Scientific research project evaluation expert precise matching method based on big data of science and technology

Abstract

The present application relates to the technical field of data processing, and discloses a scientific research project review expert precise matching method based on scientific and technological big data, comprising: extracting characteristic elements of a target expert's historical business text and a project declaration text to construct an orthogonal semantic tensor; calculating the correlation density of the characteristic elements to generate a metric tensor matrix to establish a non-Euclidean manifold; calculating the Christoffel symbol and the covariant acceleration in combination with the metric tensor matrix to obtain a predicted evolution trajectory of the expert during project execution; constructing an implicit reciprocity network graph based on historical interaction data between nodes, and removing candidate experts with interest associations by identifying strongly connected components; setting a manifold radius in combination with the semantic information entropy of the declaration, calculating the instantaneous geodesic distance to construct an instantaneous collision strength function; and calculating the definite integral of the function to select the final expert according to the space-time collision integral value. The present application solves the problems of matching lag and interest conflict.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method for accurately matching experts for scientific research project review based on big data. Background Technology

[0002] Research project review is a core component of science and technology resource allocation, and the quality of the selection of review experts directly impacts the project's approval rate. With the continuous accumulation of big data in science and technology, using information technology to automatically match review experts through text similarity calculations has become a common technique in this field.

[0003] Existing expert matching systems typically only extract experts' historical papers or past project texts for static keyword comparison. This mechanism fixes the experts' research characteristics at a past point in time, failing to consider the dynamic evolution of researchers' research directions over time. Since research projects are usually executed within a future period, matching based on static historical data will result in a time lag, causing the research direction of the selected experts during the actual project execution period to deviate from the project's forward-looking needs.

[0004] Existing expert recusal systems primarily rely on explicit recusal lists manually submitted by applicants or basic intra-institutional recusal rules. This conventional filtering method fails to uncover hidden conflicts of interest within massive amounts of historical collaboration and review data. In practice, some experts and applicants form implicit, cross-institutional networks of mutual benefit through long-term co-signing or historical review ratings. Current technology lacks the means to detect and measure such complex data topologies, making it impossible to identify and eliminate expert nodes within closed-loop benefit-transfer relationships, thus hindering the objectivity and fairness of the review process.

[0005] With the development of technology, modern scientific research projects often involve the intersection of multiple disciplines, placing different demands on the breadth of technical expertise required of review experts. Conventional matching algorithms mostly use fixed similarity thresholds for single-point distance calculations, failing to transform the interdisciplinary breadth of a specific project into a dynamic matching space, nor quantifying the continuous overlap between expert research directions and project needs throughout the entire execution cycle on a continuous time scale. This rigid matching logic is prone to local biases and cannot accurately assess the comprehensive suitability of experts throughout the entire lifecycle of a research project. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a precise matching method for scientific research project review experts based on big data. This method solves the technical problems in existing technologies, such as the time lag in expert matching due to static feature comparison, the lack of topological measurement and avoidance of implicit conflicts of interest, and the inability to assess the dynamic fit between experts and scientific research projects throughout the entire lifecycle when using fixed threshold single-point matching.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for accurately matching experts for scientific research project review based on big data in science and technology, including the following steps: Step S100: Obtain the set of historical scientific research business texts of the target expert and the text of scientific research project application forms to be matched, extract feature elements and construct corresponding orthogonal semantic tensors through tensor product operation respectively; Step S200: Calculate the correlation density based on the joint co-occurrence probability and marginal distribution probability of feature elements, and construct a metric tensor matrix based on the correlation density to establish the non-Euclidean geometric manifold; Step S300: Combine Christopher symbol to calculate the covariant acceleration of the target expert, and use the covariant acceleration to calculate the corrected feature coordinates in order to obtain the expert time series prediction evolution trajectory during the execution of the scientific research project. Step S400: Construct an implicit reciprocity network graph. By identifying strongly connected components in the implicit reciprocity network graph, candidate review experts with conflicting interests are eliminated, and candidate review expert nodes that pass the topological independence detection are obtained. Step S500: Set the manifold radius based on semantic information entropy and calculate the instantaneous geodesic distance, construct the instantaneous collision intensity function and perform definite integral calculation, and select scientific research project review experts based on the spatiotemporal collision integral value.

