43 results about "Graph classification" patented technology

The invention discloses an abnormal transaction dynamic detection method and system fusing multi-scale analysis and information entropy, and the method comprises the steps: S01, constructing a dynamic heterogeneous graph which comprises account nodes, equipment nodes and IP nodes; s02, extracting node features of each node, extracting a spectrum feature matrix of the dynamic heterogeneous graph, and forming a node feature vector of each node; step S03, clustering the dynamic heterogeneous graph to screen out candidate detection sub-graphs, inputting node feature vectors of nodes in the candidate detection sub-graphs into a pre-trained graph classification model, and identifying an abnormal account node set; and S04, calculating a corresponding dynamic risk score in real time according to the memory state vector of each abnormal account node in the time sequence diagram network and the diagram information entropy change rate so as to evaluate a real-time risk and identify a core abnormal account. According to the method, the group type abnormal transaction account can be quickly and accurately identified, and the core abnormal account can be positioned.

This application relates to the field of integrated circuit design automation technology, and provides a method and system for verifying analog circuit structures based on graph neural networks. The method includes: converting a SPICE netlist into a graph structure representation, where nodes correspond to devices and edges correspond to electrical connections, and extracting feature vectors containing device type, parameters, functional roles, connection types, and signal directions; encoding the circuit diagram using a graph attention network to obtain node embeddings and graph embeddings; performing multi-task prediction based on the embedding results, including node classification, graph classification, structure scoring, and error localization; and finally generating a verification report containing an error list, location information, and repair suggestions. This application achieves automatic detection and precise localization of errors in analog circuit structures, improving detection accuracy by a significant margin and possessing strong generalization capabilities.

The invention provides a node identification method and device based on graph classification and a related product, and relates to the technical field of data processing. The method comprises the following steps: calling a preset connected sub-graph recognition algorithm to recognize one or more connected sub-graphs of which the node number is greater than a set threshold value in a social network to obtain a connected sub-graph set; extracting a user node in the connected sub-graph, and extracting a local structure of the user node in the global network, namely a user case sub-graph taking the user node as a center; then classifying the user case sub-graphs by using a pre-trained graph convolutional network dichotomy model; in this way, user node recognition is converted into recognition of the local network where the user node is located, the user case sub-graphs with the user node as the center are classified through the pre-trained graph convolutional network dichotomy model, and therefore the classification result of the user node is obtained, the calculation amount is greatly reduced, and the recognition accuracy and efficiency are improved.

The application discloses a graph classification method based on a saliency regularization graph neural network and related devices. The method comprises obtaining graph structure data corresponding to a target to be classified, and inputting the graph structure data into a saliency regularization graph neural network; determining a classification category of the target to be classified through the saliency regularization graph neural network; wherein the saliency regularization graph neural network learns a node feature matrix through a skeleton network, extracts the node feature matrix into a compact graph feature representation through a graph neural memory network, determines a saliency distribution vector based on the compact graph feature representation and the node feature matrix, and finally normalizes the aggregation weight of the skeleton network through the saliency distribution vector. In this way, the saliency regularization graph neural network focuses on the nodes more related to graph classification by measuring the compatibility between the entire compact graph feature representation and the node feature matrix, and can learn a more effective representation for the entire graph, thereby improving the classification effect of the graph classification task.

The invention discloses a graph classification method based on multi-view adaptive attention and feature integration enhancement, and aims to improve the representation capability and classification performance of graph structure data. An existing image classification method generally has the problems of single feature extraction view angle and inflexible fusion mechanism, and multi-level structure information in an image is difficult to fully mine. Therefore, the method has the following characteristics and contributions: firstly, a feature integration module fusing multi-graph convolution and one-dimensional convolution is constructed, graph node features are extracted from different view angles through multiple graph convolution, and self-adaptive integration of cross-view-angle features is realized by means of the one-dimensional convolution; secondly, designing a multi-angle adaptive channel attention mechanism, dynamically calculating weight distribution of each channel, and emphatically strengthening information expression of key semantic channels; finally, experimental verification on eight public graph classification data sets shows that the average classification accuracy of the method is improved by about 10% compared with that of a current advanced method, and the superiority and practicability of the method in a graph classification task are effectively proved.

