Paper classification method and apparatus, storage medium, and electronic device

By constructing topological diagrams of papers, citing papers, authors and institutions and using graph neural network models, the problem of insufficient classification accuracy of papers in the existing technology is solved, and the classification accuracy of papers is improved.

WO2024229974A1PCT designated stage expired Publication Date: 2024-11-14ZHEJIANG LAB
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2023/107529
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-10
Filing Date
2023-07-14
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

Existing paper classification methods are prone to classification errors when processing complex text content, which reduces the accuracy of paper classification, especially because papers may involve keywords on multiple topics, resulting in insufficient accuracy in classification by existing machine learning models. .

Method used

Using the paper classification method based on graph neural network, the topological graphs including papers, cited papers, authors and institutions are constructed, and the pre-trained paper category prediction model is used to generate each node's The final feature, thereby improving the classification accuracy of the paper.

Benefits of technology

By comprehensively considering the relationship between the paper and the cited paper, the author and the institution, the accuracy of the feature expression of the paper nodes is improved, thereby improving the accuracy of the paper classification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2023107529_14112024_PF_FP_ABST
    Figure CN2023107529_14112024_PF_FP_ABST
Patent Text Reader

Abstract

The present disclosure discloses a paper classification method and apparatus, a storage medium, and an electronic device. According to embodiments of the present disclosure, by means of a paper category prediction model, for each node in a topological graph comprising association information of a paper, a preset number of feature adjustments are performed on an initial node feature of the node on the basis of a node feature of the node, a node feature of a neighbor node and an attention weight between the node and the neighbor node, to obtain a final node feature of the node, and the category of the paper is predicted on the basis of the final node feature of a paper node. The association information comprises the text of the paper, the text of a cited paper, the author of the paper, the institution of the author and the like.
Need to check novelty before this filing date? Find Prior Art

Description

A paper classification method, device, storage medium and electronic device

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This disclosure claims priority to Chinese patent application No. 202310536784.9, filed on May 10, 2023, which is incorporated herein by reference. Technical Field

[0003] The present disclosure relates to the field of graph neural network technology, and in particular to a paper classification method, device, storage medium and electronic device. Background Art

[0004] With the development of science and technology, the number of papers published each year has increased exponentially. Most of the thematic categories of papers are manually determined by the authors. Machine learning models for classifying paper topics can not only alleviate the enormous burden of manual topic labeling, but can also be used to classify the large number of papers with unlabeled topic categories, thereby improving the search and organization of academic papers.

[0005] In the existing technology, the text content of the paper can be matched with keywords in the text content to classify the subject categories of the paper. However, since the text content of the paper is relatively complex and may involve keywords from multiple topics, the existing paper classification method may cause classification errors, thereby reducing the accuracy of paper classification.

[0006] Therefore, the present disclosure provides a paper classification method based on a graph neural network model to improve the accuracy of paper classification.

[0007] Summary of the Invention

[0008] The embodiments of the present disclosure provide a paper classification method, device, storage medium and electronic device to partially solve the problems existing in the above-mentioned prior art.

[0009] The embodiments of the present disclosure adopt the following technical solutions:

[0010] The present disclosure provides a paper classification method, comprising:

[0011] Obtaining related information of the paper to be classified, wherein the related information includes: the text of the paper, the text of the papers cited by the paper, the author of the paper, and the institution to which the author of the paper belongs;

[0012] Based on the association information, a topological graph is constructed, wherein the topological graph includes paper nodes, citing paper nodes, author nodes, and institution nodes, and each node in the topological graph corresponds to an initial node feature;

[0013] The topological map is input into a pre-trained paper category prediction model to determine, for each node in the topological map, neighbor nodes connected to the node through edges through the paper category prediction model; for each feature adjustment of the node, the node features of the node under the feature adjustment are adjusted according to the node features of the neighbor nodes under the feature adjustment and the attention weights between the node and the neighbor nodes under the feature adjustment, to obtain the adjusted features of the node, which are used as the node features of the node under the next feature adjustment; and based on the adjusted features of the node and the node features of the neighbor nodes under the feature adjustment, the attention weights between the node and the neighbor nodes under the feature adjustment are adjusted, to obtain the adjusted attention weights, which are used as the attention weights between the node and the neighbor nodes under the next feature adjustment, and the node features of the node are continuously adjusted for the next feature adjustment of the node until the number of feature adjustments reaches a preset number; the initial node features of the node are adjusted according to the adjusted features of the node obtained under each feature adjustment of the node, to obtain the final node features of the node;

[0014] The paper category prediction model is used to predict the category of the paper based on the final node features of the paper nodes in the topological graph.

[0015] Optionally, constructing a topology map based on the association information specifically includes:

[0016] A topological graph is constructed using the text of the paper, the text of the paper's cited papers, the paper's authors, and the institutions to which the paper's authors belong as nodes, and using the citation relationship between the paper and the cited papers, the affiliation relationship between the author and the institution, and the authorship relationship between the paper and the author as edges;

[0017] Based on the text of the paper, the initial node characteristics of the paper node in the topological diagram are determined; based on the text of the cited paper, the initial node characteristics of the cited paper node in the topological diagram are determined; based on the text of all papers written by the author, the initial node characteristics of the author node in the topological diagram are determined; based on the text of all papers written by all authors under the institution, the initial node characteristics of the institution node in the topological diagram are determined.

[0018] Optionally, adjusting the node feature of the node under the secondary feature adjustment according to the node feature of the neighbor node under the secondary feature adjustment and the attention weight between the node and the neighbor node under the secondary feature adjustment to obtain the adjusted feature of the node includes:

[0019] Determine an aggregate feature for the node based on the node features of the neighboring nodes under the secondary feature adjustment and the attention weight between the node and the neighboring nodes under the secondary feature adjustment; determine a product between a preset parameter and the node features of the node under the secondary feature adjustment to obtain a smooth feature;

[0020] The aggregated feature and the smoothed feature are added to obtain an adjusted feature of the node.

[0021] Optionally, based on the adjusted feature of the node and the node feature of the neighboring node under the feature adjustment, adjusting the attention weight between the node and the neighboring node under the feature adjustment to obtain the adjusted attention weight includes:

[0022] Processing the adjusted feature of the node and the node feature of the neighboring node under the feature adjustment to obtain a relationship feature; wherein the processing includes: one of a concatenation process, a dot product process, and an addition process;

[0023] According to the relationship feature, the attention weight between the node and the neighboring node under this feature adjustment is adjusted to obtain the adjusted attention weight.

[0024] Optionally, adjusting the attention weight between the node and the neighboring node under the feature adjustment according to the relationship feature to obtain the adjusted attention weight includes:

[0025] Inputting the relationship features into a multilayer perceptron in the paper category prediction model to process the relationship features through the multilayer perceptron to obtain adjustment parameters; wherein the multilayer perceptron includes at least: an attention layer, a pooling layer, a dropout layer, a regularization layer, and a normalization layer;

[0026] Based on the adjustment parameter, the attention weight between the node and the neighboring node under this feature adjustment is adjusted to obtain the adjusted attention weight.

