Standard content consistency dynamic comparison method and system based on cloud computing
By constructing a standard knowledge graph and generating enhanced semantic vectors using graph attention networks, and combining vector commitment and zero-knowledge proof techniques, the problem of limited accuracy in standard document consistency comparison in existing technologies is solved, and reliable consistency comparison in complex scenarios is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-03-10
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing technologies lack a structured understanding of the inherent logic and terminology definitions of standards when processing standard document consistency comparisons, resulting in limited accuracy in complex scenarios.
By constructing a standard knowledge graph for compliance intent reasoning, using graph attention networks to generate enhanced semantic vectors, and combining vector commitment and zero-knowledge proof techniques, data integrity is ensured, and a reliable consistency comparison report is generated.
It improves the depth of semantic understanding and the credibility of computation results in complex scenarios, and generates a reliable standard content consistency comparison report.
Smart Images
Figure CN121638205A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for dynamic comparison of standard content consistency based on cloud computing. Background Technology
[0002] With the surge in the number of standard documents, automated content comparison technology has become a core requirement for enterprise compliance and intelligent auditing. Existing technologies typically employ cloud computing and natural language processing methods: first, the documents to be compared and the standard text library are preprocessed and parsed, transforming them into structured text; then, a pre-trained semantic model is used to convert the text into high-dimensional semantic vectors, and by calculating the cosine similarity between vectors and generating a consistency report based on a preset threshold, the comparison is automated and scalable, effectively improving processing efficiency.
[0003] Existing technologies mainly rely on the surface semantic similarity of texts, failing to deeply understand the inherent compliance intent of standard clauses. The binding force of standard clauses is not only reflected in keyword matching, but also in the logical relationships between clauses and the accurate definition of terms. Existing methods lack the ability to integrate and reason about the systematized knowledge of standard structures, and their accuracy and reliability face challenges when dealing with scenarios with diverse expressions or complex dependencies between clauses. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a cloud computing-based dynamic comparison method for standard content consistency, which solves the problem that the accuracy of existing technologies is limited in complex scenarios due to the lack of a structured understanding of the inherent logic and terminology definitions of standards.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a standard content consistency dynamic comparison method based on cloud computing, which includes receiving a document to be compared and calling a standard text library, preprocessing the document to be compared and the standard text library to obtain the structured text of the document to be compared and the structured text of the standard text library. The semantic model is used to vectorize the structured text of the document to be compared and the structured text of the standard text library, generating a set of semantic vectors for the document to be compared and a set of semantic vectors for the standard text library. Based on a pre-built standard knowledge graph, compliance intent reasoning is performed on the semantic vector set of the standard text library to generate an enhanced semantic vector set of the standard text library. By calculating the similarity between the semantic vector set of the document to be compared and the dynamically aligned enhanced semantic vector set of the standard text library, a preliminary similarity matrix is generated. The user selects the pair to be verified from the preliminary similarity matrix and sends a request to the cloud service to obtain the generation proof of the corresponding pair to be verified. Based on the generated proof, the similarity result of the verification pair is calculated by the semantic vector set of the document to be compared and the enhanced semantic vector set of the standard text library after dynamic alignment. Based on the verification result, the final consistency comparison report is generated according to the preliminary similarity matrix.
[0007] As a preferred embodiment of the cloud computing-based dynamic content consistency comparison method of the present invention, the method includes the following steps: receiving the document to be compared and calling the standard text library, preprocessing the document to be compared and the standard text library to obtain the structured text of the document to be compared and the structured text of the standard text library: The system receives the document to be compared and calls the standard text from the standard text library. It performs optical character recognition processing on the document to be compared and the standard text in the standard text library to generate the plain text content of the document to be compared and the standard text in the standard text library. The plain text content of the document to be compared and the standard text in the standard text library is segmented into words to generate word sequences of the document to be compared and the standard text in the standard text library. Stop word filtering and word form restoration are performed on the word sequences of the document to be compared and the standard text in the standard text library to obtain normalized word sequences of the document to be compared and the standard text in the standard text library. The word sequences of the normalized document to be compared and the standard text in the standard text library are encapsulated according to a predefined structured format to obtain the structured text of the document to be compared and the structured text of the standard text library.
[0008] As a preferred embodiment of the cloud computing-based dynamic content consistency comparison method of the present invention, the method includes the following steps: using a semantic model to vectorize the structured text of the document to be compared and the structured text of the standard text library to generate a semantic vector set of the document to be compared and a semantic vector set of the standard text library. Input the structured text of the document to be compared and the structured text of the standard text library into the semantic model; The semantic model processes the structured text of the documents to be compared, generating a set of semantic vectors for the documents to be compared. The semantic model also processes the structured text of the standard text library, generating a set of semantic vectors for the standard text library.
[0009] As a preferred embodiment of the cloud computing-based dynamic comparison method for standard content consistency described in this invention, the method involves: performing compliance intent reasoning on the semantic vector set of the standard text library based on a pre-built standard knowledge graph to generate an enhanced semantic vector set of the standard text library, including the following steps: By performing structured parsing on the original text of standard documents, standard clauses, terms and semantic relationships are automatically identified and extracted. The standard clauses, terms and semantic relationship entities and relationships are stored in the form of a graph structure to construct a standard knowledge graph. Based on the standard clause number in the semantic vector set of the standard text library, a local knowledge subgraph centered on the standard clause containing directly related neighbor nodes is extracted from the pre-constructed standard knowledge graph. The local knowledge subgraph and the corresponding semantic vectors in the standard text library semantic vector set are input into the graph attention network. By calculating the attention coefficients between nodes and aggregating the information of neighboring nodes, a graph enhancement vector is generated for the central node. The semantic vectors in the standard text library semantic vector set are concatenated with the graph augmentation vectors. The intent gating value is calculated using the intent gating formula. The semantic vectors and graph augmentation vectors are then weighted and fused using the intent gating value to generate the standard text library augmented semantic vector set.
