A method and system for assisting in the review of scientific and technological literature

By performing variable substitution and difference analysis on the feature dimensions of scientific and technological documents, and combining clustering algorithms, the problem of inaccurate identification of abnormal feature dimensions in existing technologies has been solved, realizing an automated and reliable review assistance system for scientific and technological documents.

CN123115704APending Publication Date: 2026-11-17INNOVATION & INNOVATION CENT OF STATE GRID ZHEJIANG ELECTRIC POWER CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611619201.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies lack objective means to automatically identify abnormal features in the evaluation of scientific and technological literature quality, resulting in low review efficiency and high subjectivity, and failing to quantify the degree of deviation and provide actionable verification guidelines.

Method used

By performing variable permutation operations on each feature dimension of the target scientific and technological literature, permutation feature data is generated. The quality score is predicted using a literature evaluation model, and key abnormal feature dimensions are identified through difference analysis and clustering algorithms to generate graded early warning and verification guidelines.

Benefits of technology

It enables the automatic and objective identification of significant document feature dimensions, quantifies the degree of deviation, and provides clear warnings and verification suggestions, thereby improving the efficiency and reliability of the review process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115704A_ABST
    Figure CN123115704A_ABST
Patent Text Reader

Abstract

This invention provides a method and system for assisting in the review of scientific and technological literature. The method includes: acquiring feature data of the target scientific and technological literature to be analyzed, feature data of historical scientific and technological literature, and quality score values; performing variable substitution operations on each feature dimension of the target scientific and technological literature, inputting each substituted feature data into a preset literature evaluation model to obtain substitution analysis results; inputting the feature data of the target scientific and technological literature into the literature evaluation model to obtain original analysis results; determining the feature dimensions of the target literature based on the degree of difference between the original analysis results and the substitution analysis results, generating comprehensive information on the feature dimensions of the target literature, and using the comprehensive information to assist in the review of the target scientific and technological literature. This invention can identify dimensions that have an abnormal impact on the evaluation results from multiple feature dimensions of a single document, realize quantitative analysis and graded early warning of abnormal features, and provide technical support for the assisting in the review of scientific and technological literature.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of literature analysis technology, and more specifically, to an auxiliary review method and system for scientific and technological literature. Background Technology

[0002] With the rapid growth in the number of scientific and technological documents and the in-depth application of machine learning and natural language processing technologies in the field of literature analysis, automated quality assessment of scientific and technological documents based on computer models has become an important development direction in the field of literature review. Literature evaluation models typically use multiple document characteristic dimensions as input variables, such as the number of references, abstract length, number of keywords, number of authors, and average citation count of references, to output a quality score for a single document. However, when the evaluation model gives a score to a document that deviates from expectations, reviewers often find it difficult to intuitively determine which document characteristic dimensions caused the abnormal score, thus hindering targeted verification of the abnormal document.

[0003] In the existing technology, there are already some technical solutions for intelligent processing of engineering documents. For example, Chinese invention patent CN117251582B discloses a method for constructing and optimizing a substation knowledge graph based on multi-view learning. This method analyzes and describes multi-source heterogeneous data of substations across the entire region, integrates all structured data into an expert knowledge base using text information extraction technology, extracts entities, attributes, and connections from the structured data to construct a knowledge graph data layer, and constructs relationship views and attribute views respectively. On this basis, the data is further layered and fused to construct a tag view, establishes a power system entity relationship database, and merges the external and internal knowledge graphs of the substation into a large knowledge graph based on the Neo4j graph database. At the same time, with accuracy, consistency, simplicity, and coverage as evaluation objectives, a multi-objective gradient descent approach is used to optimize the constructed knowledge graph from multiple views. This solution realizes the structured organization and entity association of multi-source heterogeneous data in the power field.

[0004] For example, Chinese invention patent CN121561116B discloses a method and system for constructing and retrieving a knowledge graph for power grid and substation engineering. This method dynamically and adaptively divides relevant documents into blocks to obtain semantically coherent and independent text blocks. Based on these text blocks, it concurrently calls a large language model and combines domain-specific prompts to extract entities and relationships, then completes the implicit entities and relationships in the text through multiple rounds of refinement. The extracted and completed entities and relationships are then fused and disambiguated to form a unified knowledge graph. The knowledge graph is then hierarchically clustered using the Leiden community detection algorithm to generate multi-granularity community structures and corresponding community reports. For ambiguous user queries, intent resolution and predictive guidance are performed to optimize the retrieval strategy. Based on the optimized retrieval strategy, multi-hop semantic retrieval is executed to recall relevant knowledge and generate answers.