The application discloses a personalized federated graph learning method and system suitable for cross-domain graph classification, and general graph spectrum knowledge is shared in view of the influence of structural heterogeneity in global cooperation, the general spectrum knowledge is shared in a global spectrum encoder including a global feature value encoder and a global filtering encoder, and customers benefit from the cooperation; wherein, other components representing non-general knowledge are reserved locally, each customer customizes personalized graph convolution for its own graph characteristics, and negative influence of spectrum bias is avoided; in view of the influence of structural heterogeneity in local application, a learnable preference is configured for each customer, a personalized preference module is used to cooperate with the work of the global spectrum encoder, personalized graph preference adjustment is performed, and the personalized graph preference adjustment is adapted to the unique graph structure of each customer; wherein, a regularization term is used to limit the personalized preference module to focus on local preference, and an over-reliance problem caused by the regularization term is solved.

The present application relates to a kind of decoupling knowledge distillation hardware target detection method and system, method includes: according to data set mark frame in first model and second model respectively, decoupling is carried out to mark data set, and the foreground target area and background area are obtained, to determine the target kind probability of hardware, and according to target kind probability determine common feature score mask;According to common feature score mask and feature map classification score graph determine first distillation loss function;Based on global context module, according to the determination feature map, construct second distillation loss function;According to each distillation loss function, determine total loss function, and according to total loss function, the second model is trained, to carry out target detection to the hardware image to be measured.The present application uses the common shape feature of different kinds of hardware, adopts decoupling knowledge distillation method, for foreground target area, the common feature belonging to the unique hardware between is used for the information integration of target classification to migrate to small model, improves hardware detection precision.

This invention discloses a method for tracking and tracing sensitive network behaviors, relating to the field of network security technology. The specific steps of this method are as follows: collecting and preprocessing multi-source heterogeneous data in the network environment and storing it in a distributed database; constructing a cross-modal semantic graph using a large language model; training an intent classification model to predict sensitive behavior intents and forming an association set; locating intent trigger points and constructing a complete tracing chain through an intent trigger point causal contribution algorithm; integrating data to generate a standardized tracing report, and presenting it through a visualization tool, supporting interaction and export. This invention first integrates and analyzes multi-source heterogeneous data to construct a cross-modal semantic graph, accurately identifies sensitive behavior intents through model training and filtering, then locates intent trigger points through algorithms, constructs a tracing chain, and finally outputs a visualized report, forming an easy-to-understand report, thus improving the comprehensiveness, accuracy, and convenience of sensitive behavior identification and tracing.

The application discloses a sea surface target detection method based on data enhancement and recursive graph classification of a Transform technology and belongs to the field of radar target detection. The sea surface target detection method comprises the following steps: obtaining a sea surface echo amplitude sequence by using a radar, pre-processing the sea surface echo amplitude sequence to obtain radar measured target echo amplitude data, and processing the radar measured target echo amplitude data into a data format suitable for being input into a Transform; generating historical item data and future item data of the measured target amplitude by using a Transform bidirectional prediction mechanism according to the pre-processed measured target amplitude data, and splicing the historical item data and the future item data with the measured target amplitude data to obtain enhanced target amplitude data; constructing the enhanced measured target amplitude data into a recursive graph data set; training the recursive graph data set by using a convolutional neural network to obtain a classifier model; and classifying a to-be-detected recursive graph by using the classifier model and outputting a classification result.

The application belongs to the field of artificial intelligence, and discloses an instance-level graph prompt and label attention distillation graph classification method, comprising S1, data preparation: obtaining a graph data set, extracting node features and graph structures of each graph sample, and dividing the data set into a training set, a validation set and a test set; S2, teacher model training: constructing a teacher graph neural network model with double prompts and training; S3, student model distillation training: the training target includes fitting the real label of the classification result output by the student model, and approximating the teacher graph-level embedding output by the teacher model to the graph-level embedding generated by the student model; S4, performance evaluation: using the test set to evaluate the graph classification performance of the trained student model. The application adaptively fuses multi-source information through a weight learning module, and adopts a pre-training-teacher-student three-stage training framework, which significantly improves the classification accuracy, model generalization ability and deployment efficiency.

