Retrieval and tracking method and device for scientific research information

Abstract

The invention discloses a method and a device for retrieving and tracking scientific research information. The method comprises the following steps: S1, acquiring retrieval keyword information; s2, performing retrieval processing on the retrieval keyword information to obtain preliminary retrieval information; s3, performing tracking retrieval processing on the preliminary retrieval information to obtain a retrieval tracking information set; the S3 comprises the steps of setting a retrieval time interval; in each retrieval time interval, performing semantic expansion on the retrieval keyword information by utilizing a large model tool to obtain first expanded keyword information; performing structure expansion on the retrieval keyword information by utilizing a knowledge graph to obtain second expanded keyword information; based on a scientific research literature database, performing retrieval processing on the first expansion keyword information and the second expansion keyword information to obtain a tracking literature information set; and carrying out fusion processing on the tracking document information set and the preliminary retrieval information to obtain a retrieval tracking information set in the retrieval time interval.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of text data processing, graph data network processing and artificial intelligence technology, and in particular to a scientific research information retrieval and tracking method and device. BACKGROUND

[0002] In scientific research activities, timely and accurate acquisition and tracking of relevant field scientific research information is the basis for carrying out research work. With the development of information technology, scientific research literature data presents an explosive growth, and the traditional scientific research information retrieval method gradually exposes many deficiencies.

[0003] Currently, scientific researchers mainly retrieve in academic databases through keyword matching, but this method highly depends on the accuracy and comprehensiveness of the initial keywords, and is prone to missing due to keyword setting limitations. At the same time, the terminology in the field of scientific research has rich semantic associations, such as synonyms, hypernyms, and application scenario words, and simple keyword matching cannot cover these semantic associations, resulting in insufficient relevance and completeness of the retrieval results. Traditional database retrieval relies on precise keywords, and is prone to miss relevant literature due to terminology differences (such as "deep learning" and "deep neural network").

[0004] In addition, scientific research information has the characteristics of dynamic updating, and new research results and literature will continue to emerge, and scientific researchers need to track the progress of the field regularly. However, the existing tracking methods mostly rely on manual repeated retrieval, which is not only time-consuming and laborious, but also difficult to ensure the timeliness and continuity of tracking. Although some tools introduce simple keyword expansion functions, they lack deep mining of semantic associations and knowledge structures, and the expansion effect is limited, and still cannot meet the needs of scientific researchers for comprehensive and accurate tracking of scientific research information. SUMMARY

[0005] The present application mainly solves the problem of how to accurately and timely retrieve and track the relevant literature information of scientific research keywords based on semantic information association, and discloses a scientific research information retrieval and tracking method and device.

[0006] In a first aspect, a scientific research information retrieval and tracking method is disclosed, comprising:

[0007] S1, acquiring retrieval keyword information;

[0008] S2, performing retrieval processing on the retrieval keyword information to obtain preliminary retrieval information;

[0009] S3, performing tracking retrieval processing on the preliminary retrieval information to obtain a set of retrieval tracking information. The retrieval processing on the retrieval keyword information to obtain preliminary retrieval information comprises:

[0010] S21, using a scientific literature database, performing retrieval processing on the retrieval key information to obtain a retrieval literature information set;

[0011] S22, performing first mapping processing on the retrieval literature information set to obtain a first matrix set;

[0012] S23, performing second mapping processing on the retrieval literature information set to obtain a second matrix set;

[0013] S24, performing fusion clustering on the first matrix set and the second matrix set to obtain a fusion matrix set;

[0014] S25, performing vector text conversion processing on the fusion matrix set to obtain preliminary retrieval information.

[0015] The tracking retrieval processing on the preliminary retrieval information obtains a retrieval tracking information set, comprising:

[0016] S31, setting a retrieval time interval;

[0017] S32, in each retrieval time interval, using a large model tool to perform semantic expansion on the retrieval keyword information to obtain first expanded keyword information;

[0018] S33, using a knowledge graph to perform structural expansion on the retrieval keyword information to obtain second expanded keyword information;

[0019] S34, based on a scientific literature database, performing retrieval processing on the first expanded keyword information and the second expanded keyword information to obtain a tracking literature information set;

[0020] S35, performing fusion processing on the tracking literature information set and the preliminary retrieval information to obtain a retrieval tracking information set in the retrieval time interval;

[0021] S36, in each retrieval time interval, performing S32 to S35 to obtain a retrieval tracking information set for each retrieval time interval.

