Big data hypergraph network modeling method and device, computer equipment and product

Abstract

The invention discloses a big data hypergraph network modeling method and device, computer equipment and a product, aims to solve the problem that efficiency and accuracy are difficult to consider when large-scale data of intellectual property is processed in the prior art, and aims at the characteristics of high increment, high timeliness and multiple modes of the intellectual property data. The method comprises the following steps: analyzing and concluding passed intellectual property files, classifying according to corresponding professions, and establishing a corresponding intellectual property text library; constructing a text corpus based on intellectual property by using a BERT algorithm model, performing capsule graph neural network processing on the image information of the corresponding category, obtaining a correlation between image objects and text attributes thereof, and constructing an image database; capturing complex correlation characteristics in the text by utilizing the characteristics of the graph convolutional neural network; and constructing a hypergraph neural network model. The invention provides a big data hypergraph network modeling method capable of processing high-increment, high-timeliness and multi-modal data.

Description

Technical Field

[0001] The present application relates to the technical field of intellectual property detection, and in particular to a big data hypergraph network modeling method. Background Art

[0002] With the advent of the information age, the protection and duplication detection of intellectual property rights have become increasingly important. However, with the surge in the number of intellectual property applications, traditional duplication detection methods face huge challenges, especially when dealing with large-scale data, it is difficult to achieve both efficiency and accuracy. Summary of the invention

[0003] The present application provides a big data hypergraph network modeling method, apparatus, computer equipment and product, aiming to solve the problem in the prior art of difficulty in balancing efficiency and accuracy when processing large-scale data of intellectual property rights. At the same time, in view of the high-increment, high-efficiency and multi-modal characteristics of intellectual property data, a big data hypergraph network modeling method capable of processing high-increment, high-efficiency and multi-modal data is proposed.

[0004] In a first aspect, a big data hypergraph network modeling method is provided, the method comprising steps A1-A5.

[0005] Step A1: Analyze and summarize the approved intellectual property documents, classify them according to profession, and establish the corresponding intellectual property text library and image database.

[0006] Step A2, using the BERT algorithm model to process and analyze the intellectual property text library, generate deep sentence vectors, and perform dimensionality reduction processing on high-dimensional features to generate a text corpus, retain core information to improve model efficiency, and achieve rapid detection of new application-type texts.

[0007] Step A3, based on the text corpus, capsule graph neural network processing is performed on the image information of the corresponding category in the image database to obtain the relationship between the image object and its text attributes, build an attribute-associated image database, and realize the association between images and texts.

[0008] Step A4, using the characteristics of the graph convolutional neural network, analyzes and learns the attribute-associated image database, extracts corresponding structural information, aggregates neighbor information of nodes in the graph, captures complex association characteristics in the text, and generates an attribute database.

[0009] Step A5, using a hypergraph neural network, by learning the attribute database and the text corpus, updating the representation of nodes and hyperedges to capture the complex relationships between entities, using the nodes in the hypergraph model to perform similarity search, and quickly retrieve similar intellectual property content.

[0010] In the above scheme, optionally, the step A1 includes: collecting the required patent data set from the intellectual property text library, the patent data set including publication number, title, abstract, IPC classification number, and number of citations; collecting professional terms and technical terms in specific fields for rapid annotation; and screening out applicable features in the same field as evaluation criteria.

[0011] In the above scheme, optionally, step A2 includes: further training the intellectual property text library using the BERT algorithm, the training set is derived from the patent efficacy word set and the technical word set, randomly extracting the same number of words from the two word sets, and combining them into efficacy phrases and technical phrases respectively, and forming a training sample in pairs in the vector form of "phrase A, phrase B, label Label", Label is 0, indicating that phrase A and phrase B belong to different categories, and Label is 1, indicating that phrase A and phrase B belong to the same category; through the loss function and back propagation mechanism, the parameters are continuously iterated and updated to optimize the model and improve its generalization ability; wherein, the patent efficacy word set is a vocabulary set that describes the technical achievements or advantages of the patent, and the technical word set is a term set that constitutes the core of the patent technology.

