A paper text classification method and system based on graph neural network

Abstract

The present invention discloses a method, system, electronic device and computer-readable storage medium for paper text classification based on graph neural network, which belongs to the field of text classification technology. The method comprises an edge generator, an adaptive label propagator and a GNN classifier. The edge generator is used to obtain graph structure data constructed based on labeled paper text data and perform noise reduction processing on the graph structure data. The adaptive label propagator is used to obtain the graph structure data after noise reduction processing and perform densification processing on the graph structure data after noise reduction processing. The GNN classifier is used to obtain the graph structure data after densification processing to train the GNN classifier and obtain the graph structure data constructed based on the paper text data to be classified for classification. The accuracy and robustness of paper text classification are improved, and the problem of "inaccurate and poor robustness of paper text classification caused by sparse labels and high noise in paper text data" existing in the prior art is solved.

Description

Technical Field

[0001] The present application relates to the technical field of text classification, and in particular to a method and system for paper text classification based on graph neural networks. Background Art

[0002] The statements in this section merely mention the background technology related to this application and do not necessarily constitute the prior art.

[0003] Paper-related data is complex and diverse. Each paper has different content, authors, abstracts, and other information. However, it is difficult to obtain data with relevant labels, and the obtained labels are sparse and may contain certain noise. Therefore, classifying papers is challenging and requires solving the problems of sparse and noisy labels.

[0004] In the existing technology, the solution to the shortage of labels is usually to use semi-supervised learning. The working principle of semi-supervised learning is to predict a large number of unlabeled samples based on a small number of labels. Semi-supervised learning algorithms can be roughly divided into three types. First, the collaborative learning method is to complete the learning of unlabeled samples through the collaboration of two or more learners. Representative methods include joint training and three-level training. Second, the boundary-based method finds an optimal decision boundary in the sample to maximize the margin between the boundary and the samples on each side. The main methods include semi-supervised SVM (S3VM), mean S3VM and safe S3VM (S4VM). Third, the graph-based method has a flow shape for the entire sample set, and the sample label changes as the flow shape changes. The main algorithms are GraphSAGE and Co-teaching. The above algorithms work well in the case of insufficient labels, but do not work well in noisy labels.

[0005] There has been some research on learning from noisy labels. This noise typically originates from annotation errors by human annotators and subjectivity by observers. Existing solutions include loss correction and sample selection. Loss correction methods perform loss correction on noisy labeled training samples. Goldberger et al. proposed an additional softmax layer to model the noise to correct the loss, where the softmax layer concatenates the noise and the correct labels. Patrini et al. used two forward correction and backward correction methods to correct the corrupted label matrix. However, when the number of classes is large, the noise transfer matrix of these methods is difficult to estimate. Sample selection methods involve finding clean samples through training. For example, Mentornet et al. improved the generalization of the model by training a pre-trained teacher network on samples with reweighted noisy labels during training of a student network. Han et al. proposed joint teaching to train two neural networks to learn in a given mini-batch and select samples with smaller loss. Existing loss correction and sample selection methods are used to learn from noisy labels, but these methods cannot directly learn from noisy labels.

[0006] In addition to the above methods, there are some other deep learning methods to deal with noisy labels. For example, Li et al. proposed a unified framework using "edge" information, which leverages a small clean dataset and knowledge graph label relationships to mitigate the impact of noisy labels. Veit et al. reduced the noise on a large-scale dataset by pre-training on large-scale noisy data and then fine-tuning using small-scale clean data. Tanaka et al. proposed a joint optimization framework to simultaneously update network parameters and estimate ground truth labels during training. Ren Mengjuan et al. used a meta-learning algorithm to perform meta-gradient descent on small batches of example weights to redistribute weights to training examples. Rodrigues et al. added a universal crowd layer after the output layer and trained a deep neural network on noisy labels using only backpropagation. However, all methods require additional resources or more complex networks. Summary of the Invention

[0007] In order to address the shortcomings of the existing technology, the present application provides a paper text classification system, method, electronic device and computer-readable storage medium based on graph neural network, which is conducive to better propagation of effective labels, improves the connection density between graph nodes, reduces the impact of noise labels, and increases the number of labels in training data, thereby improving the accuracy and robustness of paper text classification.

[0008] In the first aspect, this application provides a paper text classification system based on graph neural network;

[0009] A paper text classification system based on graph neural network, including:

[0010] An edge generator, configured to obtain graph structure data constructed based on labeled paper text data and perform noise reduction on the graph structure data;

[0011] Adaptive label propagator, used to obtain the graph structure data after denoising and perform densification on the graph structure data after denoising;

[0012] The GNN classifier is used to obtain the densified graph structure data to train the GNN classifier and obtain the graph structure data constructed based on the paper text data to be classified for classification.

