Information extraction processing method and system applied to data sharing service

Abstract

The invention provides an information extraction processing method and system applied to a data sharing service, and the method comprises the steps: obtaining a historical interaction text set from a cross-department business data sharing system, carrying out the entity relation network construction of the historical interaction text set, generating a multi-dimensional entity relation network comprising entity nodes, relation edges and attribute tags, and carrying out the data sharing service through the multi-dimensional entity relation network. Performing semantic topological graph construction based on the multi-dimensional entity relationship network to obtain a semantic topological graph containing entity association strength and a semantic evolution path; calling a preset zero-knowledge proof protocol to perform access control vector generation on the semantic topological graph to obtain an access control vector containing a topological weight threshold and a node access permission; and performing joint processing on the access control vector and the semantic topological graph through a space-time correlation analysis model to generate a business event template containing a standardized field and a correlation index. According to the invention, the accuracy, security and scene adaptability of information extraction in the data sharing service can be improved.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to an information extraction and processing method and system applied to data sharing services. Background Technology

[0002] With the deepening of business informatization, cross-departmental business data sharing services are increasingly becoming a key support for improving business efficiency. Information extraction and processing technology, as a fundamental link in data sharing, plays a crucial role in extracting key information from massive amounts of unstructured text interaction records. Currently, information extraction methods in business data sharing typically rely on traditional natural language processing techniques. These techniques use entity recognition and relationship extraction to extract entity information such as business terms, organization names, and business identifiers from text, and establish simple relationships between entities. However, in practical applications, these methods are often limited to surface-level parsing of text content, making it difficult to transform scattered text interaction records into structured entity relationship networks. This results in a lack of systematic and dynamic organization of entity relationships. Furthermore, due to the special requirements for data security and access control in cross-departmental data sharing scenarios, existing methods struggle to achieve fine-grained access control management based on the semantic strength of relationships between entities, easily leading to data leakage risks. In addition, the formats of text interaction records vary significantly across different business scenarios, and existing information extraction results lack a unified standardized template, resulting in low data reuse efficiency. This makes it difficult to meet the actual needs of cross-departmental business collaboration for information extraction results, hindering the improvement of the quality and efficiency of business data sharing services. Summary of the Invention

[0003] This invention provides a method and system for information extraction and processing applied to data sharing services.

[0004] In a first aspect, embodiments of the present invention provide an information extraction and processing method for data sharing services. The method includes: obtaining a historical interaction text set from a cross-departmental business data sharing system, the historical interaction text set containing text interaction records with timestamps generated under different business scenarios; constructing an entity relationship network on the historical interaction text set to generate a multi-dimensional entity relationship network containing entity nodes, relationship edges, and attribute labels, the attribute labels being used to characterize the contextual association attributes of entity nodes in the text interaction records; constructing a semantic topology graph based on the multi-dimensional entity relationship network to obtain a semantic topology graph containing entity association strength and semantic evolution path; calling a preset zero-knowledge proof protocol to perform access control vector generation on the semantic topology graph to obtain an access control vector containing topology weight thresholds and node access permissions; and jointly processing the access control vector and the semantic topology graph through a spatiotemporal association analysis model to generate a business event template containing standardized fields and association indexes.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the information extraction and processing method for data sharing services as described above.

[0006] The information extraction and processing method for data sharing services provided by this invention, after obtaining a set of historical interactive texts containing timestamps from a cross-departmental business data sharing system, transforms the text interaction records into a multi-dimensional entity relationship network containing entity nodes, relationship edges, and attribute labels through entity relationship network construction. This transforms unstructured text data into a structured entity association topology, effectively improving the structuring degree of data organization and the traceability of entity relationships. Based on this, a semantic topology graph is constructed, generating a semantic topology graph containing entity association strength and semantic evolution paths. This dynamically depicts the changing trajectory of entity relationships over time and quantifies the semantic dependence between entities, enabling the information extraction results to adapt to the dynamic evolution of entity relationships in cross-departmental data interaction scenarios. (The last sentence appears to be incomplete and requires further context.) The knowledge proof protocol generates access control vectors for the semantic topology graph, resulting in access control vectors that include topology weight thresholds and node access permissions. This enables fine-grained access control based on the statistical characteristics of entity association strength, ensuring operational security and permission compliance for different sharing entities during data sharing. Finally, the access control vectors and semantic topology graph are jointly processed by a spatiotemporal association analysis model to generate business event templates containing standardized fields and association indexes. This unifies the text interaction record format under different business scenarios and accurately locates the mapping position of entity nodes in the template, effectively solving the problem of low data reuse efficiency caused by inconsistent formats in cross-departmental data sharing. Overall, it improves the accuracy, security, and scenario adaptability of information extraction in data sharing services. Attached Figure Description

[0007] Figure 1 This is a flowchart of an information extraction and processing method for data sharing services provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 The flowchart illustrates an information extraction and processing method for data sharing services provided in this embodiment of the invention. This method can be executed by a computer system and includes: Step S100: Obtain a set of historical interaction texts from the cross-departmental business data sharing system. The set of historical interaction texts contains text interaction records with timestamps generated under different business scenarios.

[0010] A cross-departmental business data sharing system is a comprehensive system designed to facilitate data flow and sharing between different business departments. This system possesses functions such as data storage, management, and interaction, integrating scattered data from different departments and breaking down data silos. The historical interaction text collection refers to a set of text interaction records stored in the cross-departmental business data sharing system. These records originate from different business scenarios and cover various types of business data from multiple departments. A timestamp is a time identifier added to each text interaction record to indicate the specific time the record was generated. Text interaction records refer to text information exchanged between different departments or between departments and users during the cross-departmental business data sharing process, such as email correspondence, chat logs, and approval comments.

[0011] In practice, historical interaction text sets can be obtained by calling the interface of a cross-departmental business data sharing system. For example, by using the system's API (Application Programming Interface), a request can be sent to the system, specifying conditions such as the time range and business scenario for data retrieval. The system will then return a set of historical interaction text sets that meet the conditions.

[0012] Step S200: Construct an entity relationship network for the historical interaction text set, generating a multi-dimensional entity relationship network containing entity nodes, relationship edges, and attribute labels. The attribute labels are used to characterize the contextual association attributes of entity nodes in the text interaction records.

[0013] An entity-relationship network is a graph structure used to represent the relationships between entities. Entity nodes represent various entities in the text, such as business domain-specific terms, cross-departmental organizational identifiers, and business process nodes. Relationship edges represent the associations between entities, such as an organization participating in a business process or a term being semantically related to another term. Attribute labels are additional information added to each entity node to describe the contextual attributes of the entity in the text interaction record, such as the time when the entity first appears, the business scenario it belongs to, and the role it plays in the interaction.

[0014] A multidimensional entity relationship network means that the network not only includes the relationships between entities, but also considers various attribute information of entities, describing entities and their relationships from multiple perspectives.

[0015] In one implementation, step S200 may include the following steps S210 to S260: Step S210: Divide the historical interactive text set into blocks in ascending order of timestamps to obtain a text block sequence with continuous time series characteristics. Each text block contains text interaction records within a preset time interval, and there is a preset proportion of time overlap between adjacent text blocks.

[0016] Timestamp ascending order sorts the historical interactive text collection according to its timestamps from smallest to largest, ensuring that the text records are arranged in chronological order. Chunking is the process of dividing the sorted historical interactive text collection into multiple text blocks, each containing text interaction records generated within a preset time interval. The preset time interval can be set according to specific business needs and analysis objectives, such as a day, a week, or a month. Continuous time series characteristics indicate that each text block in the text block sequence is continuous in time, forming an ordered time series, which allows for better capture of the changing patterns of text interaction records over time.

[0017] A preset proportion of time overlap exists between adjacent text blocks to prevent information loss during the segmentation process. For example, when adjacent text blocks do not overlap, the association information of a certain entity may be fragmented at the boundary between the two text blocks, resulting in incomplete information. By setting a certain proportion of time overlap, it can be ensured that adjacent text blocks contain some of the same text interaction records, thereby more accurately capturing the relationships between entities.

[0018] Step S220: Perform entity recognition on the text interaction records in the text block sequence, and extract an entity set containing business domain-specific terms, cross-departmental organization identifiers and business process nodes. Each entity has a unique identifier and type label.

[0019] Entity recognition refers to the process of identifying entities with specific meanings from text. In this method, it mainly identifies business domain-specific terms, cross-departmental organizational identifiers, and business process nodes. Business domain-specific terms refer to professional terms used in business operations; cross-departmental organizational identifiers refer to the names or codes representing different departments; and business process nodes refer to the various stages in a business process, such as "application submission" and "approval."

[0020] An entity set is a collection of all identified entities. Each entity has a unique identifier to distinguish different entities in subsequent processing. At the same time, each entity also has a type label to indicate whether the entity belongs to a business domain-specific term, a cross-departmental organization identifier, or a business process node.

[0021] In one implementation, step S220 may include the following steps S221 to S229: Step S221: Construct a hierarchical entity recognition model that integrates a bidirectional long short-term memory network and a conditional random field. The hierarchical entity recognition model includes an input layer, a word embedding layer, a bidirectional long short-term memory layer, a multi-head attention mechanism layer, a feature fusion layer, and a conditional random field output layer.

[0022] Bidirectional Long Short-Term Memory (LSTM) networks can simultaneously consider the context of a sequence, demonstrating good performance in processing natural language sequences. Conditional Random Fields (CRFs) are probabilistic graphical models used for sequence labeling tasks, capable of predicting the label for each position based on the context of the input sequence.

[0023] The hierarchical entity recognition model combines bidirectional long short-term memory (BSSM) networks and conditional random fields (CRFs) to achieve entity recognition through multiple layers of processing. The input layer is the model's entry point, receiving input from text interaction records. The word embedding layer converts the input text into word vectors for subsequent neural network processing. The BSSM layer models the forward and backward semantic dependencies of the word vector sequences, capturing the contextual information of the text. The multi-head attention mechanism layer performs multi-scale semantic focusing on the output of the BSSM layer, extracting features at different semantic granularities. The feature fusion layer fuses the outputs of the multi-head attention mechanism layer to generate a semantically consistent feature sequence. The CRF output layer predicts the entity label for each character position based on the output of the feature fusion layer.

[0024] Step S222: Input the text interaction record into the input layer of the hierarchical entity recognition model according to the character sequence, and generate a character sequence vector through character-level segmentation. The length of the character sequence vector is consistent with the number of characters in the text interaction record.

[0025] A character sequence refers to a sequence formed by arranging each character in a text interaction record sequentially. Character-level segmentation is the process of dividing text according to its characters, treating each character as an independent unit. A character sequence vector is the result of converting a character sequence into a vector representation, where each character corresponds to an element in the vector, and the length of the vector is the same as the number of characters in the text interaction record.

[0026] In practice, the text in the text interaction record is first extracted, and then segmented by character to obtain a character sequence. Next, a corresponding index is assigned to each character, converting the character sequence into an index sequence. Finally, the corresponding word vector is searched from the word vector matrix of the word embedding layer based on the index sequence to generate the character sequence vector. For example, for the text "business approval", it is segmented into the character sequence "business", "service", "approval", and "approval", and an index is assigned to each character, such as index 1 for "business", index 2 for "service", index 3 for "approval", and index 4 for "approval". Then, the corresponding word vector is searched from the word vector matrix based on these indices to generate the character sequence vector.

[0027] Step S223: Input the character sequence vector into the word embedding layer. The character sequence is converted into a word vector sequence containing contextual semantic information through a pre-trained business domain word vector matrix. The dimension of the word vector sequence matches the preset embedding dimension parameter.

[0028] The word embedding layer is a component of the hierarchical entity recognition model. Its function is to convert character sequence vectors into word vector sequences that contain contextual semantic information. The pre-trained business domain word vector matrix is ​​obtained by training on a large amount of business text data. It can map each character or word to a low-dimensional vector space, so that characters or words with similar semantics are close to each other in the vector space.

[0029] The preset embedding dimension parameter refers to the dimension of the word vector, which can be set according to the specific task and data. By inputting the character sequence vector into the word embedding layer and using the pre-trained business domain word vector matrix, the character sequence can be converted into a word vector sequence with contextual semantic information, providing a foundation for subsequent semantic analysis.

[0030] In practice, after the character sequence vector is input into the word embedding layer, the word embedding layer will look up the corresponding word vector from the pre-trained business domain word vector matrix according to the index in the character sequence vector, and arrange these word vectors in order to obtain the word vector sequence.

