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

Abstract

The invention relates to the technical field of generalization encryption, and discloses a generalization processing method and system for text data in economic big data, and the method comprises the steps: firstly carrying out the standardization and word segmentation of original text data, and carrying out the statistics of a word pair co-occurrence relation through a preset sliding window parameter; then, according to co-occurrence counting, the correlation strength of word pairs is calculated and is not negative, and a weighted word graph with the correlation strength as the edge weight is constructed; calculating a normalized graph matrix and performing feature analysis to obtain a second feature value and a second feature vector, thereby calculating an edge-level sensitive value and determining word-level sensitivity; screening a high-sensitivity word set according to a threshold parameter, generating a window radius in a feature vector coordinate space according to a coordinate difference between a high-sensitivity word and a peak-causing neighbor node and a radius parameter, forming a word bucket, and combining overlapped intervals; and finally, generating bucket codes for each word bucket by adopting an encryption algorithm, replacing words in the buckets, and outputting generalization encrypted text data with privacy protection and structure availability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of generalization encryption, more particularly, it relates to a generalization processing method and system for economic big data Chinese text data. BACKGROUND

[0002] With the acceleration of the process of economic digitization, a large amount of economic text data has been generated in the fields of finance, trade, industry monitoring, etc., and its sources cover various channels such as enterprise announcements, policy documents, news public opinion, transaction records, etc. In order to support application tasks such as macroeconomic prediction, risk identification, credit evaluation, etc., it is usually necessary to depict the potential semantic association between texts through a graph model, in which the nodes correspond to economic entities or keywords, and the edges represent semantic or statistical association strength. However, such text data often contains highly sensitive content such as enterprise identity and transaction information, and must be generalized to protect data security when shared across institutions.

[0003] Existing methods mostly replace and desensitize terms based on word frequency, category or superordinate / subordinate concepts, without fully considering the weight balance relationship between nodes and edges in the graph structure, which can easily cause overall information imbalance and make it difficult to effectively balance privacy protection and data usability. Economic and financial related text generalization processing often uses semantic clustering, vocabulary replacement, edge weight adjustment, etc. However, the semantics composed of economic and financial related texts exhibit a significant non-uniform distribution feature on the network: high-frequency economic terms form dense connections, and low-frequency specific entities / professional terms form sparse bridging relationships. When the generalization operation uniformly adjusts the edge weight or merges semantic nodes, the originally balanced semantic network will experience a connectivity mutation, and the structural influence of some nodes / connections will be abnormally amplified, leading to an imbalance in local information transmission capability. This not only reduces the accuracy of subsequent data analysis, but also may cause the original private information to be exposed through the residual semantic clues at the structure 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), the relevant keywords will form a high-density connected subgraph in the co-occurrence graph, with a local clustering coefficient significantly higher than the global average. If the generalization processing only maps sensitive terms to a vocabulary or generalizes concepts without considering their connectivity features in the graph structure, the degree centrality, betweenness centrality, and other topological indicators of the generalized nodes will still maintain abnormal numerical distribution. Attackers can identify structurally abnormal node clusters through graph topology analysis or spectral clustering methods, and combine public information for association reasoning, thereby breaking through the word-level generalization protection, reversing the inference of sensitive entity identity, and causing privacy leakage risks.

[0004] From the perspective of graph theory, after the economic text is constructed into a graph structure through co-occurrence or dependency relationship, the modification of edge weight, node grouping and connection mode by generalization operation will directly change the connectivity structure and mathematical properties of the graph. At the level of spectral analysis, this change is manifested as rapid fluctuations in key eigenvalues, leading to a doubling of the influence of individual connections on transmission paths. Due to the natural hierarchical and industry concentration characteristics of economic and financial related corpus, small changes in edge weight can cause significant instability in graph connectivity. As the generalization process deepens, some bridge nodes will evolve into key nodes that control the overall structure, making it possible for the encrypted text to be indirectly recognized by structural analysis methods, further exacerbating the conflict between privacy protection and information availability. SUMMARY

[0005] The present application provides a generalization processing method and system for economic big data Chinese text data, which solves the technical problems raised in the background art.

[0006] In a first aspect, a generalization processing method for economic big data Chinese text data is provided, comprising: normalizing and tokenizing the original text data to obtain a word sequence, and counting the co-occurrence of word pairs in the word sequence based on a preset sliding window parameter; calculating the correlation strength of the word pairs based on the co-occurrence count and performing non-negative processing, and constructing a weighted word graph with the correlation strength as the edge weight; calculating a normalized graph matrix based on the weighted word graph, performing feature analysis on the normalized graph matrix to obtain a second eigenvalue and a corresponding second eigenvector, and calculating the edge-level sensitivity value of each edge; for each word, taking the maximum edge-level sensitivity value in the connected edges as the word-level sensitivity of the word, and selecting the high-sensitive word set based on a preset threshold parameter; in the coordinate space of the second eigenvector, selecting the peak neighbor node for each high-sensitive word, which is the node connected to the high-sensitive word with the maximum edge-level sensitivity value; determining the window radius based on the coordinate difference between the high-sensitive word and the peak neighbor node and a preset radius parameter, generating an initial word bucket and obtaining a final bucket set by merging overlapping intervals; for each final bucket, processing the preset identifier and the bucket number using an encryption algorithm to generate a bucket code, replacing the words in the bucket with a replacement symbol in a preset format, and outputting the generalization encrypted text data.