[0008] Furthermore, feature elements of historical research business texts and research project application texts under the orthogonal semantic dimension are extracted, including: converting historical research business texts and research project application texts into standardized text sequence data; establishing a dependency parsing and semantic role labeling model using a network architecture combining a pre-trained language model and a conditional random field, and identifying the central predicate and arguments associated with the central predicate in the text sequence data through the dependency parsing and semantic role labeling model; and mapping accompanying adverbial arguments representing the tool or manner of an action to research method dimension features, mapping patient arguments representing the recipient of an action to research object dimension features, and mapping adverbial arguments representing time, space, or conditional constraints to application environment dimension features.

[0009] Furthermore, a metric tensor matrix is ​​constructed based on the correlation tightness to establish the non-Euclidean geometric manifold, including: dividing the document frequency of feature elements appearing in the global feature matrix by the total number of texts to obtain the marginal distribution probability; dividing the document frequency of two feature elements appearing simultaneously in the same text by the total number of texts to obtain the joint co-occurrence probability; dividing the joint co-occurrence probability by the product of the marginal distribution probabilities of the two feature elements, taking the logarithm of the division result to obtain the point mutual information value, and using the point mutual information value as the correlation tightness; performing an exponential decay transformation on the point mutual information value, setting the basis of the underlying flat space as the Kronecker function, and adding the Kronecker function and the result of the exponential decay transformation to obtain the metric tensor matrix.

[0010] Furthermore, the covariant acceleration of the target expert is calculated by: using the finite difference method to obtain the partial derivatives of the tensor dimension coordinates of the metric tensor matrix, and combining the inverse matrix of the metric tensor matrix with the partial derivative results to obtain the Christopher symbol; obtaining the evolution velocity vector by taking the first derivative of the expert time series evolution trajectory with respect to the time parameter, adding the Christopher symbol to the product of the dimension components of the evolution velocity vector, and summing the sum with the ordinary time derivative of the evolution velocity vector to obtain the covariant acceleration.

[0011] Furthermore, the corrected feature coordinates are calculated using covariant acceleration, including: determining the time difference based on the time node of the research project application text; and accumulating the current research feature coordinates of the target expert, the product of the evolution velocity vector and the time difference, and half of the covariant acceleration and the square of the time difference to obtain the corrected feature coordinates.

[0012] Further, the edge weights of directed edges are calculated to generate an adjacency weight matrix, including: establishing a mapping table between event types and basic weight values; assigning a first preset value to the basic weight value when the event type is a co-authored paper; and assigning a second preset value to the basic weight value when the event type is historical review data and the review score is higher than a preset excellent threshold; using a time decay function to perform weighted calculations on the time interval between each event occurrence and the current time between nodes, multiplying the weighted calculation result by the corresponding basic weight value, and accumulating the multiplication results of all events that have historical benefit output relationships between nodes to obtain the edge weights of directed edges; and generating an adjacency weight matrix based on all edge weights corresponding to each directed edge in the implicit reciprocal network graph.

[0013] Furthermore, candidate reviewers with vested interests are eliminated by identifying strongly connected components in the implicit reciprocity network graph. This includes: setting the average value of all non-zero edge weights in the adjacency weight matrix as the reciprocity strength threshold; retaining directed edges when the edge weight is greater than the reciprocity strength threshold, and deleting directed edges when the edge weight is less than or equal to the reciprocity strength threshold to obtain a pruned directed graph; traversing the pruned directed graph using a depth-first search algorithm, assigning an access order number and a traceability node number to each node during the traversal, and determining the subset of nodes with the same traceability node number as strongly connected components.

[0014] Furthermore, candidate review experts with conflicting interests are eliminated, including: when a candidate review expert node and the applicant node are located in the same strongly connected component, the candidate review expert node is determined to have failed the topological independence detection and the associated candidate review expert node is eliminated; when multiple candidate review expert nodes are located in the same strongly connected component, only the candidate review expert node with the smallest corresponding geodesic distance among the multiple candidate review expert nodes is retained, and the remaining candidate review expert nodes located in the same strongly connected component are eliminated.

[0015] Furthermore, an instantaneous collision intensity function is constructed, including: statistically analyzing the feature word frequencies of the research project application text in each dimension; dividing the feature word frequencies by the total word frequencies of the corresponding dimensions to obtain the feature probabilities; calculating the semantic information entropy of each dimension by summing the negative form of the product of the feature probabilities and the natural logarithms of the feature probabilities; summing the semantic information entropies of each dimension to obtain the total information entropy; setting the product of the preset regional diffusion coefficient and the total information entropy as the manifold radius; when the instantaneous geodesic distance is less than or equal to the manifold radius, dividing the negative of the square of the instantaneous geodesic distance by twice the square of the manifold radius to obtain the division result; taking the exponential function of the division result to obtain the instantaneous collision intensity function; when the instantaneous geodesic distance is greater than the manifold radius, setting the value of the instantaneous collision intensity function to zero.