[0012] In the above scheme, optionally, step A2 includes: inputting the intellectual property text library into a Transformer encoder, combining a self-attention mechanism to learn deep features of the text data, given a query key Q, a key K and a value V, and using a multi-head attention mechanism to capture the associated information of different text segments and related text segments in the model.

[0013] Q=X*W i (q)

[0014] K=X*W i (k)

[0015] V=X*W i (v)

[0016]

[0017] Z MH =Concat(Z1,Z2,......Z i )W0

[0018] In the formula, Q represents the query vector Query, which represents the information that the element currently being processed wants to obtain from other elements; K represents the key vector Key, which represents the identity or characteristics of other elements and is used to match the query vector; V represents the value vector Value, which represents the information content that is allowed to be obtained by the query vector after matching with the key vector; W i (q),W i (k) ,W i (v) represents the learnable weight matrix; represents the dot product attention score between Query and Key, which is used to measure the correlation between Q and K; dk represents the dimension of Key; Z MH represents the final multi-head attention convergence output; W0 represents the learnable parameter; i represents the number of detection heads.

[0019] In the above scheme, optionally, the step A3 includes steps A301 to A303.

[0020] Step A301, processing the input image and converting it into primitive capsules.

[0021] Step A302: converting the original capsule into a digital capsule through a dynamic routing mechanism.

[0022] Step A303: pass the digital capsule to the classification layer to obtain the classification result.

[0023] Among them, the feature vector extracted from the image database is called a capsule, and a routing mechanism is used to generate high-level features. The routing mechanism retains all information from low-level capsules and allows each image to be modeled using multiple embeddings. Each embedding reflects different attributes of the image. The main calculation process is as follows:

[0024]

[0025]

[0026]

[0027]

[0028] In the formula, u i represents the i-th capsule unit of the lower-level capsule unit; W ij represents the connection weight matrix between the i-th capsule unit and the j-th prediction capsule unit; U j|i represents the predicted capsule unit; b ij represents the consistency parameter of the i-th capsule unit in the L-th layer linking the j-th capsule in the (L+1)-th layer, where L starts from 0 and increases; c ij represents the coupling coefficient of each capsule unit; sj represents the weighted sum of the jth capsule unit; v j represents the j-th capsule unit in the (L+1)-th layer.

[0029] Get the capsule vector output by the capsule layer, determine the probability of the professional list of intellectual property rights to which the capsule vector belongs according to the result of the capsule vector, and construct the corresponding loss function for reverse optimization:

[0030] L k =T k max(0,m + -v k ) 2 +λ(1-T k )max(0,v k -m - ) 2

[0031] In the formula, k represents the classification category; T k Indicator function representing the classification; v k The vector unit representing the dynamic routing clustering into category k; m + , m - Represent the upper and lower boundaries respectively; λ represents the weight to reduce the loss value; ||v k || represents the vector modulus of category k in dynamic routing clustering.

[0032] In the above scheme, optionally, step A4 includes the following contents.

[0033] As a basic data structure, a graph consists of a set of nodes and a set of edges, denoted as G = (V, E), where V represents the set of nodes in the graph, E represents the set of edges connecting nodes, and the nodes have features that represent the information of the nodes. After graph convolution processing, the embedded representation of the nodes is obtained:

[0034]

[0035] In the formula, H (l) Represents the node feature matrix of the lth layer, with a size of N*F (l) , N represents the number of nodes, F (l) Represents the feature dimension of this layer; W (l) Represents the weight matrix of the lth layer, with size F (l) ×F (l+1) , used to transform features from dimension F (l) Mapping to F (l +1) ; A represents the matrix containing the node information after node aggregation, that is, A plus the identity matrix I: A = A + I; D represents the degree matrix of A, that is, D i is the sum of the i-th row of A; σ represents a nonlinear activation function, and at the same time, it is normalized by the graph to coordinate the feature information of all nodes:

[0036]

[0037]

[0038] In the formula, G k represents the set of all nodes in the Kth graph, G k represents the number of all nodes, Represents the mean and standard deviation vectors of the graph; the graph neural network transmits and aggregates information through neighborhood edges to update the embedding representation of all nodes, improves model performance by capturing the characteristic relationships of high-order neighborhoods, effectively processes structural data, and captures complex relationships between nodes.