[0013] Secondly, this application provides a paper text classification method based on graph neural network;

[0014] A paper text classification method based on graph neural network, including:

[0015] Obtain paper text data, and obtain graph structure data based on the paper text data and its corresponding label values;

[0016] Inputting the graph structure data into an edge generator to obtain the denoised graph structure data;

[0017] Input the denoised graph structure data into the adaptive label propagator to obtain the densified graph structure data;

[0018] Input the densified graph structure data into the GNN classifier to train the GNN classifier;

[0019] Input the paper text data to be classified into the trained GNN classifier to obtain the paper text classification results.

[0020] In a third aspect, the present application provides an electronic device;

[0021] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of the above-mentioned paper text classification method based on graph neural network are completed.

[0022] In a fourth aspect, the present application provides a computer-readable storage medium;

[0023] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-mentioned graph neural network-based paper text classification method.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] First, to address the problem of noisy labels in paper text data, an edge generator is used to generate embedded representations of graph nodes. If the feature similarity between two nodes is low, the weight of the edge between the two nodes is reduced, thereby reducing the impact of noisy labels. If there is no edge connection between graph nodes with high feature similarity generated by the edge generator, a new edge is generated to transmit information between the nodes, allowing valid labels to be better propagated to influence unlabeled labels. This not only increases the connection density between nodes but also further reduces the impact of noisy labels. By increasing edges and reducing the weight of noisy edges through the edge generator, the topology of the graph-structured data becomes relatively compact. Second, to address the problem of sparse labels in paper text data, an adaptive label propagator assigns a pseudo-label to unlabeled nodes to increase the number of labels in the training data. This improves the robustness and accuracy of paper text data classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings in the specification, which constitute a part of this application, are used to provide further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute improper limitations on this application.

[0027] Figure 1 A schematic diagram of the system architecture provided in an embodiment of the present application;

[0028] Figure 2 A schematic diagram of the process provided in the embodiment of the present application. DETAILED DESCRIPTION

[0029] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by those skilled in the art to which this application belongs.

[0030] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0032] Example 1

[0033] like Figure 1 As shown, this application provides a paper text classification system based on graph neural network.

[0034] A paper text classification system based on graph neural network, including:

[0035] The edge generator is used to obtain graph structure data constructed based on the paper text data and perform noise reduction on the graph structure data;

[0036] The accuracy of machine learning models is log-linearly proportional to the amount of training data. The presence of noisy labels in the training dataset can lead to reduced performance of graph neural networks on classification problems. First, graph neural networks will overfit to noisy labels, resulting in poor generalization. Noisy labels, as an extension of graph neural networks, can also lead to performance degradation. Second, for graph-structured data, the network topology amplifies the impact of noisy information, propagating noise to adjacent unlabeled nodes. Therefore, edge generators are needed to denoise the graph-structured data acquired from paper text data.

[0037] Specifically, the edge generator is configured as follows:

[0038] If a graph is noisy and contains few edges, adding edges to the original graph is a common method for adding directed edges and removing noise. The added edges result in a relatively tighter connection and fewer noisy edges. The construction of this new graph depends on the similarity between the graph nodes. Generally speaking, two nodes connected by noisy edges have different characteristics, while two nodes with similar characteristics are likely to have similar labels, providing a basis for judgment.

[0039] Since graph neural networks aggregate information from adjacent graph nodes, which will make the impact of noise information greater, MLP is used to learn the representation of graph nodes, z i =MLP(x i ), the weight formula between directed edges is specifically expressed as:

[0040]

[0041] Where, the activation function is used to keep the gradient from vanishing, and z is the learned node representation. If the edge generator can effectively reconstruct the adjacency matrix A, then it can predict the missing links well.

[0042] Therefore, based on existing work, we use adjacency matrix reconstruction as a loss. However, most elements in A are 0, which may dominate the loss function and make the edge generator only output 0. To avoid this, we use negative sampling, that is, for each positive sample a ij= 1, we randomly sample nodes with disconnected nodes as negative samples. In the case of negative sampling, the loss function can be formally written as:

[0043]

[0044] Among them, θ e is η e Parameter set, W is the weight of the directed edge, N(v i ) is the node v i Neighbors, is the node v in the graph i Distribution of unconnected nodes.