[0027] Optionally, adjusting the initial node feature of the node according to the adjusted feature of the node obtained in each feature adjustment of the node to obtain the final node feature of the node specifically includes:

[0028] The adjusted features of the node obtained in each feature adjustment of the node and the initial node features of the node are weighted summed to obtain the final node features of the node.

[0029] Optionally, pre-training the paper category prediction model includes:

[0030] Obtaining sample association information of the sample paper to be classified, the sample association information including: the text of the sample paper, the text of the sample cited paper of the sample paper, the author of the sample paper, and the institution to which the author of the sample paper belongs;

[0031] Based on the sample association information, construct a topological graph for the sample paper as a sample topological graph, wherein each sample node in the sample topological graph corresponds to an initial node feature;

[0032] The sample topology graph is input into the paper category prediction model to be trained, so as to determine, for each sample node in the sample topology graph, a sample neighbor node connected to the sample node through an edge through the paper category prediction model to be trained; for each feature adjustment of the sample node, according to the node feature of the sample neighbor node under the feature adjustment and the attention weight between the sample node and the sample neighbor node under the feature adjustment, the node feature of the sample node under the feature adjustment is adjusted to obtain the adjusted feature of the sample node, which is used as the node feature of the sample node under the next feature adjustment, and based on the adjusted feature of the sample node and the node of the sample neighbor node under the feature adjustment feature, adjusting the attention weight between the sample node and the sample neighbor node under this feature adjustment to obtain the adjusted attention weight, which is used as the attention weight between the sample node and the sample neighbor node under the next feature adjustment, and continuing to adjust the node feature of the sample node for the next feature adjustment of the sample node until the number of feature adjustments reaches a preset number; adjusting the initial node feature of the sample node according to the adjusted feature of the sample node obtained under each feature adjustment of the sample node to obtain the final node feature of the sample node; predicting the category of the sample paper based on the final node feature of the sample node corresponding to the sample paper in the sample topology graph;

[0033] The paper category prediction model to be trained is trained with minimizing the difference between the category of the sample paper and the true category of the sample paper as the optimization goal.

[0034] The present disclosure provides a paper classification device, comprising:

[0035] An acquisition module is used to acquire related information of the paper to be classified, wherein the related information includes: the text of the paper, the text of the paper's cited papers, the author of the paper, and the institution to which the author of the paper belongs;

[0036] A construction module is used to construct a topological map based on the association information, wherein the topological map includes paper nodes, citing paper nodes, author nodes, and institution nodes, and each node in the topological map corresponds to an initial node feature;

[0037] A feature expression module is used to input the topological map into a pre-trained paper category prediction model, so as to determine, for each node in the topological map, a neighbor node connected to the node through an edge through the paper category prediction model; for each feature adjustment of the node, according to the node feature of the neighbor node under the feature adjustment and the attention weight between the node and the neighbor node under the feature adjustment, adjust the node feature of the node under the feature adjustment to obtain the adjusted feature of the node, which is used as the node feature of the node under the next feature adjustment; and based on the adjusted feature of the node and the node feature of the neighbor node under the feature adjustment, adjust the attention weight between the node and the neighbor node under the feature adjustment to obtain the adjusted attention weight, which is used as the attention weight between the node and the neighbor node under the next feature adjustment; continue to adjust the node feature of the node for the next feature adjustment of the node until the number of feature adjustments reaches a preset number; according to the adjusted feature of the node obtained under each feature adjustment of the node, adjust the initial node feature of the node to obtain the final node feature of the node;

[0038] The paper classification module is used to predict the category of the paper based on the final node features of the paper nodes in the topological graph through the paper category prediction model.

[0039] Optionally, the construction module is specifically used to construct a topological graph with the text of the paper, the text of the cited paper of the paper, the author of the paper, and the institution to which the author of the paper belongs as nodes, and the citation relationship between the paper and the cited paper, the affiliation relationship between the author and the institution, and the authorship relationship between the paper and the author as edges; based on the text of the paper, determine the initial node characteristics of the paper node in the topological graph; based on the text of the cited paper, determine the initial node characteristics of the cited paper node in the topological graph; based on the text of all papers written by the author, determine the initial node characteristics of the author node in the topological graph; based on the text of all papers written by all authors under the institution, determine the initial node characteristics of the institution node in the topological graph.

[0040] Optionally, the feature expression module is specifically used to process the adjusted features of the node and the node features of the neighboring nodes under the feature adjustment to obtain relationship features; wherein, the processing includes: one of splicing processing, dot multiplication processing, and addition processing; according to the relationship features, the attention weight between the node and the neighboring nodes under the feature adjustment is adjusted to obtain the adjusted attention weight.

[0041] Optionally, the feature expression module is specifically used to input the relationship features into the multi-layer perceptron in the paper category prediction model, so as to process the relationship features through the multi-layer perceptron to obtain adjustment parameters; wherein, the multi-layer perceptron includes at least: an attention layer, a pooling layer, a Dropout layer, a regularization layer and a normalization layer; based on the adjustment parameters, the attention weight between the node and the neighboring node under this feature adjustment is adjusted to obtain the adjusted attention weight.

[0042] Optionally, the feature expression module is specifically configured to perform weighted summation on the adjusted feature of the node obtained in each feature adjustment of the node and the initial node feature of the node to obtain the final node feature of the node.

[0043] Optionally, it further includes: a training module;

[0044] The training module is specifically used to obtain sample association information of sample papers to be classified, and the sample association information includes: the text of the sample paper, the text of the sample cited paper of the sample paper, the author of the sample paper, and the institution to which the author of the sample paper belongs; based on the sample association information, a topological map for the sample paper is constructed as a sample topological map, wherein each sample node in the sample topological map corresponds to an initial node feature; the sample topological map is input into the paper category prediction model to be trained, so as to determine, for each sample node in the sample topological map, a sample neighbor node connected to the sample node through an edge through the paper category prediction model to be trained; for each feature adjustment of the sample node, the node feature of the sample node under the feature adjustment is adjusted according to the node feature of the sample neighbor node under the feature adjustment and the attention weight between the sample node and the sample neighbor node under the feature adjustment, so as to obtain the adjusted feature of the sample node as the next feature adjustment. The node features of the sample node under the feature adjustment are adjusted, and based on the adjusted features of the sample node and the node features of the sample neighbor nodes under the feature adjustment, the attention weight between the sample node and the sample neighbor nodes under the feature adjustment is adjusted to obtain the adjusted attention weight, which is used as the attention weight between the sample node and the sample neighbor nodes under the next feature adjustment, and the node features of the sample node are adjusted for the next feature adjustment of the sample node until the number of feature adjustments reaches a preset number; according to the adjusted features of the sample node obtained under each feature adjustment of the sample node, the initial node features of the sample node are adjusted to obtain the final node features of the sample node; based on the final node features of the sample node corresponding to the sample paper in the sample topology graph, the category of the sample paper is predicted; with minimizing the difference between the category of the sample paper and the true category of the sample paper as the optimization goal, the paper category prediction model to be trained is trained.