The application discloses a kind of based on multilayer perceptron's graph classification model construction method and graph classification method, the former includes: the classification task of required application scene is acquired to several graph data and constitutes training set;At least part of graph data of training set has class label;Using training set, selected graph neural network is trained, and the graph neural network that training is completed is used as teacher network and saves the classification result corresponding to graph neural network;Selected multilayer perceptron is used as student model, based on training set, teacher network and corresponding classification result, student model is trained using the method of knowledge distillation, and the student model that training is completed is used as graph classification model for the remaining graph data classification of required application scene.The graph classification model of the application combines the respective advantages of multilayer perceptron and graph neural network, without the dependency of graph in reasoning process, ensure higher accuracy while can greatly reduce model calculation complexity, improve reasoning speed, can be used for time limited engineering deployment.

The invention provides a sub-graph reconstruction method for unbalanced graph classification, which can be applied to the technical field of graph neural networks. The method comprises the following steps: identifying minority class nodes in an original graph, and clustering the minority class nodes to obtain at least one local dense cluster; at least one hub node is generated for each cluster, the hub nodes are added into the graph, and each hub node is in-cluster connection with the nodes in the cluster which the hub node belongs to and is in cross-cluster connection with the nodes of other clusters or the hub nodes; selecting seed nodes from the minority class nodes based on preset significance measurement; at least one synthesis node is generated for each seed node, the synthesis nodes are added into the graph, and each synthesis node is connected with a corresponding source seed node and a node adjacent to the source seed node; taking the graph added with the hub nodes and the composite nodes as an expansion graph, and carrying out graph neural network training; wherein a joint loss function is adopted in the training process.

The invention provides an explainable graph neural network method and system based on evidence sub-graph extraction, which are used for improving the explainability and robustness of a graph model while keeping the task prediction performance. In order to solve the problem that information loss and shortcut residues are easily caused due to the fact that single-stage decomposition is too rough in an existing graph evidence technology, the invention provides a two-step decomposition and feature alignment graph evidence framework. According to the framework, firstly, an input graph is divided into an important sub-graph and a noise sub-graph through a first decomposition module, and it is ensured that the important sub-graph can independently complete tasks; on the basis, the second decomposition module further divides the important sub-graph into an evidence sub-graph and a shortcut sub-graph, so that shortcut associated information influencing the generalization ability of the model is eliminated, and an evidence structure supporting a prediction decision is obtained. Meanwhile, a feature mask is introduced to carry out information compensation, and InfoNCE alignment loss between a mask pattern and evidence representation is calculated, so that information loss caused by multi-level decomposition is reduced. And finally realizing compact expression of the evidence sub-graph through sparsity constraint. The method can effectively obtain evidence sub-graph representation with strong discrimination ability, high robustness and high interpretability, is suitable for tasks such as graph classification, and can be deployed in application scenes such as judicial analysis and drug discovery requiring high-credibility interpretation.

The invention relates to a small sample intention recognition method based on multi-view logistic regression, belongs to the technical field of artificial intelligence and natural language processing, and solves the problems that an existing small sample intention classification method is insufficient in example selection, semantic similar intentions are difficult to distinguish and the model generalization ability is limited. According to the technical scheme, the method comprises the steps that semantic embedding processing is conducted on training data and query statements, cosine similarity is calculated, the size K of an initial support set is dynamically determined based on semantic uncertainty, first K examples are selected to form the initial support set, a multi-view logistic regression model is applied to conduct embedding space discriminative conversion, similarity is recalculated, and the initial support set is selected. And carrying out diversity constraint optimization on the final support set by adopting a maximum and minimum distance selection algorithm, and constructing large language model input for intention recognition. According to the method, the most discriminative and representative examples can be automatically screened, and the classification accuracy, robustness and generalization ability are improved.

The application provides a kind of based on betweenness centrality and geometric consistency graph contrast learning method and device, it is related to graph classification technical field.The method includes: according to the betweenness centrality of graph data node and edge, enhanced graph data is constructed to delete probability;Get the node feature of graph data and enhanced graph data, obtain graph representation by aggregating node feature based on Gaussian kernel reading method;According to the topological structure and node feature of graph data, construct multi-scale geometric consistency loss, construct normalized temperature scale cross-entropy loss according to graph representation, and then construct total loss to realize graph classification.The application introduces graph betweenness centrality into graph contrast learning, so as to retain important structural information of graph in graph enhancement process;Introducing multi-scale geometric consistency loss makes the structure information of graph can better guide the generation of graph feature;Optimize aggregated node feature based on Gaussian kernel reading method to obtain complete graph representation, to obtain more discriminative graph-level features.

