial word bucket is a temporary bucket formed by grouping nodes whose one-dimensional coordinates are within a window range for a single high-sensitivity word, as follows: Centered on the one-dimensional coordinates of the highly sensitive word, and within the range of the window radius, all nodes whose one-dimensional coordinates fall within the interval of [highly sensitive word coordinates - window radius, highly sensitive word coordinates + window radius] are all assigned to the initial word bucket corresponding to that highly sensitive word.

[0067] For example, if the high-sensitivity term "monetary policy" has a coordinate of 0.32 and a window radius of 0.015, with a window range of [0.305, 0.335], then nodes with one-dimensional coordinates between 0.305 and 0.335 (such as monetary policy 0.32 and interest rate derivatives 0.318) are all included in this initial term bucket.

[0068] The initial word bucket is generated from a single high-sensitivity word, and the initial word buckets of different high-sensitivity words may have overlapping coordinate ranges. For example, the initial word bucket range of the high-sensitivity word interest rate is [0.33, 0.355], which overlaps with the range of monetary policy [0.305, 0.335]. Nodes within the overlapping range (such as interest rate 0.35) will belong to both initial word buckets simultaneously.

[0069] Merging overlapping intervals involves integrating the coordinate ranges of all initial word buckets into non-overlapping maximally connected subintervals, as detailed below: First, the window range of each initial word bucket ([high-sensitivity word coordinates - radius, high-sensitivity word coordinates + radius]) is converted into a feature interval, and each initial word bucket corresponds to a feature interval.

[0070] Merge all feature intervals to obtain a total interval that includes the range of all initial intervals (which may consist of multiple non-contiguous sub-intervals).

[0071] Extracting maximally connected subintervals: Select non-overlapping, gap-free continuous subintervals from the total intervals; these are the maximally connected subintervals. For example, after merging the initial feature intervals [0.305, 0.335] and [0.33, 0.355], the maximally connected subinterval is [0.305, 0.355]. If there is also an interval [0.8, 0.82], then that interval is another maximally connected subinterval.

[0072] To resolve overlapping conflicts in the initial word buckets (avoiding the same node belonging to multiple buckets), nodes with close semantic associations and continuous coordinates are grouped into the same interval, ensuring that the final buckets generated subsequently can match the natural cluster structure of the word graph, rather than artificially segmented scattered groups of nodes.

[0073] The final bucket set is a set of nodes that fall within a maximally connected subinterval, grouped together to form a non-overlapping, semantically related set of nodes, as follows: Each maximally connected subinterval corresponds to a final bucket. All nodes whose one-dimensional coordinates fall within this maximally connected subinterval are assigned to the corresponding final bucket. If the node coordinates are not within any maximally connected subinterval (not a high-sensitivity word and weakly associated with a high-sensitivity word), they are not assigned to any word bucket.

[0074] For example, the maximally connected sub-interval [0.305, 0.355] corresponds to final bucket 1, which includes monetary policy, interest rates, and interest rate derivatives within the interval; the maximally connected sub-interval [0.8, 0.82] corresponds to final bucket 2, which includes exchange rate futures and cross-border settlement within the interval, ultimately forming a final bucket set containing the two buckets.

[0075] Nodes (words) within the same final bucket will be uniformly replaced with the same encrypted token. This achieves privacy protection for sensitive words, and because the nodes within the bucket are closely related semantically, the overall structural pattern of the word graph can still be preserved after the replacement, thus balancing privacy and usability.

[0076] In one embodiment of the present invention, for each final bucket, an encryption algorithm is used to process a preset identifier and bucket number to generate a bucket code, and words in that bucket are replaced with replacement tokens of a preset format, outputting generalized encrypted text data, including: For the Each final bucket constructs its corresponding message to be encrypted through string concatenation operations. ;in, This indicates string concatenation. This indicates that a preset identifier has been set. Using the HMAC-SHA256 encryption algorithm and Base32 encoding, the first... The message to be encrypted in each final bucket is used to construct the corresponding replacement token; For the first word in the original word sequence Word Perform replacement: If the first Word Belongs to the final bucket set Then the first Word Replace with a replacement symbol; otherwise, keep the first one. Word constant; The original word sequence after replacement is used as the result of generalization processing of text data.

[0077] The message to be encrypted is a string formed by concatenating a preset identifier with the bucket number for a single final bucket, as follows: The preset identifier is a fixed string (such as ECON_TEXT) set in advance, and its purpose is to distinguish the buckets of different generalization tasks. If multiple economic texts are processed at the same time, the preset identifiers of different tasks are different, which can avoid the duplication of bucket codes of different tasks and ensure the uniqueness of the identifier.