[0045] The present disclosure provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned paper classification method is implemented.

[0046] The present disclosure provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned paper classification method when executing the program.

[0047] At least one of the above technical solutions adopted in the embodiments of the present disclosure can achieve the following beneficial effects:

[0048] In the disclosed embodiment, a topological map containing the association information of the paper is input into a pre-trained paper category prediction model, so that for each node in the topological map, based on the node features of the node, the node features of the neighboring nodes, and the attention weight between the node and the neighboring nodes, the paper category prediction model performs a preset number of feature adjustments on the initial node features of the node to obtain the final node features of the node, and predicts the category of the paper based on the final node features of the paper node. The association information includes: the text of the paper, the text of the cited paper, the author of the paper, the author's organization, etc. In this method, the relationship between the paper and the cited paper, the relationship between the paper and the author, and the relationship between the author and the organization are comprehensively considered to determine the final node features of the paper node, which can improve the accuracy of the feature expression of the paper node, and then classification is performed based on the final node features of the paper node to improve the accuracy of paper classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The drawings described herein are used to provide a further understanding of the present disclosure and constitute a part of the present disclosure. The exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation of the present disclosure. In the drawings:

[0050] FIG1 is a flow chart of a paper classification method provided by an embodiment of the present disclosure;

[0051] FIG2 is a schematic diagram of the model structure of the paper category prediction model provided by an embodiment of the present disclosure;

[0052] FIG3 is a schematic diagram of feature adjustment provided by an embodiment of the present disclosure;

[0053] FIG4 is a schematic diagram of the structure of a paper classification device provided by an embodiment of the present disclosure;

[0054] FIG5 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0055] The paper classification method disclosed herein aims to aggregate paper-related features using a graph neural network model to obtain final features for the paper, and then classify the paper based on the final features. When aggregating paper-related features, different attention weights can be used to aggregate the paper-related features multiple times to obtain the final features for the paper.

[0056] To make the objectives, technical solutions, and advantages of the present disclosure more clear, the technical solutions of the present disclosure will be clearly and completely described below in conjunction with the specific embodiments of the present disclosure and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present disclosure.

[0057] The technical solutions provided by various embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0058] FIG1 is a flow chart of a paper classification method provided by an embodiment of the present disclosure. The paper classification method can be applied to a server or terminal device deployed with a paper category prediction model, including:

[0059] S100: Obtaining related information of a paper to be classified, wherein the related information includes: the text of the paper, the text of the cited papers of the paper, the author of the paper, and the institution to which the author of the paper belongs.

[0060] S102: Construct a topology map based on the association information, wherein the topology map includes paper nodes, cited paper nodes, author nodes, and institution nodes, and each node in the topology map corresponds to an initial node feature.

[0061] In the disclosed embodiments, the paper category prediction model is mainly used to represent the features of the paper, and based on the feature representation of the paper, the paper's subject category is classified. The paper category prediction model can be a graph neural network model, and the subject category can refer to the subject category involved in the paper, such as mathematics, geography, physics, chemistry, etc. The subject category here refers to the category of the paper.

[0062] Before using the paper category prediction model, it is necessary to build a topological map for the paper.

[0063] Specifically, you can first obtain the associated information of the paper to be classified. Among them, the associated information of the paper refers to information related to the category of the paper. The associated information may include: the text of the paper, the text of the paper's cited papers, the author of the paper, and the institution to which the author of the paper belongs. The text of the paper can identify the category of the paper. The text of the cited paper can indirectly identify the category of the paper. The academic field of the author of the paper is generally fixed, and the category of the paper can be indirectly identified through the author's academic field. The institution to which the author belongs can refer to the company or research institute where the author works, or it can be the professional college where the author studies. The academic field of the institution to which the author belongs can indirectly identify the category of the paper.

[0064] After obtaining the associated information of the papers to be classified, a topology graph can be constructed based on the associated information of the papers to be classified. The topology graph includes nodes corresponding to the paper text, nodes corresponding to the text of the citing papers, nodes corresponding to the authors, and nodes corresponding to the institutions. Each node in the topology graph has an initial node feature. Furthermore, nodes corresponding to the paper text are used as paper nodes, nodes corresponding to the text of the citing papers are used as citing paper nodes, nodes corresponding to the authors are used as author nodes, and nodes corresponding to the institutions are used as institution nodes.

[0065] When constructing a topological graph, the text of the paper, the text of the paper's citation paper, the paper's author, and the institution to which the paper's author belongs are used as nodes, and the citation relationship between the paper and the cited paper, the affiliation relationship between the author and the institution, and the publication relationship between the paper and the author are used as edges to construct the topological graph.

[0066] In the constructed topology graph, the initial node features of the paper nodes in the topology graph can be determined based on the text of the paper. Specifically, feature extraction is performed on the text of the paper to obtain text features specific to the paper, which are used as the initial node features of the paper nodes in the topology graph.

[0067] Based on the text of the cited paper, the initial node features of the cited paper node in the topological graph can be determined. Specifically, feature extraction is performed on the text of the cited paper to obtain text features specific to the cited paper, which are used as the initial node features of the cited paper node in the topological graph.

[0068] The initial node features of the author node in the topology graph can be determined based on the text of all papers written by the author. Specifically, all papers written by the author can be identified. For each of these papers, feature extraction is performed on the text of the paper to obtain the text features of the paper. The text features of all papers are weighted and summed to obtain a comprehensive feature, which is used as the initial node feature of the author node.

[0069] The initial node features of the institution node in the topology graph can be determined based on the text of all papers written by all authors under the institution. Specifically, all authors under the institution can be determined first. The comprehensive features corresponding to each of the authors are weighted summed to obtain a fusion feature, which serves as the initial node feature of the institution node in the topology graph. The method for determining the comprehensive features of each author can include: for each paper of an author, performing feature extraction on the text of the paper to obtain the text features of the paper. The text features of all papers of this author are weighted summed to obtain the comprehensive features corresponding to the author.

[0070] S104: Input the topological map into a pre-trained paper category prediction model to determine, for each node in the topological map, neighbor nodes connected to the node through edges through the paper category prediction model; for each feature adjustment of the node, adjust the node features of the node under the feature adjustment according to the node features of the neighbor nodes under the feature adjustment and the attention weights between the node and the neighbor nodes under the feature adjustment to obtain the adjusted features of the node, which are used as the node features of the node under the next feature adjustment; and adjust the attention weights between the node and the neighbor nodes under the feature adjustment based on the adjusted features of the node and the node features of the neighbor nodes under the feature adjustment to obtain the adjusted attention weights, which are used as the attention weights between the node and the neighbor nodes under the next feature adjustment, and continue the next feature adjustment for the node until the number of feature adjustments reaches a preset number; adjust the initial node features of the node according to the adjusted features of the node obtained under each feature adjustment to obtain the final node features of the node.