[0045] Adaptive label propagator, used to obtain the graph structure data after denoising and perform densification on the graph structure data after denoising;

[0046] In sparsely labeled graph data, erroneous information caused by noise is almost impossible to correct, and even if the label information is correct, these noisy labels will not be affected. In addition, when the graph nodes are limited, it is almost impossible for unlabeled graph nodes to aggregate information from labeled graph nodes. Due to the small number of labels in the paper data, it is necessary to increase the number of labels in the graph data constructed based on the paper text data through an adaptive label propagator to better train the GNN classifier.

[0047] Adaptive label propagator, configured to:

[0048] Since the graph structure data processed by the edge generator becomes more compact and contains fewer noisy edges, this provides the premise for label propagation. Labels are propagated through directed edges between graph nodes. The larger the weight on the directed edge, the more similar the two graph nodes are, and the easier it is for labels to propagate. Specifically, the goal of label propagation is to find a prediction matrix that is consistent with the label matrix. The specific formula is as follows:

[0049]

[0050] Where, Y(0)=Y, K represents the number of power iteration steps, is the predicted label, and the transposed matrix is denoted by T, which can be set to the normalized adjacency matrix. After propagating the label K times, the predicted label can obtain the neighborhood label information of K hops away. By using the teleportation probability, the smoothed label does not lose focus on the local domain. Most existing label propagation algorithms cannot adaptively balance the different neighborhoods of different nodes. To this end, we design an adaptive label propagation algorithm, which can be expressed as follows:

[0051]

[0052] where γ ik represents the influence of k-hop neighborhood information on the predicted label, γ ik It can be expressed as:

[0053]

[0054] in The attention vector W is the weight matrix, and ReLU is the activation function. The adaptive label propagation operator adjusts its own propagation strategy for each node by setting the attention vector and weight matrix as learnable parameters. The final smooth label can capture the rich structural information in the input graph.

[0055] The objective function of adaptive label propagation is as follows:

[0056]

[0057] in, is the node v i The prediction of y i is the original label, l() is the cross entropy loss.

[0058] The GNN classifier is used to obtain the densified graph structure data to train the GNN classifier and obtain the graph structure data constructed based on the paper text data to be classified for classification.

[0059] Specifically, the GNN classifier is configured as:

[0060] Most existing graph convolution operations use neighborhood aggregation (or message passing) methods to learn graph node representations by propagating neighborhood representations and then applying transformations. The l-th layer convolution of general graph structured data can be described as:

[0061]

[0062]

[0063] in represents the node i in the first layer, Initialized to node feature x i In order to extract the information of graph nodes in graph structure data, a k-nearest neighbor graph (knn graph) G is constructed using the feature matrix X of the graph nodes. f =(A f ,X), where A f Is the adjacency matrix of the kNN graph. First calculate the similarity matrix M∈R between n nodes n×n , and then use the cosine distance to select the similarity matrix M, and then set the edge of each node by selecting k similar nodes to obtain the adjacency matrix. The cosine distance can be expressed as:

[0064]

[0065] where x i and x j It is v i and v j Then, for the input of the feature structure, the output of the lth layer can be expressed as:

[0066]

[0067] Where W (l) is the weight matrix of the first layer of GCN, ReLU is the activation function, the initial yes The diagonal matrix, S F is the embedding representation of the last layer output.

[0068] When the paper text data to be classified is input into the GNN classifier, equation (11) is used for the classification task, Represents the classification prediction of n nodes, and is calculated as the probability that the node belongs to class c The calculation is as follows:

[0069]

[0070] in, is the normalization formula for all classes, and b is the bias vector.

[0071] The dataset used for training is denoted as L and the true label of the node is denoted as Y l , the predicted label is The loss is expressed as:

[0072]

[0073] Combining equation (2) and equation (6), the final objective function is as follows:

[0074]

[0075] where β and is a hyperparameter for the balanced edge generator and the adaptive label propagation operator.

[0076] Next, simulation experiments demonstrate the effectiveness of our graph neural network-based paper text classification system on classification tasks. We use accuracy as a metric for classification results. We demonstrate the role and necessity of each component of the model, including the edge generator and adaptive label propagator. Across three public datasets, 500 nodes were used for validation, 1000 nodes for testing, and the remaining nodes for training. A subset of nodes in the training set were labeled: 3.6% for CiteSeer, 5.2% for Cora, and 0.3% for PubMed. See Table 1 for dataset details.

[0077]

[0078] Table 1

[0079] We compare our LNS-GNN approach with the following methods, which are typical noisy learning methods:

[0080] GCN: GCN is a semi-supervised learning method. It is a variant of convolutional neural network.

[0081] Co-teaching: Co-teaching is to train two neural networks, each neural network selects the sample with smaller loss to update its corresponding network.