[0078] The bucket number is a unique identifier for each final bucket (e.g., 1, 2, 3, etc.), determined by the final bucket set according to the generation order or coordinate range, ensuring that the bucket number is not repeated within the same task.

[0079] The message to be encrypted is formed by concatenating the preset identifier and the bucket number into a complete string. For example, if the preset identifier is ECON_TEXT and the number of the 3rd final bucket is 3, then the message to be encrypted is ECON_TEXT3.

[0080] This ensures that the encrypted input of each final bucket is unique. Even if different buckets have the same sequence number (in cross-task scenarios), the messages to be encrypted are still different due to different preset identifiers. Within the same task, the sequence number is unique, further ensuring the uniqueness of the message.

[0081] The substitution token is a unique identifier used to replace words within a bucket, obtained after processing the message to be encrypted using a specific encryption algorithm and encoding. It employs the HMAC-SHA256 encryption algorithm and Base32 encoding, as detailed below: Step 1: HMAC-SHA256 Encryption. HMAC-SHA256 is a keyed hash algorithm that requires the message to be encrypted and a preset key (held only by the processor to ensure security). It outputs a fixed-length binary hash value. The advantage of this algorithm is its tamper-proof nature; even if the messages differ by only one character (e.g., ECON_TEXT3 and ECON_TEXT4), the output hash values ​​will be completely different. Furthermore, without the key, the original message cannot be deduced in reverse, ensuring privacy and security.

[0082] Step 2: Base32 encoding. The binary hash value output by HMAC-SHA256 is converted into a string consisting of uppercase letters and numbers (0-9) (Base32 encoding rules). The purpose of this encoding is to standardize the format, meaning that the binary hash value cannot be directly used for text replacement (containing invisible characters). The string obtained after Base32 encoding has no special characters, is easy to recognize, and has a controllable length (such as truncating the first 8 characters), making it suitable as a replacement mark in text.

[0083] Notation format: Typically, a preset format marker (such as angle brackets + prefix G_) is added to the encoded string to form the final replacement notation. For example, if the encoded string is J2K3L4M5, the replacement notation would be...<G_J2K3L4M5> This allows for quick differentiation between substitution marks and original words.

[0084] The replacement token balances security and usability. HMAC-SHA256 encryption ensures that the token cannot be reverse-engineered (protecting the privacy of sensitive words within the bucket), Base32 encoding ensures that the token format is standardized (adapting to text replacement scenarios), and formatting tags improve the token's recognizability.

[0085] Word sequence replacement involves traversing the original word sequence and determining whether to replace each word with the corresponding replacement token based on whether it belongs to the final bucket, as detailed below: If the t-th word in the original word sequence (e.g., monetary policy) belongs to a certain final bucket (e.g., the 3rd bucket, corresponding to the symbol)<G_J2K3L4M5> If the word is ), then replace it with the corresponding replacement symbol; If the t-th word does not belong to any final bucket (such as ordinary goods, non-structural sensitive words), then the word remains unchanged and is not replaced.

[0086] For example, if the original word sequence is "monetary policy affects interest rates, and the prices of ordinary goods remain stable," and both monetary policy and interest rates belong to the third bucket (notation), then...<G_J2K3L4M5> If a regular product does not belong to any bin, the replacement sequence will be:<G_J2K3L4M5> Influence<G_J2K3L4M5> Prices of ordinary commodities remained stable.

[0087] By replacing only the structurally sensitive words (highly sensitive words and related nodes) in the final bucket, their recognizability is eliminated; non-sensitive words remain unchanged to avoid overgeneralization that could lead to loss of text semantics, ensuring that the replaced text can still support economic data analysis (such as semantic association recognition and trend judgment).