The application discloses a few-shot graph classification method based on track enhancement and residual prompt, relates to the technical field of artificial intelligence and graph deep learning, and is suitable for molecular property prediction, drug screening and other graph learning tasks. The method obtains node structure roles through graph element track statistics, constructs a track correlation matrix and performs random walk in a track topological space, generates a track enhanced view to extract high-order topological semantics, introduces a learnable global prototype in the pre-training stage, uniformly constrains the prototype probability distribution of the original view and the track enhanced view, thereby learning a more stable graph encoder for the structure mode, in the downstream stage, track anchor representations are aggregated from a small number of class samples, and the anchor representations are parameterized to generate a probability prompt representation, a graph prompt vector is obtained in a differentiable manner, and the weighted aggregation and classification prediction of the node-level representation are guided. The application can adapt to the distribution difference within the class, and improve the generalization performance and robustness under the few-shot condition.

The invention discloses a graph neural network migration method, device and system based on prompt learning, and the method comprises the steps: arranging original business data into an original graph containing nodes and edges, and adding a prompt module to obtain an enhanced graph; establishing a comprehensive model according to the pre-training graph neural network of the freezing parameters, the enhanced graph and the optimized prompt module; and performing task prediction based on the comprehensive model to obtain a node classification, graph classification or link prediction result. The invention provides a graph neural network migration method, device and system based on prompt learning, and the method comprises the steps: converting original business data into an original graph, adding an optimization prompt module to obtain an enhanced graph, fusing a frozen parameter to pre-train a graph neural network to construct a comprehensive model, can adapt to model input, translates downstream task demands, achieves the feature alignment, and improves the migration efficiency. According to the method, the high feature extraction capacity can be reserved, and the problems that under the condition that parameters are frozen, the adaptability of a pre-training model and a downstream prediction task is poor, and cross-task migration prediction is inaccurate can be solved.

The present disclosure provides a kind of training method of graph classification model, prediction method, training device and prediction device, training method includes: obtaining the training graph data including information structure graph, wherein, information structure graph includes molecular structure graph or sentiment text graph;First label predictive subgraph and second label predictive subgraph are extracted from training graph data using first subgraph extractor and second subgraph extractor respectively;First label and second label corresponding to first label predictive subgraph and second label subgraph are predicted using first subgraph classifier and second subgraph classifier respectively;According to first label, second label, first subgraph extractor and second subgraph extractor, the first mask matrix and the second mask matrix learned by the first mask matrix and the second mask matrix are used to calculate comprehensive loss, and the parameters of first graph classifier and second graph classifier are adjusted to train to obtain graph classification model with the minimum target of comprehensive loss, wherein, first mask matrix and second mask matrix are different from each other.

The invention belongs to the technical field of material nondestructive testing, and discloses an ultrasonic intelligent detection method for GH4169 bar black spots, which comprises the following steps: acquiring a GH4169 bar bottom wave A scanning signal; performing digital signal processing on the bottom wave A-scan signal to obtain a bottom wave A-scan signal characteristic time-frequency graph; preprocessing the data; constructing a bottom wave A-scan signal feature time-frequency graph classification model by using a convolutional neural network technology; model training; and loading the bottom wave A scanning signal feature time-frequency graph of the GH4169 bar to be detected into the trained bottom wave A scanning signal feature time-frequency graph classification model, detecting whether the bar has black spots or not, and finally obtaining an ultrasonic detection result of the black spots of the GH4169 bar. According to the method, special materials are not needed, intelligent detection of the black spots of the GH4169 bar can be rapidly carried out under the condition that the GH4169 bar is not damaged, a reliable qualitative method is provided for an ultrasonic detection technology of the GH4169 bar, operability is high, analysis results are rapid and visual, and quality control and fault analysis can be effectively carried out on the high-temperature alloy bar.

Various embodiments of the present disclosure provide data storage, processing, and prediction techniques for providing predictive insights within large data prediction domains. The techniques may include generating, using a plurality of source tables for a prediction domain, a global graph for the prediction domain. The techniques may include generating, using a graph-based machine learning model, a plurality of node-level weights for the plurality of graph nodes based on a plurality of node attributes corresponding to the plurality of graph nodes. The techniques may include generating, using the graph-based machine learning model, a plurality of semantic-level weights for the plurality of weighted edges based on a designated predictive task for the global graph. The techniques may include generating plurality of graph node embeddings and initiating the performance of the designated predictive task based on the plurality of graph node embeddings.