[0039] In the above scheme, optionally, step A5 includes the following contents.

[0040] Design a hypergraph sampling module and a computational acceleration module based on a single-stage hypergraph convolution to aggregate the acquired features and the hyperedge information related to the nodes.

[0041]

[0042] Where N i Represents the set of adjacent nodes of node i, j∈N i Indicates a range of the network; H ij Represents the association weight between nodes; maps the aggregated features to a new feature space for similarity scoring or classification; X i The node representing the i-th feature, X j The node representing the jth feature.

[0043] Construct the corresponding triple H = (V, E, W), where V represents the aggregate database node X i The set of hypergraphs is E, and W is the set of weights. The hypergraph sampling module is used to reduce the computational complexity:

[0044] V b = RandomSample(V b ,W)

[0045] H' b =Slice(V b ',E,W)

[0046]

[0047] Where V b ' represents the sampled set, H b ' represents the hypergraph correlation matrix, X b 'Given the final correlation expression, in the intellectual property duplication check, the hypergraph model is used to capture the complex relationship between documents and update the feature representation of each document, so as to perform effective duplication check.

[0048] In a second aspect, a big data hypergraph network modeling device is provided, the device comprising: a data preprocessing module, a text feature extraction module, an image feature extraction module, a structural information feature extraction module and a hypergraph model building module.

[0049] The data preprocessing module is used to analyze and summarize the intellectual property documents that have been approved, classify them according to profession, and establish the corresponding intellectual property text library and image database.

[0050] The text feature extraction module is used to use the BERT algorithm model to process and analyze the intellectual property text library, generate deep sentence vectors, and perform dimensionality reduction processing on high-dimensional features to generate a text corpus, retain core information to improve model efficiency, and achieve rapid detection of new application-type texts.

[0051] The image feature extraction module is used to perform capsule graph neural network processing on the image information of the corresponding category in the image database based on the text corpus, obtain the relationship between the image object and its text attributes, build an attribute-related image database, and realize the correlation between images and texts.

[0052] The structural information feature extraction module is used to analyze and learn the attribute-related image database by utilizing the characteristics of the graph convolutional neural network, and to extract the corresponding structural information, aggregate the neighbor information of the nodes in the graph, capture the complex association characteristics in the text, and generate an attribute database.

[0053] A hypergraph model module is constructed to utilize a hypergraph neural network to learn the attribute database and the text corpus, update the representation of nodes and hyperedges to capture the complex relationships between entities, and use the nodes in the hypergraph model to perform similarity searches to quickly retrieve similar intellectual property content.

[0054] According to a third aspect, a computer device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described above.

[0055] In a fourth aspect, a computer program product comprises a computer program / instruction, which implements the steps of the method described above when executed by a processor.

[0056] Compared with the prior art, this application has at least the following beneficial effects:

[0057] This application is based on further analysis and research of existing technical problems, and recognizes that with the rapid growth in the number of intellectual property applications, traditional duplicate checking methods have problems with balancing efficiency and accuracy when processing large-scale data. This application achieves efficient processing and analysis of intellectual property data by constructing a hypergraph network structure, combining a high-incremental data processing algorithm, a high-efficiency data update mechanism, a multimodal data processing module, and an attention mechanism. It supports all-round needs from knowledge discovery to legal compliance analysis by integrating multi-source data, providing real-time updates, and deeply mining data patterns, thereby promoting technological innovation and business development. Through the big data hypergraph network modeling method, the efficiency and accuracy problems faced by traditional duplicate checking methods when processing large-scale data are effectively solved, providing a more efficient and accurate solution for the protection and duplicate checking of intellectual property rights.

[0058] This application also improves the detection speed of different types of intellectual property rights through a hypergraph model. Its flexible data structure effectively represents and manages complex relationships between entities, allowing multimodal data and many-to-many interactive relationships in intellectual property rights to be expressed intuitively, meeting real-time requirements. This is crucial for the field of intellectual property rights that needs to respond quickly to market changes and legal updates.