[0007] In a second aspect, a generalization processing system for economic big data Chinese text data is provided, which can be applied to any of the generalization processing methods for economic big data Chinese text data, comprising: a data processing module for normalizing and tokenizing the original text data to obtain a word sequence, and counting the co-occurrence of word pairs in the word sequence based on a preset sliding window parameter; An association strength module calculates the association strength of a word pair according to the co-occurrence count and performs non-negative processing, and constructs a weighted word graph with the association strength as the edge weight; An edge-level sensitivity module calculates a normalized graph matrix based on the weighted word graph, obtains a second eigenvalue and a corresponding second eigenvector through feature analysis of the normalized graph matrix, and calculates an edge-level sensitivity value of each edge; A high-sensitive word module takes the maximum edge-level sensitivity value in the connected edges of each word as the word-level sensitivity of the word, and screens the word-level sensitivity based on a preset threshold parameter to obtain a high-sensitive word set; A final bucket module selects a peak neighbor node for each high-sensitive word in the coordinate space of the second eigenvector, the peak neighbor node being a node connected to the high-sensitive word and having the maximum corresponding edge-level sensitivity value; determines a window radius based on the coordinate difference between the high-sensitive word and the peak neighbor node and a preset radius parameter, generates an initial word bucket, and obtains a final bucket set by merging overlapping intervals; A generalization encryption module processes a preset identifier and a bucket number using an encryption algorithm to generate a bucket code for each final bucket, replaces the words in the bucket with a replacement symbol in a preset format, and outputs generalization encrypted text data.

[0008] The beneficial effects of the present application include: based on the weighted word graph constructed from economic text data, combining the feature analysis of the normalized graph matrix to realize the sensitive recognition of the structure level, which can accurately locate the key nodes and edges affecting the overall connectivity at the spectral level; on this basis, the adaptive window radius is generated by the coordinate difference between the high-sensitive word and the peak neighbor node, the semantic related word bucket set is automatically formed, and the bucket-level replacement symbol is generated by using the encryption algorithm, so as to realize semantic generalization at the word level, maintain connectivity balance at the structure level, and guarantee privacy security at the encryption level. The present application can eliminate the recognizability of sensitive information while maintaining the statistical regularity of the original graph structure after generalization processing of economic text, effectively balancing data security and analysis availability, and is particularly suitable for privacy protection applications of cross-institution sharing and large-scale economic corpus. BRIEF DESCRIPTION OF DRAWINGS

[0009] Fig. 1 is a flowchart of the generalization processing method for economic big data Chinese text data of the present application; Fig. 2 is a module diagram of the generalization processing system for economic big data Chinese text data of the present application. DETAILED DESCRIPTION

[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 Figs. 1-2 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 each position, ; Set preset sliding window parameters For each position in the word sequence Define the forward window set : ; wherein, is a window position offset; for any two words and word in the word sequence, an ordered co-occurrence count is computed as ; wherein, is an indicator function that takes the value 1 when and and 0 otherwise; for any two words and word in the word sequence, an unordered co-occurrence count is obtained by symmetrizing the ordered co-occurrence count as .

[0013] The source of the original text data includes, but is not limited to, the following types: Enterprise-related text: such as the enterprise announcements (including financial disclosure, major investment / cooperation announcements) released by listed companies, enterprise internal operation reports (such as supply chain management text, text description corresponding to customer cooperation agreement); Policy and regulatory text: such as economic regulation policy documents (such as monetary policy details, industry support policies) released by the government and regulatory agencies, financial regulatory rules (such as text descriptions of bank credit approval specifications); Market and public opinion text: such as news reports (including industry trends, market trend analysis) released by financial media, economic event discussion texts (such as comments on changes in supply and demand in a certain industry) in social media or professional forums; Transaction-related text: such as transaction record text (such as clause descriptions in loan contracts, text information in cross-border trade documents) of financial institutions, and explanation text corresponding to transaction credentials between enterprises (such as order notes, economic-related descriptions in logistics information).

[0014] The original text data has two properties: Strong economic relevance: the text content closely revolves around economic activities and contains a large number of economic professional terms, such as GDP growth, monetary policy, supply chain finance, specific economic entity identifiers, such as full name / abbreviation of enterprises, financial institution codes, and transaction detail information, such as text representations of transaction amounts, and names of cooperation parties.

[0015] High privacy sensitivity: the text contains sensitive information that needs to be strictly protected, such as enterprise internal transaction data, operation strategy text of specific institutions, and client-related identity / trade information. Direct cross-institution sharing of such information may easily lead to privacy leakage, which is the motivation for implementing privacy protection through generalization encryption.

[0016] Normalization and tokenization, including: Normalization: Perform standardization operations on the original economic text (such as enterprise announcements, financial public opinion, etc.), such as unifying character encoding to NFKC format, converting all characters to lowercase, and eliminating the same word different forms caused by encoding differences and case inconsistencies. For example, avoid misjudging GDP and gdp as two different words.

[0017] Tokenization: Use a tokenization tool adapted to Chinese economic text, such as a tokenizer optimized for professional terms, to split the normalized text into a continuous word sequence, i.e., a list composed of individual words in the original text order.

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

[0019] The preset sliding window parameter is a positive integer set in advance, representing the maximum distance allowed between two words when counting co-occurrence. This parameter is set to avoid counting distant word pairs that have no actual semantic association. For example, in a 1000-word text, if the word "bank" at the beginning and the word "exchange rate" at the end are more than the sliding window distance, their co-occurrence has no actual meaning.

[0020] For each position t in the word sequence, the forward window set refers to all positions within the sliding window parameter range offset from position t, for example, if the sliding window parameter is 5 and position t = 3, the forward window set is positions 4, 5, 6, 7, 8, and if the total length is insufficient, it is taken to the total length. The forward window set only counts the co-occurrence of the word at position t and the words within its forward window, ensuring that co-occurrence statistics focus on word pairs with closer semantic associations.

[0021] Ordered co-occurrence refers to the explicit order of the appearance of two words, i.e., only counting the case where word a appears at position t and word b appears within the forward window of position t (not counting the case where word b appears first and word a appears later).

[0022] Iterate through all possible positions t and all possible forward window offsets, and add up the results of the indicator function (1 or 0) for each condition that meets the criteria. The final sum is the ordered co-occurrence count of words a and b, representing the total number of times word a appears before word b within the sliding window.

