A quantitative method for correlation degree of literature research fields based on multi-dimensional feature fusion

Abstract

The present invention discloses a method for quantifying the degree of association in a literature research field based on multi-dimensional feature fusion, and belongs to the technical field of literature research association calculation. The method of the present invention comprehensively captures the relationship between literature features by constructing multiple undirected graphs, and enhances the ability to express the association of literature; through experimental comparative analysis of multiple feature aggregation operations, the optimal feature fusion scheme is selected to extract the association features between documents, and the prediction accuracy of the quantitative association degree is improved; through the minimum dimensional sphere model, a cross-domain scientific research capability assessment method is performed to help scientific researchers identify the contributions and influence of different scholars in multidisciplinary cross-disciplinary research; at the same time, the accuracy rate with error tolerance is used as an evaluation indicator to improve the flexibility and adaptability under different accuracy requirements. The method of the present invention can help researchers quickly lock in the most relevant information from a large number of documents, so that literature research can better meet actual scientific research needs, thereby promoting interdisciplinary cooperation and innovation.

Description

Technical Field

[0001] The present invention relates to the technical field of document research association degree calculation, and in particular to a document research field association degree quantification method based on multi-dimensional feature fusion. Background Art

[0002] The methods for quantifying associations in the field of literature research are mainly divided into methods based on single relationship modeling and multi-dimensional feature enhancement. Traditional methods based on single relationship modeling mainly focus on single-type associations such as co-authorship or citation relationships. Such methods are difficult to fully capture the multi-dimensional cross-correlations between documents, especially in the context of highly cross-integrated research, and cannot meet the needs of multi-level analysis, limiting their application in literature retrieval and interdisciplinary research.

[0003] Existing methods based on multi-dimensional feature enhancement mostly rely on basic node attributes and their connection relationships based on knowledge graphs, but the feature representation depth is insufficient and it is difficult to reveal the feature differences between nodes, resulting in low accuracy in correlation calculations, lack of empirical data research support, and insufficient interpretability of the results. In recent years, graph neural networks (GNNs) have gradually been applied to graph data analysis, with the advantage of enhancing feature expression capabilities.

[0004] However, most existing GNN methods use a single graph structure, focusing more on the direct field feature extraction of documents for document field classification tasks, and fail to deeply explore the complex correlation features between documents for more detailed document research field correlation calculation tasks. In addition, existing technologies have not yet achieved effective differentiation of scholar characteristics in interdisciplinary scientific research capability assessment, resulting in difficulty in accurately reflecting the contributions and influence of different scholars in cross-disciplinary research in document retrieval and multidisciplinary field research. Summary of the invention

[0005] The problem to be solved by the present invention is to provide a method for quantifying the degree of association in the field of literature research based on multidimensional feature fusion, which is used to quickly lock the most relevant information from a large number of documents, reduce operation time, accelerate the scientific research process, enable literature research to better meet actual scientific research needs, and effectively distinguish the characteristics of scholars.

[0006] The present invention adopts the following technical solution: a method for quantifying the degree of association of literature research fields based on multi-dimensional feature fusion, comprising the following steps:

[0007] Step 1: Obtain target document data from the document database, extract document metadata and perform preprocessing, generate initial document vectors and document research field relevance labels, and balance the target document data set;

[0008] Step 2: construct multiple undirected graphs and adjacency matrices based on the document metadata to represent the multidimensional association relationships between documents;

[0009] Step 3: Build a graph convolutional neural network model, perform multi-dimensional feature fusion on document nodes, select the optimal aggregation strategy to enhance the association representation ability of document nodes, and obtain enhanced document node features;

[0010] Step 4: Based on the enhanced document node features, calculate the domain correlation between documents, and dynamically update the document node feature vector through iterative training model to evaluate the effectiveness of quantifying the degree of correlation between document research fields;

[0011] Step 5: Combine the minimum The dimensional ball model quantitatively compares the cross-disciplinary scientific research capabilities of scholars and effectively distinguishes the characteristics of scholars.