[0010] As a preferred embodiment of the cloud computing-based dynamic content consistency comparison method of the present invention, the method includes the following steps: Calculating the similarity between the semantic vector set of the documents to be compared and the dynamically aligned enhanced semantic vector set of the standard text library to generate a preliminary similarity matrix. Iterate through each semantic vector of the document to be compared in the set of semantic vectors of the document to be compared. For each semantic vector of the document to be compared in the set of semantic vectors of the document to be compared, calculate the cosine similarity between the semantic vector of the document to be compared and each standard text library enhanced semantic vector in the set of standard text library enhanced semantic vectors after dynamic alignment. The cosine similarity calculation results of each semantic vector of the document to be compared in the set of semantic vectors to be compared with each enhanced semantic vector of the standard text library in the set of enhanced semantic vectors of the standard text library after dynamic alignment are organized into a preliminary similarity matrix.
[0011] As a preferred embodiment of the cloud computing-based dynamic content consistency comparison method described in this invention, the user terminal selects a comparison pair to be verified from the preliminary similarity matrix and sends a request to the cloud service to obtain the generation proof of the corresponding comparison pair to be verified, including the following steps: The cloud service extracts the semantic vector of the document to be compared from the set of semantic vectors of the document to be compared based on the coordinates of the document to be verified and compares, and extracts the enhanced semantic vector of the standard text library from the set of enhanced semantic vectors of the standard text library after dynamic alignment. The cloud service uses the Merkle tree structure of the set of semantic vectors of the documents to be compared to generate Merkle tree membership certificates of the semantic vectors of the documents to be compared. The cloud service utilizes the Merkle tree structure of the standard text library enhanced semantic vector set after dynamic alignment to generate Merkle tree membership proofs for the standard text library enhanced semantic vectors. The cloud service calculates the dot product and magnitude of the semantic vector of the document to be compared and the augmented semantic vector of the standard text library to obtain the similarity result of the pair to be verified. The cloud service generates proofs of computational correctness for dot product and modulus operations based on the vector commitments of the semantic vector set of the documents to be compared and the vector commitments of the enhanced semantic vector set of the standard text library after dynamic alignment. The cloud service combines the Merkle tree membership proof of the semantic vectors of the documents to be compared, the Merkle tree membership proof of the enhanced semantic vectors of the standard text library, and the proof of computational correctness to generate a generation proof of the comparison pair to be verified.
[0012] As a preferred embodiment of the cloud computing-based dynamic content consistency comparison method of the present invention, the following steps are included: based on the generation proof, the similarity result of the comparison pair to be verified is calculated by the semantic vector set of the document to be compared and the dynamically aligned standard text library enhanced semantic vector set; based on the verification pass result, a final consistency comparison report is generated according to the preliminary similarity matrix: The user end uses the Merkle tree membership proof in the generated proof and the locally stored Merkle root to verify that the semantic vector of the document to be compared and the enhanced semantic vector of the standard text library belong to the semantic vector set of the document to be compared and the dynamically aligned set of the enhanced semantic vector of the standard text library, respectively. Using the computational correctness proof in the generated proof and the locally stored vector commitment, verify the dot product and modulus-length operation correct dot product results on which the similarity results of the pairs to be verified depend. The cosine similarity is recalculated using the verified dot product and modulus results, and the calculated results are compared with the similarity results of the pair to be verified to verify consistency. When the vector membership verification, calculation correctness verification, and result consistency verification pass, the preliminary similarity matrix is deemed reliable. Based on the preliminary similarity matrix deemed credible and the statistical analysis of the comparison results between mandatory requirement clauses in the standard knowledge graph and historical compliance documents, a final consistency comparison report is generated by calculating the quantiles of the semantic similarity distribution and the preset consistency threshold.
[0013] Secondly, the present invention provides a standard content consistency dynamic comparison system based on cloud computing, including a data processing module, which receives a document to be compared and calls a standard text library, preprocesses the document to be compared and the standard text library, and obtains the structured text of the document to be compared and the structured text of the standard text library. The semantic vectorization module uses a semantic model to vectorize the structured text of the document to be compared and the structured text of the standard text library, generating a set of semantic vectors for the document to be compared and a set of semantic vectors for the standard text library. The similarity matrix module performs compliance intent reasoning on the semantic vector set of the standard text library based on a pre-built standard knowledge graph, generates an enhanced semantic vector set of the standard text library, and generates a preliminary similarity matrix by calculating the similarity between the semantic vector set of the document to be compared and the dynamically aligned enhanced semantic vector set of the standard text library. In the proof generation module, the user selects the pair to be verified from the preliminary similarity matrix and sends a request to the cloud service to obtain the generated proof for the corresponding pair to be verified. The report generation module, based on the generated proof, verifies that the similarity results of the comparison pairs to be verified are calculated from the semantic vector set of the documents to be compared and the enhanced semantic vector set of the standard text library after dynamic alignment. Based on the verification results, the module generates the final consistency comparison report according to the preliminary similarity matrix.
[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the cloud computing-based standard content consistency dynamic comparison method as described in the first aspect of the present invention.
[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the cloud computing-based dynamic content consistency comparison method as described in the first aspect of the present invention.