[0031] Step S224: Perform forward and backward semantic dependency modeling on the word vector sequence through a bidirectional long short-term memory layer to generate a forward hidden state sequence containing historical context information and a backward hidden state sequence containing future context information. The time steps of the forward and backward hidden state sequences correspond one-to-one.

[0032] The bidirectional long short-term memory layer consists of two independent long short-term memory networks: one processes the sequence from beginning to end (forward processing), and the other processes the sequence from end to beginning (backward processing). Forward processing captures the historical context information of the sequence, while backward processing captures the future context information of the sequence.

[0033] The forward hidden state sequence is a sequence of hidden states output by the forward long short-term memory network at each time step, containing historical context information from the beginning of the sequence to the current time step. The backward hidden state sequence is a sequence of hidden states output by the backward long short-term memory network at each time step, containing future context information from the end of the sequence to the current time step. The time steps of the forward and backward hidden state sequences correspond one-to-one; that is, the i-th element of the forward hidden state sequence and the i-th element of the backward hidden state sequence correspond to the same character position.

[0034] In practice, the word vector sequence is sequentially input into the forward and backward long short-term memory networks of the bidirectional long short-term memory layer. For the forward long short-term memory network, the hidden state at each time step is calculated sequentially, starting from the first character of the sequence; for the backward long short-term memory network, the hidden state at each time step is calculated sequentially, starting from the last character of the sequence.

[0035] Step S225: Concatenate the forward hidden state sequence and the backward hidden state sequence into a bidirectional hidden state sequence, input it into the multi-head attention mechanism layer, and perform multi-scale semantic focusing processing on the bidirectional hidden state sequence through a preset number of attention heads to generate attention-enhanced feature sequences containing different semantic granularities.

[0036] Concatenation is the operation of merging the forward and backward hidden state sequences along the feature dimension to obtain a bidirectional hidden state sequence. Multi-head attention is a mechanism that can simultaneously focus on different positions in a sequence, using multiple attention heads to perform semantic focusing processing on the input sequence at different scales.

[0037] The preset number of attention heads refers to the number of attention heads set in the multi-head attention mechanism. Each attention head can focus on the input sequence from different perspectives and extract features at different semantic granularities. The attention-enhanced feature sequence is the output of the multi-head attention mechanism layer, containing feature information extracted from different semantic granularities.

[0038] In practice, the forward and backward hidden state sequences are concatenated along the feature dimension to obtain a bidirectional hidden state sequence. Then, this bidirectional hidden state sequence is input into a multi-head attention mechanism layer. Each attention head performs independent attention calculations on the bidirectional hidden state sequence. Finally, the outputs of each attention head are concatenated to obtain the attention-enhanced feature sequence.

[0039] Step S226: Input the attention-enhanced feature sequence into the feature fusion layer, and integrate the output features of different attention heads through residual connections and layer normalization operations to generate a fused feature sequence with semantic consistency.

[0040] Residual connections can solve the vanishing gradient problem in deep neural networks, while allowing the network to learn more complex features. Layer normalization is a normalization method that normalizes the features of each sample, ensuring that the features have the same scale and distribution.

[0041] The function of the feature fusion layer is to integrate the features from different attention heads output by the multi-head attention mechanism layer, generating a fused feature sequence with semantic consistency. Through residual connections, the input attention-enhanced feature sequence is directly added to the output of the fusion layer, allowing the network to learn the residuals between the input features and the fused features. Layer normalization operations are then used to normalize the fused features, ensuring their stability.

[0042] In practice, the attention-enhanced feature sequence is input into the feature fusion layer. First, a linear transformation is performed on the attention-enhanced feature sequence to obtain intermediate features. Then, the intermediate features are joined with the input attention-enhanced feature sequence using a residual connection, i.e., an addition operation. Finally, the result of the addition is subjected to layer normalization to obtain a fused feature sequence with semantic consistency.

[0043] Step S227: Input the fused feature sequence into the conditional random field output layer, and calculate the state transition probability matrix of the entity label set corresponding to each character position. The entity label set includes the entity start label, the entity internal label, and the non-entity label.

[0044] The state transition probability matrix is ​​a parameter in a conditional random field model, representing the probability of transitioning from one entity label to another. The entity label set contains three types of labels: entity start labels, which indicate the starting position of an entity; entity interior labels, which indicate the character positions within an entity; and non-entity labels, which indicate character positions that do not belong to any entity.

[0045] In practice, after the fused feature sequence is input into the output layer of the Conditional Random Field (CRF), the CRF model calculates the state transition probability matrix for each character position's entity label set based on the feature information of the fused feature sequence. Specifically, for each character position, the model calculates the probability of transitioning from a label at the current position to each label at the next position. For example, for a character in a text, the model calculates the probability of transitioning from a non-entity label of that character to the entity's starting label, internal entity label, or non-entity label of the next character, as well as the probability of transitioning from the entity's starting label to other labels, ultimately obtaining the state transition probability matrix.

[0046] Step S228: Based on the state transition probability matrix, use the Viterbi dynamic programming algorithm to search for the optimal path, determine the entity label category for each character position, and generate a label sequence containing entity boundary information.

[0047] The Viterbi dynamic programming algorithm is used to find the optimal path in Markov models. In conditional random field models, it can find the most probable sequence of entity labels based on the state transition probability matrix. The optimal path is the path that maximizes the probability of the entire sequence among all possible label sequences.

[0048] In practice, based on the state transition probability matrix, the Viterbi algorithm starts from the first character of the sequence and calculates the maximum probability path for each label at each character position. Specifically, for each character position and each label, the algorithm calculates the maximum probability of transitioning from each label at the previous character position to that label at the current character position and records the path. Finally, starting from the last character of the sequence, it backtracks based on the recorded paths to find the most likely label sequence.

[0049] Step S229: Extract character fragments corresponding to consecutive entity tags from the text interaction record based on the tag sequence, and generate an entity set containing the entity's unique identifier, text content, type, and start and end coordinates in the text interaction record by combining the entity type annotation.

[0050] A continuous entity tag's corresponding character fragment refers to a segment of characters composed of consecutive entity start tags and internal entity tags within a tag sequence; these fragments constitute a complete entity. Entity type annotation refers to predefined entity types, such as business domain-specific terms, cross-departmental organizational identifiers, business process nodes, etc.

[0051] The entity unique identifier is a unique number assigned to each entity to distinguish different entities in subsequent processing; the text content is the specific text of the character fragment corresponding to the entity; the entity type is the type to which the entity belongs, determined by the entity type label; the start and end position coordinates are the coordinates of the start and end positions of the entity in the text interaction record, used to locate the entity's position in the text.

[0052] In practice, the tag sequence is traversed to find consecutive character fragments corresponding to entity start tags and entity internal tags. For example, for the tag sequence "non-entity tag, entity start tag, entity internal tag, entity internal tag, non-entity tag", the corresponding text is "This is a business approval process", so the character fragment "business approval" is extracted. Then, based on the predefined entity type annotation, the type of the entity is determined to be a business domain proprietary term. A unique identifier, such as "E001", is assigned to the entity, and its start and end position coordinates in the text are recorded, such as the start position being 3 and the end position being 6. Finally, an entity set is generated containing the entity unique identifier "E001", the text content "business approval", the type "business domain proprietary term", and the start and end position coordinates (3,6).

[0053] Step S230: Based on the location coordinates of the entity set in the text interaction record and the context, calculate the spatial distance parameter and semantic association coefficient between entity pairs, and generate a co-occurrence feature matrix to characterize the potential association between entities.

[0054] Spatial distance parameter refers to the spatial distance between two entities in a text interaction record, reflecting the physical distance relationship between the entities. Semantic association coefficient refers to the degree of semantic association between two entities, reflecting the semantic similarity or relevance between the entities.

[0055] The co-occurrence feature matrix is ​​a matrix where rows and columns correspond to entities in the entity set, and matrix element values ​​represent the potential associations between entity pairs. By calculating the spatial distance parameter and semantic association coefficient between entity pairs and combining them, the co-occurrence feature matrix can be generated, providing a basis for subsequently constructing relationship edges between entities.

[0056] In one implementation, step S230 may include the following steps S231 to S237: Step S231: Traverse the entity set in the text interaction record, extract the start and end coordinates of each entity in the text, and calculate the text coverage length parameter of the entity. The text coverage length parameter is the difference between the end coordinate and the start coordinate.

[0057] The start and end position coordinates represent the beginning and end positions of the entity in the text interaction record, respectively. The text coverage length parameter is the length occupied by the entity in the text, reflecting the size of the entity within the text.

[0058] In practice, each entity in the entity set is traversed, and its start and end coordinates in the text are extracted from the entity's information.

[0059] Step S232: Pair entities in the same text interaction record into pairs to generate a set of entity pairs. Each entity pair contains the identifiers of two different entities and their position coordinates in the text.

[0060] An entity pair set is a collection of pairs of entities from the same text interaction record. Each entity pair contains the identifiers of two different entities and their position coordinates in the text. By pairing entities, it is convenient to subsequently calculate the spatial distance parameter and semantic association coefficient between entity pairs.

[0061] In practice, two nested loops are used to traverse the entity set in the same text interaction record, pairing each entity with other entities. For example, for the entity set {E1,E2,E3}, the generated entity pair set is {(E1,E2),(E1,E3),(E2,E3)}, where each entity pair contains the entity identifier and its position coordinates in the text.

[0062] Step S233: Calculate the spatial distance parameter between the two entities in the entity pair. The spatial distance parameter is the weighted average of the absolute value of the difference between the starting position coordinates of the two entities and the text coverage length parameter.

[0063] The spatial distance parameter is calculated by comprehensively considering the difference in the starting position coordinates of two entities and their text coverage length. By using a weighted average, the spatial distance relationship between the two entities can be measured more reasonably.

[0064] Step S234: Extract the context fragment of the text interaction record where the entity pair is located. The context fragment contains a preset number of character sequences before and after the entity pair.

[0065] Contextual fragments refer to text segments containing a certain number of characters before and after an entity pair. They provide semantic information about the surrounding entity pair and help calculate the semantic association coefficient between them. The preset number can be set according to the specific task and data conditions.

[0066] Step S235: Calculate the semantic association coefficient between two entities in the context segment using a pre-trained semantic similarity model. The semantic association coefficient ranges from [0,1], and the larger the value, the stronger the semantic association.

[0067] The pre-trained semantic similarity model is a model trained on a large amount of text data, which can calculate the semantic similarity between two text segments. The semantic association coefficient is calculated based on the pre-trained semantic similarity model, and its value ranges from 0 to 1. 0 indicates that there is no semantic association between the two entities, and 1 indicates that the two entities are semantically identical.

[0068] In practice, two entities from the extracted context fragment are taken as input and fed into a pre-trained semantic similarity model. The model outputs a value between 0 and 1 as the semantic association coefficient between the two entities.

[0069] Step S236: Construct a co-occurrence feature matrix with entity identifiers as row and column indices. The matrix element values ​​are the product of the inverse of the spatial distance parameter of the entity pair and the semantic association coefficient. The diagonal element values ​​are set to preset constants.

[0070] The rows and columns of the co-occurrence feature matrix correspond to entity identifiers in the entity set, and the matrix element values ​​represent the potential associations between entity pairs. By multiplying the reciprocal of the spatial distance parameter by the semantic association coefficient, the spatial distance and semantic association of entity pairs can be considered comprehensively. Diagonal elements refer to elements in the matrix with the same row and column indices; these are set as preset constants, usually to avoid division by zero or for uniform processing.

[0071] In practice, first create a matrix with entity identifiers as row and column indices. For each off-diagonal element in the matrix, calculate the product of the reciprocal of the spatial distance parameter of the corresponding entity pair and the semantic association coefficient; for diagonal elements, set it to a preset constant, such as 1.

[0072] Step S237: Perform row normalization on the co-occurrence feature matrix so that the sum of the elements in each row is 1, and generate a normalized co-occurrence feature matrix for constructing relation edges.

[0073] Row normalization is a process of normalizing each row of a matrix so that the sum of the elements in each row is 1. Row normalization eliminates scale differences in element values ​​between different rows, making matrix element values ​​comparable and facilitating the construction of subsequent relational edges.

[0074] In practice, for each row of the co-occurrence feature matrix, the sum of the elements in that row is calculated; then, each element in that row is divided by the sum of the elements in that row to obtain the normalized element values. This ultimately yields the standardized co-occurrence feature matrix used for constructing relational edges.