[0071] S106: Predicting the category of the paper based on the final node features of the paper nodes in the topological graph using the paper category prediction model.

[0072] In an embodiment of the present disclosure, after constructing a topological graph containing paper nodes, the topological graph can be input into a pre-trained paper category prediction model. The paper category prediction model can be used to adjust the initial node features of the paper nodes to obtain the final node features of the paper nodes, and then predict the category of the paper based on the final node features of the paper nodes. The paper category prediction model can be a graph neural network model.

[0073] An embodiment of the present disclosure provides a schematic diagram of the model structure of a paper category prediction model, as shown in Figure 2. In Figure 2, the paper category prediction model may include a graph feature processing layer, a multi-layer perceptron, and a classification layer. The graph feature processing layer is used to perform feature processing on the node features of all nodes in the topology graph for a preset number of times. There may be multiple graph feature processing layers, and the topology graph in each graph feature processing layer can determine the attention weight between any two connected nodes in the topology graph through the multi-layer perceptron. Figure 2 takes a graph feature processing layer as an example. The multi-layer perceptron is used to determine the attention weight between any two connected nodes based on the node features of all nodes in the topology graph. The classification layer is used to classify the paper into paper categories based on the final node features of the paper nodes in the topology graph.

[0074] Next, a method for processing node features of all nodes in the topology graph is described.

[0075] Although only the node features of paper nodes are used when classifying papers, each node in the topological graph affects each other. Therefore, when expressing the final node features of the paper nodes, it is necessary to consider the node features of all nodes in the topological graph and the relationship features between any two connected nodes.

[0076] Therefore, after inputting the topology map into the pre-trained paper category prediction model, the paper category prediction model first determines, for each node in the topology map, the neighboring nodes connected to the node via edges, as the node's neighbor nodes. The node's neighbor nodes may be one or more. After determining the node's neighbor nodes, the node features of the node can be continuously adjusted or updated.

[0077] Using the paper category prediction model, for each feature adjustment of the node, the node features of the node under this feature adjustment are adjusted based on the node features of the neighboring nodes under this feature adjustment and the attention weights between the node and the neighboring nodes under this feature adjustment, obtaining the adjusted features of the node, which serve as the node features of the node under the next feature adjustment. Subsequently, based on the adjusted features of the node and the node features of the neighboring nodes under this feature adjustment, the attention weights between the node and the neighboring nodes under this feature adjustment are adjusted, obtaining the adjusted attention weights, which serve as the attention weights between the node and the neighboring nodes under the next feature adjustment. The node features of the node are adjusted again for the next feature adjustment of the node until the number of feature adjustments reaches a preset number. Finally, after the preset number of feature adjustments, the adjusted features of the node obtained under each feature adjustment are obtained, and the initial node features of the node are adjusted based on the adjusted features of the node obtained under each feature adjustment to obtain the final node features of the node.

[0078] After traversing all nodes in the topology graph, the final node features of all nodes in the topology graph can be obtained.

[0079] In the process of obtaining the final node features of a node, the node features of the node are adjusted multiple times, and the adjusted features after multiple adjustments are fused, which can more fully extract the key features in the node features of neighboring nodes to improve the accuracy of the feature expression of the node.

[0080] The method of adjusting the node features of the node under the secondary feature adjustment according to the node features of the neighboring nodes under the secondary feature adjustment and the attention weight between the node and the neighboring nodes under the secondary feature adjustment to obtain the adjusted features of the node may include:

[0081] Based on the node features of the neighboring nodes under this feature adjustment and the attention weights between the node and its neighbors under this feature adjustment, an aggregated feature for the node is determined. Simultaneously, the product of the preset parameters and the node features under this feature adjustment is determined to obtain a smoothed feature. Finally, the aggregated feature and the smoothed feature are added together to obtain the adjusted feature for the node, as shown in Figure 3.

[0082] Among them, the formula for calculating the adjusted features is: h′ j is the adjusted feature of the jth node in the topology graph. j represents the initial node feature of the jth node in the topology graph, h i Represents the node characteristics of the i-th neighbor node of the j-th node in the topology graph. N means that the j-th node in the topology graph has N neighbor nodes. α ij Represents the attention weight between the j-th node and the i-th neighbor node in the topology graph. Expressed as aggregate features, (1+∈)h j It is represented as a smooth feature, (1+∈) is a preset parameter, which plays a smoothing role and reduces the feature mutation caused by feature aggregation. ∈ is a constant.

[0083] In Figure 3, we take the paper node as an example and calculate the adjusted features of the paper node after one feature adjustment. Assume that the initial node feature of the paper node is h1, and the paper node has three neighbor nodes, namely neighbor node 1, neighbor node 2, and neighbor node 3. The node features of neighbor node 1 are h2, the node features of neighbor node 2 are h3, and the node features of neighbor node 3 are h4. The adjusted features of the paper node are:

[0084] When adjusting the node features of a node for the first time, the node features of the node are used as the initial node features. Based on the initial node features of the node and the node features of its neighboring nodes, the attention weights between the node and its neighboring nodes can be determined as the initial attention weights. It should be noted that the node features of the neighboring nodes may be either the initial node features of the neighboring nodes or the final node features of the neighboring nodes.

[0085] When determining the attention weights between the node and its neighboring nodes based on the initial node features of the node and the node features of its neighboring nodes, the initial node features of the node and the node features of the neighboring nodes can be processed to obtain initial relationship features. Then, based on the initial relationship features, the initial attention weights between the node and the neighboring nodes are determined. The processing includes one of concatenation, dot product, and addition.

[0086] When determining the initial attention weights between the node and its neighboring nodes based on the initial relationship features, the initial relationship features can be input into a multilayer perceptron in the paper category prediction model to output the initial attention weights between the node and its neighboring nodes. The multilayer perceptron includes at least an attention layer, a pooling layer, a dropout layer, a regularization layer, and a normalization layer.

[0087] Similar to the first adjustment of the node feature of the node, after the feature of the node is adjusted once, the attention weight between the node and the neighboring nodes can be re-determined based on the adjusted feature of the node.

[0088] That is, after adjusting the attention weights between the node and the neighboring nodes under the secondary feature adjustment based on the adjusted features of the node and the node features of the neighboring nodes under the secondary feature adjustment, the adjusted attention weights can be processed with the node features of the neighboring nodes under the secondary feature adjustment to obtain the relationship features. The processing includes one of concatenation, dot product, and addition. Then, based on the relationship features, the attention weights between the node and the neighboring nodes under the secondary feature adjustment are adjusted to obtain the adjusted attention weights.