[0016] Furthermore, the corresponding candidate review expert nodes are selected as research project review experts based on the spatiotemporal collision integral value. This includes: performing a definite integral operation on the instantaneous collision intensity function within the interval from the project start time stamp to the project completion time stamp corresponding to the research project application text to obtain the spatiotemporal collision integral value; sorting the candidate review expert nodes in descending order according to the spatiotemporal collision integral value, and extracting the top-ranked candidate review expert nodes in the descending order sequence whose number is equal to the threshold of the number of review experts, and outputting them as the final research project review experts.

[0017] This invention provides a method for accurately matching experts for scientific research project review based on big data in science and technology. It has the following beneficial effects: 1. This invention, by constructing a non-Euclidean geometric manifold and calculating the covariant acceleration of expert evolution trajectories, can predict the corrected feature coordinates of experts during the execution of research projects. This technique overcomes the lag problem caused by traditional methods relying solely on static historical text comparison. It can infer the future research direction of experts based on changes in their past research trajectories, thereby aligning the time benchmark for expert matching with the actual execution cycle of research projects and improving the alignment between the technical evolution direction of review experts and the forward-looking needs of projects.

[0018] 2. This invention extracts historical collaboration data and historical review data to construct an implicit reciprocal network graph, and uses a depth-first search algorithm to identify strongly connected components in the topology, thereby eliminating candidate experts with vested interests. This process combines a time decay function and event type to quantify the intensity of benefit transfer between nodes, objectively detecting expert groups with closed-loop reciprocal relationships from the underlying data structure. This blocks potential interference from personal relationships at the source, ensuring the independence and impartiality of the research project review list.

[0019] 3. This invention dynamically sets the radius of the required manifold region using the semantic information entropy of the research project application text, and performs interval definite integral calculation on the instantaneous collision intensity function to obtain the spatiotemporal collision integral value. This matching logic not only transforms the interdisciplinary breadth of the project into an objective spatial range parameter, but also quantifies the degree of overlap between the expert's research direction and the technical requirements throughout the entire project execution cycle through definite integral operations, achieving dynamic matching on a continuous time scale and avoiding local errors caused by single-point matching. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This invention provides a method for accurately matching experts for scientific research project review based on big data, including the following steps: Step S100: Obtain the set of historical scientific research business texts of the target expert and the text of scientific research project application forms to be matched, extract feature elements and construct corresponding orthogonal semantic tensors through tensor product operation respectively; Step S200: Calculate the correlation density based on the joint co-occurrence probability and marginal distribution probability of feature elements, and construct a metric tensor matrix based on the correlation density to establish the non-Euclidean geometric manifold; Step S300: Combine Christopher symbol to calculate the covariant acceleration of the target expert, and use the covariant acceleration to calculate the corrected feature coordinates in order to obtain the expert time series prediction evolution trajectory during the execution of the scientific research project. Step S400: Construct an implicit reciprocity network graph. By identifying strongly connected components in the implicit reciprocity network graph, candidate review experts with conflicting interests are eliminated, and candidate review expert nodes that pass the topological independence detection are obtained. Step S500: Set the manifold radius based on semantic information entropy and calculate the instantaneous geodesic distance, construct the instantaneous collision intensity function and perform definite integral calculation, and select scientific research project review experts based on the spatiotemporal collision integral value.

[0023] In one specific embodiment, the process of extracting orthogonal semantic tensors is described in detail, and the extraction of orthogonal semantic tensors specifically includes the following sub-steps: Step S110: Collect multi-source science and technology business texts. These texts originate from heterogeneous databases of various science and technology management departments. The historical research business text collection includes text data of research project applications, project completion and acceptance reports, published academic papers, and authorized patents previously participated in by the target expert. The system is configured with a data extraction, transformation, and loading tool to extract and clean the structured fields and unstructured text in the heterogeneous databases, removing stop words and invalid characters, and generating standardized text sequence data. For the underlying data flow control of data synchronization and text cleaning in heterogeneous databases, those skilled in the art can use a distributed message queue combined with a regular expression filtering script. The underlying data flow control is a well-known technology in this field and will not be elaborated here.