[0075] Step S240: Construct a set of directed relation edges between entities based on the co-occurrence feature matrix. Each relation edge contains a source entity identifier, a target entity identifier, and a dynamic weight value. The dynamic weight value is determined by the product of the inverse of the spatial distance parameter and the semantic association coefficient.

[0076] A set of directed relation edges is a collection of directed edges between entities, where each directed edge represents a relationship from one entity to another. The source entity identifier is the identifier of the entity that initiates the relation edge, and the destination entity identifier is the identifier of the entity that terminates the relation edge. Dynamic weight values ​​represent the strength of the relation edge; they are determined by the product of the reciprocal of the spatial distance parameter and the semantic association coefficient, reflecting the degree of association between the two entities.

[0077] In practice, the standardized co-occurrence feature matrix is ​​traversed, and for each non-zero element in the matrix, a directed relation edge is constructed. For example, for the element mij in the i-th row and j-th column of the matrix (i≠j), a directed relation edge is constructed from entity Ei to entity Ej, with the source entity identifier Ei, the target entity identifier Ej, and the dynamic weight value mij. This ultimately yields a set of directed relation edges between entities.

[0078] Step S250: Assign a set of attribute tags to each entity node, which includes time attribute, domain attribute and interaction role attribute. The time attribute corresponds to the timestamp of the first appearance of the entity in the text block sequence, and the domain attribute is determined based on the business scenario keywords of the text interaction record in which the entity is located. The attribute tag set is an additional set of information added to each entity node, including time attributes, domain attributes, and interaction role attributes. The time attribute records the timestamp of the entity's first appearance in the text block sequence, reflecting the chronological order of the entities; the domain attribute is determined based on the business scenario keywords of the text interaction record in which the entity is located, and is used to indicate the business domain to which the entity belongs; the interaction role attribute describes the role played by the entity in the text interaction record, such as initiator, participant, etc.

[0079] In practice, for each entity node, the first step is to find the text interaction record in which the entity first appears in the text block sequence and record its corresponding timestamp as a time attribute. Then, the business scenario keywords of the text interaction record are extracted, and the domain attribute of the entity is determined based on the keywords. Finally, the role and function of the entity in the text interaction record are analyzed to determine its interaction role attribute.

[0080] Step S260: Construct a multi-dimensional entity relationship network by using the entity set as the node layer, the relationship edge set as the relationship layer, and the attribute label set as the attribute layer. The node layer and the relationship layer are associated through edge weight values, and the relationship layer and the attribute layer are associated through attribute mapping rules.

[0081] The node layer is a layer composed of a set of entities, with each entity serving as a node; the relationship layer is a layer composed of a set of relationship edges, which represent the connections between nodes; and the attribute layer is a layer composed of a set of attribute labels, which provide additional information to the nodes.

[0082] Edge weights represent the strength of relation edges. These weights connect the node layer and the relation layer, allowing the degree of association between nodes to be measured. Attribute mapping rules associate attribute labels with relation edges or nodes. By linking the relation layer and the attribute layer, attribute labels can be used to further analyze the characteristics of nodes and relationships.

[0083] In practice, firstly, each entity in the entity set is treated as a node to build a node layer; then, each relationship edge in the relationship edge set is connected to the corresponding node to build a relationship layer; finally, according to the attribute mapping rules, the attribute labels in the attribute label set are assigned to the corresponding node or relationship edge to build an attribute layer.

[0084] Step S300: Construct a semantic topology graph based on a multi-dimensional entity relationship network to obtain a semantic topology graph that includes entity association strength and semantic evolution path. Entity association strength is used to quantify the degree of semantic dependence between entity nodes, and semantic evolution path is used to describe the trajectory of entity relationship changes with timestamp.

[0085] A semantic topology graph is a graph structure used to represent semantic relationships between entities. It is built on a multi-dimensional entity relationship network and includes two pieces of information: entity association strength and semantic evolution path. Entity association strength is a numerical value used to measure the degree of semantic dependence between two entity nodes; the larger the value, the stronger the semantic dependence. The semantic evolution path is the trajectory of entity relationships over time, reflecting the dynamic changes in the relationships between entities at different points in time.

[0086] By constructing semantic topology graphs, we can more intuitively display the semantic relationships between entities and their changes over time, providing a powerful tool for subsequent information extraction and analysis.

[0087] In one implementation, step S300 may include the following steps S310 to S360: Step S310: Calculate the centrality index for the node layer of the multi-dimensional entity relationship network to generate a comprehensive centrality index set that includes degree centrality, betweenness centrality and proximity centrality. Degree centrality considers the dynamic weight value of the relationship edge, and betweenness centrality considers the number of shortest paths between entity pairs. Centrality metrics are used to measure the importance of a node in a network. Degree centrality refers to the number of relational edges a node has. In this method, degree centrality considers the dynamic weights of relational edges, meaning it considers not only the number of relational edges but also the strength of each relational edge. Betweenness centrality refers to the number of times a node appears in the shortest paths between all entity pairs, reflecting the node's mediating role in the network. Proximity centrality is the reciprocal of the average shortest path length from a node to all other nodes, reflecting the proximity of nodes in the network. The comprehensive centrality metric set is a weighted sum of degree centrality, betweenness centrality, and proximity centrality according to a certain weight ratio, which can more comprehensively measure the importance of a node in a multi-dimensional entity relationship network.

[0088] In one implementation, step S310 may include the following steps S311 to S318: Step S311: Traverse the set of relation edges in the multi-dimensional entity relation network, and count the number of in-degree relation edges and the number of out-degree relation edges for each entity node. The number of in-degree relation edges is the total number of relation edges with the entity as the target, and the number of out-degree relation edges is the total number of relation edges with the entity as the source.

[0089] An in-degree edge is a relation edge that has a destination node as its destination, while an out-degree edge is a relation edge that has a source node as its source. By counting the number of in-degree and out-degree edges for each entity node, we can understand the connectivity of that node in the network.

[0090] In practice, the set of relation edges in the multi-dimensional entity relationship network is traversed. For each relation edge, its target entity and source entity are checked. If the target entity is node Ei, the number of in-degree relation edges of node Ei is incremented by 1; if the source entity is node Ei, the number of out-degree relation edges of node Ei is incremented by 1. For example, for the relation edge set {(E1,E2),(E2,E3),(E3,E1)}, the number of in-degree relation edges of node E1 is 1 (relation edges from E3), and the number of out-degree relation edges is 1 (relation edges pointing to E2).

[0091] Step S312: Calculate the weighted degree centrality index of each entity node. The weighted degree centrality index is the average of the sum of the dynamic weights of the in-degree edges and the sum of the dynamic weights of the out-degree edges.

[0092] The weighted degree centrality index takes into account the dynamic weight values ​​of the in-degree and out-degree relationships of a node, and can more accurately measure the importance of a node in the network.

[0093] In practice, for each entity node Ei, the sum of the dynamic weights Wi of its in-degree edges is calculated. in The sum of dynamic weights Wi of edges with out-degree relationships out Then calculate the weighted degree centrality index Ci=(Wi) in +Wi out ) / 2.

[0094] Step S313: Normalize the weighted degree centrality index to generate degree centrality components. The normalization is achieved by dividing the weighted degree centrality index by the maximum weighted degree centrality value in the network.

[0095] Normalization scales the weighted degree centrality index to a value between 0 and 1, facilitating subsequent comprehensive calculations. By dividing the weighted degree centrality index of each node by the maximum weighted degree centrality value in the network, the scale difference in the weighted degree centrality index between different nodes can be eliminated. In practice, the maximum value C among all weighted degree centrality indices of all nodes in the network is first identified. maxThen, for the weighted degree centrality index Ci of each node, calculate the degree centrality component Di = Ci / C. max .

[0096] Step S314: Use the Floyd-Worscher algorithm to calculate the set of shortest paths between all entity pairs in the multidimensional entity relationship network. The length of each shortest path is the sum of the dynamic weights of all relation edges on the path.

[0097] The Floyd-Warshall algorithm is used to compute the shortest path between all pairs of nodes in a graph, and it can handle weighted directed graphs. In this method, the algorithm is used to compute the shortest path between all pairs of entities in a multi-dimensional entity relationship network. The length of the shortest path is the sum of the dynamic weights of the relationship edges along the path.

[0098] In practice, a distance matrix is ​​first constructed, where each element represents the initial distance between two nodes. If a relation edge exists between the two nodes, the initial distance is the dynamic weight of that edge; otherwise, it is infinity. Then, the Floyd-Worscher algorithm is used to iteratively update the distance matrix until the shortest path between all pairs of nodes is found.

[0099] Step S315: Count the number of times each entity node is intermediated in the shortest path set, and generate the original value of betweenness centrality. The number of intermediaries is the number of shortest paths passing through the entity node.

[0100] The betweenness centrality value refers to the number of shortest paths between all entity pairs that pass through a particular entity node. The original value of betweenness centrality is calculated based on the betweenness centrality value, reflecting the mediating role of a node in the network. In practice, the set of shortest paths between all entity pairs is traversed, and for each shortest path, it is checked whether it passes through a particular entity node. If it does, the betweenness centrality value of that node is incremented by 1.

[0101] Step S316: Standardize the original value of betweenness centrality to generate betweenness centrality components. The standardization is achieved by dividing the original value of betweenness centrality by the total number of possible shortest paths in the network.

[0102] Standardization involves scaling the original betweenness centrality value to a range between 0 and 1, facilitating subsequent comprehensive calculations. By dividing the original betweenness centrality value by the total number of possible shortest paths in the network, the scale difference in betweenness centrality values ​​between different nodes can be eliminated.

[0103] In practice, first calculate the total number of possible shortest paths N in the network; then, for the original value of betweenness centrality Bi of each node, calculate the betweenness centrality component Mi = Bi / N.

[0104] Step S317: Calculate the average of the shortest path lengths from each entity node to all other entity nodes, and generate the average distance parameter. The near-centrality component is the reciprocal of the average distance parameter.

[0105] The average distance parameter is the average of the shortest path lengths from a given entity node to all other entity nodes, reflecting the average distance of that node in the network. The proximity centrality component is the reciprocal of the average distance parameter, reflecting the proximity of nodes in the network; a larger value indicates that the node is closer to other nodes in the network.

[0106] In practice, for each entity node Ei, calculate the sum of the shortest path lengths Si to all other entity nodes; then, calculate the average distance parameter Ai = Si / (n-1), where n is the total number of nodes in the network; finally, calculate the proximity centrality component Ci = 1 / Ai.

[0107] Step S318: The degree centrality component, betweenness centrality component, and proximity centrality component are weighted and summed according to a preset weight ratio to generate a comprehensive centrality index for each entity node. The preset weight ratio is determined by the analytic hierarchy process.

[0108] The Analytic Hierarchy Process (AHP) is a method used to determine weight ratios. It determines the relative importance of multiple factors by comparing them pairwise, thereby obtaining the weight ratios.

[0109] The comprehensive centrality index is obtained by weighting and summing the degree centrality component, betweenness centrality component, and proximity centrality component according to a preset weight ratio. It can more comprehensively measure the importance of a node in a multi-dimensional entity relationship network.

[0110] In practice, the weight ratios of degree centrality, betweenness centrality, and proximity centrality are first determined using the analytic hierarchy process (AHP), and set as w1, w2, and w3 (w1+w2+w3=1), respectively. Then, for each entity node Ei, the comprehensive centrality index Zi is calculated as Zi=w1*Di+w2*Mi+w3*Ci, where Di is the degree centrality component, Mi is the betweenness centrality component, and Ci is the proximity centrality component.

[0111] Step S320: Sort the entity nodes in descending order according to the comprehensive centrality index set, and select the entity nodes of the preset proportion before sorting as the core node set of the semantic topology graph. The preset proportion is dynamically adjusted based on the total number of nodes in the multi-dimensional entity relationship network.

[0112] Descending order sorting refers to arranging entity nodes in descending order of their comprehensive centrality index. Preset ratio refers to a pre-defined proportion of core nodes to be selected, which can be dynamically adjusted based on the total number of nodes in the multi-dimensional entity relationship network to ensure that the selected core node set reflects the main structure of the network without becoming excessively large.

[0113] The core node set is a subset of nodes selected from the entity nodes. These nodes are of high importance in the network and will constitute the core structure of the semantic topology graph.