[0016] The beneficial effects of this invention are as follows: compliance intent reasoning is realized by constructing a standard knowledge graph, enhanced semantic vectors are generated by using graph attention networks to improve the depth of semantic understanding, verifiable computation is realized by combining vector commitment and zero-knowledge proof techniques, data integrity is ensured by constructing a Merkle tree structure, and a dynamic threshold is set based on the statistical quantile of historical compliance data to form a credible consistency comparison report, thus solving the technical limitations of traditional text comparison methods in terms of understanding the logic of complex clauses and the credibility of calculation results. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of a cloud-based standard content consistency dynamic comparison method.
[0019] Figure 2 This is a schematic diagram of a file encryption system.
[0020] Figure 3 This is a flowchart of the document preprocessing process.
[0021] Figure 4 A flowchart for semantic vector generation and enhancement. Detailed Implementation
[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a cloud computing-based standard content consistency dynamic comparison method, including the following steps: S1. Receive the document to be compared and call the standard text library. Preprocess the document to be compared and the standard text library to obtain the structured text of the document to be compared and the structured text of the standard text library.
[0026] S1.1 Receive the document to be compared and call the standard text from the standard text library. Perform optical character recognition processing on the document to be compared and the standard text in the standard text library to generate plain text content of the document to be compared and the standard text in the standard text library.
[0027] Furthermore, the system receives the document to be compared and retrieves relevant standard text from the standard text library. It then uses an optical character recognition engine to scan and recognize both the document to be compared and the standard text in the standard text library. The system converts the image format or layout file content contained in the document to be compared and the standard text in the standard text library into a unified plain text character sequence, generating the plain text content of the document to be compared and the plain text content of the standard text in the standard text library. The optical character recognition engine adopts an end-to-end text recognition model based on a combination of convolutional neural networks and recurrent neural networks. It is trained with a large amount of labeled document image data and can accurately recognize printed and handwritten characters.
[0028] S1.2. Perform word segmentation on the plain text content of the document to be compared and the standard text in the standard text library to generate word sequences of the document to be compared and the standard text in the standard text library.
[0029] Furthermore, a word segmentation algorithm combining dictionary and statistical word segmentation is adopted to perform word segmentation on the plain text content of the document to be compared and the plain text content of the standard text in the standard text library. Joint word segmentation is performed by the forward maximum matching algorithm and the hidden Markov model to identify the word boundaries in the text and remove spaces and punctuation marks, thereby generating the word sequence of the document to be compared and the word sequence of the standard text in the standard text library.
[0030] It should be noted that the Baum-Welch algorithm is used for parameter estimation. The parameter values of the initial state probability distribution, state transition probability matrix, and observation symbol probability matrix are continuously adjusted through iterative calculation. The forward algorithm is used to calculate the forward probability of being in each state at each time step under a given observation sequence, and then the backward algorithm is used to calculate the corresponding backward probability. The state occupancy probability and the expected count of state transitions are calculated based on the product of the forward and backward probabilities. The parameters of the Hidden Markov Model are re-estimated using statistics, so that the likelihood function value of the Hidden Markov Model generating the observation sequence gradually increases until it converges to a local optimum. The entire process is implemented through the expectation-maximization algorithm framework to obtain the Hidden Markov Model parameters that best match the training data.
[0031] S1.3. Stop word filtering and word form restoration are performed on the word sequences of the document to be compared and the standard text in the standard text library to obtain the normalized word sequences of the document to be compared and the standard text in the standard text library.
[0032] Furthermore, based on the stop word list, the word sequences of the documents to be compared and the word sequences of the standard text in the standard text library are filtered to remove stop words such as prepositions, conjunctions, and modal particles that have no actual semantic contribution. Then, a dictionary-based lemmatization algorithm is used to restore verbs in the word sequence to their base form and nouns to their singular form, eliminating the influence of tense, voice, and number of words, and obtaining the normalized word sequence of the documents to be compared and the normalized word sequence of the standard text in the standard text library.
[0033] S1.4. Encapsulate the word sequences of the normalized document to be compared and the standard text in the standard text library according to a predefined structured format to obtain the structured text of the document to be compared and the structured text of the standard text library.
[0034] Furthermore, following a predefined JSON structured format, the word sequences of the normalized documents to be compared and the word sequences of the standard text in the normalized standard text library are encapsulated. Part-of-speech tags, positional information, and dependency relationship tags are added to each word, generating structured text of the documents to be compared and structured text of the standard text library with hierarchical relationships. The structured format includes metadata fields such as document identifiers, paragraph identifiers, sentence identifiers, and word identifiers to ensure the integrity and parsability of the text information.
[0035] S2. Use a semantic model to vectorize the structured text of the document to be compared and the structured text of the standard text library to generate a set of semantic vectors for the document to be compared and a set of semantic vectors for the standard text library.
[0036] S2.1 Input the structured text of the document to be compared and the structured text of the standard text library into the semantic model.
[0037] Furthermore, the structured text of the document to be compared and the structured text of the standard text library are used as input data and loaded into the pre-trained semantic model. The structured text of the document to be compared and the structured text of the standard text library are organized in JSON format and contain complete word sequences and part-of-speech and position information.
[0038] It should be noted that the semantic model adopts a pre-trained language model based on the Transformer architecture. Its training process includes two stages: pre-training on a large-scale unsupervised text corpus through a masked language modeling task to learn a general language representation; and then supervised fine-tuning on standard text and compliant document data in a specific domain. A contrastive learning loss function is used to make the vector representations of similar terms closer in the semantic space, thereby optimizing the semantic model's adaptability to the contrastive task. The gradient descent algorithm is used to optimize the semantic model parameters during the training process, ultimately obtaining a semantic model that can accurately capture the semantic features of the text.