[0059] This application also establishes a comprehensive intellectual property text library by classifying and summarizing the intellectual property documents that have been approved, providing a solid foundation for subsequent duplicate checking. The BERT algorithm model is used to build a text corpus based on intellectual property. By generating deep sentence vectors and performing high-dimensional feature dimensionality reduction processing, it not only retains the core information, but also significantly improves the detection speed and accuracy of the model when processing new application texts.

[0060] This application also constructs an image database by performing capsule graph neural network processing on image information of corresponding categories, and realizes the correlation between images and texts. This helps to comprehensively consider various forms of intellectual property rights in the process of duplicate checking and improve the comprehensiveness of duplicate checking.

[0061] This application also analyzes and learns the image database through a graph convolutional neural network, extracts the corresponding structural information, and aggregates the neighbor information of the nodes in the graph, thereby capturing the complex correlation characteristics in the text, which is particularly important for identifying highly similar but slightly modified intellectual property content.

[0062] This application also proposes to build a hypergraph neural network to conduct in-depth learning of the database, continuously update the representation of nodes and hyperedges, and capture the complex relationship between entities. Using the nodes in the hypergraph model for similarity search can quickly retrieve similar intellectual property content, which not only improves the efficiency of duplicate checking, but also ensures the accuracy of duplicate checking. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A flowchart of a big data hypergraph network modeling method provided for one embodiment of the present application.

[0064] Figure 2 A flowchart of database preprocessing provided for one embodiment of the present application.

[0065] Figure 3 A schematic diagram of the detection process of image information provided for one embodiment of the present application.

[0066] Figure 4 A schematic diagram of a hypergraph algorithm modeling structure provided for one embodiment of the present application.

[0067] Figure 5 A block diagram of the module architecture of a big data hypergraph network modeling device provided for one embodiment of the present application.

[0068] Figure 6 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0070] In the description of this application: unless otherwise specified, "plurality" means two or more. "Including", "comprising", "having" and other expressions also mean "not limited to" (certain units, components, materials, steps, etc.).

[0071] In one embodiment, reference Figure 1 , a big data hypergraph network modeling method is provided, which includes steps A1-A5.

[0072] Step A1: Analyze and summarize the approved intellectual property documents, classify them according to profession, and establish the corresponding intellectual property text library and image database.

[0073] Step A2: Use the BERT algorithm model to process and analyze the intellectual property text library, generate deep sentence vectors, and perform dimensionality reduction processing on high-dimensional features to generate a text corpus, retain core information to improve model efficiency, and achieve rapid detection of new application texts.

[0074] Step A3, based on the text corpus, capsule graph neural network processing is performed on the image information of the corresponding category in the image database to obtain the relationship between the image object and its text attributes, build an attribute-related image database, and realize the correlation between images and texts.

[0075] Step A4, using the characteristics of graph convolutional neural networks, analyze and learn the attribute-related image database, extract the corresponding structural information, aggregate the neighbor information of the nodes in the graph, capture the complex association characteristics in the text, and generate an attribute database.

[0076] Step A5, using the hypergraph neural network, by learning the attribute database and text corpus, updating the representation of nodes and hyperedges to capture the complex relationship between entities, using the nodes in the hypergraph model to perform similarity search, and quickly retrieve similar intellectual property content.

[0077] In one embodiment, reference Figure 2 ,Step A1 includes: collecting the required patent data set from the intellectual property text library, the patent data set includes publication number, title, abstract, IPC classification number, and number of citations; collecting professional terms and technical terms in specific fields for rapid annotation; screening out applicable features in the same field as evaluation criteria.

[0078] In one embodiment, step A2 includes: further training the intellectual property text library using the BERT algorithm to generate a text corpus. The training set is derived from the patent efficacy word set and the technical word set. The same number of words are randomly extracted from the two word sets, and the words are respectively combined into efficacy phrases and technical phrases. A training sample is formed in pairs in the form of a vector of "phrase A, phrase B, label Label". Label is 0, indicating that phrase A and phrase B belong to different categories, and Label is 1, indicating that phrase A and phrase B belong to the same category. Through the loss function and back-propagation mechanism, the parameters are continuously iterated and updated to optimize the model and improve its generalization ability. Among them, the patent efficacy word set is a vocabulary set that describes the technical achievements or advantages of the patent, and the technical word set is a term set that constitutes the core of the patent technology.