[0012] Preferably, the method of step 1 is as follows:

[0013] Step 1.1: Obtain target document data from the literature database and extract key metadata, including: title, abstract, keywords, author, and publication information of the document;

[0014] Step 1.2: Preprocess the document metadata to obtain the initialized document vector:

[0015] Eliminate documents that lack key metadata; apply a stop word list and regular expressions to remove common stop words in document titles and abstracts; use the jieba word segmentation tool to perform Chinese word segmentation on document titles and abstracts, and use the Doc2Vec model to train the segmented text. The model parameter size is the document vector dimension, and generate a semantic feature vector for each document;

[0016] Step 1.3: Preprocess the document metadata, extract the field classification number field from the document metadata, and uniformly process it into a fixed-length string to represent the field category and sub-research field. Perform prefix matching on the processed field classification number to generate a label of the degree of relevance of the document research field;

[0017] Step 1.4: Perform stratified sampling during the balanced sampling process. For each prefix category, if the number of samples is less than or equal to the set number of samples, all of them are retained; if the number of samples is greater than the set number of samples, a random number seed is set to randomly extract a specified number of samples to form a balanced data set; for the balanced data set, group them by prefix category and divide them into test set and training set so that the proportion of each prefix category in the training set and test set is consistent.

[0018] Preferably, the method of step 2 is as follows:

[0019] Step 2.1: Construct several different undirected graphs. Each undirected graph structure is represented as , used to capture the multi-dimensional relationship between documents, nodes Represents a document that carries document metadata. represents the association between documents, and the edge weight reflects the strength of the association between document types;

[0020] Step 2.2: Generate the corresponding adjacency matrix for each undirected graph structure ,initialization The zero matrix of ,in, represents the correlation strength between document i and document j in the graph structure dimension, Represents the number of documents, traverses each pair of documents, and fills in the corresponding elements in the adjacency matrix.

[0021] Preferably, the method of step 3 is as follows:

[0022] Step 3.1: The semantic feature vector generated in step 1 and the adjacency matrix information of different undirected graphs in step 2 are used together to form the initial feature vector of each document node as the input of the graph convolutional neural network model;

[0023] Step 3.2: Aggregate neighbor nodes in the graph convolutional neural network layer to update the nodes;

[0024] Step 3.3: Select the optimal aggregation strategy to enhance the association representation capability of the document node, wherein the aggregation strategy includes a summation strategy, an average strategy, and a maximum value pooling strategy to obtain enhanced document node features.

[0025] Preferably, the method of step 4 is as follows:

[0026] Step 4.1, calculate the domain relevance between documents, calculate the Euclidean distance from the enhanced document node set as the predicted value of the quantitative method of the degree of association between document research fields, and calculate the Euclidean distance between each pair of document vectors to represent the predicted value of the quantified degree of association between documents;

[0027] Step 4.2: Compare the predicted value with the preset label, introduce the accuracy with error tolerance as the evaluation indicator, and evaluate the degree of relevance of literature research fields under different precisions by setting the error threshold;

[0028] Step 4.3, repeat steps 4.1 and 4.2 until the model converges. By continuously iteratively training and updating the feature vectors of document nodes, the accuracy and effectiveness of the quantitative method for the degree of association in the research field of documents can be improved.

[0029] Preferably, step 5 combines the minimum The dimensional ball model is used to quantitatively compare the cross-disciplinary scientific research capabilities of scholars. The method is as follows:

[0030] Step 5.1: Integrate the feature vectors of all documents into a set , each eigenvector Represents the characteristics of a document. is the number of feature vectors of the document, is the dimension of the document vector, using the eigenvector mean as the minimum The center of the ball ;

[0031] Step 5.2: Calculate the maximum Euclidean distance from all eigenvectors to the eigenvector mean as the radius, which represents the minimum radius covering the entire eigenvector set. Dimensional ball;

[0032] Step 5.3: Extract all published papers for each scholar and repeat step 5.1 to generate the minimum dimensional ball model, comparing the minimum Dimensional sphere volume, minimum The larger the volume of the dimensional sphere, the broader the scientific research field of the corresponding scholars and the stronger the cross-field scientific research capabilities.

[0033] The technical solution of the present invention also provides: an electronic device, comprising:

[0034] one or more processors;

[0035] a storage device having one or more programs stored thereon;

[0036] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-mentioned methods for quantifying the degree of association in literature research fields based on multi-dimensional feature fusion.