[0023] The purpose of the symmetrization processing: in economic texts, the semantic association of many word pairs does not depend on the order of the sequence, for example, monetary policy and interest rate, no matter who is in front of who, it represents the relevant concept. The ordered co-occurrence count only counts the single order, and cannot fully reflect the actual association strength of the word pair, so it is necessary to eliminate the order influence through symmetrization processing.

[0024] The unordered co-occurrence count represents the total number of co-occurrence of word a and word b in the sliding window range (without considering the order before and after), which is more in line with the actual statistical needs of the semantic association of word pairs in economic texts.

[0025] In an embodiment of the present application, the association strength of the word pair is calculated according to the co-occurrence count and non-negative processing is performed, and a weighted word graph with the association strength as the edge weight is constructed, including: Each word in the word sequence is taken as a node set of the weighted word graph ; The total co-occurrence amount of all unordered word pairs is calculated based on the unordered co-occurrence count : ; For any two words and word in the word sequence, the joint probability of the word pair and the marginal probability of the word are calculated respectively: , ; wherein, is the joint co-occurrence probability of word and word , is the marginal co-occurrence probability of word ; The association strength original value of word and word is calculated and non-negative processing is performed: ; ; Wherein, is the natural logarithm function, is the final association strength after non-negative processing; Based on the node set and the final association strength , a weighted undirected word graph is constructed: Indicates the edge set, including: ; Indicates the weight matrix, including: When , ; when Time, or Time, .

[0026] Each word in the word sequence as a node set of a weighted word graph refers to each independent word in the word sequence after the original text is split as a node of the weighted word graph (i.e. the basic unit of the graph). For example, after the economic text is split, the words such as monetary policy, interest rate, GDP, and inflation are obtained, and each word corresponds to a node. The node set is the total of all independent words.

[0027] The total co-occurrence of all unordered word pairs is calculated based on the unordered co-occurrence count, which needs to be combined with the unordered co-occurrence count: All unordered word pairs refer to pairs composed of all 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 as the same pair, which is not repeated). The total co-occurrence is the sum of the unordered co-occurrence counts of all 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, so the total co-occurrence is 8+5+200=213.

[0028] Joint probability of word pair: For any two words, such as monetary policy and interest rate, the joint probability is the unordered co-occurrence count of the two words divided by the total co-occurrence. For example, the unordered co-occurrence count of monetary policy-interest rate is 8, and the total co-occurrence is 213, so the joint probability is 8 / 213≈0.037. The joint probability reflects the frequency of the co-occurrence of the two words in the sliding window, and the larger the value, the more frequent the association of the two words.

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

[0030] Calculate the original value of the association strength: use the natural logarithm function, and use the product of the joint probability of the word pair and the marginal probability of the two words as the input of the logarithm. If the frequency of the co-occurrence of the two words is much higher than the product of the independent appearance frequency of each word (i.e. joint probability> product of marginal probability), 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≈product of marginal probability, the original value is close to 0, indicating that the association between the two words is weak. If the joint probability<product of marginal probability, the original value is negative, indicating that the two words have little association, such as monetary policy and ice cream.

[0031] Non-negative processing: all negative values and 0 in the original association strength are adjusted to 0, only positive values are retained, and the final association strength is obtained. The reason is that in the semantic graph of economic text, the negative association or no association between word pairs is meaningless for subsequent structure analysis, such as the negative association between monetary policy and ice cream does not need to be reflected in the graph. Non-negative processing can simplify the graph structure and only retain the strength values of the word pairs with actual semantic association.

[0032] The edge set is the total of the connected word pairs in the graph, and only contains 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 between monetary policy and interest rate is 2.1 (> 0), so there is an edge between the two nodes; the strength between monetary policy and ice cream is 0, so there is no connection, which ensures that only meaningful semantic connections are retained 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 the pair of a word with itself, the value at the corresponding position in the matrix is 0.

[0034] Since all word pairs are based on unordered co-occurrence counts and unordered association strength, the connection of the edge has no direction (the edge of word A-word B is the same as the edge of word B-word A, with the same strength), which conforms to the symmetry of semantic association between word pairs in economic text, such as the association between monetary policy affecting interest rate and interest rate reflecting monetary policy is bidirectional.

[0035] In one embodiment of the present application, a normalized graph matrix is calculated based on the weighted word graph, a second eigenvalue and a corresponding second eigenvector are obtained by performing feature analysis on the normalized graph matrix, and an edge level sensitivity value of each edge is calculated, including: Based on the weighted undirected word graph , a diagonal matrix is constructed , the dimension of the diagonal matrix is equal to the number of nodes in the node set , and satisfies: ; ; ; A unit matrix with a dimension equal to the number of nodes in the node set is defined, and a normalized graph matrix is calculated: ; wherein, is the inverse square root matrix of the diagonal matrix ; The normalized graph matrix Eigenvalue decomposition is performed to solve the second eigenvalue of the following constraint and the corresponding second eigenvector , comprising: a first constraint ; a second constraint ; a third constraint ; wherein is a full one vector with a dimension equal to the number of nodes in the node set ; For each edge in the weighted undirected word graph , the edge-level sensitive value is calculated: ; wherein denotes the component of the second eigenvector at node .

[0036] The angle matrix is a diagonal matrix (only the diagonal position has a value in the matrix, and the non-diagonal position is 0) used to describe the total connection strength of each node in the word graph, which is as follows: Matrix dimension: The dimension is equal to the number of nodes in the weighted word graph, that is, each node corresponds to an element of a diagonal line in the matrix, ensuring that the connection strength of each node can be quantified separately.

[0037] Diagonal element value: The value of a diagonal element of the diagonal matrix is equal to the sum of the associated strengths of all connected edges of the node corresponding to the element. For example, the edge association strengths of the node monetary policy with interest rate, inflation, and exchange rate are 2.1, 1.8, and 1.5, respectively, so the corresponding diagonal element value is 2.1+1.8+1.5=5.4; if a node has no connected edges (extreme case), the diagonal element value is 0.