[0089] After adjusting the attention weights between the node and neighboring nodes under the feature adjustment based on the relationship features to obtain the adjusted attention weights, the relationship features can be input into a multilayer perceptron in the paper category prediction model to process the relationship features through the multilayer perceptron to obtain adjustment parameters. Based on the adjustment parameters, the attention weights between the node and neighboring nodes under the feature adjustment can be adjusted to obtain the adjusted attention weights. The adjustment parameters can refer to parameters used to adjust the attention weights between the node and neighboring nodes under the feature adjustment.

[0090] When the attention weight between the node and the neighboring nodes under the feature adjustment is adjusted based on the adjustment parameter and the adjusted attention weight is obtained, the adjustment parameter and the attention weight between the node and the neighboring nodes under the feature adjustment can be XORed to obtain the adjusted attention weight.

[0091] Among them, the formula for calculating the adjusted attention weight is: α ij Represents the attention weight between the jth node and the ith neighbor node of the topology graph. MLP() represents the processing of multi-layer perceptron. concat(h i , h j′) represents the relationship feature obtained by concatenating the adjusted feature of the j-th node in the topological graph and the node feature of the i-th neighbor node of the node.

[0092] After a preset number of feature adjustments are performed on a node's node features, the adjusted features of the node obtained during each feature adjustment can be determined. The adjusted features obtained during each feature adjustment and the initial node features of the node are then weighted summed to obtain the final node features of the node.

[0093] For example: a secondary feature adjustment is performed on a node, with a preset number of times being 2. During the first feature adjustment, the initial node feature of the node, the node features of the neighboring nodes of the node, and the initial attention weights between the node and the neighboring nodes are first determined. Then, the node features of the neighboring nodes are aggregated to the node through the attention weights, and combined with the initial node features of the node, the adjusted features of the node are obtained as the first adjusted features. According to the first adjusted features of the node and the node features of the neighboring nodes, the initial attention weights are adjusted to obtain the adjusted attention weights between the node and the neighboring nodes. During the second feature adjustment, the node features of the neighboring nodes remain unchanged, and the adjusted features of the node are used as the node features of the node under the second feature adjustment, and the adjusted attention weights are used as the attention weights between the node and the neighboring nodes under the second feature adjustment. That is, the node features of the neighboring nodes are aggregated to the node through the adjusted attention weights, and combined with the first adjusted features of the node, the second adjusted features obtained after adjusting the first adjusted features of the node are obtained. Finally, the first adjusted feature, the second adjusted feature, and the initial node feature are weighted and summed to obtain the final node feature of the node.

[0094] After determining the final node features of each node in the topology graph, that is, after determining the final node features of the paper node, the category of the paper can be predicted based on the final node features of the paper node through the paper category prediction model.

[0095] Specifically, the final node features of the paper node are input into the classification layer to predict the category of the paper through the classification layer.

[0096] In addition, before using the paper category prediction model, the paper category prediction model needs to be trained first.

[0097] Specifically, the sample association information of the sample papers to be classified is first obtained. The sample association information includes: the text of the sample paper, the text of the sample paper's sample cited paper, the sample author of the sample paper, and the sample organization to which the author of the sample paper belongs. Then, based on the sample association information, a topology graph for the sample paper is constructed as a sample topology graph. The sample topology graph includes sample paper nodes (nodes corresponding to the text of the sample paper), sample cited paper nodes (nodes corresponding to the text of the sample cited paper), sample author nodes, sample organization nodes, and other sample nodes. Each sample node in the sample topology graph corresponds to an initial node feature.

[0098] A topological graph is constructed using the text of the sample papers, the text of the sample citing papers, the sample authors, and the sample institutions as nodes, and the citation relationships between the sample papers and the sample citing papers, the affiliation relationships between the sample authors and the sample institutions, and the authorship relationships between the sample papers and the sample authors as edges. Based on the text of the sample papers, the initial node features of the sample paper nodes in the sample topological graph are determined. Based on the text of the sample citing papers, the initial node features of the sample citing paper nodes in the sample topological graph are determined. Based on the text of all papers written by the sample authors, the initial node features of the sample author nodes in the sample topological graph are determined. Based on the text of all papers written by all authors under the sample institution, the initial node features of the sample institution nodes in the sample topological graph are determined.

[0099] The sample topology graph is input into the paper category prediction model to be trained, so that the paper category prediction model to be trained determines, for each sample node in the sample topology graph, a sample neighbor node connected to the sample node by an edge. For each feature adjustment of the sample node, the node feature of the sample node under the feature adjustment is adjusted according to the node feature of the sample neighbor node under the feature adjustment and the attention weight between the sample node and the sample neighbor node under the feature adjustment, and the adjusted feature of the sample node is obtained as the node feature of the sample node under the next feature adjustment. Based on the adjusted feature of the sample node and the node feature of the sample neighbor node under the feature adjustment, the attention weight between the sample node and the sample neighbor node under the feature adjustment is adjusted to obtain the adjusted attention weight as the attention weight between the sample node and the sample neighbor node under the next feature adjustment. The node feature of the sample node is adjusted continuously for the next feature adjustment of the sample node until the number of feature adjustments reaches a preset number. According to the adjusted features of the sample node obtained after each feature adjustment of the sample node, the initial node features of the sample node are adjusted to obtain the final node features of the sample node, and the category of the sample paper is predicted based on the final node features of the sample node corresponding to the sample paper in the sample topology graph.

[0100] The paper category prediction model to be trained is trained with the optimization goal of minimizing the difference between the category of the sample paper and the true category of the sample paper. That is, the model parameters of the paper category prediction model to be trained are adjusted.

[0101] The cross entropy loss is determined based on the difference between the category of the sample paper and the true category of the sample paper. The paper category prediction model to be trained is trained with the minimum cross entropy loss as the optimization goal.

[0102] In addition, when obtaining a sample topology map for a sample paper, there is no need to construct a sample topology map. A sub-topology map containing sample association information of the sample paper can be segmented from the existing complete topology map as the sample topology map.

[0103] Specifically, we first determine the target topology graph, then segment it into a sub-topology graph containing the sample papers to be classified, which serves as the sample topology graph. The target topology graph is constructed with the text of each paper, the text of each citing paper, each author, and each institution as nodes, and the citation relationships between each paper and each citing paper, the affiliation relationships between each author and each institution, and the authorship relationships between each paper and each author as edges.

[0104] Methods for segmenting the target topology graph include entity segmentation, relationship segmentation, neighbor sampling, and black-box segmentation. Overall segmentation goals include maintaining a balanced distribution of nodes within the sub-topology graph, ensuring that the sub-topology graph can fit into computer memory, and evenly distributing high-degree entity nodes across different sub-graphs. These goals cannot be met simultaneously in a single segmentation algorithm; the appropriate goal and algorithm should be selected based on the actual problem or experimental results.