[0079] In one embodiment, step A2 includes: inputting the intellectual property text library into the Transformer encoder, combining the self-attention mechanism to learn the deep features of the text data, given the query Q key, K key and value V, and using the multi-head attention mechanism to capture the associated information of different text segments and related text segments in the model.

[0080] Q=X*W i (q)

[0081] K=X*W i (k)

[0082] V=X*W i (v)

[0083]

[0084] Z MH =Concat(Z1,Z2,......Z i )W0

[0085] In the formula, Q represents the query vector Query, which represents the information that the element currently being processed wants to obtain from other elements; K represents the key vector Key, which represents the identity or characteristics of other elements and is used to match the query vector; V represents the value vector Value, which represents the information content that is allowed to be obtained by the query vector after matching with the key vector; W i (q) ,W i (k) ,W i (v) represents the learnable weight matrix; represents the dot product attention score between Query and Key, which is used to measure the correlation between Q and K; dk represents the dimension of Key; Z MH represents the final multi-head attention convergence output; W0 represents the learnable parameter; i represents the number of detection heads.

[0086] In this embodiment, the BERT model can efficiently learn the deep semantic representation of text and perform effective classification prediction.

[0087] In one embodiment, reference Figure 3 , step A3 performs feature learning of the preprocessed image database, including steps A301 to A303.

[0088] Step A301, processing the input image and converting it into primitive capsules.

[0089] Step A302: converting the original capsule into a digital capsule through a dynamic routing mechanism.

[0090] Step A303: pass the digital capsule to the classification layer to obtain the classification result.

[0091] The feature vector extracted from the image database is called a capsule, and a routing mechanism is used to generate high-level features. The routing mechanism retains all information from low-level capsules and allows each image to be modeled using multiple embeddings. Each embedding reflects different attributes of the image. The main calculation process is as follows:

[0092]

[0093]

[0094]

[0095]

[0096] In the formula, u i represents the i-th capsule unit of the lower-level capsule unit; W ij represents the connection weight matrix between the i-th capsule unit and the j-th prediction capsule unit; U j|i represents the predicted capsule unit; b ij represents the consistency parameter of the i-th capsule unit in the L-th layer linking the j-th capsule in the (L+1)-th layer, where L starts from 0 and increases; c ij Represents the coupling coefficient of each capsule unit; s j represents the weighted sum of the j-th capsule unit; v j represents the j-th capsule unit in the (L+1)-th layer.

[0097] Get the capsule vector output by the capsule layer, determine the probability of the professional list of intellectual property rights to which the capsule vector belongs based on the result, and construct the corresponding loss function for reverse optimization:

[0098] L k =T k max(0,m + -||v k ||) 2 +λ(1-T k )max(0,||v k ||-m - ) 2

[0099] In the formula, k represents the classification category; T k Indicator function representing the classification; v k represents the vector unit of dynamic routing clustering into category k; m + , m - Represent the upper and lower boundaries respectively; λ represents the weight to reduce the loss value; ||v k || represents the vector modulus of category k in dynamic routing clustering.

[0100] In the above scheme, optionally, step A4 includes the following contents.

[0101] As a basic data structure, a graph consists of a set of nodes and a set of edges, denoted as G = (V, E), where V represents the set of nodes in the graph, and E represents the set of edges connecting nodes. The nodes have features that represent the information of the nodes. After graph convolution processing, the embedded representation of the nodes is obtained:

[0102]

[0103] In the formula, H (l) Represents the node feature matrix of the lth layer, with a size of N*F(l) , N represents the number of nodes, F (l) Represents the feature dimension of this layer; W (l) Represents the weight matrix of the lth layer, with size F (l) ×F (l+1) , used to transform features from dimension F (l) Mapping to F (l +1) ; A represents the matrix containing the node information after node aggregation, that is, A plus the identity matrix I: A = A + I; D represents the degree matrix of A, that is, D i is the sum of the i-th row of A; σ represents a nonlinear activation function, and at the same time, it is normalized by the graph to coordinate the feature information of all nodes:

[0104]

[0105]

[0106] In the formula, G k represents the set of all nodes in the Kth graph, G k represents the number of all nodes, Represents the mean and standard deviation vectors of the graph; the graph neural network transmits and aggregates information through neighborhood edges to update the embedding representation of all nodes, improves model performance by capturing the characteristic relationships of high-order neighborhoods, effectively processes structural data, and captures complex relationships between nodes, and is applied to various graph analysis tasks.