[0039] S2.2 The semantic model processes the structured text of the document to be compared and generates a set of semantic vectors for the document to be compared. The semantic model also processes the structured text of the standard text library and generates a set of semantic vectors for the standard text library.
[0040] Furthermore, the semantic model processes the structured text of the document to be compared through a multi-layer Transformer encoder, extracts the contextual semantic features of each text unit, and generates the corresponding set of semantic vectors for the document to be compared; the semantic model processes the structured text of the standard text library with the same architecture, extracts the deep semantic representation of the standard terms, and generates the set of semantic vectors for the standard text library.
[0041] S3. Based on the pre-built standard knowledge graph, perform compliance intent reasoning on the semantic vector set of the standard text library to generate an enhanced semantic vector set of the standard text library.
[0042] S3.1 By performing structured parsing on the original text of the standard document, standard clauses, terms and semantic relationships are automatically identified and extracted. The standard clauses, terms and semantic relationship entities and relationships are stored in the form of a graph structure to construct a standard knowledge graph. Based on the standard clause number in the semantic vector set of the standard text library, a local knowledge subgraph centered on the standard clause containing directly related neighbor nodes is extracted from the pre-constructed standard knowledge graph.
[0043] Furthermore, based on the original text of the standard document, a structured parsing is performed to automatically identify and extract standard clauses, terms, and semantic relationships. The identified standard clauses and terms are used as nodes, and semantic relationships are used as edges, and a standard knowledge graph is constructed by storing them in a graph structure. Based on the standard clause numbers in the semantic vector set of the standard text library, a query is performed in the pre-constructed standard knowledge graph to extract a local knowledge subgraph containing the standard clause node and its directly related term nodes, reference clause nodes, and other neighboring nodes, forming a local knowledge subgraph centered on the standard clause.
[0044] S3.2 Input the local knowledge subgraph and the corresponding semantic vectors in the standard text library semantic vector set into the graph attention network. By calculating the attention coefficients between nodes and aggregating the information of neighboring nodes, a graph enhancement vector is generated for the central node.
[0045] Furthermore, the local knowledge subgraph and the corresponding semantic vectors from the standard text library's semantic vector set are input into the graph attention network. The importance of different neighboring nodes is measured by calculating the attention coefficients between the central node and each neighboring node. Based on these attention coefficients, the feature vectors of the neighboring nodes are weighted and aggregated to generate a graph enhancement vector that integrates graph structure information. The attention coefficients in the graph attention network are calculated using the LeakyReLU activation function, and the node features are transformed and scored using trainable attention vectors.
[0046] The expression for the attention coefficient is: ;
[0047] in, Neighboring nodes For the central node The weight, As the central node The set of neighboring nodes, This is the transposed attention vector. For attention vectors, For font commands, It is a non-linear activation function. For iteration variables For neighboring nodes, As the central node, For any neighbor node in the set of neighbor nodes of the central node The input feature vector, As the central node The input feature vector, Neighboring nodes The input feature vector.
[0048] The enhancement vector expression is: ; in, As the central node The augmented vector after processing by a graph attention network.
[0049] S3.3. Concatenate the semantic vectors in the standard text library semantic vector set with the graph augmentation vectors, calculate the intent gating value using the intent gating formula, and use the intent gating value to perform weighted fusion of the semantic vectors and graph augmentation vectors to generate the standard text library augmented semantic vector set.
[0050] Furthermore, the original semantic vectors in the standard text library semantic vector set are concatenated with the graph augmentation vectors generated by the graph attention network. The intent gating value is calculated using a trainable weight matrix and bias terms. The intent gating value is constrained between 0 and 1 using the Sigmoid function. Based on the intent gating value, the original semantic vectors and graph augmentation vectors are weighted and fused to generate a standard text library augmented semantic vector set that contains both textual semantic information and graph structure information.
[0051] The intent gating value expression is: ;
[0052] in, For intent-gated values, The weights of the intent-gated values, This represents the weight of the bias term.
[0053] The standard text library's enhanced semantic vector expression is: ; in, Enhance semantic vectors for standard text libraries.
[0054] It should be noted that the graph attention network adopts an end-to-end training approach, and supervised learning is performed on standard clause classification tasks or similarity calculation tasks. The weight parameters in the attention mechanism and the parameters in the intent gating mechanism are optimized through the backpropagation algorithm, so that the network can adaptively learn how to extract effective information from the knowledge graph and achieve optimal information fusion. The training data uses a labeled standard clause dataset, and the loss function adopts cross-entropy loss or mean squared error loss.
[0055] S4. Calculate the similarity between the semantic vector set of the documents to be compared and the dynamically aligned standard text library to enhance the semantic vector set, and generate a preliminary similarity matrix.
[0056] S4.1. Traverse each semantic vector of the document to be compared in the set of semantic vectors of the document to be compared. For each semantic vector of the document to be compared in the set of semantic vectors of the document to be compared, calculate the cosine similarity between the semantic vector of the document to be compared and each standard text library enhanced semantic vector in the set of standard text library enhanced semantic vectors after dynamic alignment.