[0024] Step S120: Extract orthogonal dimension features. In this embodiment, a pre-trained language model combined with a conditional random field network architecture is used to construct a dependency parsing and semantic role labeling model. After inputting standardized text sequence data, the dependency parsing and semantic role labeling model identifies the central predicate and the arguments associated with the central predicate in the text sequence data. The system configures preset argument mapping rules to classify the identified arguments into three independent orthogonal semantic dimensions. In the argument mapping rules, accompanying adverbial arguments representing the tool or manner of an action are mapped to research method dimension features; patient arguments representing the recipient of an action are mapped to research object dimension features; and adverbial arguments representing time, space, or conditional constraints are mapped to... The application environment dimension features and research method dimension features cover the algorithm names, experimental techniques, and detection methods described in the text sequence data. The research object dimension features cover the names of the entities involved, disease models, and target substances described in the text sequence data. The application environment dimension features cover the application scenarios, boundary conditions, and experimental environment constraints described in the text sequence data. Through orthogonal decomposition, the features of the above three independent dimensions are extracted. For the calculation of the self-attention mechanism of the pre-trained language model and the sequence label decoding of the conditional random field, those skilled in the art can call existing open-source natural language processing framework libraries to build it. The calculation of the self-attention mechanism and the sequence label decoding are well-known technologies in this field and will not be described in detail here.

[0025] Step S130: Construct an orthogonal semantic tensor. Following the feature extraction in step S120, obtain feature vectors corresponding to three independent dimensions. Let the extracted feature vector for the research method dimension be... The extracted feature vector of the research object dimension is The extracted feature vector of the application environment dimension is Research method dimension feature vector The dimension size is , Research object dimension feature vector The dimension size is Application environment dimension feature vector The dimension size is Third-order orthogonal semantic tensor The calculation formula is: In the formula, Represents the tensor product operator, a third-order orthogonal semantic tensor. The spatial dimension is Third-order orthogonal semantic tensor Establish the fundamental coordinate representation of non-Euclidean geometric manifolds. Each historical research business text and research project application text in the historical research business text collection uniquely corresponds to a third-order orthogonal semantic tensor. In one specific embodiment, the process of constructing a non-Euclidean geometric manifold and a metric tensor is described in detail. The construction of a non-Euclidean geometric manifold and a metric tensor specifically includes the following sub-steps: Step S210: Calculate the joint co-occurrence probability and marginal distribution probability. Construct a global feature matrix from the extracted sets of all third-order orthogonal semantic tensors. Using a single historical research business text or a single research project application text as the statistical granularity, traverse the global feature matrix and count the feature elements. The document frequency appearing in the global feature matrix will be used to determine the feature elements. The feature element is obtained by dividing the document frequency appearing in the global feature matrix by the total number of texts. Marginal distribution probability Statistical characteristic elements The document frequency appearing in the global feature matrix will be used to determine the feature elements. The feature element is obtained by dividing the document frequency appearing in the global feature matrix by the total number of texts. Marginal distribution probability Statistical characteristic elements With characteristic elements The frequency of documents appearing simultaneously in the same text will be considered as a feature element. With characteristic elements The joint co-occurrence probability is obtained by dividing the frequency of documents appearing simultaneously in the same text by the total number of texts. For the storage and high-concurrency traversal retrieval of the global feature matrix, those skilled in the art can use a distributed columnar database combined with an inverted index structure. Distributed columnar databases and inverted index structures are well-known technologies in this field and will not be elaborated here.

[0026] Step S220: Calculate the point mutual information value based on the feature elements output in step S210. With characteristic elements joint co-occurrence probability Characteristic elements Marginal distribution probability and characteristic elements Marginal distribution probability The mutual information value of points represents the degree of correlation between dimensional feature elements. The calculation formula is: When feature elements With characteristic elements The joint co-occurrence probability is higher than that of feature elements. With characteristic elements When the probabilities of mutually independent elements are multiplied, the mutual information value of the points is positive, indicating that the feature elements... With characteristic elements There is a positive correlation in technological evolution between them.

[0027] Step S230: Construct the metric tensor matrix, setting the basis of the underlying flat space as the Kronecker function. Introducing set scale adjustment parameters Perform an exponential decay transformation on the point mutual information values ​​to measure the tensor matrix. The calculation formula is: in and Summation index for tensor dimensions, metric tensor matrix Reflecting the local spatial curvature of a non-Euclidean geometric manifold, when the characteristic element With characteristic elements The higher the correlation, the better the metric tensor matrix. The smaller the corresponding component value, the better the non-Euclidean geometric manifold is in terms of its characteristic features. With characteristic elements The corresponding spatial dimension shrinks, establishing a non-Euclidean geometric manifold that reflects the maturity of the evolution of each technology field.