[0105] Entity partitioning is a method that considers which subgraph entity nodes should be placed in one by one. Common algorithms include random partitioning, METIS, and LDG. Taking random partitioning as an example, the number of partitions is set, and entity nodes are evenly and randomly assigned to each partition (i.e., sub-topology graph).

[0106] Relationship partitioning is a method of placing nodes with high degrees in different sub-topology graphs as much as possible to maintain the balance of the sub-topology graphs.

[0107] Neighbor sampling is a method that starts from a single entity node and adds first-degree neighbors, second-degree neighbors, etc. to the sub-topology graph.

[0108] Black-box algorithm segmentation is to train a black-box algorithm model, such as a neural network model, to optimize the segmentation method.

[0109] It should be noted that all actions of acquiring signals, information or image data in this application are carried out in compliance with the relevant data protection laws and policies of the country where they are located and with the authorization given by the owner of the corresponding device.

[0110] It can be seen from the method shown in Figure 1 above that the present disclosure inputs a topological map containing the association information of the paper into a pre-trained paper category prediction model, so that for each node in the topological map, based on the node features of the node, the node features of the neighboring nodes, and the attention weight between the node and the neighboring nodes, the initial node features of the node are adjusted a preset number of times through the paper category prediction model to obtain the final node features of the node, and based on the final node features of the paper node, the category of the paper is predicted. Wherein, the association information includes: the text of the paper, the text of the cited paper, the author of the paper, the institution of the author, etc. In this method, while comprehensively considering the relationship between the paper and the cited paper, the relationship between the paper and the author, and the relationship between the author and the institution, the node features of other nodes are fused multiple times to adjust the node features of the paper node to determine the final node features of the paper node, which can improve the accuracy of the feature expression of the paper node, and then classify based on the final node features of the paper node to improve the accuracy of paper classification.

[0111] The above is the paper classification method provided by the embodiment of the present disclosure. Based on the same idea, the present disclosure also provides corresponding devices, storage media and electronic devices.

[0112] FIG4 is a schematic diagram of the structure of a paper classification device provided by an embodiment of the present disclosure, wherein the device includes:

[0113] The acquisition module 401 is used to acquire related information of the paper to be classified, wherein the related information includes: the text of the paper, the text of the paper's cited papers, the author of the paper, and the institution to which the author of the paper belongs;

[0114] A construction module 402 is configured to construct a topology map based on the association information, wherein the topology map includes paper nodes, citing paper nodes, author nodes, and institution nodes, and each node in the topology map corresponds to an initial node feature;

[0115] The feature expression module 403 is used to input the topological map into a pre-trained paper category prediction model, so as to determine, for each node in the topological map, a neighbor node connected to the node through an edge through the paper category prediction model; for each feature adjustment of the node, according to the node feature of the neighbor node under the feature adjustment and the attention weight between the node and the neighbor node under the feature adjustment, adjust the node feature of the node under the feature adjustment to obtain the adjusted feature of the node, which is used as the node feature of the node under the next feature adjustment; and based on the adjusted feature of the node and the node feature of the neighbor node under the feature adjustment, adjust the attention weight between the node and the neighbor node under the feature adjustment to obtain the adjusted attention weight, which is used as the attention weight between the node and the neighbor node under the next feature adjustment; continue to adjust the node feature of the node for the next feature adjustment of the node until the number of feature adjustments reaches a preset number; according to the adjusted feature of the node obtained under each feature adjustment of the node, adjust the initial node feature of the node to obtain the final node feature of the node;

[0116] The paper classification module 404 is used to predict the category of the paper based on the final node features of the paper nodes in the topological graph through the paper category prediction model.

[0117] The apparatus further includes a training module 405;

[0118] The training module 405 is used to obtain sample association information of the sample paper to be classified, and the sample association information includes: the text of the sample paper, the text of the sample cited paper of the sample paper, the author of the sample paper, and the institution to which the author of the sample paper belongs; based on the sample association information, a topological map for the sample paper is constructed as a sample topological map, wherein each sample node in the sample topological map corresponds to an initial node feature; the sample topological map is input into the paper category prediction model to be trained, so as to determine, for each sample node in the sample topological map, a sample neighbor node connected to the sample node through an edge through the paper category prediction model to be trained; for each feature adjustment of the sample node, the node feature of the sample neighbor node under the feature adjustment and the attention weight between the sample node and the sample neighbor node under the feature adjustment are adjusted to obtain the adjusted feature of the sample node as the feature for the next feature adjustment. The node features of the sample node under the feature adjustment are adjusted, and based on the adjusted features of the sample node and the node features of the sample neighbor nodes under the feature adjustment, the attention weight between the sample node and the sample neighbor nodes under the feature adjustment is adjusted to obtain the adjusted attention weight, which is used as the attention weight between the sample node and the sample neighbor nodes under the next feature adjustment, and the node features of the sample node are adjusted for the next feature adjustment of the sample node until the number of feature adjustments reaches a preset number; according to the adjusted features of the sample node obtained under each feature adjustment of the sample node, the initial node features of the sample node are adjusted to obtain the final node features of the sample node; based on the final node features of the sample node corresponding to the sample paper in the sample topology graph, the category of the sample paper is predicted; with minimizing the difference between the category of the sample paper and the true category of the sample paper as the optimization goal, the paper category prediction model to be trained is trained.

[0119] Optionally, the construction module 402 is specifically used to construct a topological graph with the text of the paper, the text of the cited paper of the paper, the author of the paper, and the institution to which the author of the paper belongs as nodes, and the citation relationship between the paper and the cited paper, the affiliation relationship between the author and the institution, and the authorship relationship between the paper and the author as edges; based on the text of the paper, determine the initial node characteristics of the paper node in the topological graph; based on the text of the cited paper, determine the initial node characteristics of the cited paper node in the topological graph; based on the text of all papers written by the author, determine the initial node characteristics of the author node in the topological graph; based on the text of all papers written by all authors under the institution, determine the initial node characteristics of the institution node in the topological graph.

[0120] Optionally, the feature expression module 403 is specifically used to process the adjusted features of the node and the node features of the neighboring nodes under the feature adjustment to obtain relationship features; wherein the processing includes: one of splicing processing, dot multiplication processing, and addition processing; according to the relationship features, the attention weight between the node and the neighboring nodes under the feature adjustment is adjusted to obtain the adjusted attention weight.

[0121] Optionally, the feature expression module 403 is specifically used to input the relationship features into the multi-layer perceptron in the paper category prediction model, so as to process the relationship features through the multi-layer perceptron to obtain adjustment parameters; wherein, the multi-layer perceptron includes at least: an attention layer, a pooling layer, a Dropout layer, a regularization layer and a normalization layer; based on the adjustment parameters, the attention weight between the node and the neighboring node under this feature adjustment is adjusted to obtain the adjusted attention weight.

[0122] Optionally, the feature expression module 403 is specifically configured to perform weighted summation on the adjusted feature of the node obtained in each feature adjustment of the node and the initial node feature of the node to obtain the final node feature of the node.