[0057] Furthermore, an iterative algorithm is used to process each semantic vector of the document to be compared in the set of semantic vectors to be compared sequentially. For the currently processed semantic vector of the document to be compared, the cosine similarity between it and each standard text library enhanced semantic vector in the dynamically aligned set of standard text library enhanced semantic vectors is calculated in turn. The cosine similarity calculation process is as follows: Calculate the dot product between the semantic vector of the document to be compared and the enhanced semantic vector of the standard text library, which is the sum of the products of the corresponding dimensional component values; calculate the magnitude of the semantic vector of the document to be compared and the magnitude of the enhanced semantic vector of the standard text library, which is the square root of the sum of the squares of each vector component value; divide the dot product result by the product of the two magnitudes to obtain the cosine similarity value, which ranges from -1 to 1. Dimension indexing is used in the calculation process. Traversing from 1 to vector dimension This ensures that all vector components participate in the calculation. The cosine similarity expression is: ; in, Enhance the cosine similarity between semantic vectors of standard text libraries and semantic vectors of documents to be compared. The semantic vector of the document to be compared. Enhance semantic vectors for standard text libraries. To enhance the dimensionality of semantic vectors in standard text libraries and semantic vectors of documents to be compared, For the semantic direction of the document to be compared Component values on the dimension To enhance the number of component values in the k-dimensional semantic vector of the standard text library, For dimensional indexing.
[0058] S4.2. Organize the cosine similarity calculation results of each semantic vector of the document to be compared in the set of semantic vectors of the document to be compared with each enhanced semantic vector of the standard text library in the set of enhanced semantic vectors of the standard text library after dynamic alignment into a preliminary similarity matrix.
[0059] Furthermore, all cosine similarity values are organized according to a specific row-column correspondence. The vector order in the semantic vector set of the documents to be compared is used as the row index, and the vector order in the dynamically aligned standard text library enhanced semantic vector set is used as the column index. The cosine similarity value of each semantic vector of the documents to be compared and each enhanced semantic vector of the standard text library is filled into the corresponding position in the matrix, thus forming a preliminary similarity matrix. The number of rows in the preliminary similarity matrix is equal to the number of vectors in the semantic vector set of the documents to be compared, and the number of columns is equal to the number of vectors in the dynamically aligned standard text library enhanced semantic vector set. Each element value in the matrix represents the degree of semantic similarity between a specific document fragment to be compared and a specific standard clause.
[0060] S5. The user selects the pair to be verified from the preliminary similarity matrix and sends a request to the cloud service to obtain the generation proof of the corresponding pair to be verified.
[0061] S5.1 The cloud service extracts the semantic vector of the document to be compared from the set of semantic vectors of the document to be compared based on the coordinates of the document to be verified and compares, and extracts the enhanced semantic vector of the standard text library from the set of enhanced semantic vectors of the standard text library after dynamic alignment.
[0062] Furthermore, the cloud service receives the coordinates of the document to be verified and compared from the user, including row and column index information. Based on the row index, it locates and extracts the corresponding semantic vector of the document to be compared from the set of semantic vectors of the document to be compared, and based on the column index, it locates and extracts the corresponding standard text library enhanced semantic vector from the dynamically aligned set of standard text library enhanced semantic vectors.
[0063] S5.2 The cloud service uses the Merkle tree structure of the semantic vector set of the documents to be compared to generate Merkle tree membership certificates of the semantic vectors of the documents to be compared.
[0064] Furthermore, the cloud service accesses the Merkle tree structure of the pre-generated set of semantic vectors of the documents to be compared, and generates a Merkle tree membership proof containing vector hash paths based on the position information of the semantic vectors of the documents to be compared. This proof consists of the hash values of all sibling nodes on the path from the leaf node where the semantic vector of the document to be compared is located to the root node of the Merkle tree.
[0065] S5.3 The cloud service utilizes the Merkle tree structure of the standard text library enhanced semantic vector set after dynamic alignment to generate Merkle tree membership proofs for the standard text library enhanced semantic vectors.
[0066] Furthermore, the cloud service accesses the Merkle tree structure of the pre-generated dynamically aligned standard text library enhanced semantic vector set, and generates a Merkle tree membership proof containing vector hash paths based on the position information of the standard text library enhanced semantic vectors. This proof consists of the hash values of all sibling nodes on the path from the leaf node where the standard text library enhanced semantic vector is located to the root node of the Merkle tree.
[0067] S5.4 The cloud service calculates the dot product and magnitude of the semantic vector of the document to be compared and the enhanced semantic vector of the standard text library to obtain the similarity result of the pair to be verified.
[0068] Furthermore, the cloud service calculates the dot product of the semantic vector of the document to be compared and the enhanced semantic vector of the standard text library, which is the sum of the products of the corresponding dimensional component values. The magnitudes of the semantic vector of the document to be compared and the enhanced semantic vector of the standard text library are then calculated separately, which are the square roots of the sum of the squares of each vector component value. Finally, the dot product result is divided by the product of the two magnitudes to obtain the dot product calculation expression for the similarity result of the pair to be verified: ; in, The result of the dot product operation. In the first Component values on the dimension; The expression for calculating the modulus is: ; in, The magnitude of the semantic vector of the document to be compared; ; in, Enhance the magnitude of semantic vectors in standard text libraries; The expression for calculating cosine similarity is: ; in, The similarity results are for the pairs to be verified. S5.5, the cloud service generates a proof of the correctness of the calculation of the dot product and modulus length operations based on the vector commitment of the semantic vector set of the documents to be compared and the vector commitment of the enhanced semantic vector set of the standard text library after dynamic alignment.