[0107] In one embodiment, reference Figure 4 To achieve high effectiveness of step A5, it also includes: designing a hypergraph sampling module and a computing acceleration module based on a single-stage hypergraph convolution to aggregate the features already acquired and aggregate hyperedge information related to the nodes.

[0108]

[0109] Where N i Represents the set of adjacent nodes of node i, j∈N i Indicates a range of the network; H ij Represents the association weight between nodes; maps the aggregated features to a new feature space for similarity scoring or classification; X i The node representing the i-th feature, X j The node representing the jth feature.

[0110] Construct the corresponding triple H = (V, E, W), where V represents the aggregate database node X i The set of hypergraphs is E, and W is the set of weights. The hypergraph sampling module is used to reduce the computational complexity:

[0111] V b = RandomSample(V b ,W)

[0112] H' b =Slice(V b ',E,W)

[0113]

[0114] Where V b ' represents the sampled set, H b represents the hypergraph correlation matrix, X b 'Given the final correlation expression, in the intellectual property duplication check, the hypergraph model is used to capture the complex relationship between documents and update the feature representation of each document, so as to perform effective duplication check.

[0115] In one embodiment, reference Figure 5 , provides a big data hypergraph network modeling device, which includes the following program modules: data preprocessing module, text feature extraction module, image feature extraction module, structural information feature extraction module and hypergraph model building module.

[0116] The data preprocessing module is used to analyze and summarize the intellectual property documents that have been approved, classify them according to profession, and establish the corresponding intellectual property text library and image database.

[0117] The text feature extraction module is used to use the BERT algorithm model to process and analyze the intellectual property text library, generate deep sentence vectors, and reduce the dimensionality of high-dimensional features to generate a text corpus, retain core information to improve model efficiency, and achieve rapid detection of new application texts.

[0118] The image feature extraction module is used to perform capsule graph neural network processing on the image information of the corresponding category in the image database based on the text corpus, obtain the relationship between the image object and its text attributes, build an attribute-related image database, and realize the correlation between images and texts.

[0119] The structural information feature extraction module is used to analyze and learn the attribute-related image database by utilizing the characteristics of the graph convolutional neural network, extract the corresponding structural information, aggregate the neighbor information of the nodes in the graph, capture the complex association characteristics in the text, and generate an attribute database.

[0120] Build a hypergraph model module to use the hypergraph neural network to learn from the attribute database and text corpus, update the representation of nodes and hyperedges to capture the complex relationships between entities, use the nodes in the hypergraph model to perform similarity search, and quickly retrieve similar intellectual property content.

[0121] The specific implementation content of each module can be found in the above definition of a big data hypergraph network modeling method, which will not be repeated here.

[0122] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, a communication interface, and a human-computer interaction interface (such as a combination of a display, a keyboard, and a mouse, and a touch screen, etc.) connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities, and the communication interface is used to communicate with an external terminal in a wired or wireless manner. The wireless method can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The computer device implements the above-mentioned big data hypergraph network modeling method by loading and running a computer program.

[0123] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0124] In one embodiment, a computer program product is also provided, including a computer program / instruction, which implements the steps of the above-mentioned embodiment method when executed by a processor.