[0022] The large model tool is used to perform semantic expansion on the retrieval keyword information to obtain first expanded keyword information, comprising:

[0023] S321, using a context understanding module of a large model tool to generate word information related to the semantic of the retrieval keyword information;

[0024] S322, encoding the retrieval keyword information and the related word information into vector information respectively through a word encoding method;

[0025] S323, cluster all the vector information to obtain category information containing the largest number of vector information;

[0026] S324, convert all the vector information of the category information into vector text to obtain first extended keyword information.

[0027] The structure expansion of the search keyword information by using the knowledge graph to obtain second extended keyword information, comprising:

[0028] S331, using a path traversal algorithm, starting from the entity corresponding to the search keyword information in the knowledge graph, traversing to obtain an entity set within N1-hop relationships;

[0029] S332, filtering the traversed entity set to obtain second extended keyword information.

[0030] The filtering of the traversed entity set to obtain second extended keyword information, comprising:

[0031] S3321, using a knowledge graph embedding model, converting the entity corresponding to the search keyword information in the knowledge graph into a standard entity vector, and converting the relationship information set of the entity corresponding to the search keyword information in the knowledge graph into a standard relationship matrix; Each row vector of the standard relationship matrix is obtained by converting each relationship information of the relationship information set of the entity;

[0032] S3322, using a knowledge graph embedding model, converting the traversed entity set to obtain a traversal entity matrix, and converting the relationship information set of each entity of the traversed entity set to obtain a corresponding traversal relationship matrix;

[0033] S3323, filtering the standard entity vector, the standard relationship matrix, the traversal entity matrix and the traversal relationship matrix to obtain a filtered entity set;

[0034] S3324, using the filtered entity set and the search keyword information to construct second extended keyword information.

[0035] The filtering of the standard entity vector, the standard relationship matrix, the traversal entity matrix and the traversal relationship matrix to obtain a filtered entity set, comprising:

[0036] Subtracting each row vector of the standard entity vector and the traversal entity matrix to obtain a corresponding difference vector;

[0037] Using each difference vector as a row vector to construct a difference matrix C;

[0038] performing multivariate variational modal decomposition transformation on the row vectors of the difference matrix to obtain corresponding transformation vectors;

[0039] constructing a transformation matrix F by using each transformation vector as a row vector;

[0040] performing feature calculation on the transformation matrix and the difference matrix to obtain a feature matrix T;

[0041] The expression of the feature calculation is:

[0042] T = C 1 / 2 FC -1 / 2 ,

[0043] wherein T is the feature matrix;

[0044] calculating rank value α, norm value β, core norm γ and singular value from the feature matrix T;

[0045] judging whether each singular value satisfies a first discrimination inequality to obtain a first discrimination result;

[0046] The expression of the first discrimination inequality is:

[0047]

[0048] wherein N is the row dimension of the feature matrix T;

[0049] constructing a first screening entity set by using entities of row vectors corresponding to singular values whose first discrimination results are yes;

[0050] performing deep screening processing on a traversal relationship matrix and a standard relationship matrix corresponding to the first screening entity set to obtain a screening entity set.

[0051] In a second aspect of the embodiment of the present application, a scientific research information retrieval and tracking device is disclosed, which comprises:

[0052] a memory storing executable program codes;

[0053] a processor coupled with the memory;

[0054] the processor invokes the executable program codes stored in the memory to execute the scientific research information retrieval and tracking method.

[0055] In a third aspect of the embodiment of the present application, a computer storage medium is disclosed, which stores computer instructions, and the computer instructions are invoked by a computer to execute the scientific research information retrieval and tracking method.

[0056] In a fourth aspect of this invention, an information data processing terminal is disclosed, which is used to implement the aforementioned method for retrieving and tracking scientific research information.