[0037] The technical solution of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in any of the above-mentioned methods for quantifying the degree of association in literature research fields based on multi-dimensional feature fusion are implemented.

[0038] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0039] 1. The method for quantifying the degree of association in the field of literature research of the present invention constructs an undirected graph of multiple association dimensions, combines it with a graph neural network for training, and obtains the optimal multi-dimensional feature aggregation scheme through experimental comparative analysis. It discovers how to conduct beneficial training on the extraction of associated features from multiple dimensions and according to the influence of the feature range, thereby improving the accuracy of quantification of association in the field of literature research.

[0040] 2. The method for quantifying the degree of association between research fields of documents in the present invention, through multi-dimensional feature fusion and combined with the training of multiple text association graphs, improves the adaptability of association patterns between texts in different research fields, balances the influence of the joint action of multiple association dimensions, has a certain stability, and has a significant advantage in refined association calculation.

[0041] 3. The method for quantifying the degree of correlation in the research field of literature of the present invention more scientifically quantifies the contribution and influence of scholars in multidisciplinary cross-disciplinary research by combining the minimum n-dimensional ball model, and captures the comprehensive performance of scholars in different disciplines by constructing a multidimensional feature space, thereby providing a solution to the problem of how to compare the cross-disciplinary scientific research capabilities of cross-disciplinary scholars, which is more comprehensive and explainable. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the method for quantifying the degree of association in the research field of literature of the present invention;

[0043] Figure 2 A comparison diagram of different feature aggregation operations provided by an embodiment of the present invention;

[0044] Figure 3 A graph showing the quantification results of literature research relevance at different precisions provided by an embodiment of the present invention;

[0045] Figure 4 A diagram of the quantification results of model performance under single-dimensional modeling provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0047] In one embodiment of the present invention, a method for quantifying the degree of association in literature research fields based on multi-dimensional feature fusion is used, with Wanfang.com data as the source. Figure 1 As shown, the following steps are included:

[0048] Step 1: Obtain document data and perform metadata preprocessing to generate an initial document feature vector and automatically generate a document research field relevance label, while ensuring the balance of the document data set;

[0049] Specifically:

[0050] Step 1.1, obtain the target document data from the Wanfang document database, and extract key metadata including the document title, abstract, keywords, author, publication information, etc.

[0051] Step 1.2: Preprocess the document metadata to obtain the initialized document vector:

[0052] First, remove documents that are missing key metadata to ensure data integrity; then, apply a custom stop word list and regular expressions to remove common stop words in document titles and abstracts to improve text quality; then, use the jieba word segmentation tool to perform Chinese word segmentation on the document title and abstract to generate the text format after word segmentation; finally, use the Doc2Vec model to train the text after word segmentation, and the model parameter size corresponds to the dimension of the document vector , generate the semantic feature vector of each document.

[0053] Step 1.3: Preprocess the document metadata and automatically generate labels for the degree of relevance of the document research field:

[0054] The field classification number field is extracted from the document metadata, and the content format of this field is not fixed. In this embodiment, it is uniformly processed into a string of length 4, such as G420, where the first letter represents the field category, and the last three digits represent the sub-research field.

[0055] Prefix matching is performed on the processed field classification number: Assuming that there are document i and document j, character matching is performed bit by bit respectively. If they are the same, 1 is added, otherwise 0. The prefix matching length represents the distance D(i, j) between document i and document j. The distance is converted into the corresponding discrete association level label for subsequent quantitative evaluation and analysis. The larger the number, the higher the degree of association between the research fields of the documents.

[0056] Step 1.4, stratified sampling technology is used in the balanced sampling process. Specifically, for each prefix category, if the number of samples is less than or equal to the set number of samples, all of them are retained; if the number of samples is greater than the set number of samples, a random number seed is set to randomly select a specified number of samples from them to form a balanced data set.

[0057] For the balanced data set, group them by prefix category; extract 20% of the samples from each group as the test set, and the remaining samples as the training set; the above steps ensure that the proportion of each prefix category in the training set and the test set is consistent during the sampling process, thereby maintaining a balanced distribution of sample types.