[0069] Furthermore, the cloud service uses vector commitments based on the semantic vector set of the documents to be compared and the vector commitments of the enhanced semantic vector set of the dynamically aligned standard text library to generate computational correctness proofs using a zero-knowledge proof protocol. By constructing polynomial constraint relationships, it verifies the correct execution of dot product and modulus operations, while ensuring that the vectors used are consistent with the committed values, without revealing the specific numerical values of the vectors. The expression proving the correctness of the calculation is: ; in, To prove the correctness of the calculation, For the vector commitment of the set of semantic vectors of the documents to be compared, Vector commitments to enhance the semantic vector set of a dynamically aligned standard text library. For public reference strings, The vector commitment blinding factor is the set of semantic vectors of the documents to be compared. Enhance the vector commitment blinding factor of the semantic vector set for dynamically aligned standard text libraries.
[0070] S5.6 The cloud service combines the Merkle tree membership proof of the semantic vectors of the documents to be compared, the Merkle tree membership proof of the enhanced semantic vectors of the standard text library, and the proof of computational correctness to generate the generation proof of the comparison pair to be verified.
[0071] Furthermore, the cloud service structurally combines the Merkle tree membership proof of the semantic vectors of the documents to be compared, the Merkle tree membership proof of the enhanced semantic vectors of the standard text library, and the proof of computational correctness. It encapsulates these into a complete generation proof of the pair to be verified according to a predefined proof format specification. The generation proof contains three independent and verifiable components: the first part is the membership proof of the semantic vectors of the documents to be compared in the set of semantic vectors of the documents to be compared; the second part is the membership proof of the enhanced semantic vectors of the standard text library in the dynamically aligned set of enhanced semantic vectors of the standard text library; and the third part is the cryptographic proof of the correctness of the dot product and modulo length operations. The three parts of the proof are linked together by digital signatures to form a logically complete chain of evidence, generating a generation proof data package of the pair to be verified that can be verified by the user.
[0072] S6. Based on the generated proof, the similarity result of the pair to be verified is calculated by the semantic vector set of the document to be compared and the enhanced semantic vector set of the standard text library after dynamic alignment. Based on the verification result, the final consistency comparison report is generated according to the preliminary similarity matrix.
[0073] S6.1 The user end uses the Merkle tree membership proof in the generated proof and the locally stored Merkle root to verify that the semantic vector of the document to be compared and the enhanced semantic vector of the standard text library belong to the semantic vector set of the document to be compared and the dynamically aligned enhanced semantic vector set of the standard text library, respectively.
[0074] Furthermore, the user end extracts the Merkle tree membership proof of the semantic vector of the document to be compared and the Merkle tree membership proof of the standard text library enhanced semantic vector from the generated proof. It then uses the Merkle root of the locally stored semantic vector set of the document to be compared and the Merkle root of the dynamically aligned standard text library enhanced semantic vector set to perform verification. The verification process involves recalculating the hash path provided in the membership proof and checking whether the obtained root hash completely matches the locally stored Merkle root, thereby confirming that the semantic vector of the document to be compared does indeed belong to the semantic vector set of the document to be compared, and that the standard text library enhanced semantic vector does indeed belong to the dynamically aligned standard text library enhanced semantic vector set.
[0075] S6.2 Using the computational correctness proof in the generated proof and the locally stored vector commitment, verify the dot product and modulus length operation correct dot product results on which the similarity results of the pairs to be verified depend.
[0076] Furthermore, the user end extracts the computational correctness proof from the generated proof, combines the vector commitment of the semantic vector set of the documents to be compared stored locally with the vector commitment of the enhanced semantic vector set of the standard text library after dynamic alignment, and runs a zero-knowledge proof verification algorithm. The verification algorithm checks whether the cryptographic evidence in the computational correctness proof can confirm that the dot product operation and the modulus operation are indeed correctly executed and that the input vector used is consistent with the vector commitment, thereby verifying the authenticity and correctness of the dot product result and the modulus result.
[0077] S6.3. Recalculate the cosine similarity using the verified correct dot product and modulus results, and compare the calculated results with the similarity results of the pair to be verified to verify consistency.
[0078] Furthermore, the user end uses the verified dot product and modulus results to recalculate the cosine similarity value. The recalculated cosine similarity value is then compared with the similarity result of the pair to be verified returned from the cloud. A very small floating-point calculation error is allowed during the comparison. If the absolute difference between the two values is less than the preset fault tolerance threshold, the consistency verification of the result is deemed successful.
[0079] The similarity result expression for the pair to be verified returned by the cloud is: ; in, This refers to the similarity results of the pairs to be verified, returned from the cloud. For the verified correct dot product result, The magnitude of the semantic vector of the document to be compared, which has been verified to be correct. Enhance the magnitude of semantic vectors in the already verified standard text library. S6.4. When vector membership verification, calculation correctness verification, and result consistency verification pass, the preliminary similarity matrix is deemed reliable.
[0080] Furthermore, if only the vector membership verification, calculation correctness verification, and result consistency verification pass, the user terminal determines that the calculation process of the comparison pair to be verified is credible. Based on the statistical principle of random sampling, if multiple randomly sampled comparison pairs to be verified pass the verification, then the entire preliminary similarity matrix is determined to be credible.
[0081] S6.5. Based on the preliminary similarity matrix determined to be credible and the comparison results between mandatory requirement clauses in the standard knowledge graph and historical compliance documents, perform statistical analysis, calculate the quantile of the semantic similarity distribution and the preset consistency threshold, and generate the final consistency comparison report.
[0082] Furthermore, based on the preliminary similarity matrix deemed credible, and combined with the consistency threshold calculated by analyzing the similarity data distribution between historical compliance documents and standard clauses, a final consistency comparison report is generated.