[0057] The beneficial effects of this invention are as follows:

[0058] This invention constructs an integrated scientific research information retrieval and tracking system by combining scientific literature databases, large-scale model tools, and knowledge graphs, effectively solving the problems of keyword limitations and insufficient semantic association mining in traditional retrieval methods. By semantically and structurally expanding the search keywords, this invention can cover a wider range of related terms and knowledge associations, significantly improving the comprehensiveness of the retrieved information and reducing the number of missed documents due to insufficient keyword settings.

[0059] This invention converts retrieved literature information into matrix form through first and second mapping processes, and then performs fusion clustering, achieving in-depth integration and structured processing of literature information. This makes the initial retrieved information more systematic and relevant, facilitating researchers to quickly grasp the core content of the field.

[0060] This invention sets retrieval time intervals and automatically performs tracking retrieval. Combining the semantic understanding capabilities of large models and the structured knowledge of knowledge graphs, it dynamically generates extended keywords, realizing automated and periodic tracking of scientific research information. This ensures the timeliness and continuity of the tracked information and reduces the manual tracking burden on researchers.

[0061] This invention introduces vector encoding and clustering processing in semantic expansion, and uses knowledge graph embedding models and multi-dimensional matrix filtering in structural expansion to ensure a high degree of relevance between the expanded keywords and the original keywords, avoid interference from irrelevant information, and improve the accuracy of tracking and retrieval.

[0062] This invention achieves the accumulation and updating of scientific research information by integrating and processing tracking literature information with preliminary search information, forming a complete set of search and tracking information. It provides researchers with full-process information support from initial search to dynamic tracking, which helps to improve scientific research efficiency and innovation capabilities. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention. Detailed Implementation

[0064] To better understand the content of this invention, an embodiment is provided here.

[0065] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention.

[0066] In a first aspect, this invention discloses a method for retrieving and tracking scientific research information, comprising:

[0067] S1, Obtain search keyword information;

[0068] S2, perform retrieval processing on the search keyword information to obtain preliminary retrieval information;

[0069] S3, perform tracking retrieval processing on the preliminary retrieval information to obtain a set of retrieval tracking information;

[0070] The process of retrieving the keyword information to obtain preliminary retrieval information includes:

[0071] S21, Using a scientific research literature database, the key information to be retrieved is processed to obtain a set of retrieved literature information;

[0072] S22, Perform a first mapping process on the retrieved document information set to obtain a first matrix set;

[0073] S23, perform a second mapping process on the retrieved document information set to obtain a second matrix set;

[0074] S24, perform fusion clustering on the first matrix set and the second matrix set to obtain a fusion matrix set;

[0075] S25, Perform vector text conversion processing on the fusion matrix set to obtain preliminary retrieval information;

[0076] Both the first and second mapping processes are implemented through a method of converting text to graph data, including:

[0077] S2201, Preprocess the retrieved document information set to obtain a preprocessed text dataset;

[0078] S2202, Entity extraction processing is performed on the data subset of each document in the preprocessed text dataset to obtain the corresponding entity node set;

[0079] S2203, perform relation extraction processing on the data subset of each document in the preprocessed text dataset to obtain the corresponding set of edge nodes;

[0080] S2204, Based on the set of entity nodes and the set of edge nodes, construct the graph object data corresponding to the document;

[0081] S2205, Perform Euclidean space network embedding representation on the graph object data to obtain the first matrix corresponding to the document; Perform hyperbolic space network embedding representation on the graph object data to obtain the second matrix corresponding to the document;

[0082] S2206. Using the first matrix of all documents, construct the first matrix set; using the second matrix of all documents, construct the second matrix set.