[0125] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A big data hypergraph network modeling method, characterized in that: include: Step A1: Analyze and summarize the approved intellectual property documents, classify them according to profession, and establish the corresponding intellectual property text library and image database; Step A2, using the BERT algorithm model to process and analyze the intellectual property text library, generate deep sentence vectors, and perform dimensionality reduction processing on high-dimensional features to generate a text corpus, retain core information to improve model efficiency, and achieve rapid detection of new application texts; Step A3, based on the text corpus, capsule graph neural network processing is performed on the image information of the corresponding category in the image database to obtain the relationship between the image object and its text attributes, build an attribute-associated image database, and realize the association between the image and the text; Step A4, using the characteristics of graph convolutional neural networks, analyze and learn the attribute-related image database, extract corresponding structural information, aggregate neighbor information of nodes in the graph, capture complex association characteristics in the text, and generate an attribute database; Step A5, using a hypergraph neural network, by learning the attribute database and the text corpus, updating the representation of nodes and hyperedges to capture the complex relationships between entities, using the nodes in the hypergraph model to perform similarity search, and quickly retrieve similar intellectual property content.

2. The big data hypergraph network modeling method according to claim 1, characterized in that: The step A1 comprises: Collect the required patent data set from the intellectual property text library, the patent data set includes publication number, title, abstract, IPC classification number, and number of citations; collect professional terms and technical terms in specific fields for rapid annotation; and screen out applicable features in the same field as evaluation criteria.

3. The big data hypergraph network modeling method according to claim 1, characterized in that: The step A2 comprises: The BERT algorithm is used to further train the intellectual property text library. The training set comes from the patent efficacy word set and the technical word set. The same number of words are randomly extracted from the two word sets and combined into efficacy phrases and technical phrases respectively. A training sample is formed in pairs in the vector form of "phrase A, phrase B, label Label". Label is 0, indicating that phrase A and phrase B belong to different categories, and Label is 1, indicating that phrase A and phrase B belong to the same category. Through the loss function and back propagation mechanism, the parameters are continuously iterated and updated to optimize the model and improve its generalization ability. Among them, the patent efficacy word set is a vocabulary set that describes the technical achievements or advantages of the patent, and the technical word set is a term set that constitutes the core of the patent technology.

4. The big data hypergraph network modeling method according to claim 1, characterized in that: The step A2 comprises: Input the intellectual property text library into the Transformer encoder, combine the self-attention mechanism to learn the deep features of the text data, and use the multi-head attention mechanism to capture the correlation information of different text segments and related text segments in the model given the query key Q, key K and value V; Q=X*W i (q) K=X*W i (k) V=X*W i (v) From MH =Concat(Z1,Z2,......Z i )W0 In the formula, Q represents the query vector Query, which represents the information that the element currently being processed wants to obtain from other elements; K represents the key vector Key, which represents the identity or characteristics of other elements and is used to match the query vector; V represents the value vector Value, which represents the information content that is allowed to be obtained by the query vector after matching with the key vector; W i (q) ,W i (k) ,W i (v) represents the learnable weight matrix; represents the dot product attention score between Query and Key, which is used to measure the correlation between Q and K; dk represents the dimension of Key; Z MH represents the final multi-head attention convergence output; W0 represents the learnable parameter; i represents the number of detection heads.

5. The big data hypergraph network modeling method according to claim 1, characterized in that: The step A3 comprises: Step A301, processing the input image and converting it into primitive capsules; Step A302, converting the original capsule into a digital capsule through a dynamic routing mechanism; Step A303, passing the digital capsule to the classification layer to obtain the classification result; Among them, the feature vector extracted from the image database is called a capsule, and a routing mechanism is used to generate high-level features. The routing mechanism retains all information from low-level capsules and allows each image to be modeled using multiple embeddings. Each embedding reflects different attributes of the image. The main calculation process is as follows: In the formula, u i represents the i-th capsule unit of the lower-level capsule unit; W ij represents the connection weight matrix between the i-th capsule unit and the j-th prediction capsule unit; U j|i represents the predicted capsule unit; b ij represents the consistency parameter of the i-th capsule unit in the L-th layer linking the j-th capsule in the (L+1)-th layer, where L starts from 0 and increases; c ij Represents the coupling coefficient of each capsule unit; s j represents the weighted sum of the j-th capsule unit; v j represents the j-th capsule unit in the (L+1)-th layer; Get the capsule vector output by the capsule layer, determine the probability of the professional list of intellectual property rights to which the capsule vector belongs according to the result of the capsule vector, and construct the corresponding loss function for reverse optimization: L k =T k max(0,m + -||v k ||) 2 +λ(1-T k )max(0,||v k ||-m - ) 2 In the formula, k represents the classification category; T k Indicator function representing the classification; v k The vector unit representing the dynamic routing clustering into category k; m + represents the upper boundary, m - represents the lower boundary; λ represents the weight to reduce the loss value; ||v k || represents the vector modulus of category k in dynamic routing clustering.