[0038] Non-diagonal element value: The element values of all non-diagonal positions are 0, which is an inherent property of the diagonal matrix, focusing only on the total connection of the node itself and not involving the direct association between nodes.

[0039] The normalized graph matrix is the matrix obtained by standardizing the weight matrix of the weighted word graph, the core of which is to eliminate the bias caused by the number of connections of the node itself (i.e., the diagonal value of the diagonal matrix), so that the edge weight more objectively reflects the real semantic association between nodes. The normalized graph matrix can depict the structural characteristics of the word graph and avoid distortion of the edge weight of high-frequency nodes due to excessive connections.

[0040] By decomposing the normalized graph matrix, a series of eigenvalues and corresponding eigenvectors are obtained, among which the first eigenvalue and the first eigenvector usually correspond to the case of no difference among all nodes (such as the all-1 vector), which has no structural distinction; while the second eigenvalue and the second eigenvector can reflect the cluster structure difference of the word graph. That is, which nodes belong to the same semantic cluster, and which nodes are the key nodes connected across clusters.

[0041] The first constraint (eigen equation relationship): the normalized graph matrix multiplied by the second eigenvector is equal to the second eigenvalue multiplied by the second eigenvector. This is the basic definition of eigenvalue and eigenvector, which ensures that the vector is a characteristic vector that can reflect the structure of the matrix, and the value is the corresponding eigenvalue.

[0042] The second constraint (unit norm): the norm of the second eigenvector is equal to 1. Thus, the influence of the difference in the scale of the vector is eliminated. If the norm is not limited, the same eigenvalue may correspond to numerous eigenvectors of different scales. After the norm is unified, the value range of the eigenvector is fixed.

[0043] The third constraint (orthogonal to the all-1 vector): the inner product of the second eigenvector and the all-1 vector (a vector with all elements equal to 1 and the same dimension as the number of nodes) is 0 (i.e. orthogonal). The all-1 vector represents the state of no difference among all nodes, and the orthogonality means that the second eigenvector can break through the state of no difference and accurately reflect the difference between nodes. That is, the nodes within the cluster have similar values on the second eigenvector, and the values of the nodes across the clusters differ significantly.

[0044] The size of the second eigenvalue reflects the overall connectivity of the word graph. The larger the value, the stronger the connectivity of the word graph (the connection between clusters is tight); the smaller the value, the weaker the connectivity (the connection between clusters is sparse, and there are bottlenecks). The value distribution of the second eigenvector reflects the cluster attribution of the nodes, and the nodes with similar values belong to the same semantic cluster, and the nodes with large value differences belong to different clusters.

[0045] The edge-level sensitive value is an indicator that measures the sensitivity of each edge to the connectivity of the word graph (i.e. the second eigenvalue), and the value of the edge is greater, the greater the influence on the overall connectivity of the word graph, i.e. the structural sensitive edge.

[0046] The edge-level sensitive value is equal to the correlation strength of the edge multiplied by the square of the difference in components of the two nodes on the second eigenvector. Among them: The square of the difference in components: if the two nodes belong to the same cluster, the difference in components is small, and the value after squaring is smaller; if they belong to different clusters (cross-cluster edges), the difference in components is large, and the value after squaring is larger, i.e. this part directly quantifies the cross-cluster property of the edge, and the stronger the cross-cluster property, the greater the influence on the connectivity.

[0047] Multiply the correlation strength of the edge: the edge with large correlation strength itself contributes greatly to the connection between nodes, and after combining the cross-cluster property, its sensitivity to the overall connectivity will be further amplified.

[0048] The edge level sensitivity value can directly screen the edges that are most critical to the connectivity of the word graph, that is, these edges are usually the core bridging edges across clusters, such as the edge between monetary policy and GDP in economic text.

[0049] In an embodiment of the present application, for each word, the maximum edge level sensitivity value in the connected edges of the word is taken as the word level sensitivity of the word, and a high-sensitive word set is obtained by screening the word level sensitivity based on a preset threshold parameter, including: determining the set of adjacent nodes of the node ; ; selecting the maximum edge level sensitivity value in the set of adjacent nodes of the node as the word level sensitivity of the node ; if the word level sensitivity of the node is greater than or equal to a preset word level sensitivity, marking the node as a high-sensitive word, and merging all high-sensitive words to obtain a high-sensitive word set.

[0050] The set of adjacent nodes is a set of all other nodes connected to a single node in a weighted undirected word graph, and is specifically as follows: Source of the set: based on a weighted undirected word graph, the edges in the word graph represent semantic association (association strength > 0) between nodes (words), so the adjacent nodes of a node are all nodes connected to the node through edges with an association strength > 0.

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

[0052] The word level sensitivity is an index for measuring the degree of influence of a single node on the connectivity of the word graph, and its calculation logic is based on the maximum edge level sensitivity value of the associated edges of the node, and is specifically as follows: From all the edge level sensitivity values of the adjacent edges of the current node, the maximum value is selected as the word level sensitivity of the node. For example, the edge level sensitivity values of the nodes monetary policy, interest rate, and inflation are 5.2, 3.8, and 4.1, respectively, and the maximum value 5.2 is the word level sensitivity of the node monetary policy.

[0053] ​The influence of the node on the word graph structure is determined by the most sensitive edge of the node, that is, the edge is the key branch point connecting the node and other parts of the word graph, and if the weight of the edge changes, the connectivity of the word graph will be most affected, therefore, the sensitivity of the node is characterized by the sensitivity value of the edge, which can reflect the structural importance of the node.