[0088] Example 2 A system for generalizing text data in economic big data, applied in any of the methods for generalizing text data in economic big data, characterized in that it includes: The data processing module performs normalization processing on the raw text data and segments it into word sequences. Based on preset sliding window parameters, it counts the co-occurrence of word pairs in the word sequences. The association strength module calculates the association strength of word pairs based on co-occurrence counts and performs non-negation processing to construct a weighted word graph with association strength as the edge weight; The edge-level sensitivity module calculates a normalized graph matrix based on a weighted word graph, performs feature analysis on the normalized graph matrix to obtain the second eigenvalue and the corresponding second eigenvector, and calculates the edge-level sensitivity value for each edge. The high-sensitivity word module takes the largest edge-level sensitivity value among the connected edges of each word as the word-level sensitivity, and filters the word-level sensitivity based on preset threshold parameters to obtain a set of high-sensitivity words. In the final bucket module, in the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitivity word. The peak neighbor node is the node connected to the high-sensitivity word and has the largest corresponding edge sensitivity value. The window radius is determined based on the coordinate difference between the high-sensitivity word and the peak neighbor node and the preset radius parameter. An initial word bucket is generated and the final bucket set is obtained by merging overlapping intervals. The generalized encryption module uses an encryption algorithm to process the preset identifier and bucket number to generate a bucket code for each final bucket. It then replaces the words in the bucket with a preset replacement token and outputs generalized encrypted text data.

[0089] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for generalizing text data in economic big data, characterized in that, include: The original text data is normalized and segmented to obtain word sequences. The co-occurrence count of word pairs in the word sequences is counted based on preset sliding window parameters. The association strength of word pairs is calculated based on co-occurrence counts and then nonnegated to construct a weighted word graph with association strength as the edge weight. The normalized graph matrix is ​​calculated based on the weighted word graph. Feature analysis is performed on the normalized graph matrix to obtain the second eigenvalue and the corresponding second eigenvector. The edge-level sensitivity value of each edge is also calculated. For each word, the largest edge-level sensitivity value among its connected edges is taken as the word-level sensitivity of that word. A set of highly sensitive words is obtained by filtering the word-level sensitivity based on a preset threshold parameter. In the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitivity word. The peak neighbor node is the node connected to the high-sensitivity word and has the largest corresponding edge sensitivity value. The window radius is determined based on the coordinate difference between the high-sensitivity word and the peak neighbor node and the preset radius parameter. An initial word bucket is generated and the final bucket set is obtained by merging overlapping intervals. For each final bucket, an encryption algorithm is used to process the preset identifier and bucket number to generate a bucket code. The words in the bucket are replaced with a replacement token in a preset format, and the generalized encrypted text data is output.

2. The method for generalizing text data in economic big data according to claim 1, characterized in that, The original text data is normalized and segmented to obtain a word sequence. Based on preset sliding window parameters, the co-occurrence count of word pairs in the word sequence is calculated, including: After performing normalization on the original text data, word segmentation is performed to obtain a word sequence: ;in, Indicates the total length of the word sequence. Indicates the first word in the word sequence Words in each position, ; Set preset sliding window parameters For each position in the word sequence Define the forward window set : ;in, This is the offset within the window; Any two words in the word sequence Harmony Calculate the ordered co-occurrence count : ;in, Indicates an indicator function, when and The value is 1 if the condition is met, and 0 otherwise. Any two words in the word sequence Harmony Symmetricization of ordered co-occurrence counts yields disordered co-occurrence counts. : 。 3. The method for generalizing text data in economic big data according to claim 2, characterized in that, The association strength of word pairs is calculated based on co-occurrence counts and then nonnegated. A weighted word graph with association strength as edge weights is constructed, including: Each word in the word sequence is used as a set of nodes in a weighted word graph. ; Calculate the total co-occurrence of all unordered word pairs based on unordered co-occurrence counting. : ； Any two words in the word sequence Harmony Calculate the joint probability of word pairs and the marginal probability of words respectively: , ;in, For words Harmony The joint co-occurrence probability, For words The probability of edge co-occurrence; Calculation words Harmony The original value of the correlation strength is then nonnegated: ； ； in, It is the natural logarithm function. The final correlation strength after nonnegation; Based on node set and final correlation strength Constructing a weighted undirected word graph : Representing edge sets, including: ； The weight matrix includes: when hour, ;when Sometimes, or hour, .

4. The method for generalizing text data in economic big data according to claim 3, characterized in that, Based on the weighted word graph, a normalized graph matrix is ​​calculated. Feature analysis is performed on the normalized graph matrix to obtain the second eigenvalue and the corresponding second eigenvector. The edge-level sensitivity value for each edge is also calculated, including: Based on weighted undirected word graph Construct the angle matrix For angle matrix The dimension is equal to the set of nodes. The number of nodes, and satisfying: ； ； ； Define a set of dimension nodes The identity matrix of the number of nodes Calculate the normalized graph matrix : ;in, For the angle matrix The inverse square root matrix; For normalized graph matrix Perform eigenvalue decomposition to solve for the second eigenvalue of the following constraint. and the corresponding second eigenvector ,include: First constraint ; Second constraint ; Third constraint ; in, For a dimension equal to the set of nodes A vector of all 1s representing the number of nodes; Weighted undirected word graph Each edge Calculate the edge sensitivity value : ;in, Represents the second eigenvector At the node The component at that location.