[0028] In one specific embodiment, the process of constructing the expert dynamic trajectory and performing affine connection correction is described in detail. The process of constructing the expert dynamic trajectory and performing affine connection correction specifically includes the following sub-steps: Step S310: Establish the expert time series evolution trajectory, extract the third-order orthogonal semantic tensor corresponding to the target expert at each time node, and treat each time node as a continuous time parameter. By treating the research history of the target expert as a parametric curve on a non-Euclidean geometric manifold, the time-series evolution trajectory of the target expert is established. Time parameters Obtained by normalizing the publication timestamps of historical scientific research documents, each point on the parametric curve corresponds to a specific time parameter of the target expert. The research characteristic coordinates below show the transformation of target experts across different research fields over time, which can be represented as the time series evolution trajectory of experts. Position movement on a non-Euclidean geometric manifold.

[0029] Step S320: Calculate the Christopher symbol on the non-Euclidean geometric manifold based on the metric tensor matrix. Christopher symbol Characterizing the intrinsic curvature of the technological evolution manifold and the evolutionary tendency forces between technological fields, Christopher's notation. The calculation formula is: In the formula, To measure the inverse of a tensor matrix, , as well as Representing the coordinates of the tensor dimension respectively Tensor dimensional coordinates and tensor dimension coordinates The partial derivatives are obtained by using the finite difference method to approximate the derivatives of discrete third-order orthogonal semantic tensor feature data. For the matrix inversion operation and the underlying algorithm operation of the finite difference method, those skilled in the art can call the linear algebra module in the numerical computing library to implement them. The matrix inversion operation and the finite difference method are well-known technologies in this field and will not be described in detail here.

[0030] Step S330: Calculate the covariant acceleration, specifically the covariant acceleration of the target expert moving on the non-Euclidean geometric manifold. Through the evolution trajectory of expert time series Regarding time parameters The first derivative yields the evolution velocity vector of the target expert. Combined with Christopher symbol And the calculation of covariant acceleration using evolution velocity vectors Covariant acceleration The calculation formula is: In the formula, Let be the ordinary time derivative of the evolution velocity vector. and Covariant acceleration is the dimensional component corresponding to the evolutionary velocity vector. Covariant acceleration is the degree to which the actual scientific research evolution path of a target expert deviates from the geodesic of technological evolution. The higher the value, the greater the degree of deviation of the target expert in the process of interdisciplinary innovation or research direction shift.

[0031] Step S340: Perform feature coordinate correction and matching, utilizing covariant acceleration. Affine connection correction is performed on the current research characteristic coordinates of the target expert, and the time difference is determined based on the time point in the research project application text. Combined with the evolution velocity vector With covariant acceleration Predicting the corrected feature coordinates of target experts during project execution. Correct feature coordinates The calculation formula is: In the formula, Calculate the corrected feature coordinates for the current research feature coordinates of the target expert. The geodesic distance of the third-order orthogonal semantic tensor corresponding to the research project proposal text on a non-Euclidean geometric manifold is obtained by calculating the minimum path length between two points on the non-Euclidean geometric manifold along the integral of the metric tensor matrix. The geodesic distance is used as a quantitative indicator to evaluate the technical matching degree between the target expert and the research project proposal. The target expert with the minimum geodesic distance is output as the matched research project review expert.

[0032] In one specific embodiment, the process of constructing an implicit reciprocal network graph and performing topology independence detection is described in detail. The construction of the implicit reciprocal network graph and the performance of topology independence detection specifically include the following sub-steps.

[0033] Step S410: Establish an implicit reciprocity network graph, extract historical collaboration data and historical review data of candidate review experts and applicants corresponding to research project applications within a preset historical time window, and treat candidate review experts and applicants as the node set of the implicit reciprocity network graph. Historical cooperation data and historical review data are used as a set of nodes. The set of directed edges between each node in the middle Directed edge set Each directed edge in Represents a node For nodes There is a history of vested interests involved; historical collaboration data includes jointly authored papers, jointly invented patents, and jointly participated in research projects; historical review data covers key milestones. As a review expert, I have focused on key points in the review of past research projects. The given review scores indicate that the implicit reciprocal network graph is represented as a directed graph. .