[0054] The high-sensitive word set is a set of nodes selected from all nodes, and the word-level sensitivity of the nodes reaches a preset threshold, and the specific process is as follows: The meaning of the preset threshold: the preset word-level sensitivity is a value set in advance according to the privacy protection demand of economic text and the structure characteristics of the word graph (according to the generalization experience in the past, the threshold is set to 4.0), and the function is to distinguish nodes with strong structural influence and nodes with weak structural influence, that is, the higher the threshold, the fewer the high-sensitive words selected, and all the nodes are core nodes with extremely strong structural influence; the lower the threshold, the more the high-sensitive words, and the nodes cover more nodes with certain structural influence.

[0055] The screening rule: compare the word-level sensitivity of each node with the preset threshold one by one, if the word-level sensitivity of the node is greater than or equal to the threshold, the node is marked as a high-sensitive word; if it is less than the threshold, it is not marked. All nodes marked as high-sensitive words are collected to obtain a high-sensitive word set. For example, if the preset threshold is 4.0, the nodes of monetary policy (sensitivity 5.2) and GDP (sensitivity 4.5) meet the standard and are included in the set; the node of inflation (sensitivity 3.8) does not meet the standard and is not included.

[0056] The screening purpose: the high-sensitive word is a key node in the word graph, that is, the change of semantic association of these nodes will significantly affect the connectivity of the word graph, and if the generalization is not performed, it is easy to become a breakthrough for privacy leakage (for example, through the association mode of these nodes, the sensitive information of the original text is deduced).

[0057] In an embodiment of the present application, in the coordinate space of the second feature vector, a peak neighbor node is selected for each high-sensitive word, the peak neighbor node is a node connected to the high-sensitive word and having the maximum edge-level sensitivity value; a window radius is determined based on the coordinate difference between the high-sensitive word and the peak neighbor node and a preset radius parameter, 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 , a one-dimensional coordinate corresponding to the node is defined . The adjacent node set corresponding to the first high-sensitive word and the edge-level sensitivity value are determined, and the peak neighbor node of the first high-sensitive word is calculated ; wherein the peak neighbor node is the node adjacent to the first high-sensitive word and having the maximum edge-level sensitivity value; For the first a corresponding window radius is calculated for the high-sensitive word ; wherein, a preset radius parameter, represents the one-dimensional coordinate of the high-sensitive word, represents the component of the second feature vector at the node ; an initial word bucket for the high-sensitive word is generated; the coordinate range corresponding to each initial word bucket is mapped to a feature interval ; the union set of all feature intervals is calculated, and all maximal connected sub-intervals in the union set are extracted ; based on the maximal connected sub-intervals , a final bucket set is generated: ; wherein, the number of final buckets, is the index of , is the index of , represents the one-dimensional coordinate of the high-sensitive word.

[0058] Based on the second feature vector, the one-dimensional coordinate of the node is defined, which converts the abstract nodes (words) in the word graph into specific positions on the one-dimensional number axis, so that the structural association (in-cluster / cross-cluster) between nodes can be reflected through the coordinate distance, as follows: The role of the second feature vector is to depict the cluster structure of the word graph (in-cluster node vector components are similar, and cross-cluster node components differ greatly), so the component value corresponding to each node in the second feature vector is directly taken as the one-dimensional coordinate of the node.

[0059] For example, if the component value of the node monetary policy in the second feature vector is 0.32, the interest rate is 0.35, and the GDP is -0.28, then the one-dimensional coordinates of the three are 0.32, 0.35, and -0.28, respectively. The coordinates of monetary policy and interest rate are close (belonging to the same semantic cluster), and the coordinates of GDP differ greatly from them (belonging to different semantic clusters). Reducing the high-dimensional word graph structure (the association between nodes) to one-dimensional coordinates can divide the word buckets through coordinate intervals, so that the generation logic of the word buckets is transformed from semantic association to geometric position proximity, which simplifies the grouping operation and accurately matches the cluster structure of the word graph.

[0060] The peak neighbor node is the most influential adjacent node in the high-sensitive word association edge on the connectivity of the word graph, as follows: ​For the i-th high-sensitive word, first determine its set of adjacent nodes (all nodes connected to the high-sensitive word), and then filter out the adjacent node with the maximum edge level sensitivity value from the set, which is the peak-causing neighbor node.

[0061] For example, the adjacent nodes of the high-sensitive word monetary policy are interest rate, inflation, and exchange rate, and the corresponding edge level sensitivity values are 5.2, 3.8, and 4.1, respectively. The maximum value of 5.2 corresponds to the adjacent node interest rate, so interest rate is the peak-causing neighbor node of monetary policy.

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

[0063] The window radius is a parameter that controls the size of the initial word bucket corresponding to a single high-sensitive word, and is as follows: The preset radius parameter: a positive number less than or equal to 1 (such as 0.5) that is set in advance to adjust the size of the window radius and balance the generalization range and semantic accuracy.

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

[0065] The window radius is dynamically determined by the coordinate difference + preset parameter, i.e., the window radius is small (the generalization range is concentrated) for high-sensitive words with close structural associations and peak-causing neighbors, and the radius can be adjusted to expand for slightly looser associations, ensuring that the size of each initial word bucket is adapted to its structural association strength, avoiding excessive or insufficient generalization caused by a uniform radius.

[0066] The initial word bucket is a temporary bucket formed by grouping nodes with one-dimensional coordinates within the window range of a single high-sensitive word, and is as follows: All nodes with one-dimensional coordinates falling within the interval [high-sensitive word coordinate - window radius, high-sensitive word coordinate + window radius] are included in the initial word bucket corresponding to the high-sensitive word.

[0067] For example, the high-sensitive word monetary policy has a coordinate of 0.32 and a window radius of 0.015, and the window range is [0.305, 0.335]. Therefore, nodes with one-dimensional coordinates between 0.305 and 0.335 (such as monetary policy 0.32 and interest rate derivative 0.318) are included in the initial word bucket.

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

[0069] Merging overlapping intervals is to integrate the coordinate ranges of all initial word buckets into non-overlapping maximal connected sub-intervals, as follows: First, convert the window range of each initial word bucket ([high-sensitive word coordinate - radius, high-sensitive word coordinate + radius]) into a feature interval, and each initial word bucket corresponds to a feature interval.