[0082] Self-Training: Self-Training is the first semi-supervised training model that introduces pseudo labels for training and has achieved good results.

[0083] GAT: GAT is a graph neural network that introduces an attention mechanism to assign different weights to different nodes by focusing on a hierarchy of neighborhood features.

[0084] RLVECN: RLVECN is a unique hybrid model that utilizes edge sampling techniques to learn the context of each node and its connected nodes for node classification.

[0085] Decoupling: Decoupling trains two neural network classifiers and updates the networks when there is a disagreement between them.

[0086] UnionNET-GCN: UnionNET-GCN is a network that estimates node class distribution through label aggregation to guide sample re-

[0087] New weighting and label correction framework.

[0088] The classification accuracy of different baselines under two noise conditions, symmetrical noise and paired noise, is compared. The classification accuracy results of different baselines under different noise interferences are shown in Table 2.

[0089] Table 2 Experimental results (%)

[0090]

[0091] GCN performs poorly on both types of noisy graphs. In contrast, self-training with pseudo-labels achieves better classification results, demonstrating that pseudo-labels provide effective information for learning embeddings. Using pseudo-labels on noisy and sparsely labeled graphs improves model performance. Compared to self-training with pseudo-labels, LNS-GCN achieves superior performance. This is because LNS-GCN uses an adaptive label propagator and edge generator, which adds useful edges and removes noisy edges. The adaptive label propagator propagates labels across the modified graph, generating higher-quality pseudo-labels for unlabeled nodes. Co-teaching, a loss correction-based method, offers limited benefits for semi-supervised learning with small datasets. LNS-GCN's adaptive label propagator is not limited by dataset size, allowing labels to propagate across the graph. Compared to self-training with pseudo-labels, LNS-GCN achieves superior performance. By increasing connections between nodes, it propagates effective pseudo-labels, influencing neighboring nodes while reducing the impact of noisy labels on unlabeled nodes. UnionNET-GCN is a competitive baseline because it can simultaneously train GNNs and perform label correction. However, due to the optimized edge generator, LNS-GCN outperforms UnionNET-GCN on various classification results.

[0092] Example 2

[0093] like Figure 2 As shown, this embodiment provides a paper text classification method based on graph neural network, which is based on the paper text classification system based on graph neural network described in Example 1.

[0094] A paper text classification method based on graph neural network, including:

[0095] S1. Obtain paper text data, and based on the paper text data and its corresponding label values, obtain graph structure data; wherein the graph structure data includes a graph node set and a graph directed edge set;

[0096] S2. Input the graph structure data into the edge generator to obtain the denoised graph structure data; specifically, it includes:

[0097] S201. Obtain a directed edge weight matrix based on the graph node set and the graph directed edge set of the graph structure data; wherein the weight formula of the directed edge is specifically expressed as:

[0098]

[0099] Among them, f is the activation function used to keep the gradient from disappearing, and z is the learned node representation.

[0100] S202. Obtain a negative sample set with disconnected nodes based on the graph node set; wherein most elements in the adjacency matrix A are 0, and nodes with disconnected nodes are randomly sampled as negative samples.

[0101] S203, according to the directed edge weight matrix and the negative sample set, obtain the adjacency matrix. Among them, the adjacency matrix is reconstructed through the loss function, and the loss function is

[0102]

[0103] Among them, θ e is η e Parameter set, W is the weight of the directed edge, N(v i ) is the node v i Neighbors, is the node v in the graph i Distribution of unconnected nodes.

[0104] S3. Input the denoised graph structure data into the adaptive label propagator to obtain the densified graph structure data. Specifically, it includes:

[0105] S301. According to the directed edge weight matrix and the neighborhood label information, the influence degree of the neighborhood label information on the predicted label is obtained; wherein the influence degree of the neighborhood label information on the predicted label is

[0106]

[0107] in, Note that the vector W is the weight matrix and ReLU is the activation function.

[0108] S302, according to the graph node set in the graph structure data, the original label information and the influence of the neighborhood label information on the predicted label, obtain the predicted label. Among them, the predicted label is obtained by the adaptive label propagation algorithm, and the formula is

[0109]

[0110] Among them, γ ik Represents the influence of k-hop neighborhood information on the predicted label;

[0111] The objective function of adaptive label propagation is

[0112]

[0113] in, is the node v i The prediction of y i is the original label, l() is the cross entropy loss.