[0034] Step S420: Construct the adjacency weight matrix and calculate the set of directed edges. The weights of each directed edge are calculated by introducing a time decay function to weight historical collaboration data and historical review data. Corresponding edge weight The calculation formula is: In the formula For nodes With nodes The total number of events in which historical interest transfer relationships occurred between them. For the first The basic weight value of the sub-event. The set time decay coefficient, For the first The time interval between the occurrence of the secondary event and the current time, and the basic weight value. The system assigns values ​​based on event type. To quantify the strength of benefits from different event types, a mapping table between event types and basic weight values ​​is pre-built. For example, when the event type is a co-authored paper, the basic weight value is... The value is assigned to the first preset value, such as 1.0. This is the base weight value when the event type is historical review data and the review score is higher than the preset excellent threshold. Assign a value to the second preset value, such as 0.8, as the edge weight. Quantitative characterization nodes With nodes The strength of implicit reciprocity between nodes, if nodes With nodes If there are no historical events between them, then the edge weight is... The value is zero, based on the set of directed edges in the implicit reciprocal network graph. Corresponding edge weights Generate the adjacency weight matrix.

[0035] Step S430: Perform topological independence detection and analyze the directed graph based on the adjacency weight matrix. Perform topology analysis to identify strongly connected components in the implicit reciprocity network graph, and set a reciprocity strength threshold. The reciprocity strength threshold is set as the average of the weights of all non-zero edges in the adjacency weight matrix. When the edge weight... When the reciprocity strength threshold is greater than 1, the directed edge is retained. When edge weight From the directed graph when the reciprocity strength threshold is less than or equal to the threshold. Delete directed edges The pruned directed graph is obtained, and a depth-first search algorithm is used to traverse the pruned directed graph. During the traversal, each node is assigned an access order number and a traceability node number. For the underlying traversal mechanism of the depth-first search algorithm and the update logic of the node number, those skilled in the art can use the Tayan algorithm to implement it. The Tayan algorithm is a well-known technology in this field and will not be described in detail here. When a subset of nodes with the same traceability node number is identified, the subset of nodes with the same traceability node number is determined to be a strongly connected component. A strongly connected component indicates that there is a closed-loop benefit transfer relationship in the network topology.

[0036] Step S440: Perform independence screening and list update. Based on the identification results of strongly connected components, perform expert elimination operation. Determine whether the candidate review expert node and the applicant node are located in the same strongly connected component. If the candidate review expert node and the applicant node are located in the same strongly connected component, it is determined that the candidate review expert node has failed the topological independence detection and is removed from the list of matched scientific research project review experts. Determine whether different candidate review expert nodes are located in the same strongly connected component. If multiple candidate review expert nodes are located in the same strongly connected component, only the candidate review expert node with the smallest geodesic distance to the scientific research project application is retained, and the remaining candidate review expert nodes located in the same strongly connected component are eliminated. The candidate review expert nodes that pass the topological independence detection are output as the final scientific research project review experts.

[0037] In one specific embodiment, the process of target region deduction and spatiotemporal collision integral matching is described in detail. The target region deduction and spatiotemporal collision integral matching specifically includes the following sub-steps: Step S510: Construct the target requirement region and extract the third-order orthogonal semantic tensor coordinates corresponding to the research project application text. The coordinates of the third-order orthogonal semantic tensor As the regional center point on a non-Euclidean geometric manifold, the semantic information entropy of the research project proposal text is calculated in the dimensions of research methodology, research object, and application environment. The feature word frequencies of the research project proposal text in each dimension are statistically analyzed. The feature word frequencies are divided by the total word frequencies of the corresponding dimension to obtain the feature probabilities. The semantic information entropy of each dimension is calculated by summing the negative forms of the products of the feature probabilities and their natural logarithms. The total information entropy is obtained by summing the semantic information entropies of the research methodology, research object, and application environment dimensions. Introducing a preset regional diffusion coefficient Set the target demand manifold region manifold radius manifold radius The calculation formula is: Target demand manifold region The enclosed spatial extent represents the breadth of technological crossover required in the research project proposal, and the total information entropy. A larger radius indicates a more pronounced interdisciplinary nature in the research project proposal, and a correspondingly larger manifold radius. The larger the target demand manifold region The wider the manifold volume covered, the more appropriate the underlying calculation method for semantic information entropy can be. Those skilled in the art can use the discrete probability distribution logarithmic summation algorithm to implement it. The discrete probability distribution logarithmic summation algorithm is a well-known technology in this field and will not be elaborated here.