[0070] Merge all feature intervals to obtain a total interval containing the ranges of all initial intervals (which may consist of multiple discontinuous sub-intervals).

[0071] Extracting maximal connected sub-intervals: From the total interval, filter out the continuous sub-intervals without overlap and gaps, which are the maximal connected sub-intervals. For example, after merging the initial feature intervals [0.305, 0.335] and [0.33, 0.355], the maximal connected sub-interval is [0.305, 0.355]; if there is another interval [0.8, 0.82], this interval is another maximal connected sub-interval.

[0072] Solving the overlap conflict of initial word buckets (avoiding the same node belonging to multiple buckets), while grouping nodes with close semantic association and continuous coordinates into the same interval, to ensure that the final buckets generated later can match the natural cluster structure of the word graph, rather than artificial segmentation of scattered nodes.

[0073] The final bucket set is a non-overlapping, semantically associated node set formed by grouping nodes within the maximal connected sub-interval, as follows: Each maximal connected sub-interval corresponds to a final bucket, and all nodes with one-dimensional coordinates within the maximal connected sub-interval are grouped into the corresponding final bucket; if the node coordinates are not within any maximal connected sub-interval (non-high-sensitive and weakly associated with high-sensitive words), they are not grouped into any word bucket.

[0074] For example, the maximal connected sub-interval [0.305, 0.355] corresponds to final bucket 1, and the nodes with coordinates within this interval, such as monetary policy, interest rate, and interest rate derivatives, are grouped into it. The maximal connected sub-interval [0.8, 0.82] corresponds to final bucket 2, and the nodes with coordinates within this interval, such as exchange rate futures and cross-border settlement, are grouped into it. Finally, a final bucket set containing two buckets is formed.

[0075] The nodes (words) in the same final bucket will be uniformly replaced by the same encryption symbol, which not only realizes the privacy protection of sensitive words, but also can still retain the overall structure rule of the word graph after replacement because the semantic association of nodes in the bucket is close.

[0076] In an embodiment of the application, for each final bucket, a preset identifier and a bucket serial number are processed by an encryption algorithm to generate a bucket code, and the words in the bucket are replaced by a replacement symbol in a preset format, and the output generalized encrypted text data includes: The first final bucket is used to construct a corresponding encrypted message to be encrypted by string concatenation operation. represents string concatenation, represents setting a preset identifier; The HMAC-SHA256 encryption algorithm and Base32 encoding are used to construct a corresponding replacement symbol for the encrypted message to be encrypted of the first final bucket; The first word in the original word sequence is replaced: if the first word belongs to the final bucket set , the first word is replaced by the replacement symbol; otherwise, the first word is unchanged; The replaced original word sequence is used as the generalization processing result of the text data.

[0077] The encrypted message to be encrypted is a string formed by concatenating the preset identifier and the bucket serial number for a single final bucket, which is as follows: The preset identifier is a fixed string (such as ECON_TEXT) set in advance, which is used 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 repetition of bucket codes of different tasks and ensure the uniqueness of the symbol.

[0078] The bucket serial number is the unique number of each final bucket (such as 1, 2, 3…), which is determined by the final bucket set according to the generation order or coordinate range, which ensures that the serial number of each bucket in the same task is not repeated.

[0079] The preset identifier and the bucket serial number are combined into a complete string by string concatenation, which is the encrypted message to be encrypted. For example, the preset identifier is ECON_TEXT, and the serial number of the third final bucket is 3, then the encrypted message to be encrypted is ECON_TEXT3.

[0080] Thus, the encryption input of each final bucket is ensured to be unique, even if the sequence numbers of different buckets are the same (cross-task scenario), because the preset identifiers are different, the messages to be encrypted are still different; the sequence numbers are unique within the same task, further ensuring the uniqueness of the messages.

[0081] The replacement symbol is obtained by processing the message to be encrypted by a specific encryption algorithm and encoding, and is used to replace the words in the bucket. The HMAC-SHA256 encryption algorithm and Base32 encoding are used, as follows: First step: HMAC-SHA256 encryption processing. HMAC-SHA256 is a keyed hash algorithm that requires input of the message to be encrypted and a preset key (only the processor has it to ensure security), and outputs a fixed-length binary hash value. The advantage of this algorithm is tamper resistance, that is, even if the message to be encrypted differs by only one character (such as ECON_TEXT3 and ECON_TEXT4), the output hash value is completely different; and without the key, the original message cannot be deduced in reverse, ensuring privacy and security.

[0082] Second step: Base32 encoding processing. The binary hash value output by HMAC-SHA256 is converted into a string composed of uppercase letters and numbers (0-9) (Base32 encoding rules). The purpose of this encoding is to standardize the format, that is, binary hash values cannot be directly used for text replacement (containing invisible characters), and the string obtained after Base32 encoding has no special characters, is easy to identify, and the length is controllable (such as the first 8 characters), which is suitable for use as a replacement symbol in text.

[0083] Symbol format: A preset format marker (such as angle brackets + prefix G_) is usually added to the encoded string to form the final replacement symbol, for example, if the encoded string is J2K3L4M5, the replacement symbol is <G_J2K3L4M5>, so that the replacement symbol and the original word can be quickly distinguished.

[0084] The replacement symbol takes into account both security and usability. HMAC-SHA256 encryption ensures that the symbol cannot be cracked in reverse (protecting the privacy of sensitive words in the bucket), Base32 encoding ensures that the symbol format is standardized (adapted to the text replacement scenario), and the format marker improves the recognition of the symbol.

[0085] Word sequence replacement is to traverse the original word sequence and decide whether to replace it with the corresponding replacement symbol according to whether each word belongs to the final bucket, as follows: If the tth word (such as monetary policy) in the original word sequence belongs to a certain final bucket (such as the 3rd bucket, corresponding to the symbol <G_J2K3L4M5>), replace the word with the corresponding replacement symbol; If the tth word does not belong to any final bucket (such as ordinary goods, non-structural sensitive words), keep the word unchanged without replacement.