6. The big data hypergraph network modeling method according to claim 1, characterized in that: The step A4 comprises: As a basic data structure, a graph consists of a set of nodes and a set of edges, denoted as G = (V, E), where V represents the set of nodes in the graph, E represents the set of edges connecting nodes, and the nodes have features that represent the information of the nodes. After graph convolution processing, the embedded representation of the nodes is obtained: In the formula, H (l) Represents the node feature matrix of the lth layer, with a size of N*F (l) , N represents the number of nodes, F (l) Represents the feature dimension of this layer; W (l) Represents the weight matrix of the lth layer, with size F (l) ×F (l+1) , used to transform features from dimension F (l) Mapping to F (l+1) ; A represents the matrix containing the node information after node aggregation, that is, A plus the identity matrix I: A = A + I; D represents the degree matrix of A, that is, D i is the sum of the i-th row of A; σ represents a nonlinear activation function, and at the same time, it is normalized by the graph to coordinate the feature information of all nodes: In the formula, G k represents the set of all nodes in the Kth graph, |G k | represents the number of all nodes, Represents the mean and standard deviation vectors of the graph; the graph neural network transmits and aggregates information through neighborhood edges to update the embedding representation of all nodes, improves model performance by capturing the characteristic relationships of high-order neighborhoods, effectively processes structural data, and captures complex relationships between nodes.

7. The big data hypergraph network modeling method according to claim 1, characterized in that: The step A5 comprises: Design a hypergraph sampling module and a computational acceleration module based on a single-stage hypergraph convolution to aggregate the acquired features and aggregate the hyperedge information related to the nodes. Where N i Represents the set of adjacent nodes of node i, j∈N i Indicates a range of the network; H ij Represents the association weight between nodes; maps the aggregated features to a new feature space for similarity scoring or classification; X i The node representing the i-th feature, X j The node representing the jth feature; Construct the corresponding triple H = (V, E, W), where V represents the aggregate database node X i The set of hypergraphs is E, and W is the set of weights. The hypergraph sampling module is used to reduce the computational complexity: In' b =RandomSample(V b ,W) H' b =Slice(V’ b ,E,W) Where V' b represents the sampled set, H' b represents the hypergraph correlation matrix, X' b Given the final relevance expression, in intellectual property duplication checking, a hypergraph model is used to capture the complex relationships between documents and update the feature representation of each document, thereby performing effective duplication checking.

8. A big data hypergraph network modeling device, characterized in that: include: The data preprocessing module is used to analyze and summarize the approved intellectual property documents, classify them according to profession, and establish the corresponding intellectual property text library and image database; A text feature extraction module is used to process and analyze the intellectual property text library using the BERT algorithm model to generate deep sentence vectors, and to perform dimensionality reduction processing on high-dimensional features to generate a text corpus, retain core information to improve model efficiency, and achieve rapid detection of new application texts; An image feature extraction module is used to perform capsule graph neural network processing on the image information of the corresponding category in the image database based on the text corpus, obtain the relationship between the image object and its text attributes, build an attribute-related image database, and realize the association between images and texts; A structural information feature extraction module is used to analyze and learn the attribute-related image database using the characteristics of the graph convolutional neural network, extract corresponding structural information, aggregate neighbor information of nodes in the graph, capture complex association characteristics in the text, and generate an attribute database; A hypergraph model module is constructed to utilize a hypergraph neural network to learn the attribute database and the text corpus, update the representation of nodes and hyperedges to capture the complex relationships between entities, and use the nodes in the hypergraph model to perform similarity searches to quickly retrieve similar intellectual property content.

9. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 1.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images