[0083] The step of fusing and clustering the first matrix set and the second matrix set to obtain a fused matrix set includes:

[0084] A set of clustering method models is constructed; the set of clustering method models includes K-means clustering method, spectral clustering method and hierarchical clustering method;

[0085] A set of clustering methods is used to cluster the first matrix set to obtain a first clustering information set; the first clustering information includes the label information of the literature corresponding to each first matrix under the operation of three clustering methods;

[0086] The method based on the co-association matrix is ​​used to fuse the information set of the first cluster to obtain the first fused label information of the documents corresponding to each first matrix;

[0087] A set of clustering methods is used to cluster the second matrix set to obtain a second clustering information set; the second clustering information includes the label information of the literature corresponding to each second matrix obtained under the operation of three clustering methods;

[0088] The method based on the co-association matrix is ​​used to fuse the second cluster information set to obtain the second fusion label information of the documents corresponding to each second matrix;

[0089] The meta-feature clustering method is used to cluster and fuse the first and second fusion label information of all documents to obtain the fusion label information of all documents;

[0090] By using the first or second matrix of documents belonging to the same fusion tag information, a set of fusion matrices is constructed;

[0091] The process of fusing and clustering the first matrix set and the second matrix set to obtain a fused matrix set improves the robustness and anti-interference of single-space clustering by using a multi-clustering method.

[0092] The process first constructs a set of clustering method models including K-means, spectral clustering, and hierarchical clustering, and then performs clustering on the first matrix set (Euclidean space embedding) and the second matrix set (hyperbolic space embedding) respectively. K-means is good at handling linearly separable data with spherical distribution and is suitable for capturing the local density of document features; spectral clustering is more sensitive to nonlinear and complex topological structures and can identify implicit associations between documents; hierarchical clustering can preserve the hierarchical relationship between documents (such as the progression of "basic theory - derived methods - application scenarios"), and is suitable for reflecting the evolutionary logic of research topics.

[0093] By combining multiple clustering methods, the inherent limitations of single clustering algorithms are avoided (such as K-means' sensitivity to initial centers and spectral clustering's sensitivity to noise). Subsequent fusion of clustering results from the same matrix set based on the co-association matrix (statistically counting the frequency of samples being grouped into the same category in different clusters) filters out random errors, strengthens consensus information, and ultimately yields first / second fused labels that more closely approximate the true clustering structure in the literature, improving the robustness of clustering within a single space.

[0094] By integrating the complementary features of Euclidean and hyperbolic spaces, the comprehensiveness of clustering is enhanced. The first matrix set is generated based on Euclidean space network embedding, which excels at capturing local linear correlations of document features (such as correlations between documents with similar research methods or identical experimental data). The second matrix set is generated based on hyperbolic space network embedding, which is more suitable for characterizing global hierarchical structures (such as the conceptual hierarchy of documents from "machine learning" to "deep learning" to "Transformer"). By clustering and fusing the two types of matrix sets separately, the patterns of documents under different spatial characteristics can be fully explored: Euclidean space clustering focuses on "local similarity," while hyperbolic space clustering focuses on "global hierarchy." The two types of fusion labels respectively carry different dimensional features of the documents, laying the foundation for subsequent cross-space fusion.

[0095] Meta-feature clustering achieves cross-spatial information complementarity, improving the accuracy of the final clustering. After obtaining the first fusion label (Euclidean space features) and the second fusion label (hyperbolic space features), meta-feature clustering is used to perform a second fusion of the two. Meta-feature clustering integrates based on the distribution characteristics of the labels themselves (such as the overlap, difference, and hierarchical correlation of the labels), rather than directly clustering the original vectors. This can more efficiently capture the complementarity of the two types of labels: for documents with local similarities but belonging to different conceptual levels (such as "CNN image classification" and "Transformer image classification"), meta-feature clustering can preserve their local correlation in "image classification" while distinguishing their hierarchical differences in "deep learning model evolution"; for documents with different local features but belonging to the same core theme (such as "reinforcement learning algorithm design" and "application of reinforcement learning in robot control"), meta-feature clustering can integrate their global correlation in "reinforcement learning" to avoid misclassification due to differences in local features.

[0096] The final generated fusion tags can simultaneously take into account the local similarity and global hierarchy of the documents, making the document clustering results corresponding to the fusion matrix set more in line with the internal logic of scientific research information (such as topic association, method evolution, application scenarios, etc.).

[0097] The fusion matrix set, obtained through vector text conversion, yields preliminary search information with more accurate and comprehensive clustering results. This significantly reduces interference from irrelevant literature while preserving the connections between key documents (such as clusters of similar documents and the evolution of themes). This enables researchers to quickly locate core document clusters during searches and to trace research trends through clustering structures, thereby improving the efficiency and depth of information retrieval.