[0086] For example, the original word sequence is that monetary policy affects interest rate, and general commodity price is stable. If monetary policy and interest rate belong to the third bucket (denoted as <G_J2K3L4M5>), and general commodity does not belong to any bucket, then the replaced sequence is <G_J2K3L4M5> affects <G_J2K3L4M5>, and general commodity price is stable.

[0087] By replacing only the structure-sensitive words (high-sensitive words and associated nodes) in the final bucket, the recognizability of the structure-sensitive words is eliminated; the non-sensitive words remain unchanged, so as to avoid excessive generalization and loss of text semantics, and ensure that the replaced text can still support economic data analysis (such as semantic association recognition and trend judgment).

[0088] Embodiment Two The generalization processing system for economic big data Chinese text data is applied to any one of the generalization processing methods for economic big data Chinese text data, and has the characteristics that the generalization processing system comprises: The data processing module performs normalization processing on the original text data and obtains a word sequence by word segmentation, and counts the co-occurrence of word pairs in the word sequence based on a preset sliding window parameter; The association strength module calculates the association strength of the word pairs according to the co-occurrence count and performs non-negative processing, and constructs a weighted word graph with the association strength as the edge weight; The edge-level sensitive module calculates a normalized graph matrix based on the weighted word graph, obtains a second eigenvalue and a corresponding second eigenvector by performing feature analysis on the normalized graph matrix, and calculates an edge-level sensitive value of each edge; The high-sensitive word module takes the maximum edge-level sensitive value in the connected edges of each word as the word-level sensitivity of the word, and selects a high-sensitive word set based on a preset threshold parameter; The final bucket module selects a peak neighbor node for each high-sensitive word in the coordinate space of the second eigenvector, and the peak neighbor node is a node connected to the high-sensitive word and having the maximum edge-level sensitive value; determines a window radius based on the coordinate difference between the high-sensitive word and the peak neighbor node and a preset radius parameter, generates an initial word bucket, and obtains a final bucket set by merging overlapping intervals; The generalization encryption module generates a bucket code by processing a preset identifier and a bucket number using an encryption algorithm for each final bucket, replaces the words in the bucket with a replacement mark in a preset format, and outputs the generalization encrypted text data.

[0089] The above describes the embodiments of the present embodiment, but the present embodiment is not limited to the specific embodiments described above, and the specific embodiments described above are only illustrative but not limiting, and a person of ordinary skill in the art can make many forms under the inspiration of the present embodiment, which all belong to the protection of the present embodiment.