[0114] Step S330: Perform time-series tracking of the changes in relation edges of the core node set in the text block sequence, extract the creation timestamp, termination timestamp, and weight value change sequence of the relation edges, and generate dynamic evolution features of entity relations.

[0115] Temporal tracking refers to tracking and recording the changes in relational edges of the core node set within a text block sequence in chronological order. The creation timestamp is the time when the relational edge first appears; the termination timestamp is the time when the relational edge disappears; and the weight value change sequence refers to the changes in the weight values ​​of the relational edges over time.

[0116] Dynamic evolution features are generated based on the creation timestamp, termination timestamp, and weight value change sequence of relation edges, and can reflect the dynamic changes of entity relations over time.

[0117] In practice, the text block sequence is traversed, and the creation timestamp, termination timestamp, and weight value changes of the relationship edges between the core node sets are recorded.

[0118] Step S340: Construct a relation evolution prediction model based on dynamic evolution features, input the relation edge state of the current text block, predict the relation edges that may be added or disappear in the next text block, and generate a set of potential relation edges.

[0119] The relation evolution prediction model is a model that makes predictions based on historical data. It can predict the relation edges that may appear or disappear in the next text block based on the dynamic evolution characteristics of the relation edges between core node sets.

[0120] The relation edge state of the current text block refers to the existence and weight values ​​of relation edges between the core node set in the current text block. The potential relation edge set is a set of relation edges that may be added or disappear in the predicted next text block.

[0121] In practice, machine learning or deep learning algorithms such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) are used to construct relation evolution prediction models. Dynamic evolution features are used as training data to train the model. After training, given the relation edge state of the current text block, the model outputs the relation edges that may be added or disappear in the next text block, generating a set of potential relation edges.

[0122] Step S350: Calculate the entity association strength between the core node sets. The entity association strength is determined by the dynamic weight value of the relation edge, the normalized element value in the co-occurrence feature matrix, the time decay factor, and the weighted sum of the domain similarity. The time decay factor decreases as the interval between the last co-occurrence timestamp of the entity pair and the current timestamp increases.

[0123] Entity association strength is an indicator used to measure the degree of semantic dependence between core node sets. It comprehensively considers the dynamic weight values ​​of relation edges, the normalized element values ​​in the co-occurrence feature matrix, the time decay factor, and domain similarity. The time decay factor is a factor that changes over time, reflecting the trend of weakening association between entity pairs over time; domain similarity refers to the degree of similarity between the domains to which two entities belong.

[0124] By combining these factors through a weighted sum, the strength of entity associations between core node sets can be calculated more accurately.

[0125] As one implementation method, step S350, calculating the entity association strength between the core node sets, may specifically include the following steps S351~S356: Step S351: Extract the dynamic weight value of the relationship edge between any two entity nodes in the core node set. If there is no direct relationship edge between the two entities, the dynamic weight value is set to the preset minimum value.

[0126] The dynamic weight value of a relational edge refers to the weight value of the relational edge between two entity nodes in the core node set, reflecting the strength of the direct association between the two entities. The preset minimum value is to assign a smaller value when there is no direct relational edge between the two entities, avoiding zero values ​​that could affect subsequent calculations.

[0127] In practice, for any two entity nodes Ei and Ej in the core node set, check whether there is a direct relationship edge between them. If it exists, extract the dynamic weight value of the relationship edge; if it does not exist, set the dynamic weight value to a preset minimum value, such as 0.01.

[0128] Step S352: Extract the normalized element values ​​corresponding to the two entities from the co-occurrence feature matrix as co-occurrence association components.

[0129] The normalized element values ​​in the co-occurrence feature matrix reflect the co-occurrence of two entities and their potential association in the text interaction record. The co-occurrence association component is the normalized element value corresponding to the two entities extracted from the co-occurrence feature matrix, and serves as a component in calculating the entity association strength.

[0130] In practice, for entity nodes Ei and Ej in the core node set, the normalized element value in the i-th row and j-th column of the co-occurrence feature matrix is ​​found and used as the co-occurrence association component.

[0131] Step S353: Calculate the domain similarity parameter based on the domain attributes in the attribute label sets of the two entities. The domain similarity parameter is determined by the cosine similarity of the domain attribute keywords.

[0132] Domain attribute is an attribute within the entity attribute tag set, used to represent the business domain to which the entity belongs. Domain similarity parameter is obtained by calculating the cosine similarity of the domain attribute keywords of two entities, reflecting the degree of similarity between the domains to which the two entities belong.

[0133] Cosine similarity is a commonly used method for calculating vector similarity. It measures the similarity between two vectors by calculating the cosine of the angle between them.

[0134] In practice, for entity nodes Ei and Ej in the core node set, their domain attribute keywords are extracted and converted into vector representations. Then, the cosine similarity between the two vectors is calculated to obtain the domain similarity parameter.

[0135] Step S354: Obtain the timestamp of the last co-occurrence of the two entities in the text block sequence, and calculate the time decay factor. The time decay factor is an exponential function of the difference between the current timestamp and the last co-occurrence timestamp.

[0136] The time decay factor is a time-varying factor that reflects the weakening trend of the association between entity pairs over time. The current timestamp refers to the current processing time; the last co-occurrence timestamp refers to the time when the two entities last appeared together in a text block sequence. An exponential function can cause the time decay factor to decrease rapidly as the time interval increases.

[0137] In practice, the timestamp of the last co-occurrence of two entities is found in the text block sequence, and the difference between the current timestamp and the last co-occurrence timestamp is calculated. Then, this difference is substituted into the exponential function to calculate the time decay factor.

[0138] Step S355: The entity association strength is obtained by weighted summation using the dynamic weight value as the basic component, the co-occurrence association component as the enhancement component, the domain similarity parameter as the adjustment component, and the time decay factor as the correction component. The weight coefficients of the weighted summation are optimized and determined by the grid search method.

[0139] The basic component refers to the dynamic weight values ​​of the relation edges, which form the basis of entity association strength. The enhancement component refers to the co-occurrence association component, which can strengthen entity association strength. The adjustment component refers to the domain similarity parameter, which can adjust the association strength according to the similarity between the entities' domains. The correction component refers to the time decay factor, which can correct the association strength according to time intervals. The grid search method is a parameter optimization technique that finds the optimal parameter values ​​for the objective function by traversing all possible combinations of parameter values. In this method, the grid search method is used to optimize the weight coefficients of the weighted summation to obtain a more accurate entity association strength.

[0140] In practice, the first step is to determine the range of values ​​for the weight coefficients of the basic component, the enhancement component, the adjustment component, and the correction component. Then, a grid search method is used to traverse all possible combinations of values ​​and calculate the entity association strength under each combination. Finally, the combination of weight coefficients that makes the entity association strength most consistent with the actual situation is selected.

[0141] Step S356: The calculated entity association strength is truncated to upper and lower limits to limit the value range of the entity association strength to a preset value range, thereby generating a standardized entity association strength for edge weights in the semantic topology graph.

[0142] Upper and lower limit truncation refers to limiting the calculated entity association strength to a preset numerical range to avoid excessively large or small values. The preset numerical range is a pre-defined range of entity association strength values, such as [0,1].

[0143] The standardized entity association strength is the entity association strength after upper and lower bound truncation, which will be used as the edge weight of the semantic topology graph to represent the association strength between core node sets.

[0144] In practice, for the calculated entity association strength, if it is less than the lower limit of the preset numerical range, it is set as the lower limit value; if it is greater than the upper limit of the preset numerical range, it is set as the upper limit value.

[0145] Step S360: Using the core node set as vertices, entity association strength as edge weights, dynamic evolution features as time attributes, and the set of potential relation edges as prediction attributes, a semantic topology graph containing static topology structure and dynamic evolution information is constructed. The static topology structure consists of the core node set and the current relation edges, while the dynamic evolution information consists of the weight value change sequence and the set of potential relation edges.

[0146] Vertices are nodes in the semantic topology graph, consisting of a set of core nodes; edge weights are the weights of edges in the semantic topology graph, represented by the standardized entity association strength; temporal attributes are one attribute of edges in the semantic topology graph, provided by dynamic evolution features, reflecting the temporal changes of relational edges; predictive attributes are another attribute of edges in the semantic topology graph, provided by the set of potential relational edges, used to predict relational edges that may appear or disappear in the future.

[0147] The static topology is the basic structure of the semantic topology graph, consisting of a set of core nodes and current relation edges, reflecting the current state of the network; the dynamic evolution information is the dynamic part of the semantic topology graph, consisting of a sequence of weight changes and a set of potential relation edges, reflecting the dynamic changes of the network.

[0148] In practice, the core node set is first used as vertices to construct the node structure of the semantic topology graph. Then, edge weights are set for each edge according to the standardized entity association strength. Next, dynamic evolution features are added as time attributes to the edge attributes. Finally, the set of potential relation edges is added as prediction attributes to the graph, thus constructing a semantic topology graph that contains both static topology structure and dynamic evolution information.

[0149] Step S400: Invoke the preset zero-knowledge proof protocol to generate access control vectors for the semantic topology graph, and obtain access control vectors containing topology weight thresholds and node access permissions. The topology weight thresholds are determined based on the statistical characteristics of entity association strength, and the node access permissions are used to limit the scope of operation of different sharing subjects on entity nodes.

[0150] Zero-knowledge proof protocols are cryptographic methods that prove a statement is true without revealing its specific content. In this method, the pre-configured and optimized zero-knowledge proof protocol is used to process the semantic topology graph to generate access control vectors. Topology weight thresholds are parameters used to control access, determined based on the statistical characteristics of entity association strength. Entity association strength reflects the tightness of association between nodes in the semantic topology graph. By statistically analyzing these strength values, such as calculating the mean, median, and standard deviation, topology weight thresholds that effectively classify access permission levels can be obtained. Node access permissions clarify the scope of operations that different sharing entities, such as different business departments or users, can perform on entity nodes in the semantic topology graph, such as whether they can query, modify, or delete node information.

[0151] In practice, when invoking a pre-defined zero-knowledge proof protocol, the relevant information in the semantic topology graph must first be encrypted to ensure data security and privacy. Then, the protocol calculates a topology weight threshold based on the statistical characteristics of entity association strength. For example, by sorting all entity association strength values ​​in the semantic topology graph, a suitable quantile is selected as the topology weight threshold. Determining node access permissions requires combining the identity authentication information of the sharing entity and the description of business requirements. Assuming a sharing entity is a housing and construction bureau in a certain district, and its business requirement is mainly to query node information related to building approvals, then based on its identity and requirements, it can be assigned corresponding node access permissions, such as only allowing querying information related to building approval processes, but not allowing modification or deletion.

[0152] In one implementation, step S400 may include the following steps S410 to S460: Step S410: Perform topological structure preprocessing on the semantic topology graph, extract topological feature vectors containing core node sets, entity association strength, and semantic evolution paths, and match the dimension of the topological feature vectors with the size of the core node sets.

[0153] Topology preprocessing is a preliminary process for processing the semantic topology graph, aiming to extract information useful for subsequent access control vector generation. The core node set is a collection of nodes with significant importance in the semantic topology graph, playing a crucial role in the graph's structure and semantic representation. Entity association strength reflects the tightness of relationships between nodes, while the semantic evolution path records how node relationships change over time. The topological feature vector is a vector representation that integrates this key information; its dimension matches the size of the core node set, ensuring that relevant information for each core node is accurately represented in the vector.

[0154] In practice, when performing topological preprocessing on a semantic topology graph, the first step is to analyze the graph's structure and identify the core node set. Based on the comprehensive centrality index calculated in previous steps, nodes with the highest ranking can be selected as core nodes. Then, the entity association strength and semantic evolution path information between each core node are extracted. For example, for a semantic topology graph containing multiple business approval process nodes, by analyzing the relationships between nodes and historical data, the core nodes are identified as "application submission," "preliminary review," and "final review." Next, the entity association strength between these core nodes is extracted; for example, the association strength between "application submission" and "preliminary review" is 0.8. Simultaneously, the semantic evolution path of these node relationships over time is recorded; for example, the association strength between "preliminary review" and "final review" gradually increases over a certain period. Finally, this information is combined into a topological feature vector according to certain rules, with each dimension of the vector corresponding to relevant information about a core node.

[0155] Step S420: Calculate statistical feature values ​​based on the entity association strength in the topological feature vector to generate topological weight thresholds for classifying access permission levels. The statistical feature values ​​include the mean, median, and standard deviation of the entity association strength.