[0083] The method for determining the consistency threshold is as follows: collect similarity values from historical compliance cases to form a historical similarity sample set, calculate a specific quantile of the historical similarity sample set as the threshold benchmark, compare each similarity value in the preliminary similarity matrix with the consistency threshold in the report, mark the items that comply and do not comply, and generate a structured compliance analysis result.
[0084] The consistency threshold expression is: ; in, For historical similarity sample set, This is the consistency threshold. This is a quantile function. This embodiment also provides a cloud-based standard content consistency dynamic comparison system, including: a data processing module that receives the document to be compared and calls a standard text library, preprocesses the document to be compared and the standard text library to obtain the structured text of the document to be compared and the structured text of the standard text library; The semantic vectorization module uses a semantic model to vectorize the structured text of the document to be compared and the structured text of the standard text library, generating a set of semantic vectors for the document to be compared and a set of semantic vectors for the standard text library. The similarity matrix module performs compliance intent reasoning on the semantic vector set of the standard text library based on a pre-built standard knowledge graph, generates an enhanced semantic vector set of the standard text library, and generates a preliminary similarity matrix by calculating the similarity between the semantic vector set of the document to be compared and the dynamically aligned enhanced semantic vector set of the standard text library. In the proof generation module, the user selects the pair to be verified from the preliminary similarity matrix and sends a request to the cloud service to obtain the generated proof for the corresponding pair to be verified. The report generation module, based on the generated proof, verifies that the similarity results of the comparison pairs to be verified are calculated from the semantic vector set of the documents to be compared and the enhanced semantic vector set of the standard text library after dynamic alignment. Based on the verification results, the module generates the final consistency comparison report according to the preliminary similarity matrix.
[0085] This embodiment also provides a computer device applicable to the dynamic comparison method for standard content consistency based on cloud computing, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the dynamic comparison method for standard content consistency based on cloud computing as proposed in the above embodiment.
[0086] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0087] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the dynamic comparison method for standard content consistency based on cloud computing as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0088] In summary, this invention achieves compliance intent reasoning by constructing a standard knowledge graph, uses graph attention networks to generate enhanced semantic vectors to improve the depth of semantic understanding, combines vector commitment and zero-knowledge proof techniques to achieve verifiable computation, ensures data integrity by constructing a Merkle tree structure, and sets dynamic thresholds based on the statistical quantiles of historical compliance data to form a credible consistency comparison report. This invention solves the technical limitations of traditional text comparison methods in terms of understanding the logic of complex clauses and the credibility of calculation results.
[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A cloud computing-based standard content consistency dynamic comparison method, characterized in that: This includes receiving the document to be compared and calling the standard text library, preprocessing the document to be compared and the standard text library, and obtaining the structured text of the document to be compared and the structured text of the standard text library; The semantic model is used to vectorize the structured text of the document to be compared and the structured text of the standard text library, generating a set of semantic vectors for the document to be compared and a set of semantic vectors for the standard text library. Based on a pre-built standard knowledge graph, compliance intent reasoning is performed on the semantic vector set of the standard text library to generate an enhanced semantic vector set of the standard text library. By calculating the similarity between the semantic vector set of the document to be compared and the dynamically aligned enhanced semantic vector set of the standard text library, a preliminary similarity matrix is generated. The user selects the pair to be verified from the preliminary similarity matrix and sends a request to the cloud service to obtain the generation proof of the corresponding pair to be verified. Based on the generated proof, the similarity result of the verification pair is calculated by the semantic vector set of the document to be compared and the enhanced semantic vector set of the standard text library after dynamic alignment. Based on the verification result, the final consistency comparison report is generated according to the preliminary similarity matrix.
2. The cloud computing based dynamic comparison method of standard content consistency as claimed in claim 1, wherein: The process involves receiving the document to be compared and calling the standard text library, preprocessing both the document and the standard text library to obtain the structured text of the document to be compared and the structured text of the standard text library, including the following steps: The system receives the document to be compared and calls the standard text from the standard text library. It performs optical character recognition processing on the document to be compared and the standard text in the standard text library to generate the plain text content of the document to be compared and the standard text in the standard text library. The plain text content of the document to be compared and the standard text in the standard text library is segmented into words to generate word sequences of the document to be compared and the standard text in the standard text library. Stop word filtering and word form restoration are performed on the word sequences of the document to be compared and the standard text in the standard text library to obtain normalized word sequences of the document to be compared and the standard text in the standard text library. The word sequences of the normalized document to be compared and the standard text in the standard text library are encapsulated according to a predefined structured format to obtain the structured text of the document to be compared and the structured text of the standard text library.
3. The cloud computing based dynamic comparison method for standard content consistency as claimed in claim 2, wherein: The semantic model is used to vectorize the structured text of the document to be compared and the structured text of the standard text library, generating a semantic vector set for the document to be compared and a semantic vector set for the standard text library. This includes the following steps: Input the structured text of the document to be compared and the structured text of the standard text library into the semantic model; The semantic model processes the structured text of the documents to be compared, generating a set of semantic vectors for the documents to be compared. The semantic model also processes the structured text of the standard text library, generating a set of semantic vectors for the standard text library.