Claims

1. A generalization processing method for economic big data Chinese text data, characterized in that, The method comprises the following steps: normalizing and tokenizing the original text data to obtain a word sequence, and counting the co-occurrence of word pairs in the word sequence based on a preset sliding window parameter; calculating the association strength of the word pairs according to the co-occurrence count and performing non-negative processing, and constructing a weighted word graph with the association strength as the edge weight; calculating a normalized graph matrix based on the weighted word graph, performing feature analysis on the normalized graph matrix to obtain a second eigenvalue and a corresponding second eigenvector, and calculating the edge level sensitivity value of each edge; for each word, the maximum edge level sensitivity value in the connected edge is taken as the word level sensitivity of the word, and a high sensitivity word set is obtained based on a preset threshold parameter; in the coordinate space of the second eigenvector, for each high sensitivity word, a peak neighbor node is selected, which is the node connected to the high sensitivity word and has the maximum edge level sensitivity value; based on the coordinate difference between the high sensitivity 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; for each final bucket, a bucket code is generated by processing a preset identifier and a bucket serial number using an encryption algorithm, and the words in the bucket are replaced with a replacement symbol in a preset format, and the generalized encrypted text data is output. 2.The method for generalization processing of economic big data Chinese text data according to claim 1, characterized in that, The method comprises the following steps: Normalization is performed on the original text data, and then word segmentation is performed to obtain a word sequence: ; wherein, represents the total length of the word sequence, represents the word at the th position in the word sequence, ; Setting preset sliding window parameters for each position in the word sequence defining a set of forward windows : ; wherein, is a window position offset; For any two words in a sequence of words and word , compute an ordered co-occurrence count : ; wherein denotes an indicator function, which takes the value 1 when and 0 otherwise; symmetrizing the pair-ordered co-occurrence count of any two words in a word sequence and the pair of words to obtain an unordered co-occurrence count : 。 3.The method for generalization processing of economic big data Chinese text data according to claim 2, characterized in that, normalizing and tokenizing the original text data to obtain a word sequence, and counting the co-occurrence of word pairs in the word sequence based on a preset sliding window parameter; each word in the sequence of words as a node in a set of weighted word graphs ; calculating the total amount of co-occurrence of all disordered word pairs based on disordered co-occurrence count : ； For any two words in the word sequence and word , the joint probability of the word pair and the marginal probability of the word are calculated, respectively: , ; wherein, is the joint co-occurrence probability of the words and the words , is the marginal co-occurrence probability of the words ; Computing the association strength original value of the word and the word and performing non-negativity processing: ； ； wherein, is the natural logarithm function, is the final correlation strength after non-negation; Based on a set of nodes And final correlation strength Constructing a weighted undirected word graph : represents a set of edges, comprising: ； denotes a weight matrix comprising: When , ; when , or , . 4.The method for generalization processing of economic big data Chinese text data according to claim 3, characterized in that, calculating the association strength of the word pairs according to the co-occurrence count and performing non-negative processing, and constructing a weighted word graph with the association strength as the edge weight; Weighted undirected word graph based , constructing a diagonal matrix , diagonal matrix has a dimension equal to the number of nodes of the node set , and satisfies: ； ； ； defining a set of nodes of a dimension identity matrix of the number of nodes , computing a normalized graph matrix : ; wherein is the inverse square root matrix of the diagonal matrix ; to the normalized graph matrix performing eigenvalue decomposition to solve the second eigenvalue of the following constraint and the corresponding second eigenvector comprising: first constraint ; Second constraint ; Third constraint ; wherein, is a vector of all ones with dimension equal to the number of nodes in the set of nodes . Weighted undirected word graph each edge computing edge-level sensitive values : ; wherein, represents the second feature vector at the component of the node . 5.The method for generalization processing of economic big data Chinese text data according to claim 4, characterized in that, calculating a normalized graph matrix based on the weighted word graph, performing feature analysis on the normalized graph matrix to obtain a second eigenvalue and a corresponding second eigenvector, and calculating the edge level sensitivity value of each edge; determining nodes adjacent to the node ; selecting a node a set of neighboring nodes of the node a maximum edge-sensitive value among the corresponding edge-sensitive values as a word sensitivity of the node a maximum edge-sensitive value among the corresponding edge-sensitive values as a word sensitivity of the node If the word level sensitivity of the node is greater than or equal to the preset word level sensitivity, the node is marked as a high-sensitive word, and all high-sensitive words are merged to obtain a high-sensitive word set. 6.The method for generalization processing of economic big data Chinese text data according to claim 5, characterized in that, for each word, the maximum edge level sensitivity value in the connected edge is taken as the word level sensitivity of the word, and a high sensitivity word set is obtained based on a preset threshold parameter; based on the second feature vector for any node define a corresponding one-dimensional coordinate ; Determine the set of adjacent nodes corresponding to the first high-sensitive word and the edge level sensitive value, and calculate the peak neighbor node of the first high-sensitive word ; wherein the peak neighbor node is the node adjacent to the first high-sensitive word and has the maximum edge level sensitive 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; To the first high sensitivity word initial word bucket ; mapping the coordinate range corresponding to each initial word bucket to a feature interval ; computing the union of all feature intervals and extracting all maximal connected sub-intervals in the union ; Based on maximal connected sub-intervals , generating final buckets set : ; wherein, is the number of final buckets, is an index of, is an index of, denotes the one-dimensional coordinate of the th high-sensitive word. 7.The method for generalization processing of economic big data Chinese text data according to claim 6, characterized in that, in the coordinate space of the second eigenvector, for each high sensitivity word, a peak neighbor node is selected, which is the node connected to the high sensitivity word and has the maximum edge level sensitivity value; based on the coordinate difference between the high sensitivity 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; A corresponding to-be-encrypted message is constructed for the last bucket through a string concatenation operation ; wherein, ; wherein, denotes string concatenation, denotes setting a preset identifier; The HMAC-SHA256 encryption algorithm and Base32 encoding are used to construct the corresponding replacement token for the to-be-encrypted message of the first final bucket. ​ the first word in the original word sequence perform a replacement: if the first word belongs to the final bucket set replace the first word with a replacement token; otherwise, leave the first word unchanged ​​​​​​​ for each final bucket, a bucket code is generated by processing a preset identifier and a bucket serial number using an encryption algorithm, and the words in the bucket are replaced with a replacement symbol in a preset format, and the generalized encrypted text data is output.

8. A system for generalization processing of economic big data Chinese text data, applied in the method for generalization processing of economic big data Chinese text data according to any one of claims 1-7, characterized in that, The method comprises the following steps: normalizing and tokenizing the original text data to obtain a word sequence, and counting the co-occurrence of word pairs in the word sequence based on a preset sliding window parameter; calculating the association strength of the word pairs according to the co-occurrence count and performing non-negative processing, and constructing a weighted word graph with the association strength as the edge weight; calculating a normalized graph matrix based on the weighted word graph, performing feature analysis on the normalized graph matrix to obtain a second eigenvalue and a corresponding second eigenvector, and calculating the edge level sensitivity value of each edge; for each word, the maximum edge level sensitivity value in the connected edge is taken as the word level sensitivity of the word, and a high sensitivity word set is obtained based on a preset threshold parameter; in the coordinate space of the second eigenvector, for each high sensitivity word, a peak neighbor node is selected, which is the node connected to the high sensitivity word and has the maximum edge level sensitivity value; based on the coordinate difference between the high sensitivity 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; for each final bucket, a bucket code is generated by processing a preset identifier and a bucket serial number using an encryption algorithm, and the words in the bucket are replaced with a replacement symbol in a preset format, and the generalized encrypted text data is output. The method comprises the following steps: normalizing and tokenizing the original text data to obtain a word sequence, and counting the co-occurrence of word pairs in the word sequence based on a preset sliding window parameter; calculating the association strength of the word pairs according to the co-occurrence count and performing non-negative processing, and constructing a weighted word graph with the association strength as the edge weight; calculating a normalized graph matrix based on the weighted word graph, performing feature analysis on the normalized graph matrix to obtain a second eigenvalue and a corresponding second eigenvector, and calculating the edge level sensitivity value of each edge; The high-sensitive word module takes, for each word, the maximum edge-level sensitivity value in the connected edges of the word as the word-level sensitivity of the word, and filters the word-level sensitivity based on a preset threshold parameter to obtain a high-sensitive word set; The final bucket module selects, in the coordinate space of the second feature vector, a peak-causing neighbor node for each high-sensitive word, the peak-causing neighbor node being a node connected to the high-sensitive word and having the maximum corresponding edge-level sensitivity value; determines a window radius based on the coordinate difference between the high-sensitive word and the peak-causing neighbor node and a preset radius parameter, generates an initial word bucket, and obtains a final bucket set by merging overlapping intervals; The generalization encryption module processes, for each final bucket, a preset identifier and a bucket serial number using an encryption algorithm to generate a bucket code, replaces the words in the bucket with replacement symbols in a preset format, and outputs generalization encrypted text data.

Images