[0156] Statistical characteristic values ​​are numerical values ​​obtained by statistically analyzing the entity association strength data in the topological feature vector. The mean reflects the average level of entity association strength, the median represents the middle value of the data, and the standard deviation reflects the dispersion of the data. By calculating these statistical characteristic values, a comprehensive understanding of the distribution of entity association strength can be obtained, thereby generating reasonable topological weight thresholds to classify access permission levels.

[0157] In practice, the first step is to extract all entity association strength values ​​from the topological feature vector. Then, the mean, median, and standard deviation are calculated using appropriate statistical methods. Based on these statistical features, different strategies can be employed to generate topological weight thresholds. A common approach is to combine the mean and standard deviation; for example, adding a certain multiple of the standard deviation to the mean can be used as one topological weight threshold, and subtracting a certain multiple of the standard deviation from the mean can be used as another threshold. This allows for the classification of different access permission levels based on different weight thresholds.

[0158] In one implementation, step S420 may include the following steps S421 to S425: Step S421: Perform outlier detection on the entity association strength in the topological feature vector, use the quartile method to identify and remove outlier association strength values ​​that exceed a preset multiple of the upper and lower quartile distance, and generate a purified entity association strength sequence.

[0159] Outliers are extreme values ​​in data that deviate significantly from the majority of the data, potentially impacting statistical analysis results. The quartile method is an outlier detection approach. It arranges the data in ascending order and divides it into four parts, calculating the lower quartile (Q1), median (Q2), and upper quartile (Q3) for each part. The interquartile range (IQR) equals Q3 - Q1. The preset multiples are typically empirical values, such as 1.5 or 3; data points exceeding these preset multiples are considered outliers. By removing these outliers, a cleaned-up entity association strength sequence is obtained, making subsequent statistical analysis more accurate.

[0160] Step S422: Calculate the statistical characteristic values ​​of the purified entity association strength sequence, including the arithmetic mean, geometric mean, median, first quartile, and third quartile.

[0161] The arithmetic mean is the sum of all data values ​​divided by the number of data points, reflecting the average level of the data. The geometric mean is the nth root of the product of n data values, suitable for calculating the average of data with a proportional relationship. The median is the value in the middle when the data are arranged in ascending order. The first quartile (Q1) and the third quartile (Q3) divide the data into four equal parts, with data below Q1 accounting for 25% and data above Q3 accounting for 25%.

[0162] In practice, for the cleaned entity association strength sequence, the arithmetic mean is calculated by adding all values ​​in the sequence and then dividing by the sequence length. For example, if the cleaned sequence is [0.2, 0.3, 0.5, 0.6, 0.8], the arithmetic mean is (0.2 + 0.3 + 0.5 + 0.6 + 0.8) / 5 = 0.48. The geometric mean is calculated by multiplying all values ​​and then taking the fifth root (because the sequence length is 5). The median can be found by sorting the data; if the sequence length is odd, the middle value is the median; if it is even, the median is the average of the two middle values. The first and third quartiles can be calculated using a method similar to that used in outlier detection, by sorting the data.

[0163] Step S423: Construct a multi-threshold partitioning model based on statistical feature values ​​to divide the entity association strength into multiple intervals, with each interval corresponding to a permission level reference value.

[0164] The multi-threshold segmentation model uses previously calculated statistical features, such as the arithmetic mean, median, first quartile, and third quartile, to determine multiple thresholds, dividing the entity association strength into different intervals. Each interval corresponds to a permission level reference value, which is used to subsequently determine the access permission levels for different sharing subjects.

[0165] In practice, thresholds are determined based on statistical characteristic values. For example, the first quartile, median, and third quartile can be used as thresholds to divide entity association strength into four intervals. Assuming the first quartile is 0.3, the median is 0.5, and the third quartile is 0.7, then the four intervals are [0, 0.3), [0.3, 0.5), [0.5, 0.7), and [0.7, 1]. A permission level reference value is assigned to each interval; for example, the interval [0, 0.3) corresponds to permission level reference value 1, [0.3, 0.5) to 2, [0.5, 0.7) to 3, and [0.7, 1] to 4. In this way, the corresponding permission level can be initially determined by the interval where the entity association strength falls.

[0166] Step S424: Determine the optimal threshold combination through ROC curve analysis. The optimal threshold combination makes the false judgment rate of different permission levels lower than the preset threshold.

[0167] The ROC (Receiver Operating Characteristic Curve) is a tool used to evaluate the performance of classification models. It plots the relationship between the True Positive Rate and the False Positive Rate to illustrate the model's classification effectiveness. In this method, ROC curve analysis helps find the optimal combination of thresholds that keeps the false positive rate below a preset threshold for different privilege levels. The false positive rate includes cases where low privilege levels are misclassified as high privilege levels and vice versa. The preset threshold is a pre-defined upper limit for an acceptable false positive rate.

[0168] In practice, the first step is to classify entity association strength based on different threshold combinations, categorizing them into different permission levels. Then, the true positive rate and false positive rate for each threshold combination are calculated, and an ROC curve is plotted. For example, for different threshold combinations, the proportion of entities correctly classified as high-permission levels (true positive rate) and the proportion of entities incorrectly classified as low-permission levels (false positive rate) are calculated. By comparing the positions of different threshold combinations on the ROC curve, the optimal threshold combination that keeps the false positive rate below a preset threshold is found. A specialized algorithm, such as maximizing the Youden index, can be used to determine the optimal threshold combination; the Youden index equals the true positive rate minus the false positive rate.

[0169] Step S425: Use the boundary values ​​of each interval in the optimal threshold combination as topology weight thresholds to generate a threshold set containing multiple level thresholds. The number of level thresholds is consistent with the number of access permission levels.

[0170] Once the optimal combination of thresholds is determined, the boundary values ​​of each interval are used as topological weight thresholds. These thresholds constitute a threshold set containing multiple levels, with the number of level thresholds being the same as the number of access permission levels. This allows for precise division of different access permission levels based on the strength of entity associations.

[0171] In practice, assuming the optimal threshold combination obtained through ROC curve analysis divides the entity association strength into four intervals with boundary values ​​of 0.3, 0.5, and 0.7, these three boundary values ​​serve as the topology weight thresholds, generating a threshold set of {0.3, 0.5, 0.7}. The corresponding access permission levels can be categorized into four tiers: low, low-medium, medium-high, and high. By comparing the entity association strength with these thresholds, appropriate access permission levels can be accurately assigned to different entity nodes.

[0172] Step S430: Based on the identity authentication information of the sharing entity and the description of business requirements, divide the access permission levels of the core node set. Each access permission level corresponds to a different set of operation permissions, which includes query permissions, modification permissions, and deletion permissions.

[0173] The shared entity's identity authentication information includes the entity's basic identity characteristics, such as its business department and position, reflecting its status and role within the business system. The business requirement description clarifies the entity's actual business needs for core node information in the semantic topology graph, such as whether it needs to query information about specific business processes or modify information about certain nodes. By combining the identity authentication information and the business requirement description, access permission levels for the core node set can be rationally defined. Each access permission level corresponds to a different set of operational permissions: query permission allows the entity to view node information, modification permission allows the entity to change node information, and deletion permission allows the entity to delete node information.

[0174] In practice, the identity authentication information of the sharing entity is first parsed to extract key information, such as department identifiers and role identifiers. Then, keywords are extracted from its business requirement description to determine its specific business needs. For example, if the business requirement is described as "querying and reviewing information related to the financial budget approval process," then the keywords "query," "review," and "financial budget approval process" can be extracted. Based on this information, combined with the attributes of the core node set and the topological weight threshold, the access permission level of the sharing entity to the core node set is determined. For instance, for core nodes related to the financial budget approval process, based on its business needs and identity, a medium-to-high access permission level is assigned, corresponding to the ability to query and modify relevant node information, but not delete it.

[0175] In one implementation, step S430 may include the following steps S431 to S435: Step S431: Extract the role identifier, department identifier, and security level identifier from the identity authentication information of the sharing subject, and generate the subject feature vector.

[0176] Role identifiers clarify the role of the sharing entity within the business system, such as approvers or queryers; department identifiers indicate the business department to which the sharing entity belongs; and security level identifiers reflect the security permission level possessed by the sharing entity. The entity feature vector is a vector representation that integrates these key pieces of information and can serve as the basis for subsequently classifying access permission levels.

[0177] Step S432: Extract keywords from the business requirements description of the sharing entity to generate a set of requirement keywords containing business scenario words, operation type words, and data type words.

[0178] Business scenario terms clearly define the specific scenarios involved in the business needs of the sharing entity, such as "financial approval"; operation type terms indicate the operations that the entity needs to perform, such as "query," "modify," and "delete"; data type terms indicate the data types involved in the entity's needs, such as "budget data" and "student information." By extracting these keywords, the business needs of the sharing entity can be accurately grasped. In practice, natural language processing technology is used to process the descriptions of business needs.

[0179] Step S433: Query the preset role permission mapping table based on the subject feature vector to obtain the set of basic permission levels corresponding to the role.

[0180] The preset role-permission mapping table is a pre-established table that records the set of basic permission levels corresponding to different roles. The set of basic permission levels includes the range of access permissions that role typically possesses. The corresponding set of basic permission levels can be found in the mapping table using the role identifier in the subject feature vector.

[0181] In practice, the role identifier is extracted from the main feature vector and queried in a pre-defined role permission mapping table. Assuming the role identifier is "Approval Specialist," the mapping table finds the corresponding basic permission level set as {Medium, Medium-High}, indicating that this role generally has medium and medium-high level access permissions.

[0182] Step S434: Perform semantic matching based on the set of requirement keywords and the set of attribute tags of the core node set, and calculate the matching score between the requirement and the node. The matching score is determined by a weighted sum of the matching scores of business scenario words, operation type words, and data type words.

[0183] Semantic matching refers to comparing the semantic similarity between the set of keyword requirements and the set of attribute tags of the core node set. Business scenario word matching measures the similarity between the business scenario in the requirement and the business scenario to which the node belongs; operation type word matching reflects the matching between the operation type in the requirement and the allowed operation types of the node; and data type word matching reflects the consistency between the data type in the requirement and the data type of the node. These three matching scores are combined using a weighted summation method to obtain the matching score between the requirement and the node.

[0184] In practice, for each core node, the matching degree of business scenario terms, operation type terms, and data type terms is calculated separately. Semantic similarity algorithms, such as cosine similarity, can be used to calculate the matching degree. For example, for business scenario terms, the business scenario terms in the requirements and the business scenario terms in the node attribute tags are converted into vector representations, and then their cosine similarity is calculated as the business scenario term matching degree.

[0185] Step S435: Combining the basic permission level set and the matching degree score, determine the access permission level of each core node to the shared subject through the fuzzy comprehensive evaluation method. The access permission level includes multiple levels, and different levels correspond to different operation permission ranges.

[0186] Fuzzy comprehensive evaluation is a method that comprehensively considers multiple factors for evaluation and can handle fuzzy information. In this method, the fuzzy comprehensive evaluation is used to determine the access permission level of each core node to the shared subject by combining the basic permission level set and the matching degree score. The access permission level includes multiple levels, such as low, low-medium, medium, medium-high, and high. Each level corresponds to a different range of operation permissions. For example, a low level may only allow querying, while a high level may allow querying, modifying, and deleting.

[0187] In practice, the membership functions of the basic permission level set and matching score are first determined to describe their membership relationships with different access permission levels. Then, the basic permission level set and matching score are comprehensively calculated according to the fuzzy comprehensive evaluation algorithm. For example, for a core node, the basic permission level set shows that the role may have medium and medium-high levels of access permission, and the matching score is high. After calculation using the fuzzy comprehensive evaluation method, it is determined that the access permission level of this sharing subject to this core node is medium-high, meaning that query and modification operations can be performed, but deletion is not allowed.

[0188] Step S440: Map and associate the topology weight threshold with the access permission level to generate a node access permission matrix. The rows of the node access permission matrix correspond to the shared subject identifier, the columns correspond to the core node identifier, and the element values ​​are the access permission levels.

[0189] Topological weight thresholds are used to divide the range of entity association strength, thereby determining the approximate range of access permission levels. By mapping and associating topological weight thresholds with access permission levels, it can be clearly stated which access permission level corresponds to different entity association strength ranges. The node access permission matrix is ​​a two-dimensional matrix that clearly shows the access permission level of each sharing entity to each core node. The rows of the matrix correspond to different sharing entity identifiers, used to uniquely identify each sharing entity; the columns correspond to core node identifiers, used to uniquely identify each core node; and the element values ​​in the matrix are the access permission level of that sharing entity to that core node.