[0038] Step S520: Calculate the spatiotemporal collision integral and obtain the project start timestamp specified in the research project application. Project completion timestamp Obtain the project start timestamp of the target expert predicted based on the evolutionary velocity vector and covariant acceleration in the aforementioned steps. Project completion time stamp Expert time series predictions of evolutionary trajectory during the period Computational expert time series prediction of evolutionary trajectory From the point on the map to the center point of the region Instantaneous geodesic distance Construct the instantaneous collision intensity function When the instantaneous geodesic distance Less than or equal to the radius of the manifold Instantaneous collision intensity function The calculation formula is: When the instantaneous geodesic distance Greater than the radius of the manifold Instantaneous collision intensity function The value is zero for the project start timestamp. Project completion time stamp Instantaneous collision intensity function within the interval The spacetime collision integral value is obtained by performing definite integral operations. Spacetime collision integral value The calculation formula is: Spacetime collision integral value The research direction of quantitative characterization of the evolution of target experts overlaps with the technical needs of scientific research projects during the execution of research projects and the duration of continuous coverage.

[0039] Step S530: Output the list of review experts, iterate through all candidate review expert nodes that pass the topological independence detection, and calculate the spatiotemporal collision integral value corresponding to each candidate review expert node. According to the spacetime collision integral value The candidate reviewer nodes are sorted in descending order from largest to smallest. A threshold for the number of reviewers required for the research project application is pre-set. The candidate reviewer nodes that rank at the top of the descending order and whose number equals the threshold are extracted. These extracted candidate reviewer nodes are output as the final research project reviewers. The matching calculation of scientific research project reviewers is completed. The generated list of research project reviewers is stored in a dedicated relational database, and the communication module is triggered to send review invitations to the research project reviewers. For the structured storage of the dedicated relational database and the message assembly of the communication module, those skilled in the art can use Structured Query Language and Simple Mail Transfer Protocol (SMLP). Structured Query Language and SMLP are well-known technologies in this field and will not be described in detail here.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for precise matching of experts in scientific research project review based on big data in science and technology, characterized in that, Includes the following steps: Step S100: Obtain the set of historical scientific research business texts of the target expert and the text of scientific research project application forms to be matched, extract feature elements and construct corresponding orthogonal semantic tensors through tensor product operation respectively; Step S200: Calculate the correlation density based on the joint co-occurrence probability and marginal distribution probability of feature elements, and construct a metric tensor matrix based on the correlation density to establish the non-Euclidean geometric manifold; Step S300: Combine Christopher symbol to calculate the covariant acceleration of the target expert, and use the covariant acceleration to calculate the corrected feature coordinates in order to obtain the expert time series prediction evolution trajectory during the execution of the scientific research project. Step S400: Construct an implicit reciprocity network graph. By identifying strongly connected components in the implicit reciprocity network graph, candidate review experts with conflicting interests are eliminated, and candidate review expert nodes that pass the topological independence detection are obtained. Step S500: Set the manifold radius based on semantic information entropy and calculate the instantaneous geodesic distance, construct the instantaneous collision intensity function and perform definite integral calculation, and select scientific research project review experts based on the spatiotemporal collision integral value.

2. The method for precise matching of scientific research project review experts based on big data as described in claim 1, characterized in that, Extracting feature elements from the historical scientific research business text set and the scientific research project application text in the orthogonal semantic dimension, including: The historical scientific research business text set and the scientific research project application text are converted into standardized text sequence data; A dependency parsing and semantic role labeling model is established using a network architecture combining a pre-trained language model and a conditional random field. The central predicate and the arguments associated with the central predicate in the text sequence data are identified through the dependency parsing and semantic role labeling model. In the preset argument mapping rules, the accompanying adverbial arguments representing the tool or manner of the action are mapped to the research method dimension feature, the patient arguments representing the recipient of the action are mapped to the research object dimension feature, and the adverbial arguments representing time, space or conditional constraints are mapped to the application environment dimension feature.

3. The method for precise matching of scientific research project review experts based on big data as described in claim 1, characterized in that, Constructing a metric tensor matrix based on the correlation density to establish a non-Euclidean geometric manifold includes: The marginal distribution probability is obtained by dividing the document frequency of the feature element in the global feature matrix by the total number of texts, and the joint co-occurrence probability is obtained by dividing the document frequency of two feature elements appearing simultaneously in the same text by the total number of texts. The joint co-occurrence probability is divided by the product of the marginal distribution probabilities of the two feature elements, and the logarithm of the division result is taken to obtain the point mutual information value. The point mutual information value is used as the correlation tightness. The point mutual information values ​​are subjected to an exponential decay transformation, the basis of the basic flat space is set as the Kronecker function, and the Kronecker function is added to the result of the exponential decay transformation to obtain the metric tensor matrix.