The invention relates to a large-scale graph data processing method and system based on multi-view graph Mama, and the method comprises the steps: firstly, learning the attribute characteristics of a node through a multilayer perceptron; capturing topological structure information among the nodes by using a graph convolutional network; characteristic representations of sub-graph levels are obtained from global connectivity and community structures through random walk. Thirdly, integrating the features obtained from each view angle through a gating fusion mechanism; then, the fused multi-view features are directly input into a state space model, a node sequence is processed through a selective state space mechanism, and global graph information is automatically aggregated; finally, specific downstream tasks, such as node classification, link prediction or graph classification, can be completed through a simple multi-layer perceptron. By means of the design, the model can make full use of various inherent characteristics of the graph, the calculation requirement of a large-scale graph data set can be met, and the performance and generalization ability of the model are effectively improved.

The invention relates to the technical field of document processing, and discloses a document content restoration method and system based on artificial intelligence, and the method comprises the steps: collecting document image information as an input source through a camera, and then carrying out the page scanning of the preprocessed document image information through a multi-modal layout analysis model, positioning various layout elements, obtaining physical layout parameters, deducing a geometric constraint relation through a geometric constraint algorithm, and generating a structured layout model; carrying out multi-language character recognition, picture graph classification and metadata extraction, formula format conversion and table structure detection and reconstruction, and outputting corresponding structured data; distinguishing different semantic blocks and constructing a tree hierarchical structure by combining a context semantic encoder with a conditional random field and text features, and extracting and restoring text fonts and paragraph format styles; and reading the basic metadata of the document, and restoring and complementing the associated information of the document elements based on the large model to complete the information improvement work. And efficient and accurate restoration of the document content is realized.

The embodiment of the invention provides a graph classification model training method and device, and the method comprises the steps: obtaining a description text corresponding to a target node in a target sub-graph, and the node information description in the description text comprises special lexical elements indicating the target node and neighbor nodes thereof; and performing graph embedding processing on the target sub-graph by using a graph neural network to obtain node representations corresponding to the target node and the neighbor nodes thereof. And inputting the description text and each node representation into a target large model, enabling the target large model to encode each lexical element in the description text, adding each node representation to a special lexical element position indicating a corresponding node to obtain a text representation, and performing first response prediction based on the text representation. A first loss is calculated based on a first probability for each first tag lexical element in the first tag lexical element sequence in the first response prediction. Parameters of the graph neural network and the target large model are updated according to the comprehensive loss, and the comprehensive loss comprises the first loss.

The invention discloses a graph neural network training method, system and device for multi-label graph classification and a medium, and belongs to the technical field of graph neural networks, and the method comprises the steps: constructing a local private graph data set of a client in federated learning; constructing a local model according to the local private graph data set, and performing fusion processing on the local model to obtain local model parameters; aggregating the local model parameters of the client to obtain a global model; and classifying the multi-label graph according to the global model. The problems that in the prior art, feature learning is not comprehensive, the universality of the graph neural network is insufficient, and a label cannot guide training of the graph neural network in a targeted mode are solved.

The invention discloses a software row-level defect prediction method based on a hierarchical attention mechanism, and the method comprises the steps: converting each row of statements in a source code file into vector representation through employing a CodeBERT pre-training model, so as to obtain the semantic information of a code row; extracting a program dependency graph of the source code, recording a mapping relation between each node in the program dependency graph and a source code line, and calculating a network measurement index corresponding to each node; fusing the semantic vectors of the code lines and the network measurement indexes to generate fused node features so as to construct a program dependency graph after the node features are expanded; constructing a HAGLineDP model, wherein the HAGLineDP model comprises a graph feature extraction network and a dual-task classification network; the graph feature extraction network designs three-level feature evolution paths from local to global and from structure to semantic, and sequentially comprises a local structure aggregation layer, a structure generalization enhancement layer and a semantic importance weighting layer; the dual-task classification network comprises a node classification path and a graph classification path; and training the HAGLineDP model by using a joint loss function and an AdamW optimizer, outputting the defect probability of each code line for a to-be-predicted source code file through a node classification path, and sorting according to the defect probability to locate a high-risk code line. According to the method, fine positioning of code defects can be effectively completed, and efficient distribution and utilization of software testing resources are promoted.