[0058] Step 2: Construct multiple undirected graphs based on document metadata to represent the multidimensional association relationships between documents;

[0059] Specifically:

[0060] Step 2.1: Construct four different undirected graphs to capture the multi-dimensional relationships between documents and enrich the feature representation of nodes. The definition is as follows:

[0061] node : represents the document carrying document metadata; edge : Indicates the association between documents, and the weight of the edge reflects the strength of this type of association.

[0062] Preferably, the graph structures of the four different undirected graphs in this embodiment are as follows:

[0063] Title segmentation graph: If the titles of two documents have an intersection after segmentation, that is, they have at least one common segmentation, an edge is constructed between the corresponding nodes. The initial weight of the edge is set to the number of common segmentations, quantifying the similarity of the titles and extracting the domain relevance information between the documents.

[0064] Common keyword graph: When two documents have the same keyword in their keyword list, they are connected in the graph through edges. The weight of the edge is set to the number of the same keywords, indicating the importance of the keyword in the field association.

[0065] Journal name association graph: If two documents are published in the same academic journal, an edge is established and the weight of the edge is set to 1 to reflect the contribution of the journal to which the documents belong to the field association.

[0066] Field classification association graph: If two documents belong to the same field classification number, an edge is established, and the weight of the edge is also set to 1, reflecting the impact of field classification information on the degree of document association.

[0067] Step 2.2, after each graph structure is constructed, generate the corresponding adjacency matrix A.

[0068] initialization The zero matrix of ,in, It indicates the correlation strength between document i and document j in the dimension of the graph structure. Indicates the number of documents. Traverse each pair of documents and fill in the corresponding elements in the matrix according to the conditions defined in the above graph structure.

[0069] Step 3: Use graph neural networks to perform multi-dimensional feature fusion processing on document nodes, select the optimal aggregation operation, and effectively improve the association representation ability of document nodes;

[0070] Specifically:

[0071] Step 3.1: The semantic feature vector generated in step 1 and the adjacency matrix information in different undirected graphs in step 2 are combined to form the initial feature vector of each document node as the input of the graph convolutional neural network model (GCN).

[0072] In step 3.2, the convolutional neural network model performs aggregation operations on neighbor nodes in the layer to update the nodes.

[0073] For each node, aggregate the feature vectors of its neighboring nodes. The aggregation function can use different operations, such as summation, averaging, or maximum pooling. The node update process is as follows:

[0074] ;

[0075] in, Represents the node update process of each association graph, , calculated as follows:

[0076] ;

[0077] Among them, at the lth layer, the node feature matrix of different graph structures is obtained through a series of information propagation and updating: , is the adjacency matrix, for The degree matrix of , is the activation function; Represents the aggregation method symbol; the feature matrix after the feature fusion at the l+1 layer is .

[0078] Step 3.3: In order to improve the effect of feature fusion, different aggregation methods are compared experimentally to select the optimal aggregation strategy.

[0079] This implementation strives to use sum, average or maximum pooling as the aggregation function, and compare different feature aggregation operations, such as Figure 2 As shown in the figure; the quantification results of literature research relevance under different precisions are as follows: Figure 3 shown.

[0080] from Figure 3 It can be seen that Figure 3 (a) in Figure 1 uses the summation strategy as the optimal strategy for multi-graph feature fusion, showing the highest accuracy. This is because it can directly accumulate features from different graph structures, avoiding the reduction of feature amplitudes and retaining the stronger feature contributions of each graph. By adding different graph features, GCN can extract richer inter-node association information from each graph structure, allowing nodes to more accurately express the multi-dimensional associations between documents.

[0081] The comparative analysis of other strategies in this embodiment is as follows:

[0082] Average strategy: Figure 3 (b) in Figure 1 uses the average strategy as the multi-graph feature fusion strategy. The model initially has high loss values ​​on the training set and test set, and low accuracy. This is because the average strategy reduces the dynamic range of the features, and the expressive power of the features in different graph structures is diluted, making it difficult for nodes to effectively learn related features in the early stages of training. In the first 50 rounds of training, the accuracy of the model increased slowly. Although it improved to a certain extent after a period of time, the final accuracy still did not exceed 0.2, indicating that the strategy failed to fully capture the correlation between document nodes.