4. The cloud computing based dynamic comparison method for standard content consistency according to claim 3, characterized in that: Based on a pre-built standard knowledge graph, compliance intent reasoning is performed on the semantic vector set of the standard text library to generate an enhanced semantic vector set of the standard text library, including the following steps: The standard knowledge graph is constructed by structurally analyzing the original text of the standard document, automatically identifying and extracting standard clauses, terms and semantic relationships, and storing and constructing the standard clauses, terms and semantic relationship entities and relationships in the form of a graph structure. The local knowledge sub-graph is input into the graph attention network together with the corresponding semantic vector in the standard text library semantic vector set, the attention coefficients between nodes are calculated, the neighbor node information is aggregated, and the graph enhancement vector of the center node is generated. The semantic vector in the standard text library semantic vector set is spliced with the graph enhancement vector, the intention gating value is calculated through the intention gating formula, and the semantic vector and the graph enhancement vector are weighted and fused by using the intention gating value to generate the enhanced semantic vector set of the standard text library.
5. The cloud computing based dynamic comparison method for standard content consistency as claimed in claim 4, wherein: The similarity between the semantic vector set of the to-be-compared document and the enhanced semantic vector set of the dynamically aligned standard text library is calculated to generate a preliminary similarity matrix, including the following steps: Each to-be-compared document semantic vector in the to-be-compared document semantic vector set is traversed, and the cosine similarity between each to-be-compared document semantic vector in the to-be-compared document semantic vector set and each standard text library enhanced semantic vector in the dynamically aligned standard text library enhanced semantic vector set is calculated. The cosine similarity calculation results of each to-be-compared document semantic vector in the to-be-compared document semantic vector set and each standard text library enhanced semantic vector in the dynamically aligned standard text library enhanced semantic vector set are organized into a preliminary similarity matrix.
6. The cloud computing based dynamic comparison method for standard content consistency as claimed in claim 5, wherein: The user end selects a to-be-verified comparison pair in the preliminary similarity matrix and initiates a request to the cloud service to obtain the generation proof corresponding to the to-be-verified comparison pair, including the following steps: The cloud service extracts the to-be-compared document semantic vector from the to-be-compared document semantic vector set according to the to-be-verified comparison pair coordinates, and extracts the standard text library enhanced semantic vector from the dynamically aligned standard text library enhanced semantic vector set; The cloud service generates a Merkle tree member proof of the to-be-compared document semantic vector using the Merkle tree structure of the to-be-compared document semantic vector set; The cloud service generates a Merkle tree member proof of the standard text library enhanced semantic vector using the Merkle tree structure of the dynamically aligned standard text library enhanced semantic vector set; The cloud service calculates the dot product and the modulus length of the to-be-compared document semantic vector and the standard text library enhanced semantic vector to obtain the similarity of the to-be-verified comparison pair; The cloud service generates a calculation correctness proof based on the vector commitment of the to-be-compared document semantic vector set and the vector commitment of the dynamically aligned standard text library enhanced semantic vector set for the dot product and modulus length operation; The cloud service combines the Merkle tree member proof of the to-be-compared document semantic vector, the Merkle tree member proof of the standard text library enhanced semantic vector, and the calculation correctness proof to generate the generation proof of the to-be-verified comparison pair.
7. The cloud computing based dynamic comparison method for standard content consistency as claimed in claim 6, wherein: Based on the generation proof, the similarity result of the to-be-verified comparison pair is calculated from the to-be-compared document semantic vector set and the dynamically aligned standard text library enhanced semantic vector set. Based on the verification result, a final consistency comparison report is generated according to the preliminary similarity matrix, including the following steps: The user end uses the Merkle tree member proof in the generation proof and the locally stored Merkle root to verify that the to-be-compared document semantic vector and the standard text library enhanced semantic vector belong to the to-be-compared document semantic vector set and the dynamically aligned standard text library enhanced semantic vector set, respectively. The calculation correctness proof in the generation proof and the locally stored vector commitment are used to verify that the dot product and the correct dot product result of the length operation relied on by the similarity result of the to-be-verified comparison pair are correct. The verified correct dot product result and the length result are used to recalculate the cosine similarity, and the calculation result is compared with the similarity result of the to-be-verified comparison pair to verify the consistency. When the vector membership verification, calculation correctness verification, and result consistency verification pass, the preliminary similarity matrix is determined to be credible. According to the preliminary similarity matrix determined to be credible and the comparison result based on the mandatory category in the standard knowledge graph and the historical compliance document, the final consistency comparison report is generated by calculating the quantile of the semantic similarity distribution and setting the consistency threshold.
8. A cloud computing-based dynamic comparison system for standard content consistency, based on the cloud computing-based dynamic comparison method for standard content consistency according to any one of claims 1-7, characterized in that: The data processing module receives the to-be-compared document and calls the standard text library, pre-processes the to-be-compared document and the standard text library to obtain the to-be-compared document structured text and the standard text library structured text. The semantic vectorization module uses a semantic model to perform vectorization processing on the to-be-compared document structured text and the standard text library structured text to generate a to-be-compared document semantic vector set and a standard text library semantic vector set. The similarity matrix module generates a standard text library enhanced semantic vector set based on the pre-constructed standard knowledge graph, and generates a preliminary similarity matrix by calculating the similarity between the to-be-compared document semantic vector set and the dynamically aligned standard text library enhanced semantic vector set. The generation proof module selects the to-be-verified comparison pair in the preliminary similarity matrix and initiates a request to the cloud service to obtain the generation proof corresponding to the to-be-verified comparison pair. The report generation module verifies, based on the generation proof, that the similarity result of the to-be-verified comparison pair is calculated from the to-be-compared document semantic vector set and the dynamically aligned standard text library enhanced semantic vector set, and generates a final consistency comparison report based on the verification result according to the preliminary similarity matrix. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to implement the steps of the cloud computing-based standard content consistency dynamic comparison method of any one of claims 1-7.
10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the cloud computing-based standard content consistency dynamic comparison method of any one of claims 1-7.