[0098] In summary, this fusion clustering process effectively overcomes the limitations of single clustering methods and single spatial embedding through a three-layer strategy of "enhancing the robustness of a single space through multiple methods, capturing complementary features across spaces, and improving accuracy through meta-feature fusion." Ultimately, it achieves a unity of robustness, comprehensiveness, and accuracy in scientific literature clustering, laying a core foundation for high-quality scientific information retrieval and tracking.

[0099] The process of performing vector-to-text conversion on the fusion matrix set to obtain preliminary retrieval information involves converting the first or second matrix of documents with the same fusion tag information into text data to obtain the preliminary fusion retrieval information of the documents. Using the preliminary fusion retrieval information of all documents, preliminary retrieval information is constructed.

[0100] The preliminary retrieval information is then processed to obtain a set of retrieval tracking information, including:

[0101] S31, Set the search time interval;

[0102] S32, within each retrieval time interval, the retrieval keyword information is semantically expanded using a large model tool to obtain first expanded keyword information;

[0103] S33, Using a knowledge graph, the structure of the search keyword information is expanded to obtain second expanded keyword information;

[0104] S34, Based on the scientific literature database, the first extended keyword information and the second extended keyword information are retrieved and processed to obtain a set of tracked literature information;

[0105] S35, the tracking document information set and the preliminary search information are fused to obtain the search tracking information set within the search time interval;

[0106] S36. For each retrieval time interval, execute S32 to S35 to obtain the retrieval tracking information set for each retrieval time interval.

[0107] The use of large model tools to semantically expand the retrieved keyword information to obtain first expanded keyword information includes:

[0108] S321 utilizes the contextual understanding module of the large model tool to generate word information (including synonyms, hypernyms, and application scenario words) that are semantically related to the retrieved keyword information.

[0109] S322, the search keyword information and related word information are encoded into vector information using word encoding methods;

[0110] S323, perform clustering on all vector information to obtain the category information containing the largest number of vector information;

[0111] S324, Perform vector text conversion on all vector information of the category information to obtain the first extended keyword information;

[0112] The method of using large model tools to semantically expand the retrieved keyword information to obtain first expanded keyword information includes:

[0113] Call open-source LLMs (such as Llama 3 or Qwen) and generate an expanded word candidate set by inputting prompt words.

[0114] Sentence-BERT was used to calculate the semantic similarity between candidate words and original keywords (threshold ≥ 0.7), and irrelevant words were filtered out.

[0115] The method of using a knowledge graph to structurally expand the search keyword information to obtain second expanded keyword information includes:

[0116] S331, using a path traversal algorithm, starting from the entity corresponding to the search keyword information in the knowledge graph, traverse to obtain the entity set within N1 jump relationships (such as "machine learning" → "deep learning" → "convolutional neural network").

[0117] S332, the entity set obtained by the traversal is filtered to obtain the second extended keyword information.

[0118] N1 is a positive integer greater than or equal to 2;

[0119] The path traversal algorithm can be either the Floyd algorithm or the Dijkstra algorithm.

[0120] The step of filtering the entity set obtained from the traversal to obtain the second expanded keyword information includes:

[0121] Using a knowledge graph embedding model, the entities corresponding to the search keywords in the knowledge graph are converted into standard entity vectors, and the set of relation information of the entities corresponding to the search keywords in the knowledge graph is converted into a standard relation matrix; each row vector of the standard relation matrix is ​​obtained by converting each relation information of the entity's relation information set.

[0122] Using a knowledge graph embedding model, the entity set obtained by traversal is transformed to obtain a traversal entity matrix. The relation information set of each entity in the entity set obtained by traversal is transformed to obtain the corresponding traversal relation matrix.

[0123] The standard entity vector, standard relation matrix, traversed entity matrix, and traversed relation matrix are filtered to obtain a filtered entity set;

[0124] Using the filtered entity set and retrieved keyword information, a second set of expanded keyword information is constructed.