[0190] In practice, a mapping table is first established based on the correspondence between the topology weight threshold and the access permission level determined in the previous steps. Then, for each shared entity and each core node, the access permission level of the shared entity to the core node is determined based on the entity's feature vector, the matching score between the business requirements and the core node, and the topology weight threshold, and then filled into the node access permission matrix.

[0191] Step S450: Invoke the preset zero-knowledge proof protocol to encrypt the node access permission matrix and generate an encrypted permission vector containing permission proof parameters. The permission proof parameters are used to verify the legality of the permissions of the sharing subject.

[0192] The pre-defined zero-knowledge proof protocol, previously used to generate access control vectors, is used again here to encrypt the node access permission matrix, ensuring that the information within the matrix is ​​not leaked. The encrypted permission vector is a vector representation of the encrypted node access permission matrix, containing permission proof parameters. These parameters are crucial information used to verify the legitimacy of the shared subject's permissions. When a shared subject performs an access operation, the system can verify these parameters to determine whether it possesses the corresponding access rights.

[0193] In practice, the node access permission matrix information is input into a pre-defined zero-knowledge proof protocol. The protocol encrypts the matrix using advanced encryption algorithms, such as elliptic curve cryptography, converting the elements into ciphertext. Simultaneously, it generates permission proof parameters, which can be hash values, signatures, etc., generated during the encryption process. For example, encrypting the node access permission matrix using elliptic curve cryptography generates an encrypted permission vector containing a digital signature used to verify the legitimacy of permissions. When a sharing entity requests access to a node, the system requires it to provide permission proof parameters, verifying these parameters to determine the legitimacy of its access rights.

[0194] Step S460: Integrate the topological weight threshold, the encrypted permission vector, and the node identification information of the semantic topology graph to generate an access control vector containing multi-layered permission control logic. The multi-layered permission control logic includes access filtering based on weight threshold, operation restrictions based on permission level, and legality verification based on zero-knowledge proof.

[0195] Fusion is the process of integrating topological weight thresholds, encrypted permission vectors, and node identification information from the semantic topology graph. The topological weight threshold is used to initially filter out unqualified access requests; only when the entity association strength meets the threshold condition is further access allowed. The encrypted permission vector ensures the security and privacy of access permission information. The node identification information from the semantic topology graph clearly identifies each node, facilitating accurate access control. The multi-layered access control logic is the core of the access control vector, including weight threshold-based access filtering, which determines whether access is allowed by comparing entity association strength with the topological weight threshold; permission level-based operation restrictions, which restrict the operations that a sharing subject can perform on nodes based on its access permission level; and zero-knowledge proof-based legitimacy verification, which ensures the sharing subject's permissions are legitimate and valid by verifying the permission proof parameters.

[0196] In practice, the topology weight threshold, encrypted permission vector, and node identification information of the semantic topology graph are first combined according to certain rules. For example, the topology weight threshold can be used as part of the access control vector to indicate different permission levels; the encrypted permission vector can be used as the key permission information; and the node identification information of the semantic topology graph can be used as the node identification part. Then, a multi-layered permission control logic is constructed. When a shared entity initiates an access request, the system first filters the request according to the topology weight threshold. If the entity association strength does not meet the threshold requirement, access is denied. Next, the shared entity's operations are restricted according to the permission level information in the encrypted permission vector. Finally, the permission proof parameters are verified through zero-knowledge proof to ensure the legitimacy of the shared entity's permissions.

[0197] Step S500: Jointly process the access control vector and semantic topology graph through the spatiotemporal correlation analysis model to generate a business event template containing standardized fields and relation indexes. The standardized fields are used to unify the text interaction record format under different business scenarios, and the relation indexes are used to locate the mapping position of entity nodes in the business event template.

[0198] The spatiotemporal correlation analysis model is a data analysis model that comprehensively considers temporal and spatial factors. In this method, access control vectors and semantic topology graphs are jointly processed to mine spatiotemporal correlation information. Access control vectors contain multi-layered permission control logic and access permission information, while the semantic topology graph displays the relationships and semantic evolution between entity nodes. By combining these two, business event templates can be generated. Standardized fields are a component of the business event template, serving to unify the text interaction record format across different business scenarios, enabling text information from different sources and of different types to be stored and processed consistently. Relationship indexes are used to accurately locate the mapping positions of entity nodes within the business event template.

[0199] In practice, access control vectors and semantic topology graphs are input into a spatiotemporal correlation analysis model. The model first parses the permission information in the access control vectors to determine the access permissions of different sharing entities to nodes in the semantic topology graph. Then, it analyzes the spatiotemporal correlations between nodes in the semantic topology graph, such as how the strength of the correlation between nodes changes over time. Based on these analysis results, standardized fields are generated. For example, for text interaction records in different business scenarios, unified field names, data types, and length limits are determined. Simultaneously, a correlation index is generated based on the correlations between nodes in the semantic topology graph. Assuming there are nodes A, B, and C in the semantic topology graph, and there are certain correlations between them, analysis can determine the mapping position of node A in the business event template, as well as its correlation index with other nodes. Finally, the standardized fields and correlation indexes are integrated to generate the business event template.

[0200] In one implementation, step S500 may include the following steps S510 to S570: Step S510: Perform permission filtering on the access control vector, and filter the subset of core nodes and relation edges in the semantic topology graph that meet the permission requirements based on the topology weight threshold, and generate a permission-filtered sub-topology graph.

[0201] Access control filtering is a process of filtering the semantic topology graph based on topological weight thresholds in the access control vector. These topological weight thresholds are used to classify different access permission levels; only core nodes and relational edges whose entity association strength meets the threshold requirements are retained. The core node subset is the set of nodes that meet the access permission requirements selected from the core node set of the semantic topology graph; the relational edge subset is the corresponding set of relational edges that meet the access permission requirements. The filtered sub-topology graph is a new topology graph containing only the filtered core nodes and relational edges, more accurately reflecting the parts that the sharing entity has access to.

[0202] In practice, the topology weight threshold is first extracted from the access control vector. Then, the core nodes and relation edges in the semantic topology graph are traversed, and for each core node and relation edge, its corresponding entity association strength is calculated. The entity association strength is compared with the topology weight threshold; if the threshold requirement is met, the core node and relation edge are retained in the subset. Finally, based on the filtered subset of core nodes and relation edges, a permission-filtered sub-topology graph is generated.

[0203] Step S520: Extract the attribute label set of entity nodes, the entity association strength of relation edges, and the semantic evolution path from the sub-topology graph to generate a topological feature set containing spatiotemporal features. The spatiotemporal features include evolutionary features in the time dimension and association features in the spatial dimension.

[0204] The attribute tag set of an entity node contains various attribute information about the node, such as time attributes, domain attributes, and interaction role attributes. This information reflects the basic characteristics of the node and its role in the business scenario. The entity association strength of a relation edge reflects the tightness of the association between nodes, while the semantic evolution path records the changes in node relationships over time. Spatiotemporal features are the core of the topological feature set, including evolutionary features in the time dimension, such as the creation time, termination time, and weight value change sequence of relation edges, reflecting the dynamic changes in node relationships over time; and association features in the spatial dimension, such as the entity association strength between nodes, reflecting the spatial association structure of nodes.

[0205] In one implementation, step S520 may include the following steps S521 to S525: Step S521: Traverse the core node subset of the sub-topology graph, collect the attribute label set of each node, and generate a node attribute sequence by sorting them in ascending order of time attribute.

[0206] Traversing the core node subset of a sub-topology graph means visiting each core node in the sub-topology graph sequentially. The attribute label set contains various attribute information of the nodes, such as the time attribute, domain attribute, and interaction role attribute mentioned earlier. Sort by time attribute in ascending order, arranging the collected node attribute label set according to the time of the node's first appearance. This clearly shows the chronological order of the nodes, and the generated node attribute sequence can provide a temporal reference for subsequent analysis.

[0207] In practice, a loop is used to traverse the subset of core nodes in the sub-topology graph. For each core node, an attribute label set is extracted from its attribute information. After collecting the attribute label sets of all core nodes, they are sorted according to the time attribute. A sorting algorithm, such as quicksort, can be used to sort the attribute label sets according to the chronological order of the attribute attributes, ultimately generating a sequence of node attributes.

[0208] Step S522: Extract the entity association strength of the relation edge subset in the sub-topology graph, and generate a directed weight matrix by combining the direction information of the relation edges. The rows of the directed weight matrix correspond to the source nodes, and the columns correspond to the target nodes.

[0209] Entity association strength reflects the strength of relation edges in a sub-topology graph and is an indicator of the closeness of the association between nodes. The direction information of the relation edge indicates the flow of the relationship, that is, from which node to which node. The directed weight matrix is ​​a two-dimensional matrix, with rows corresponding to source nodes and columns corresponding to target nodes. The elements in the matrix are the entity association strength of the corresponding relation edge. The directed weight matrix can clearly show the directed association structure between nodes in the sub-topology graph.

[0210] In practice, the subset of relation edges in the sub-topology graph is traversed. For each relation edge, its entity association strength and direction information are extracted. For example, if a relation edge points from node A to node B and the entity association strength is 0.7, then in the directed weight matrix, the element value in the row containing node A and the column containing node B will be 0.7. If there is no relation edge between two nodes, the corresponding element value in the matrix can be set to 0. In this way, the information of all relation edges is integrated into the directed weight matrix to generate a complete directed weight matrix.

[0211] Step S523: Perform temporal segmentation on the semantic evolution path, dividing it into multiple evolution periods according to the timestamp interval of the text block sequence. Each evolution period contains records of relation edge changes within that period.

[0212] The semantic evolution path records the changes in node relationships over time. Temporal segmentation divides the semantic evolution path into multiple evolution periods according to the timestamp intervals of the text block sequence. Each evolution period contains records of changes in relation edges within that period, such as the creation, termination, and weight changes of relation edges. This allows for a more detailed analysis of the dynamic changes in node relationships across different time periods.

[0213] In practice, the timestamp interval of the text block sequence is first determined. For example, if the text block sequence is divided into monthly intervals, then the timestamp interval is one month. Then, based on the time information in the semantic evolution path, it is divided into multiple evolution periods. For each evolution period, records of changes in relation edges within that period are collected.

[0214] Step S524: Calculate the weight change rate of the relation edges in each evolution period, and generate an evolution rate sequence. The weight change rate is the ratio of the difference between the weight value of the current period and the weight value of the previous period to the weight value of the previous period.

[0215] The weight change rate reflects the degree of change in the weight value of a relation edge within each evolutionary period. By calculating the weight change rate of relation edges within each evolutionary period, we can understand the speed and trend of changes in node relationships. The evolution rate sequence is a sequence composed of the weight change rates of relation edges within each evolutionary period, which can intuitively show how the weight values ​​of relation edges change over time.

[0216] In practice, for each evolutionary period and for each relation edge, the rate of change of its weight is calculated. Assume that in the i-th evolutionary period, the weight of the relation edge is W. i In the (i-1)th evolutionary period, the weight of the relation edge is W. i-1 Then the rate of change of weight r = (W i -W i-1 ) / W i-1 After calculating the rate of change of weights of all relation edges within each evolution period, the results are organized into an evolution rate sequence.

[0217] Step S525: Perform feature fusion on the node attribute sequence, directed weight matrix, evolution time period division result and evolution rate sequence to generate a topological feature set containing time dimension features and spatial dimension features. The time dimension features correspond to the evolution rate sequence and the spatial dimension features correspond to the directed weight matrix.

[0218] Feature fusion is the process of integrating node attribute sequences, directed weight matrices, evolutionary time-segmentation results, and evolution rate sequences. The node attribute sequences provide the temporal order and basic attribute information of the nodes; the directed weight matrix displays the spatial relationship structure between nodes; and the evolutionary time-segmentation results and evolution rate sequences reflect the dynamic changes of node relationships over time. Through feature fusion, a comprehensive set of topological features can be generated, where the temporal dimension features are represented by the evolution rate sequences, reflecting the rate of change of relation edge weights over time; and the spatial dimension features are represented by the directed weight matrix, demonstrating the strength of spatial relationships between nodes.