[0083] Maximum and minimum strategies: Figure 3 (c) and (d) in the figure respectively show the effect of using maximum and minimum values ​​as feature fusion strategies. Figure 3 (c) in the figure shows that the maximum strategy can quickly improve the accuracy of the model in the early stage of training, but as the number of training rounds increases, the model suffers from serious overfitting and the accuracy gradually decreases. This is because the maximum strategy is very sensitive to local peaks and outliers, causing the model to mistake local features for global features, thereby weakening the model's generalization ability. On the contrary, Figure 3 (d) in the figure shows the performance of the minimum strategy. Although the accuracy of this strategy is low in the early stage of training, it is not sensitive to outliers and can pay more attention to the generalization pattern in the data. As the training progresses, the accuracy rate steadily increases and gradually reaches a relatively stable level. Although the minimum strategy performs poorly in the early stage, it can show good generalization ability after long-term training.

[0084] The final experimental results show that this embodiment finally selects the summation strategy as the optimal strategy for multi-graph feature fusion. This is because in this aggregation process, not only each node is allowed to integrate information from all neighboring nodes (including directly and indirectly related nodes) when updating, fusing global information to avoid information loss, but also the correlation features between nodes in different graph structures can be retained to the greatest extent, avoiding the compression of the dynamic range or the excessive influence of outliers, so that GCN can more effectively learn the multi-dimensional correlation features of document nodes.

[0085] Step 4: Based on the enhanced document node features, calculate the domain relevance between documents, and dynamically update the document node feature vector through iterative training model to evaluate the effectiveness of quantifying the degree of relevance of document research fields;

[0086] Specifically:

[0087] Step 4.1: After the multi-dimensional features of the document nodes are fused, the correlation features are enhanced. The Euclidean distance is calculated from the document node set as the predicted value of the quantitative method of the correlation degree of the document research field. The Euclidean distance between each pair of document vectors is calculated as follows:

[0088] ;

[0089] in, and are the coordinates of two points in space in the i-th dimension, is the dimension of the document vector, It is the predicted value after quantifying the degree of association between documents, which is used to provide a basis for subsequent analysis.

[0090] In step 4.2, after calculating the domain relevance of each pair of literature nodes, the predicted values ​​are compared with the preset labels to evaluate the accuracy of the quantification effect.

[0091] Since the automatically generated association degree labels in step 1 are discrete numerical levels, traditional regression indicators (such as mean square error (MSE) and mean absolute error (MAE)) can provide information on the deviation between the predicted value and the true label, but cannot directly reflect the accuracy of the model at a specific association level.

[0092] Given that actual application scenarios in the field of literature research usually do not require absolutely accurate prediction values, but are more concerned about whether the prediction values ​​fall within the correct level range, this implementation introduces the accuracy with error tolerance (TAA) as an evaluation indicator. This indicator is achieved by setting a reasonable error threshold. , making the method suitable for evaluating the degree of association in literature research fields with different precisions.

[0093] use Indicates the preset tag value. represents the predicted value, ε is the error range, and the correlation degree of literature research field , the calculation method is as follows:

[0094] ;

[0095] Among them, N is the total number of samples, and I is the indicator function, which is used to check whether the predicted value is within the allowable error range; only when and When the error of ε falls within the range of ε, it is considered a correct prediction.

[0096] Step 4.3: Repeat steps 4.1 and 4.2 until the model converges.

[0097] By continuously iteratively training and updating the feature vectors of document nodes, the accuracy and effectiveness of the quantitative method for the degree of association in the research field of documents can be improved within a certain range.

[0098] Step 5: Combine Minimum The dimensional ball model quantitatively compares scholars’ cross-disciplinary scientific research capabilities;

[0099] Specifically:

[0100] Step 5.1: integrate the feature vectors of all documents into a set, where each feature vector , where each eigenvector Represents the characteristics of a document. is the number of feature vectors of the document, the minimum The n-dimensional sphere corresponds to the dimension of the document vector .