[0125] The step of using a knowledge graph embedding model to transform the entity set obtained by traversal to obtain a traversal entity matrix involves transforming each entity in the entity set obtained by traversal to obtain a corresponding row vector, and using all the row vectors to construct the traversal entity matrix.

[0126] The process of transforming the relation information set of each entity in the traversed entity set to obtain the corresponding traversal relation matrix includes:

[0127] For each relation information set of each entity in the traversed entity set, a transformation process is performed to obtain the corresponding row vector. Using all the row vectors, the corresponding traversal relation matrix is ​​constructed.

[0128] The filtering process of the standard entity vector, standard relation matrix, traversed entity matrix, and traversed relation matrix yields a filtered entity set, including:

[0129] For the standard entity vector, subtract it from each row vector of the traversed entity matrix to obtain the corresponding difference vector;

[0130] Using each difference vector as a row vector, we construct the difference matrix C;

[0131] Perform multivariate variational mode decomposition transformation on the row vectors of the difference matrix to obtain the corresponding transformation vectors;

[0132] Using each transformation vector as a row vector, we construct the transformation matrix F;

[0133] The transformation matrix and the difference matrix are subjected to feature calculation to obtain the feature matrix T;

[0134] The expression for calculating the feature is:

[0135] T = C 1 / 2 FC -1 / 2 ,

[0136] Where T is the characteristic matrix;

[0137] The rank α, norm β, nuclear norm γ, and singular values ​​are calculated from the characteristic matrix T.

[0138] Determine whether each singular value satisfies the first discrimination inequality to obtain the first discrimination result;

[0139] The expression for the first discriminant inequality is:

[0140]

[0141] Where N is the row dimension of the feature matrix T;

[0142] The first set of filtered entities is constructed by using the entities of the row vectors corresponding to all singular values ​​whose first discrimination result is yes;

[0143] The traversal relation matrix and standard relation matrix corresponding to the first set of filtered entities are subjected to deep filtering to obtain the set of filtered entities.

[0144] The singular values ​​are obtained by performing singular value decomposition on the feature matrix T. Each singular value corresponds to a row vector of the feature matrix T.

[0145] The multivariate variational mode decomposition transformation is the MVMD transformation;

[0146] This invention achieves precise screening of singular values ​​in the feature matrix T through a first discriminant inequality. This formula combines the rank, norm, nuclear norm, and row dimension of the feature matrix, constructing a quantitative discrimination standard through trigonometric functions and proportional relationships. This effectively distinguishes the semantic and structural correlation strength between traversed entities and the entities corresponding to the original keywords. The singular values ​​reflect the importance of entity features. The threshold comprehensively considers the dimensional and structural features of the matrix, avoiding over- or under-screening issues caused by a single threshold. Entities corresponding to the singular values ​​screened by this inequality retain entities with significant features while maintaining semantic relevance to the original keywords. This provides a high-quality candidate set for subsequent structural expansion, reduces interference from irrelevant entities in retrieval tracking, and improves the accuracy of structural expansion.

[0147] The step of performing deep filtering on the traversal relation matrix and standard relation matrix corresponding to the first filtered entity set to obtain the filtered entity set includes:

[0148] For each entity in the first set of filtered entities, the traversal relation matrix corresponding to it is compared with the standard relation matrix, and the nuclear norm difference is calculated to obtain the first difference value C1 for each entity.

[0149] For each entity in the first set of filtered entities, the traversal relation matrix corresponding to it is compared with the standard relation matrix to perform a fusion difference calculation, and the second difference value C2 corresponding to each entity is obtained.

[0150] The expression for calculating the fusion difference is:

[0151]

[0152] Where N2 and M2 are the row and column dimensions of the standard relation matrix, respectively. and A represents the i-th element of the first difference vector and the second difference vector, respectively. ij and B ij These are the elements in the i-th row and j-th column of the traversal relation matrix corresponding to the standard relation matrix and the entity, respectively.

[0153] Determine whether the product of the first and second difference values ​​of each entity is less than a preset difference discrimination threshold to obtain the second discrimination result;

[0154] By using all entities whose second discrimination result is yes, a set of filtered entities is constructed.