[0219] In practice, the node attribute sequence, directed weight matrix, evolutionary time period division results, and evolution rate sequence are first organized and aligned. For example, it is ensured that each value in the evolution rate sequence corresponds to the correct evolutionary time period and relation edge. Then, this information is fused according to certain rules. They can be combined into a multidimensional array or data structure, so that the temporal and spatial dimensions can be clearly separated. Finally, a set of topological features containing both temporal and spatial dimensions is generated.

[0220] Step S530: Input the set of topological features into the feature encoding layer of the spatiotemporal correlation analysis model, and use a graph neural network to vectorize the topological features to generate a graph embedding vector containing node association information.

[0221] The feature encoding layer is a component of the spatiotemporal correlation analysis model, and its main function is to vectorize and encode the set of topological features. Graph Neural Networks (GNNs) are neural networks specifically designed for processing graph-structured data, effectively capturing the feature information of nodes and edges in the graph. Graph embedding vectors are vector representations of the encoded topological features, containing the correlation information between nodes, and can be used more efficiently for subsequent analysis and processing.

[0222] In practice, the set of topological features is input into the feature encoding layer of the spatiotemporal correlation analysis model. The graph neural network processes the topological features. First, information such as node attribute sequences and directed weight matrices is converted into node and edge features of the graph. For example, attribute information in the node attribute sequence is converted into node feature vectors, and element values ​​in the directed weight matrix are used as edge features. Then, the graph neural network extracts and transforms the graph features through multiple layers of convolution or aggregation operations. In each layer, nodes are updated based on information from their neighboring nodes, thereby capturing the correlation information between nodes. Finally, a graph embedding vector containing node correlation information is generated. This process can be implemented using common graph neural network models such as Graph Convolutional Network (GCN) or Graph Attention Network (GAT).

[0223] Step S540: Frequent itemset mining is performed on the graph embedding vector through the association rule mining layer of the spatiotemporal association analysis model to extract the set of strong association rules between entity nodes. The strong association rules include the antecedent entity, the consequent entity, and the confidence value.

[0224] The association rule mining layer is a functional module in the spatiotemporal association analysis model. Its main task is to mine association rules between entity nodes from graph embedding vectors. Frequent itemset mining is a data mining technique that identifies frequently occurring itemsets by statistically analyzing their frequency of occurrence. Strong association rules are association rules with high confidence; the confidence value represents the probability of the consequent entity appearing given that the antecedent entity is satisfied. The antecedent entity and the consequent entity are the premise and conclusion parts of the association rule, respectively.

[0225] In practice, the graph embedding vector is input into the association rule mining layer of the spatiotemporal association analysis model. First, frequent itemset mining algorithms, such as the Apriori algorithm or the FP-growth algorithm, are used to process the graph embedding vector to identify frequently occurring node combinations. For example, if nodes A and B frequently appear together, forming a frequent itemset, then association rules are generated based on these frequent itemsets, and the confidence value of each rule is calculated. Assuming an association rule is "if node A appears, then node B appears," the confidence value of this rule is obtained by statistically analyzing the proportion of times node B also appears when node A is present. Rules with high confidence values ​​are then selected to form a set of strong association rules.

[0226] Step S550: Construct a standardized field mapping table based on the set of strong association rules. The standardized field mapping table includes field names, data types, length limits, and source entity identifiers.

[0227] The standardized field mapping table is used to map the attribute information of entity nodes to standardized fields. The field name is the specific name of the standardized field, used to identify its meaning; the data type specifies the type of data stored in the field, such as integers, strings, dates, etc.; the length limit constrains the data length of the field, ensuring data consistency and standardization; and the source entity identifier clarifies which entity node the field's data originates from.

[0228] In one implementation, step S550 may include the following steps S551 to S557: Step S551: Perform confidence level stratification on the set of strongly associated rules, dividing them into multiple rule levels from high to low confidence values, with each rule level corresponding to a different association reliability level.

[0229] Confidence-based hierarchical processing is the process of sorting and dividing a set of strongly associated rules according to their confidence values. The confidence value reflects the reliability of an association rule; the higher the confidence value, the stronger the rule's reliability. Dividing rules into multiple rule levels from high to low confidence values, with each level corresponding to a different level of association reliability, allows for more granular management and application of association rules.

[0230] Step S552: Perform rule clustering on the strongly associated rules within each rule level. By calculating the attribute similarity between the antecedent and consequent entities of a rule, generate a set of rule clusters with common business scenarios. Each rule cluster contains strongly associated rules with semantic association.

[0231] Rule clustering is the process of grouping strongly related rules within each rule level, aiming to group rules with similar business scenarios and semantic connections together. The degree of similarity between rules can be determined by calculating the attribute similarity between the antecedent and consequent entities of a rule. Attribute similarity can be calculated by comparing the attribute label sets of entities, such as domain attributes or time attributes, using similarity algorithms such as cosine similarity or edit distance. A rule cluster set consists of multiple rule clusters, each containing strongly related rules with semantic connections, sharing commonalities in their business scenarios.

[0232] Step S553: ​​Extract the attribute label set of entity nodes in each rule cluster, and use the information gain algorithm to calculate the predictive contribution of each attribute label to the business field, generating an attribute importance ranking sequence. The information gain algorithm is an algorithm used to evaluate feature importance, measuring the predictive contribution of an attribute label to a business field by calculating its information gain. Business fields refer to fields that require attention in business data sharing and processing, such as approval results and business process status. The attribute importance ranking sequence is a sequence obtained by sorting attribute labels according to their predictive contribution, which helps determine which attribute labels are more important for predicting business fields.

[0233] Step S554: Construct a dynamic threshold filtering model based on the attribute importance ranking sequence, adjust the attribute filtering threshold according to the association reliability level of the rule cluster, and retain the core attribute labels that are higher than the threshold.

[0234] The dynamic threshold filtering model is a model that filters attribute labels based on the attribute importance ranking sequence and the association reliability level of rule clusters. The attribute filtering threshold is a critical value used to determine whether an attribute label is a core attribute label. The attribute filtering threshold is adjusted according to the association reliability level of the rule cluster. For rule clusters with high association reliability levels, the threshold can be appropriately increased to retain only the most core attribute labels; for rule clusters with low association reliability levels, the threshold can be decreased to retain more attribute labels.

[0235] Step S555: Perform bidirectional matching between the core attribute tags and the metadata requirements in the business data sharing standard specification, calculate the semantic similarity and structural matching degree between the attributes and the metadata, and generate a matching score matrix.

[0236] The metadata requirements in the business data sharing standard specification define the standards and specifications that business data must follow during sharing and processing, including field names, data types, and length limits. Two-way matching refers to comparing core attribute tags with metadata requirements, considering both semantic similarity (the degree of similarity in meaning) and structural matching (the degree of matching in terms of data type, length limits, etc.). The matching score matrix is ​​a two-dimensional matrix where rows correspond to core attribute tags, columns correspond to metadata requirements, and the element values ​​are the matching scores between the corresponding attribute tag and the metadata requirement.

[0237] In practice, for each core attribute tag and each metadata requirement, their semantic similarity and structural matching degree are calculated separately. Semantic similarity can be calculated using semantic similarity algorithms, such as WordMover's Distance (WMD), or similarity calculation methods based on pre-trained language models. Structural matching degree is calculated by comparing the data type, length constraints, and other structural information of the attribute tag and metadata requirement using matching rules. For example, if the data type of the attribute tag and the data type of the metadata requirement are the same, the structural matching degree is 1; otherwise, it is 0. The semantic similarity and structural matching degree are then weighted and summed to obtain the matching score. Finally, the matching scores of all core attribute tags and metadata requirements are compiled into a matching score matrix.

[0238] Step S556: Solve the optimal matching problem of the matching score matrix using the Hungarian algorithm to determine the mapping relationship between each standardized field and the core attribute label. The mapping relationship includes the attribute source entity identifier, attribute name and matching confidence.

[0239] The Hungarian algorithm is a classic algorithm for solving the maximum weighted matching problem in a bipartite graph. It finds the optimal matching scheme in the matching score matrix that maximizes the total matching score. Standardized fields are the standard fields in the business event template, while core attribute labels are important attribute labels selected from the rule clusters. The mapping relationship clarifies the correspondence between each standardized field and the core attribute label, including the attribute source entity identifier, used to identify which entity node the attribute comes from; the attribute name, i.e., the specific name of the attribute label; and the matching confidence, i.e., the reliability of the match, which can be calculated based on the matching score.

[0240] In practice, the matching score matrix is ​​input into the Hungarian algorithm. The algorithm searches for the optimal matching scheme that maximizes the total matching score. For each standardized field, the core attribute label that matches it is determined based on the algorithm's output. For example, for the standardized field "department name", the Hungarian algorithm finds the core attribute label "organization name" that matches it, with the attribute source entity identifier being node A and the attribute name being "organization name". The matching confidence can be calculated based on the corresponding element values ​​in the matching score matrix; the higher the matching score, the higher the matching confidence. Finally, these mapping relationships are organized into a set of mapping relationships containing the attribute source entity identifier, attribute name, and matching confidence.

[0241] Step S557: Perform cross-rule cluster consistency verification on the mapping relationship. If there is a conflict in the mapping relationship of the same standardized field corresponding to different rule clusters, the final mapping relationship is determined by weighted voting based on rule hierarchy priority and matching confidence, and a standardized field mapping table containing field name, data type, source attribute and verification rule is generated.

[0242] Cross-rule cluster consistency verification is a process of checking mapping relationships to ensure that the mapping relationships for the same standardized field are consistent across different rule clusters. If conflicts exist—that is, different rule clusters provide different mapping relationships for the same standardized field—they need to be resolved through weighted voting based on rule hierarchy priority and matching confidence. Rule hierarchy priority is determined by the association reliability level of the rule clusters; rule clusters with higher association reliability levels have higher priority. Matching confidence reflects the reliability of the mapping relationship. The standardized field mapping table is the final output, containing information such as field names, data types, source attributes, and verification rules, providing a basis for the standardization processing of business data.

[0243] In practice, all standardized fields are iterated through. For each standardized field, the mapping relationships provided by different rule clusters are checked. If conflicts exist, for example, rule cluster 1 provides attribute A for the standardized field "department name," while rule cluster 2 provides attribute B, a weighted vote is performed based on rule hierarchy priority and matching confidence. Assuming rule cluster 1 has a high association reliability level and a matching confidence of 0.8, and rule cluster 2 has a medium association reliability level and a matching confidence of 0.6, a weighted calculation based on rule hierarchy priority and matching confidence is performed, ultimately determining that attribute A is selected as the mapping relationship for this standardized field. The final mapping relationships of all standardized fields are then compiled into a standardized field mapping table containing field names, data types, source attributes, and validation rules.

[0244] Step S560: Generate a relationship index based on the standardized field mapping table and the entity association strength. The relationship index is used to record the mapping position and association strength between entity nodes and standardized fields.

[0245] A relationship index is a data structure used to record the relationship information between entity nodes and normalized fields. The normalized field mapping table clarifies the mapping relationship between normalized fields and core attribute labels, while the entity association strength reflects the tightness of the association between entity nodes. By combining these two, a relationship index can be generated, recording the mapping position between entity nodes and normalized fields—that is, how the attributes of entity nodes are mapped to normalized fields, and the strength of the association between them.

[0246] In practice, the mapping relationship between each standardized field and the core attribute label is first obtained from the standardized field mapping table. For example, the standardized field "Department Name" maps to the attribute "Organization Name" of entity node A. Then, the association strength between the entity node and the standardized field is determined based on the entity association strength. If entity node A has a high association strength with other related nodes, and its attribute maps to the standardized field "Department Name," then entity node A can be considered to have a high association strength with that standardized field. The mapping positions and association strengths between entity nodes and standardized fields are recorded to form an association relationship index. Data structures such as dictionaries or lists can be used to store this information for easy subsequent querying and use.

[0247] Step S570: Integrate the standardized field mapping table and the relationship index to generate a business event template containing a field definition layer and a relationship index layer. The field definition layer specifies the unified field format under different business scenarios, and the relationship index layer records the entity association path between fields.

[0248] The integration of the standardized field mapping table and the relation index is the process of combining these two data structures to generate a complete business event template. The field definition layer, part of the business event template, is based on the standardized field mapping table and specifies a unified field format for different business scenarios, including field names, data types, and length limits, ensuring consistent representation of business data across various scenarios. The relation index layer, based on the relation index, records the entity association paths between fields—that is, the relationships established between different standardized fields through entity nodes. This helps in understanding the inherent connections and dependencies between fields when processing business data.