5. The method for generalizing text data in economic big data according to claim 4, characterized in that, For each word, the largest edge-level sensitivity value among its connected edges is taken as the word-level sensitivity. A set of highly sensitive words is obtained by filtering word-level sensitivities based on a preset threshold parameter, including: Determine the node The set of adjacent nodes ; Select node The set of adjacent nodes The largest edge sensitivity value in the middle is used as the node. Word-level sensitivity; If node If the word-level sensitivity is greater than or equal to the preset word-level sensitivity, then the node will be... Mark words as highly sensitive words and merge all highly sensitive words to obtain a set of highly sensitive words.

6. The method for generalizing text data in economic big data according to claim 5, characterized in that, In the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitivity word. The peak neighbor node is the node connected to the high-sensitivity word and has the largest corresponding edge-level sensitivity value. Based on the coordinate difference between the high-sensitivity word and the peak neighbor node and a preset radius parameter, the window radius is determined, an initial word bucket is generated, and the final bucket set is obtained by merging overlapping intervals, including: Based on the second feature vector For any node Define the corresponding one-dimensional coordinates ; Determine the first The set of adjacent nodes corresponding to the nth highly sensitive words and the edge sensitivity values ​​are calculated, and the nth high-sensitivity word is calculated. The peak neighbor node of a high-sensitivity word Among them, the peak-reaching neighbor node is the _th The node adjacent to the most sensitive word and the node with the largest edge sensitivity value; Regarding the first Calculate the window radius corresponding to each highly sensitive word. ;in, For the preset radius parameter, Indicates the first One-dimensional coordinates of highly sensitive words Represents the second eigenvector At the node The component at the location; For the Initial bucket of highly sensitive words ; Map the coordinate range corresponding to each initial word bucket to a feature interval. Calculate the union of all feature intervals and extract all maximally connected subintervals from the union. ; Based on maximally connected subintervals Generate the final bucket set : ;in, For the final number of buckets, for index, for index, Indicates the first One-dimensional coordinates of highly sensitive words.

7. The method for generalizing text data in economic big data according to claim 6, characterized in that, For each final bucket, an encryption algorithm is used to process the preset identifier and bucket number to generate a bucket code. Words within that bucket are replaced with replacement tokens in a preset format, and generalized encrypted text data is output, including: For the Each final bucket constructs its corresponding message to be encrypted through string concatenation operations. ;in, This indicates string concatenation. This indicates that a preset identifier has been set. Using the HMAC-SHA256 encryption algorithm and Base32 encoding, the first... The message to be encrypted in each final bucket is used to construct the corresponding replacement token; For the first word in the original word sequence Word Perform replacement: If the first Word Belongs to the final bucket set Then the first Word Replace with a replacement symbol; otherwise, keep the first one. Word constant; The original word sequence after replacement is used as the result of generalization processing of text data.

8. A system for generalizing text data in economic big data, applied to the method for generalizing text data in economic big data as described in any one of claims 1-7, characterized in that, include: The data processing module performs normalization processing on the raw text data and segments it into word sequences. Based on preset sliding window parameters, it counts the co-occurrence of word pairs in the word sequences. The association strength module calculates the association strength of word pairs based on co-occurrence counts and performs non-negation processing to construct a weighted word graph with association strength as the edge weight; The edge-level sensitivity module calculates a normalized graph matrix based on a weighted word graph, performs feature analysis on the normalized graph matrix to obtain the second eigenvalue and the corresponding second eigenvector, and calculates the edge-level sensitivity value for each edge. The high-sensitivity word module takes the largest edge-level sensitivity value among the connected edges of each word as the word-level sensitivity, and filters the word-level sensitivity based on preset threshold parameters to obtain a set of high-sensitivity words. In the final bucket module, in the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitivity word. The peak neighbor node is the node connected to the high-sensitivity word and has the largest corresponding edge sensitivity value. The window radius is determined based on the coordinate difference between the high-sensitivity word and the peak neighbor node and the preset radius parameter. An initial word bucket is generated and the final bucket set is obtained by merging overlapping intervals. The generalized encryption module uses an encryption algorithm to process the preset identifier and bucket number to generate a bucket code for each final bucket. It then replaces the words in the bucket with a preset replacement token and outputs generalized encrypted text data.

Images