[0101] Use the eigenvector mean as the minimum The center of the ball , because these vectors are calculated by quantitative methods of correlation degree, the mean of the eigenvector is not only a single value, but also reflects the multi-dimensional field correlation relationship of its document collection:

[0102] ;

[0103] Step 5.2, calculate the maximum Euclidean distance from all eigenvectors to the eigenvector mean as the radius, which represents the smallest n-dimensional ball covering the entire eigenvector set. The radius is calculated as follows:

[0104] ;

[0105] Step 5.3: Repeat step 5.1 for each scholar by extracting all published documents to generate the minimum dimensional ball model. Comparison of the minimum The volume of the sphere is calculated as follows:

[0106] ;

[0107] in, is the gamma function, when When it is an integer, it can be converted into a factorial function. The larger the volume, the broader the scholar's scientific research field and the stronger his cross-field scientific research ability.

[0108] Furthermore, this embodiment conducts experimental verification of quantitative indicators of literature research field relevance on different models, and the experimental results are shown in the following table:

[0109] Table 1: Comparison of experimental results of different models

[0110] Model Mean Squared Error (MSE) Mean Absolute Error (MAE) Accuracy Document Vectors (Doc2Vec) 0.098±0.002 0.255±0.001 0.011±0.005 Graph Convolutional Networks (GCNs) 0.036±0.001 0.134±0.002 0.600±0.010 Graph Attention Network (GAT) 0.047±0.004 0.150±0.001 0.667±0.001 Doc2Vec+GCN+multidimensional method of the present invention 0.034±0.001 0.122±0.003 0.679±0.001

[0111] As can be seen from the table, the method of the present invention performs best in terms of MSE, MAE and accuracy. In particular, in terms of accuracy, the method of the present invention reaches 0.679, which is significantly higher than other models. This shows that the method of the present invention can more accurately quantify the relevance of the literature research field and improve the efficiency and accuracy of literature retrieval.

[0112] Furthermore, this embodiment uses an ablation experiment to verify the effect of the multi-dimensional feature fusion method of the present invention.

[0113] Through ablation experiments, the effects of single-dimensional modeling and multi-dimensional feature fusion methods were tested respectively. The results are as follows: Figure 4 shown.

[0114] Figure 4 Each sub-graph represents the effectiveness of the method under different single correlation dimensions. Figure 4 (a) in the figure represents the keyword correlation dimension between two documents, which quantifies the degree of correlation between the research fields of the documents by measuring the common keywords between the documents to reveal the relevance of the document topics; Figure 4 (b) represents the correlation dimension between two documents at the title level, which quantifies the degree of correlation between the documents by comparing the common words after the title segmentation to reflect the consistency of the research direction of the documents; Figure 4 (c) and (d) represent the correlation dimensions of the two documents in terms of journal publication name and field classification name, respectively. The consistency of their research fields is indicated by judging whether the documents are published in the same academic journal and whether the field classification name is the same through the field classification number, thereby quantifying the degree of correlation between the research fields of the documents.

[0115] It can be seen that the experimental results under single-dimensional modeling are less stable and have lower accuracy, while the multi-dimensional feature fusion method performs well in multiple indicators. The accuracy of the model is significantly reduced when only a single-dimensional structure is used to extract related features. This is because a single type of graph structure can only learn related patterns specific to that dimension, and it is difficult to learn generalized related features, resulting in information loss and a decrease in model prediction accuracy; the multi-dimensional feature fusion method can combine multiple text related graph training to improve the model's adaptability to the related patterns of different research fields between texts, balance the impact of the joint action of multiple related dimensions, and ensure that the model has a certain stability.

[0116] In an embodiment of the present invention, an electronic device is also provided, including: one or more processors; a storage device on which one or more programs are stored; when the one or more programs are executed by the one or more processors, the one or more processors implement the method for quantifying the degree of association in the literature research field based on multidimensional feature fusion described in any of the above embodiments.

[0117] In an embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the program is executed by a processor, the steps in any one of the methods for quantifying the degree of association in literature research fields based on multidimensional feature fusion in the above embodiments are implemented.

[0118] In summary, the method for quantifying the degree of association in the field of literature research of the present invention constructs an undirected graph with multiple association dimensions, combines it with graph neural network for training, and uses the optimal multi-dimensional feature aggregation scheme to discover how to conduct beneficial training on the extraction of association features from multiple dimensions and according to the influence of the feature range, thereby effectively improving the accuracy of quantification of association in the field of literature research; at the same time, the multi-dimensional feature fusion method of the present invention also has significant advantages in the refined calculation of association degree, and can capture the comprehensive performance of scholars in different disciplines, through the minimum number of scholars The volume of the dimensional sphere quantifies its interdisciplinary scientific research capabilities, and at the same time, it is more comprehensive and explainable from multiple dimensions.