[0114] S4. Input the densified graph structure data into the GNN classifier to train the GNN classifier; specifically, the following steps are performed:

[0115] S401, obtaining the densified graph structure data, and obtaining the characteristic matrix of the graph node set of the graph structure data;

[0116] S402, according to the feature matrix and the graph node set, obtain the adjacency matrix; wherein, in order to extract the information of the nodes in the graph, a k-nearest neighbor graph (knn graph) G is constructed using the node feature matrix X f =(A f ,X), where A f Is the adjacency matrix of the kNN graph. First calculate the similarity matrix M∈R between n nodes n×n , and then use the cosine distance to select the similarity matrix M, and then set the edge of each node by selecting k similar nodes to obtain the adjacency matrix. The cosine distance can be expressed as:

[0117]

[0118] where x i and x j It is v i and v j Then, for the input of the feature structure, the output of the lth layer can be expressed as:

[0119]

[0120] Where W (l) is the weight matrix of the first layer of GCN, ReLU is the activation function, the initial yes The diagonal matrix, S F is the embedding representation of the last layer output.

[0121] S403, obtain the classification result according to the feature matrix and the adjacency matrix. Represents the classification prediction of n nodes, and is calculated as the probability that the node belongs to class c The calculation is as follows:

[0122]

[0123] in is the normalization formula for all classes, and b is the bias vector.

[0124] S5. Input the paper text data to be classified into the trained GNN classifier to obtain the paper text classification results; specifically including:

[0125] S501. Obtain graph structure data based on the paper text data to be classified and its corresponding label information;

[0126] S502: Obtain a classification prediction result of the graph node according to the real label information and the predicted label information of the graph node.

[0127] S6. Obtain the accuracy of the graph node classification prediction result based on the original label information and the predicted label information of the graph node.

[0128] Example 3

[0129] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are run by the processor, the steps of the above-mentioned paper text classification method based on graph neural network are completed.

[0130] Example 4

[0131] Embodiment 4 of the present invention provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the above-mentioned graph neural network-based paper text classification method are completed.

[0132] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0133] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide the functions for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0135] The descriptions of the various embodiments in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0136] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A paper text classification system based on graph neural network, characterized by: include: An edge generator, configured to obtain graph structure data constructed based on labeled paper text data and perform noise reduction on the graph structure data; The graph structure data after denoising includes: Obtaining a directed edge weight matrix according to a graph node set and a graph directed edge set of the graph structure data; According to the graph node set, obtaining a negative sample set with non-connected nodes; Obtain the adjacency matrix based on the directed edge weight matrix and the negative sample set; The adaptive label propagator is used to obtain the graph structure data after noise reduction and perform densification on the graph structure data after noise reduction. The densified graph structure data includes: According to the directed edge weight matrix and neighborhood label information, the influence of neighborhood label information on the predicted label is obtained; Obtain the predicted label based on the influence of the graph node set, original label information and neighborhood label information on the predicted label in the graph structure data; A GNN classifier is used to obtain the densified graph structure data to train the GNN classifier and obtain the graph structure data constructed based on the paper text data to be classified for classification; wherein, inputting the densified graph structure data into the GNN classifier to train the GNN classifier includes: Obtain the densified graph structure data and obtain the feature matrix of the graph node set of the graph structure data; Obtain the adjacency matrix based on the feature matrix and the graph node set; Obtain classification results based on the feature matrix and adjacency matrix.

2. A graph neural network-based paper text classification method based on the graph neural network-based text classification system according to claim 1, characterized in that: include: Obtain paper text data, and obtain graph structure data based on the paper text data and its corresponding label values; Inputting the graph structure data into an edge generator to obtain the denoised graph structure data; Input the denoised graph structure data into the adaptive label propagator to obtain the densified graph structure data; Input the densified graph structure data into the GNN classifier to train the GNN classifier; Input the paper text data to be classified into the trained GNN classifier to obtain the paper text classification results.

3. The paper text classification method based on graph neural network as claimed in claim 2 is characterized in that include: Based on the original label information and predicted label information of the graph node, the accuracy of the graph node classification prediction result is obtained.

4. The paper text classification method based on graph neural network as claimed in claim 2 is characterized in that: The graph structure data includes a graph node set and a graph directed edge set.

5. The paper text classification method based on graph neural network as claimed in claim 2 is characterized in that: The step of inputting the paper text data to be classified into the trained GNN classifier to obtain the paper text classification result includes: Obtain graph structure data based on the paper text data to be classified and its corresponding label information; Obtain the classification prediction results of the graph nodes based on the true label information and predicted label information of the graph nodes.

6. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein the steps described in any one of claims 2 to 5 are completed when the computer instructions are executed by the processor.

7. A computer-readable storage medium, characterized in that Used to store computer instructions, which, when executed by a processor, complete the steps described in any one of claims 2 to 5.

Images