[0123] The present disclosure also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can be used to execute the paper classification method provided in Figure 1 above.

[0124] Based on the paper classification method shown in Figure 1, the present disclosure also provides a schematic diagram of the structure of an electronic device as shown in Figure 5. As shown in Figure 5, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the paper classification method described in Figure 1.

[0125] Of course, in addition to software implementation, this disclosure does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0126] In the 1990s, technological improvements could be clearly distinguished as either hardware improvements (for example, improvements to circuit structures like diodes, transistors, and switches) or software improvements (improvements to process flows). However, with the advancement of technology, many process flow improvements today can now be considered direct improvements to hardware circuit structures. Designers almost always create the corresponding hardware circuit structure by programming the improved process flow into the hardware circuit. Therefore, it cannot be said that a process flow improvement cannot be implemented using hardware modules. For example, a programmable logic device (PLD), such as a field programmable gate array (FPGA), is an integrated circuit whose logical function is determined by user programming. Designers can "integrate" a digital system on a PLD by programming it themselves, without having to hire a chip manufacturer to design and manufacture a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly performed using "logic compiler" software. This is similar to the software compiler used during program development. Before compilation, the original code must also be written in a specific programming language, called a hardware description language (HDL). There is not just one HDL, but many, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art will also understand that by simply programming the method flow in one of these hardware description languages ​​and then programming it into an integrated circuit, a hardware circuit that implements the logic method flow can be easily obtained.

[0127] The controller can be implemented in any suitable manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also know that in addition to implementing the controller in a purely computer-readable program code format, the controller can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component. Or even, the devices for implementing various functions can be considered as both software modules that implement the method and structures within the hardware component.

[0128] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0129] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing the present disclosure, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0130] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, 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 produce a device for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0132] 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 operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that 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 steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0134] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0135] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0136] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0137] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0138] Those skilled in the art will appreciate that embodiments of the present disclosure may be provided as methods, systems, or computer program products. Thus, the present disclosure may take the form of a fully hardware embodiment, a fully software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] The present disclosure may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present disclosure may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0140] The various embodiments of this disclosure are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so their description is relatively simple. For relevant portions, refer to the description of the method embodiments.

[0141] The foregoing is merely an embodiment of the present disclosure and is not intended to limit the present disclosure. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure are intended to be included within the scope of the claims of the present disclosure.

Claims

1. A paper classification method, comprising: Obtaining related information of the paper to be classified, wherein the related information includes: the text of the paper, the text of the paper cited by the paper, the author of the paper, and the institution to which the author of the paper belongs; Based on the association information, a topological map is constructed, wherein the topological map includes paper nodes, cited paper nodes, author nodes, and institution nodes, and each node in the topological map corresponds to an initial node feature; Inputting the topological graph into a pre-trained paper category prediction model, so as to determine, for each node in the topological graph, a neighboring node connected to the node through an edge through the paper category prediction model; For each feature adjustment of the node, according to the node features of the neighboring nodes under the feature adjustment and the attention weights between the node and the neighboring nodes under the feature adjustment, the node features of the node under the feature adjustment are adjusted to obtain the adjusted features of the node as the node features of the node under the next feature adjustment, and based on the adjusted features of the node and the node features of the neighboring nodes under the feature adjustment, the attention weights between the node and the neighboring nodes under the feature adjustment are adjusted to obtain the adjusted attention weights as the attention weights between the node and the neighboring nodes under the next feature adjustment, and the node features of the node are continuously adjusted for the next feature adjustment of the node until the number of feature adjustments reaches a preset number; According to the adjusted features of the node obtained in each feature adjustment of the node, the initial node features of the node are adjusted to obtain the final node features of the node; The paper category prediction model is used to predict the category of the paper based on the final node features of the paper nodes in the topological graph.

2. The method according to claim 1, characterized in that Based on the association information, a topology map is constructed, including: A topological graph is constructed with the text of the paper, the text of the paper cited by the paper, the author of the paper, and the institution to which the author of the paper belongs as nodes, and with the citation relationship between the paper and the cited paper, the affiliation relationship between the author and the institution, and the authorship relationship between the paper and the author as edges; Based on the text of the paper, determining initial node features of the paper nodes in the topological graph; Based on the text of the cited paper, determining the initial node features of the cited paper node in the topological graph; Determining initial node features of the author node in the topological graph based on the texts of all papers written by the author; Based on the texts of all papers written by all authors under the institution, the initial node characteristics of the institution node in the topological graph are determined.

3. The method according to claim 1, characterized in that According to the node features of the neighboring nodes under the secondary feature adjustment and the attention weight between the node and the neighboring nodes under the secondary feature adjustment, the node features of the node under the secondary feature adjustment are adjusted to obtain the adjusted features of the node, including: Determine an aggregate feature for the node according to the node feature of the neighbor node under the secondary feature adjustment and the attention weight between the node and the neighbor node under the secondary feature adjustment; Determine the product of the preset parameter and the node feature of the node under the feature adjustment to obtain a smooth feature; The aggregated feature is added to the smoothed feature to obtain an adjusted feature of the node.

4. The method according to claim 1, characterized in that Based on the adjusted feature of the node and the node feature of the neighboring node under the feature adjustment, the attention weight between the node and the neighboring node under the feature adjustment is adjusted to obtain the adjusted attention weight, including: Processing the adjusted feature of the node and the node feature of the neighboring node under the feature adjustment to obtain a relationship feature; wherein the processing includes: one of concatenation processing, point multiplication processing, and addition processing; According to the relationship feature, the attention weight between the node and the neighboring node under this feature adjustment is adjusted to obtain the adjusted attention weight.

5. The method according to claim 4, characterized in that According to the relationship feature, the attention weight between the node and the neighboring node under the feature adjustment is adjusted to obtain the adjusted attention weight, including: The relationship features are input into the multi-layer perceptron in the paper category prediction model to A multi-layer perceptron is provided to process the relationship features to obtain adjustment parameters; wherein the multi-layer perceptron includes at least: an attention layer, a pooling layer, a dropout layer, a regularization layer and a normalization layer; Based on the adjustment parameter, the attention weight between the node and the neighboring node under this feature adjustment is adjusted to obtain the adjusted attention weight.

6. The method according to claim 1, characterized in that According to the adjusted features of the node obtained in each feature adjustment of the node, the initial node features of the node are adjusted to obtain the final node features of the node, including: The adjusted features of the node obtained in each feature adjustment of the node and the initial node features of the node are weighted summed to obtain the final node features of the node.