4. The method for precise matching of scientific research project review experts based on big data as described in claim 1, characterized in that, Calculating the covariant acceleration of the target expert includes: The partial derivatives of the tensor dimension coordinates of the metric tensor matrix are obtained using the finite difference method, and the Christopher symbol is calculated by combining the inverse matrix of the metric tensor matrix with the partial derivative results. The evolution velocity vector is obtained by taking the first derivative of the expert time series evolution trajectory with respect to the time parameter. The product of the Christopher symbol and the dimension component of the evolution velocity vector is added together, and the summation result is performed with the ordinary time derivative of the evolution velocity vector to obtain the covariant acceleration.

5. The method for precise matching of scientific research project review experts based on big data as described in claim 4, characterized in that, Calculating the corrected characteristic coordinates using the covariant acceleration includes: The time difference is determined based on the time point in the research project application document. The corrected feature coordinates are obtained by summing the current research feature coordinates of the target expert, the product of the evolution velocity vector and the time difference, and the covariant acceleration and half of the square of the time difference.

6. The method for precise matching of scientific research project review experts based on big data as described in claim 1, characterized in that, Calculating the edge weights of the directed edges to generate an adjacency weight matrix includes: Establish a mapping table between event types and basic weight values. When the event type is a co-authored paper, assign the basic weight value to a first preset value. When the event type is historical review data and the review score is higher than a preset excellent threshold, assign the basic weight value to a second preset value. The time interval between each event occurrence and the current time is weighted using a time decay function. The weighted calculation result is multiplied by the corresponding basic weight value. The results of multiplying all events that have historical benefit output relationships between nodes are accumulated to obtain the edge weight of the directed edge. The adjacency weight matrix is ​​generated based on the weights of all edges corresponding to each directed edge in the implicit reciprocal network graph.

7. The method for precise matching of scientific research project review experts based on big data as described in claim 1, characterized in that, By identifying strongly connected components in the implicit reciprocity network graph, candidate review experts with vested interests are eliminated, including: The average value of all non-zero edge weights in the adjacency weight matrix is ​​set as the reciprocity strength threshold. The directed edge is retained when the edge weight is greater than the reciprocity strength threshold, and the directed edge is deleted when the edge weight is less than or equal to the reciprocity strength threshold to obtain the pruned directed graph. The pruned directed graph is traversed using a depth-first search algorithm. During the traversal, each node is assigned an access order number and a traceability node number. A subset of nodes with the same traceability node number is determined as the strongly connected component.

8. The method for precise matching of scientific research project review experts based on big data as described in claim 7, characterized in that, Candidate review experts with conflicting interests were excluded, including: When the candidate review expert node and the applicant node are located in the same strongly connected component, it is determined that the candidate review expert node has failed the topology independence detection and the associated candidate review expert node is removed. When multiple candidate review expert nodes are located in the same strongly connected component, only the candidate review expert node with the smallest corresponding geodesic distance among the multiple candidate review expert nodes is retained, and the other candidate review expert nodes located in the same strongly connected component are removed.

9. The method for precise matching of experts for scientific research project review based on big data as described in claim 1, characterized in that, Construct the instantaneous collision intensity function, including: The frequency of feature words in the research project application text is statistically analyzed in each dimension. The feature word frequency is divided by the total word frequency of the corresponding dimension to obtain the feature probability. The semantic information entropy of each dimension is calculated by summing the product of the feature probability and the natural logarithm of the feature probability. The total information entropy is obtained by summing the semantic information entropy of each dimension, and the product of the preset regional diffusion coefficient and the total information entropy is set as the manifold radius; When the instantaneous geodesic distance is less than or equal to the manifold radius, the negative number of the square of the instantaneous geodesic distance is divided by twice the square of the manifold radius to obtain the division result, and the instantaneous collision intensity function is obtained by taking the exponential function of the division result; when the instantaneous geodesic distance is greater than the manifold radius, the value of the instantaneous collision intensity function is set to zero.

10. The method for precise matching of scientific research project review experts based on big data as described in claim 1, characterized in that, Based on the spatiotemporal collision integral value, the corresponding candidate review expert nodes are selected as research project review experts, including: The instantaneous collision intensity function within the interval from the project start time stamp to the project completion time stamp corresponding to the research project application text is integrally calculated to obtain the spatiotemporal collision integral value. The candidate review expert nodes are sorted in descending order according to the spatiotemporal collision integral value. The candidate review expert nodes that rank at the top of the descending order and whose number is equal to the threshold of the number of review experts are extracted and output as the final review experts for scientific research projects.

Images