[0119] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for quantifying the degree of association in literature research fields based on multi-dimensional feature fusion, characterized in that: The steps include: Step 1: Obtain target document data from the document database, extract document metadata and perform preprocessing, generate initial document vectors and document research field relevance labels, and balance the target document data set; Step 2: construct multiple undirected graphs and adjacency matrices based on the document metadata to represent the multidimensional association relationships between documents; Step 3: Build a graph convolutional neural network model, perform multi-dimensional feature fusion on document nodes, select the optimal aggregation strategy to enhance the association representation ability of document nodes, and obtain enhanced document node features. The method is as follows: Step 3.1: The semantic feature vector generated in step 1 and the adjacency matrix information of different undirected graphs in step 2 are used together to form the initial feature vector of each document node as the input of the graph convolutional neural network model; Step 3.2: Aggregate the neighbor nodes in the graph convolutional neural network layer to update the nodes: Aggregate the feature vectors of each node’s neighbor nodes. The node update process is as follows: ； in, Indicates the aggregation mode symbol. Represents the node update process of each association graph, ; The feature matrix after feature fusion at the l+1 layer is ; Step 3.3, select the optimal aggregation strategy to enhance the association representation ability of the document node, the aggregation strategy includes summation strategy, average strategy, and maximum value pooling strategy, and obtain the enhanced document node features; Step 4: Based on the enhanced document node features, calculate the domain correlation between documents, and dynamically update the document node feature vector through iterative training model to evaluate the effectiveness of quantifying the degree of correlation between document research fields; Step 5: Combine the minimum The dimensional ball model quantitatively compares the cross-disciplinary scientific research capabilities of scholars and effectively distinguishes the characteristics of scholars. The method is as follows: Step 5.1: Integrate the feature vectors of all documents into a set , , where each eigenvector Represents the characteristics of a document. is the number of feature vectors of the document, Represents the dimension of the document vector, using the eigenvector mean as the minimum The center of the ball : ； Step 5.2: Calculate the maximum Euclidean distance from all eigenvectors to the eigenvector mean as the radius, which represents the minimum radius covering the entire eigenvector set. Sphere, radius The calculation is as follows: ； Step 5.3: Extract all published papers for each scholar and repeat step 5.1 to generate the minimum dimensional ball model, comparing the minimum Volume of a sphere : ； in, is the gamma function, when When it is an integer, it is converted to a factorial function; the minimum The larger the volume of the dimensional sphere, the broader the scientific research field of the corresponding scholars and the stronger the cross-field scientific research capabilities.

2. The method for quantifying the degree of association of literature research fields based on multidimensional feature fusion according to claim 1 is characterized in that: Step 1: Obtain the target document data and preprocess it as follows: Step 1.1: Obtain target document data from the literature database and extract key metadata, including: title, abstract, keywords, author, and publication information of the document; Step 1.2: Preprocess the document metadata to obtain the initialized document vector: Eliminate documents that lack key metadata; apply a stop word list and use regular expressions to remove common stop words in document titles and abstracts; use the jieba word segmentation tool to perform Chinese word segmentation on document titles and abstracts, and use the Doc2Vec model to train the segmented text; the model parameter size is the document vector dimension, and generates a semantic feature vector for each document; Step 1.3: Preprocess the document metadata, extract the field classification number field from the document metadata, and uniformly process it into a fixed-length string to represent the field category and sub-research field. Perform prefix matching on the processed field classification number to generate a label of the degree of relevance of the document research field; Step 1.4: Perform stratified sampling during the balanced sampling process. For each prefix category, if the number of samples is less than or equal to the set number of samples, all of them are retained; if the number of samples is greater than the set number of samples, set the random number seed to randomly select a specified number of samples to form a balanced data set; group by prefix category, divide the test set and training set, so that the proportion of each prefix category in the training set and test set is consistent.