7. The method according to claim 1, characterized in that Pre-training the paper category prediction model includes: Acquire sample association information of the sample paper to be classified, the sample association information including: the text of the sample paper, the text of the sample cited paper of the sample paper, the author of the sample paper, and the institution to which the author of the sample paper belongs; Based on the sample association information, construct a topological graph for the sample paper as a sample topological graph, wherein each sample node in the sample topological graph corresponds to an initial node feature; The sample topology graph is input into the paper category prediction model to be trained, so as to determine, for each sample node in the sample topology graph, a sample neighbor node connected to the sample node through an edge through the paper category prediction model to be trained; For each feature adjustment of the sample node, according to the node features of the sample neighbor nodes under the feature adjustment and the attention weights between the sample node and the sample neighbor nodes under the feature adjustment, the node features of the sample node under the feature adjustment are adjusted to obtain the adjusted features of the sample node as the node features of the sample node under the next feature adjustment, and based on the adjusted features of the sample node and the node features of the sample neighbor nodes under the feature adjustment, the attention weights between the sample node and the sample neighbor nodes under the feature adjustment are adjusted to obtain the adjusted attention weights as the attention weights between the sample node and the sample neighbor nodes under the next feature adjustment, and the node features of the sample node are continuously adjusted for the next feature adjustment of the sample node until the number of feature adjustments reaches a preset number; According to the adjusted features of the sample node obtained in each feature adjustment of the sample node, the initial node features of the sample node are adjusted to obtain the final node features of the sample node; Predicting the category of the sample paper based on the final node features of the sample nodes corresponding to the sample paper in the sample topology graph; The paper category prediction model to be trained is trained with minimizing the difference between the category of the sample paper and the true category of the sample paper as the optimization goal.

8. A paper classification device, comprising: An acquisition module, used for acquiring related information of the paper to be classified, wherein the related information includes: the text of the paper, the text of the cited paper of the paper, the author of the paper, and the institution to which the author of the paper belongs; A construction module, used to construct a topological map based on the association information, wherein the topological map includes paper nodes, cited paper nodes, author nodes, and institution nodes, and each node in the topological map corresponds to an initial node feature; A feature expression module is used to input the topological map into a pre-trained paper category prediction model, so as to determine, for each node in the topological map, a neighbor node connected to the node through an edge through the paper category prediction model; for each feature adjustment of the node, according to the node features of the neighbor nodes under the feature adjustment and the attention weights between the node and the neighbor nodes under the feature adjustment, the node features of the node under the feature adjustment are adjusted to obtain the adjusted features of the node as the node features of the node under the next feature adjustment, and based on the adjusted features of the node and the node features of the neighbor nodes under the feature adjustment, the attention weights between the node and the neighbor nodes under the feature adjustment are adjusted to obtain the adjusted attention weights as the attention weights between the node and the neighbor nodes under the next feature adjustment, and the feature adjustment is continued for the node. The next feature adjustment of the node is to adjust the node feature of the node until the number of feature adjustments reaches a preset number; according to the adjusted feature of the node obtained under each feature adjustment of the node, the initial node feature of the node is adjusted to obtain the final node feature of the node; The paper classification module is used to predict the category of the paper based on the final node features of the paper nodes in the topological graph through the paper category prediction model.

9. The device according to claim 8, characterized in that The construction module is specifically used to construct a topological graph with the text of the paper, the text of the paper cited by the paper, the author of the paper, and the institution to which the author of the paper belongs as nodes, and with the citation relationship between the paper and the cited paper, the affiliation relationship between the author and the institution, and the authorship relationship between the paper and the author as edges; Based on the text of the paper, determining initial node features of the paper nodes in the topological graph; Based on the text of the cited paper, determining the initial node features of the cited paper node in the topological graph; Determining initial node features of the author node in the topological graph based on the texts of all papers written by the author; Based on the texts of all papers written by all authors under the institution, the initial node characteristics of the institution node in the topological graph are determined.

10. The device according to claim 8, characterized in that The feature expression module is specifically used to process the adjusted features of the node and the node features of the neighboring nodes under the feature adjustment to obtain relationship features; wherein the processing includes: one of splicing processing, dot multiplication processing, and addition processing; according to the relationship features, the attention weight between the node and the neighboring nodes under the feature adjustment is adjusted to obtain the adjusted attention weight.

11. The device according to claim 10, characterized in that The feature expression module is specifically used to input the relationship features into the multilayer perceptron in the paper category prediction model, so as to process the relationship features through the multilayer perceptron to obtain adjustment parameters; wherein the multilayer perceptron includes at least: an attention layer, a pooling layer, a Dropout layer, a regularization layer and a normalization layer; based on the adjustment parameters, the attention weight between the node and the neighboring node under this feature adjustment is adjusted to obtain the adjusted attention weight.

12. The device according to claim 8, characterized in that The feature expression module is specifically used to perform weighted summation on the adjusted feature of the node obtained in each feature adjustment of the node and the initial node feature of the node to obtain the final node feature of the node.

13. The device according to claim 8, characterized in that Also includes: Training modules; The training module is specifically used to obtain sample association information of sample papers to be classified, and the sample association information includes: the text of the sample paper, the text of the sample cited paper of the sample paper, the author of the sample paper, and the institution to which the author of the sample paper belongs; based on the sample association information, a topological map for the sample paper is constructed as a sample topological map, wherein each sample node in the sample topological map corresponds to an initial node feature; the sample topological map is input into the paper category prediction model to be trained, so as to determine, for each sample node in the sample topological map, a sample neighbor node connected to the sample node through an edge through the paper category prediction model to be trained; for each feature adjustment of the sample node, the node feature of the sample node under the feature adjustment is adjusted according to the node feature of the sample neighbor node under the feature adjustment and the attention weight between the sample node and the sample neighbor node under the feature adjustment, so as to obtain the adjusted feature of the sample node as the feature for the next feature adjustment. The node features of the sample node under feature adjustment are adjusted, and based on the adjusted features of the sample node and the node features of the sample neighbor nodes under the feature adjustment, the attention weight between the sample node and the sample neighbor nodes under the feature adjustment is adjusted to obtain the adjusted attention weight, which is used as the attention weight between the sample node and the sample neighbor nodes under the next feature adjustment, and the node features of the sample node are adjusted for the next feature adjustment of the sample node until the number of feature adjustments reaches a preset number; according to the adjusted features of the sample node obtained under each feature adjustment of the sample node, the initial node features of the sample node are adjusted to obtain the final node features of the sample node; based on the final node features of the sample node corresponding to the sample paper in the sample topology graph, the category of the sample paper is predicted; and the paper category prediction model to be trained is trained with minimizing the difference between the category of the sample paper and the true category of the sample paper as the optimization goal.

14. A computer-readable storage medium storing a computer program, wherein the computer program implements the method according to any one of claims 1 to 7 when executed by a processor.

15. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method described in any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Text classification method and device

    CN112131386A

  • Quotation network node classification method and system for relation uncertainty

    CN112836050A

  • Scientific and technological paper hierarchical multi-label classification method and equipment based on graph convolutional network

    CN113312480A

  • Colored Photosensitive Resin Composition, Color Filter and Display Device

    KR1020240116639A