[0155] The difference discrimination threshold can be 3.5.

[0156] The nuclear norm difference calculation involves calculating the nuclear norm of the difference matrix between two matrices.

[0157] This invention achieves refined quantification of the differences between the standard relation matrix and the traversal relation matrix by fusing difference calculation expressions. Specifically, the first and second difference vectors extract the extreme eigenvalues ​​of each row in the standard and traversal relation matrices, respectively, highlighting significant differences between the matrices; while... This is used to normalize the difference ratio, avoiding misjudgments caused by differences in the magnitude of matrix elements. Combined with the log2 operation for non-linear amplification of the difference ratio, it enhances the discriminative power of significant differences while suppressing interference from minor fluctuations. By summing the differences in each row, the C2 value comprehensively reflects the degree of difference between the two matrices in their overall structure. This formula, combined with the first difference value C1 (nuclear norm difference), forms a multi-dimensional difference assessment system. This allows deep screening to capture both the overall structural differences of the matrices and identify significant changes in local features, ensuring that the screened entities highly match the original keywords in terms of relational structure. This further enhances the relevance and reliability of the second extended keyword information, providing more accurate keyword support for tracking and retrieval.

[0158] The row vector of the traversed entity matrix is ​​obtained by transforming each entity in the entity set obtained by the traversal;

[0159] Each row vector of the traversal relation matrix is ​​obtained by transforming each relation information of the relation information set of the entities in the entity set obtained by the traversal;

[0160] The knowledge graph embedding model (KGE) includes TransE, RotatE, etc.;

[0161] The scientific literature databases mentioned include CNKI, Wanfang, VIP, IEEE, etc.

[0162] The process of fusing the tracked document information set and the preliminary search information to obtain the search tracking information set within the search time interval includes:

[0163] The tracking document information set and the preliminary search information are semantically aligned to obtain the first fused text information;

[0164] The first fused text information is subjected to content fusion processing to obtain a set of retrieval tracking information within the retrieval time interval.

[0165] The content fusion process involves fusing the method information, conclusion information, background information, and paper abstract information from the tracked literature information set and the preliminary search information to obtain the search tracking information set. The fusion method can employ attention mechanism weighted fusion or GNN fusion method.

[0166] In the calculation expressions of this invention, the variables involved have all been dimensionless before calculation.

[0167] A second aspect of the present invention discloses a scientific research information retrieval and tracking device, the device comprising:

[0168] Memory containing executable program code;

[0169] A processor coupled to the memory;

[0170] The processor calls the executable program code stored in the memory to execute the scientific research information retrieval and tracking method.

[0171] In a third aspect, the present invention discloses a computer-storable medium storing computer instructions, which, when invoked by a computer, are used to execute the scientific research information retrieval and tracking method.

[0172] In a fourth aspect of this invention, an information data processing terminal is disclosed, which is used to implement the aforementioned method for retrieving and tracking scientific research information.

[0173] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method of searching and tracking scientific research information, characterized by, The method comprises the following steps: S1, obtaining search keyword information; S2, performing search processing on the search keyword information to obtain preliminary search information; S3, performing tracking search processing on the preliminary search information to obtain a search tracking information set.

2. The method of claim 1, wherein the research information is classified into a plurality of categories, and the research information is searched and tracked according to the categories. The search processing on the search keyword information to obtain preliminary search information comprises: S21, performing search processing on the search keyword information by using a scientific research literature database to obtain a search literature information set; S22, performing first mapping processing on the search literature information set to obtain a first matrix set; S23, performing second mapping processing on the search literature information set to obtain a second matrix set; S24, performing fusion clustering on the first matrix set and the second matrix set to obtain a fusion matrix set; S25, performing vector text conversion processing on the fusion matrix set to obtain preliminary search information.