The invention discloses an agent self-evolution training method and system based on codes and fused with graph extension. The method comprises the following steps: firstly, automatically deducing and generating a code-form reward function from an expert demonstration track through a reverse process to serve as a label function sequence; then, constructing an initial strategy graph based on the sequence, and initializing an intelligent agent and a task pool by utilizing an expert track; the intelligent agent samples a new track in an expanded task pool, after strategy graph classification, a new label function is reversely generated to expand a strategy graph, meanwhile, the task pool is expanded, and expanded data is used for supervising and finely adjusting intelligent agent parameters; finally, the trained intelligent agent can take the current environment state as input during reasoning, the optimal action is output, the action directly acts on the environment, and therefore the intelligent agent is pushed to complete the expected task target.

The present invention relates to a graph representation learning device comprising: a feature extraction unit for extracting graph-based features from an input graph; a partitioning unit which uses a pre-extracted graph-based feature and a pre-trained prediction model so as to select a community detection algorithm having the highest accuracy of a graph representation learning (GRL) model, and which applies the selected community detection algorithm so as to partition the input graph into a plurality of subgraphs; a global graph configuration unit which classifies the partitioned subgraphs into a major community and a minor community according to a predetermined node size, and which unifies the subgraphs classified into the major community and the subgraphs classified into the minor community according to a preset method so as to form a major global graph and a minor global graph; and a graph representation learning unit which learns a major graph representation from the major global graph and learns a minor graph representation from the minor global graph through the GRL model.

For the auxiliary diagnosis task of Alzheimer's disease, an Alzheimer's disease classification method based on specific brain region multi-relation reasoning is disclosed. Fully learning the multi-relation perception representation of the disease-related region in sMRI image, including spatial relationship and topological information, is the key to improve the accuracy of auxiliary diagnosis. The invention regards the distinction of disease state as a graph classification problem, and constructs a spatial graph and a semantic graph. A hollow convolution module is designed to learn specific region representation, and a multi-relation graph convolution module is used to capture various types of brain region relationship. Global reasoning is performed on the learned graph structure to select discriminative information and generate global representation for diagnosis. Based on the advantages of deep learning and the characteristics of sMRI image, the Alzheimer's disease classification network based on specific brain region multi-relation reasoning is designed, which has broad application prospects in medical image analysis, Alzheimer's disease auxiliary diagnosis and other aspects.

A compound screening method based on contrast learning adaptive edge attribute enhancement, comprising the following steps: 1) obtaining data and performing topological processing and attribute processing; 2) constructing AE-Learner, jointly analyzing subgraph network topology and original graph network attributes, assigning comprehensive edge weights to molecular graphs, and generating enhanced views; 3) training the AE encoder, and maximizing the difference between different enhanced views through the AE learning target; 4) in the same iteration process, training the GCL encoder, and maximizing the similarity between positive pairs in graph contrast learning through the GCL target; 5) repeating steps 3 and 4 until the GCL target loss tends to be stable; 6) applying the trained GCL encoder to the downstream compound screening task, and performing graph classification and property prediction through the support vector machine based on the encoded graph-level representation. The application provides an enhancement method of explicit edge learning, which combines contrast learning, solves the label scarcity problem, supplements the importance of edges, and improves the performance of compound screening.

The application provides a reference type analysis method and system based on joint multi-task learning, and relates to the technical field of semantic classification. The method comprises the following steps: configuring input data and label data of a reference intention classification task; defining a reference evaluation classification task, configuring input data and label data of the reference evaluation classification task; constructing a joint multi-task learning framework, embedding the reference intention classification task and the reference evaluation classification task as parallel classification tasks into the joint multi-task learning framework; adopting an alternating iteration strategy of the parallel classification tasks, training the joint multi-task learning framework, balancing gradient propagation between the parallel tasks, obtaining a reference type joint analysis model, and performing reference type analysis processing. Through the application, the technical problem that the deep semantic correlation between the reference intention and the reference evaluation cannot be effectively mined in the prior art, thereby affecting the intelligent analysis capability of scientific research content, can be solved, and the technical effect of improving the interpretation capability of the citation semantics can be achieved.