3. The method for quantifying the degree of association of literature research fields based on multidimensional feature fusion according to claim 1 is characterized in that: In step 1.3, prefix matching is performed on the processed field classification number based on a prefix matching algorithm; The prefix matching algorithm performs bit-by-bit character matching of the field classification codes for document i and document j respectively, and adds 1 if they are the same, otherwise it is 0. The prefix matching length is determined by character matching in the field classification code, and the distance D(i, j) between the field classification codes of document i and document j is obtained. D(i, j) is converted into a corresponding discrete association level to generate a document research field association degree label. The larger the number, the higher the degree of association between the research fields of the documents.

4. The method for quantifying the degree of association of literature research fields based on multidimensional feature fusion according to claim 2 is characterized in that: Step 2 constructs multiple undirected graphs based on the document metadata as follows: Step 2.1: Construct several different undirected graphs. Each undirected graph structure is represented as , used to capture the multidimensional relationship between documents, where the node Indicates a document that carries document metadata. It represents the association between documents, and the edge weight reflects the strength of the association between document types; Step 2.2: Generate the corresponding adjacency matrix for each undirected graph structure ,initialization The zero matrix of ,in, represents the correlation strength between document i and document j in the graph structure dimension, Represents the number of documents; traverse each pair of documents and fill in the corresponding elements in the adjacency matrix.

5. The method for quantifying the degree of association of literature research fields based on multi-dimensional feature fusion according to claim 4 is characterized in that: Step 2.1 construct several different undirected graphs, including: title segmentation graph, common keyword graph, journal name association graph, and field classification association graph; The title segmentation graph: if the titles of two documents have an intersection after segmentation, that is, at least one segmentation is the same between the two documents, then an edge is constructed between the corresponding nodes, and the initial weight of the edge is set to the number of the same segmentations to quantify the similarity of the titles; The common keyword graph: if two documents have the same keyword in the keyword list, the two documents are connected by an edge in the graph, and the weight of the edge is set to the number of the same keyword, indicating the importance of the keyword in the field association; The journal name association graph: if two documents are published in the same academic journal, an edge is established, and the weight of the edge is set to 1 to reflect the contribution of the journal to which the document belongs to the field association; The field classification association graph: if two documents belong to the same field classification number, an edge is established, and the weight of the edge is set to 1, reflecting the impact of the field classification information on the degree of document association.

6. The method for quantifying the degree of association of literature research fields based on multi-dimensional feature fusion according to claim 1 is characterized in that: In step 3.2, The calculation is as follows: ； in, Represents the node feature matrix of different graph structures obtained through information propagation and updating at layer l, is the adjacency matrix, for The degree matrix of , is the activation function.

7. The method for quantifying the degree of association of literature research fields based on multi-dimensional feature fusion according to claim 4 is characterized in that: Step 4 evaluates the effectiveness of quantifying the degree of relevance of the literature research field as follows: Step 4.1, calculate the domain relevance between documents, calculate the Euclidean distance from the enhanced document node set as the predicted value of the quantitative method of the degree of association of document research fields, and the calculation method of the Euclidean distance between each pair of document vectors is as follows: ； in, and are the coordinates of two points in space in the i-th dimension, represents the dimension of the document vector, It is the predicted value after quantifying the degree of association between documents; Step 4.2: Compare the predicted value with the preset label, introduce the accuracy with error tolerance as the evaluation indicator, and evaluate the degree of relevance of literature research fields under different precisions by setting the error threshold; use Indicates the preset tag value. represents the predicted value, ε represents the error range, and the correlation degree of the literature research field , the calculation method is as follows: ； Where N is the total number of samples, and I is the indicator function, which is used to check whether the predicted value is within the allowable error range. and When the error of falls within the range of ε, it is considered a correct prediction; Step 4.3, repeat steps 4.1 and 4.2 until the model converges. By continuously iteratively training and updating the feature vectors of document nodes, the accuracy and effectiveness of the quantitative method for the degree of association in the research field of documents can be improved.

8. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for quantifying the degree of association in literature research fields based on multidimensional feature fusion as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps in the method for quantifying the degree of association in a literature research field based on multi-dimensional feature fusion as described in any one of claims 1 to 7 are implemented.

Images