[0249] Please see Figure 2 , Figure 2This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0250] In one embodiment, the processor 101 executes the information extraction processing method for data sharing services provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. An information extraction and processing method applied to data sharing services, characterized in that, The method includes: Obtain a set of historical interaction texts from the cross-departmental business data sharing system. The set of historical interaction texts contains timestamped text interaction records generated under different business scenarios. An entity relationship network is constructed on the historical interaction text set to generate a multi-dimensional entity relationship network containing entity nodes, relationship edges and attribute labels. The attribute labels are used to characterize the contextual association attributes of entity nodes in the text interaction records. A semantic topology graph is constructed based on the multi-dimensional entity relationship network to obtain a semantic topology graph that includes entity association strength and semantic evolution path; A preset zero-knowledge proof protocol is invoked to generate an access control vector for the semantic topology graph, resulting in an access control vector that includes topology weight thresholds and node access permissions. The access control vector and the semantic topology graph are jointly processed by a spatiotemporal correlation analysis model to generate a business event template containing standardized fields and correlation indexes.

2. The method according to claim 1, characterized in that, The step of constructing an entity relationship network from the historical interaction text set to generate a multi-dimensional entity relationship network containing entity nodes, relationship edges, and attribute labels includes: The historical interactive text set is divided into blocks in ascending order of timestamps to obtain a text block sequence with continuous time series characteristics. Each text block contains text interaction records within a preset time interval, and there is a preset proportion of time overlap area between adjacent text blocks. Entity recognition is performed on the text interaction records in the text block sequence to extract an entity set containing business domain-specific terms, cross-departmental organization identifiers, and business process nodes. Each entity has a unique identifier and type label. Based on the location coordinates and context of the entity set in the text interaction record, the spatial distance parameter and semantic association coefficient between entity pairs are calculated to generate a co-occurrence feature matrix to characterize the potential association between entities. Based on the co-occurrence feature matrix, a set of directed relation edges between entities is constructed. Each relation edge contains a source entity identifier, a target entity identifier, and a dynamic weight value. The dynamic weight value is determined by the product of the reciprocal of the spatial distance parameter and the semantic association coefficient. Each entity node is assigned a set of attribute tags containing time attributes, domain attributes, and interaction role attributes. The time attribute corresponds to the timestamp of the first appearance of the entity in the text block sequence, and the domain attribute is determined based on the business scenario keywords of the text interaction record in which the entity is located. A multi-dimensional entity relationship network is constructed by using the entity set as the node layer, the relationship edge set as the relationship layer, and the attribute label set as the attribute layer. The node layer and the relationship layer are associated through edge weight values, and the relationship layer and the attribute layer are associated through attribute mapping rules.

3. The method according to claim 2, characterized in that, The step of performing entity recognition on the text interaction records in the text block sequence, and extracting an entity set containing business domain-specific terms, cross-departmental organizational identifiers, and business process nodes, includes: A hierarchical entity recognition model integrating bidirectional long short-term memory network and conditional random field is constructed. The hierarchical entity recognition model includes an input layer, a word embedding layer, a bidirectional long short-term memory layer, a multi-head attention mechanism layer, a feature fusion layer, and a conditional random field output layer. The text interaction record is input into the input layer of the hierarchical entity recognition model according to the character sequence, and a character sequence vector is generated through character-level segmentation processing. The length of the character sequence vector is consistent with the number of characters in the text interaction record. The character sequence vector is input into the word embedding layer, and the character sequence is converted into a word vector sequence containing contextual semantic information through a pre-trained business domain word vector matrix. The dimension of the word vector sequence matches the preset embedding dimension parameter. The word vector sequence is modeled with forward and backward semantic dependencies by a bidirectional long short-term memory layer, generating a forward hidden state sequence containing historical context information and a backward hidden state sequence containing future context information, with the time steps of the forward and backward hidden state sequences corresponding one-to-one. The forward hidden state sequence and the backward hidden state sequence are concatenated into a bidirectional hidden state sequence, which is then input into a multi-head attention mechanism layer. A preset number of attention heads are used to perform multi-scale semantic focusing processing on the bidirectional hidden state sequence to generate attention-enhanced feature sequences containing different semantic granularities. The attention-enhanced feature sequence is input into the feature fusion layer, and the output features of different attention heads are integrated through residual connections and layer normalization operations to generate a fused feature sequence with semantic consistency. The fused feature sequence is input into the conditional random field output layer, and the state transition probability matrix of the entity label set corresponding to each character position is calculated. The entity label set includes the entity start label, the entity internal label, and the non-entity label. Based on the state transition probability matrix, perform optimal path search to determine the entity label category for each character position and generate a label sequence containing entity boundary information; Based on the tag sequence, character fragments corresponding to consecutive entity tags are extracted from the text interaction record, and entity set is generated by combining them with entity type annotations.

4. The method according to claim 3, characterized in that, The step of calculating spatial distance parameters and semantic association coefficients between entity pairs based on the location coordinates and context of the entity set in the text interaction record, and generating a co-occurrence feature matrix to characterize potential associations between entities, includes: Traverse the entity set in the text interaction record, extract the start position coordinates and end position coordinates of each entity in the text, and calculate the text coverage length parameter of the entity, which is the difference between the end position coordinates and the start position coordinates. Pair entities in the same text interaction record to generate a set of entity pairs. Each entity pair contains the identifiers of two different entities and their position coordinates in the text. Calculate the spatial distance parameter between two entities in an entity pair. The spatial distance parameter is the weighted average of the absolute value of the difference between the starting position coordinates of the two entities and the text coverage length parameter. Extract the contextual fragment of the text interaction record in which the entity pair is located, wherein the contextual fragment contains a preset number of character sequences before and after the entity pair; The semantic association coefficient between two entities in a contextual fragment is calculated using a pre-trained semantic similarity model. Construct a co-occurrence feature matrix with entity identifiers as row and column indices. The matrix element values ​​are the product of the inverse of the spatial distance parameter of the entity pair and the semantic association coefficient, and the diagonal element values ​​are set to preset constants. The co-occurrence feature matrix is ​​row-normalized so that the sum of the elements in each row is 1, thereby generating a standardized co-occurrence feature matrix for constructing relation edges.

5. The method according to claim 1, characterized in that, The construction of a semantic topology graph based on the multi-dimensional entity relationship network yields a semantic topology graph containing entity association strength and semantic evolution paths, including: The centrality index is calculated for the node layer of the multi-dimensional entity relationship network to generate a comprehensive centrality index set of inclusion degree centrality, betweenness centrality, and proximity centrality. The entity nodes are sorted in descending order according to the comprehensive centrality index set, and a preset proportion of entity nodes before sorting are selected as the core node set of the semantic topology graph. The preset proportion is dynamically adjusted based on the total number of nodes in the multi-dimensional entity relationship network. The temporal changes of relation edges in the core node set in the text block sequence are tracked, and the creation timestamp, termination timestamp and weight value change sequence of relation edges are extracted to generate dynamic evolution features of entity relations. Based on the dynamic evolution characteristics, a relation evolution prediction model is constructed. The relation edge state of the current text block is input, and the relation edges that may be added or disappear in the next text block are predicted, generating a set of potential relation edges. Calculate the entity association strength between core node sets. The entity association strength is determined by the dynamic weight value of the relation edge, the normalized element value in the co-occurrence feature matrix, the time decay factor and the weighted sum of the domain similarity. The time decay factor decreases as the interval between the last co-occurrence timestamp of the entity pair and the current timestamp increases. By using the core node set as vertices, entity association strength as edge weights, dynamic evolution features as time attributes, and the set of potential relation edges as prediction attributes, a semantic topological graph containing both static topological structure and dynamic evolution information is constructed.

6. The method according to claim 5, characterized in that, The centrality index calculation for the node layer of the multi-dimensional entity relationship network generates a comprehensive centrality index set including degree centrality, betweenness centrality, and proximity centrality, including: Traverse the set of relation edges in the multi-dimensional entity relation network, and count the number of in-degree relation edges and the number of out-degree relation edges for each entity node. The number of in-degree relation edges is the total number of relation edges with the entity as the target, and the number of out-degree relation edges is the total number of relation edges with the entity as the source. Calculate the weighted degree centrality index for each entity node, where the weighted degree centrality index is the average of the sum of the dynamic weights of the in-degree edges and the sum of the dynamic weights of the out-degree edges; The weighted degree centrality index is normalized to generate degree centrality components; Calculate the set of shortest paths between all entity pairs in a multi-dimensional entity relationship network. The length of each shortest path is the sum of the dynamic weights of all relationship edges on the path. Count the number of times each entity node is intermediated in the shortest path set to generate the original value of betweenness centrality, where the number of intermediaries is the number of shortest paths passing through that entity node. The original values ​​of betweenness centrality are standardized to generate betweenness centrality components; Calculate the average of the shortest path lengths from each entity node to all other entity nodes to generate the average distance parameter, where the proximity centrality component is the reciprocal of the average distance parameter; The degree centrality component, betweenness centrality component, and proximity centrality component are weighted and summed according to a preset weight ratio to generate a comprehensive centrality index for each entity node.

7. The method according to claim 6, characterized in that, The calculation of entity association strength between core node sets includes: Extract the dynamic weight value of the relationship edge between any two entity nodes in the core node set. If there is no direct relationship edge between the two entities, the dynamic weight value is set to the preset minimum value. Extract the normalized element values ​​corresponding to the two entities from the co-occurrence feature matrix as co-occurrence association components; Based on the domain attributes in the attribute tag sets of the two entities, a domain similarity parameter is calculated, which is determined by the cosine similarity of the domain attribute keywords. Obtain the timestamp of the last co-occurrence of two entities in a text block sequence, and calculate the time decay factor, which is an exponential function of the difference between the current timestamp and the last co-occurrence timestamp; The entity association strength is obtained by weighted summation using dynamic weight values ​​as the basic component, co-occurrence association components as the enhancement component, domain similarity parameters as the adjustment component, and time decay factors as the correction component. The calculated entity association strength is truncated to upper and lower limits to limit the value range of the entity association strength to a preset range, thereby generating a standardized entity association strength for edge weights in the semantic topology graph.

8. The method according to claim 1, characterized in that, The method involves invoking a preset zero-knowledge proof protocol to generate access control vectors for the semantic topology graph, resulting in access control vectors that include topology weight thresholds and node access permissions, including: The semantic topology graph is preprocessed to extract a topological feature vector containing the core node set, entity association strength, and semantic evolution path. The dimension of the topological feature vector matches the size of the core node set. Statistical feature values ​​are calculated based on the entity association strength in the topological feature vector to generate topological weight thresholds for classifying access permission levels. The statistical feature values ​​include the mean, median, and standard deviation of the entity association strength. Based on the identity authentication information and business requirements of the sharing entity, the access permission levels of the core node set are divided, and each access permission level corresponds to a different set of operation permissions. The topology weight threshold is mapped and associated with the access permission level to generate a node access permission matrix. The rows of the node access permission matrix correspond to the shared subject identifier, the columns correspond to the core node identifier, and the element values ​​are the access permission levels. The node access permission matrix is ​​encrypted by calling a preset zero-knowledge proof protocol to generate an encrypted permission vector containing permission proof parameters, which are used to verify the legality of the permissions of the sharing subject. By integrating the topology weight threshold, the encrypted permission vector, and the node identification information of the semantic topology graph, an access control vector containing multi-layered permission control logic is generated.

9. The method according to claim 8, characterized in that, The step of calculating statistical feature values ​​based on the entity association strength in the topological feature vector to generate topological weight thresholds for classifying access permission levels includes: Outlier detection is performed on the entity association strength in the topological feature vector, and outlier association strength values ​​that exceed a preset multiple of the upper and lower quartiles are removed to generate a purified entity association strength sequence. Calculate the statistical characteristic values ​​of the purified entity association strength sequence; Based on the statistical feature values, a multi-threshold partitioning model is constructed to divide the entity association strength into multiple intervals, with each interval corresponding to a permission level reference value. The optimal threshold combination is determined by ROC curve analysis, and the optimal threshold combination makes the false judgment rate corresponding to different permission levels lower than the preset threshold. The boundary values ​​of each interval in the optimal threshold combination are used as topological weight thresholds to generate a threshold set containing multiple levels of thresholds.

10. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the information extraction and processing method for data sharing services as described in any one of claims 1-9.

Images