3. The method of claim 1, wherein the research information is retrieved from a database. The tracking search processing on the preliminary search information to obtain a search tracking information set comprises: S31, setting a search time interval; S32, in each search time interval, performing semantic expansion on the search keyword information by using a large model tool to obtain first expanded keyword information; S33, performing structural expansion on the search keyword information by using a knowledge graph to obtain second expanded keyword information; S34, performing search processing on the first expanded keyword information and the second expanded keyword information based on a scientific research literature database to obtain a tracking literature information set; S35, performing fusion processing on the tracking literature information set and the preliminary search information to obtain a search tracking information set in the search time interval; S36, in each search time interval, performing S32 to S35 to obtain a search tracking information set of each search time interval.

4. The method of claim 3, wherein the search information is a search result of a search for a keyword. The semantic expansion on the search keyword information by using a large model tool to obtain first expanded keyword information comprises: S321, generating keyword information related to the semantic of the search keyword information by using a context understanding module of the large model tool; S322, encoding the search keyword information and the related keyword information into vector information by using a word encoding method respectively; S323, performing clustering processing on all the vector information to obtain category information containing the maximum number of vector information; S324, performing vector text conversion on all the vector information of the category information to obtain first expanded keyword information.

5. The method for retrieving and tracking scientific research information as described in claim 3, characterized in that, The structural expansion on the search keyword information by using a knowledge graph to obtain second expanded keyword information comprises: S331, using a path traversal algorithm to start from an entity corresponding to the search keyword information in the knowledge graph, and traversing to obtain an entity set within N1-hop relationships; S332, performing screening processing on the obtained entity set to obtain second expanded keyword information.

6. The method of claim 5, wherein the search information is a search result of a search for a keyword. The screening processing on the obtained entity set to obtain second expanded keyword information comprises: S3321, using a knowledge graph embedding model, converting the entity corresponding to the search keyword information in the knowledge graph into a standard entity vector, and converting the relationship information set of the entity corresponding to the search keyword information in the knowledge graph into a standard relationship matrix; each row vector of the standard relationship matrix is obtained by converting each relationship information in the relationship information set of the entity; S3322, using a knowledge graph embedding model, converting the obtained entity set to obtain a traversal entity matrix, and converting the relationship information set of each entity in the obtained entity set to obtain a corresponding traversal relationship matrix; S3323, performing screening processing on the standard entity vector, the standard relationship matrix, the traversal entity matrix, and the traversal relationship matrix to obtain a screened entity set; S3324, using the screened entity set and the search keyword information to construct second expanded keyword information.

7. The method for retrieving and tracking scientific research information as described in claim 6, characterized in that, The screening processing on the standard entity vector, the standard relationship matrix, the traversal entity matrix, and the traversal relationship matrix to obtain a screened entity set comprises: Subtracting each row vector of the standard entity vector and the traversal entity matrix respectively to obtain a corresponding difference vector; Using each difference vector as a row vector to construct a difference matrix C; Performing multivariate variational modal decomposition transformation on the row vector of the difference matrix to obtain a corresponding transformation vector; Using each transformation vector as a row vector to construct a transformation matrix F; Performing feature calculation on the transformation matrix and the difference matrix to obtain a feature matrix T; The expression of the feature calculation is: T = C 1 / 2 FC -1 / 2 , Wherein, T is a feature matrix; Calculating the rank value α, the norm value β, the kernel norm γ, and the singular value of the feature matrix T; Determining whether each singular value satisfies a first discriminant inequality to obtain a first discriminant result; The expression of the first discriminant inequality is: Wherein, N is the row dimension of the feature matrix T; Using the entities of the row vectors corresponding to the singular values whose first discriminant results are yes to construct a first screened entity set; Performing deep screening processing on the traversal relationship matrix and the standard relationship matrix corresponding to the first screened entity set to obtain a screened entity set.

8. A scientific research information retrieval and tracking device, characterized in that, The device comprises: a memory storing executable program codes; a processor coupled with the memory; the processor invokes the executable program codes stored in the memory to execute the scientific research information retrieval tracking method according to any one of claims 1 to 7.

9. A computer storable medium, characterized by The computer storage medium stores computer instructions, which are invoked by a computer to execute the scientific research information retrieval tracking method according to any one of claims 1 to 7.

10. An information data processing terminal, characterized by The information data processing terminal is used to implement the scientific research information retrieval tracking method according to any one of claims 1 to 7.

Images