[0005] However, the aforementioned existing technical solutions are all geared towards knowledge organization, knowledge graph construction, and retrieval question answering for engineering documents. Their core lies in structuring entities and relationships in unstructured documents and supporting semantic queries, but they do not address the interpretability analysis of the quality evaluation results of individual documents. Specifically: First, after a document evaluation model provides a quality score, existing technologies lack the technical means to objectively and automatically identify target dimensions that have an abnormal impact on the score from multiple document feature dimensions. They still rely on manual experience to check each feature dimension one by one, resulting in low review efficiency and high subjectivity. Second, the clustering algorithms used in existing technologies (such as multi-view clustering and Leiden community detection algorithms) are used for entity segmentation or community discovery, rather than for screening abnormal feature dimensions from the perspective of model prediction sensitivity. Therefore, they cannot objectively quantify the degree of influence of each document feature dimension on the evaluation model's prediction results. Third, existing technologies lack quantitative analysis of the deviation direction and degree of abnormal dimensions relative to the historical document value distribution. They also lack a graded early warning mechanism based on the severity of anomalies and operational verification guidelines corresponding to the types of abnormal dimensions, making it difficult to transform model outputs into review suggestions that reviewers can directly implement.

[0006] Therefore, there is an urgent need for a scientific literature review method that can automatically identify key abnormal features at the level of a single document, quantify their degree of deviation, and generate graded warnings and verification guidelines, in order to solve the technical problems of strong subjectivity in abnormal feature identification, inaccurate positioning, and lack of operable verification suggestions in existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide an auxiliary review method and system for scientific and technological literature to improve the above-mentioned problems.

[0008] To achieve the above objectives, this application provides the following technical solution: On the one hand, embodiments of this application provide an auxiliary review method for scientific and technological literature, the method comprising: Obtain the feature data of the target scientific and technological literature to be analyzed, as well as the feature data and quality score of historical scientific and technological literature. The feature data includes the values ​​of multiple document feature dimensions; document feature dimensions include, but are not limited to: document self-attribute dimension, author association dimension, and journal association dimension. Perform variable permutation on each feature dimension of the target scientific and technological literature to generate permuted feature data corresponding to each feature dimension; input each permuted feature data into a preset literature evaluation model to obtain permutation analysis results corresponding to each permuted feature data. The literature evaluation model is used to predict the quality score of scientific and technological literature; the variable permutation operation includes at least: replacing unreasonable values ​​of literature feature dimensions with reasonable values ​​of literature feature dimensions based on the feature data and quality score of historical scientific and technological literature. The characteristic data of the target scientific and technological literature is input into the literature evaluation model to obtain the original analysis results. Based on the degree of difference between the original analysis results and the replacement analysis results, the characteristic dimensions of the target literature are determined from multiple literature characteristic dimensions, and comprehensive information on the characteristic dimensions of the target literature is generated. The comprehensive information is used to assist in the review of the target scientific and technological literature.

[0009] Secondly, this application provides an auxiliary review system for scientific and technological literature, the system comprising: The acquisition module is used to acquire the feature data of the target scientific and technological literature to be analyzed, as well as the feature data and quality score of historical scientific and technological literature. The feature data includes the values ​​of multiple feature dimensions of the literature. The permutation module is used to perform variable permutation operations on each feature dimension of the target scientific and technological literature, generating permutation feature data corresponding to each feature dimension; each permutation feature data is input into a preset literature evaluation model to obtain permutation analysis results corresponding to each permutation feature data, and the literature evaluation model is used to predict the quality score of scientific and technological literature. The generation module is used to input the feature data of the target scientific and technological literature into the literature evaluation model to obtain the original analysis results; based on the degree of difference between the original analysis results and the replacement analysis results, the target literature feature dimensions are determined from multiple literature feature dimensions, and comprehensive information on the target literature feature dimensions is generated to assist in the review of the target scientific and technological literature.

[0010] Thirdly, this application provides an auxiliary review device for scientific and technological documents, the device comprising a memory and a processor. The memory stores a computer program; the processor executes the computer program to implement the steps of the aforementioned auxiliary review method for scientific and technological documents.

[0011] Fourthly, this application provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described auxiliary review method for scientific and technological documents.

[0012] The beneficial effects of this invention are as follows: 1. This invention performs variable permutation operations on each feature dimension of the target scientific literature, inputs the permuted feature data into a literature evaluation model to obtain permutation analysis results, calculates the first difference and standard deviation corresponding to each feature dimension, and uses a density-based clustering algorithm to cluster all feature dimensions. Based on the clustering results, the target literature feature dimension is determined. This scheme can automatically and objectively identify the dimension that has the most significant impact on the evaluation result of the literature from numerous feature dimensions, avoiding the bias of subjective human judgment.

[0013] 2. In the variable replacement process, this invention uses two layers of rationality verification: statistical outlier removal based on interquartile range and value stability verification based on coefficient of variation. This rigorously screens candidate replacement values ​​to ensure that all replacement values ​​are derived from real and reasonable values ​​in historical literature. This effectively filters out data noise, improves the stability and reliability of the literature evaluation model's prediction results, and reduces the misjudgment rate caused by data anomalies.

[0014] 3. This invention calculates a comprehensive anomaly index and generates graded early warning labels based on preset thresholds, quantifying the degree of anomaly into comparable numerical indicators. This achieves precise grading from whether an anomaly exists to how severe the anomaly is, providing auditors with clear priority early warning signals and significantly improving audit efficiency.

[0015] 4. This invention combines the name, original value, historical statistical distribution, deviation direction, deviation degree, and graded warning label of the target document's feature dimensions into structured comprehensive information, and automatically matches the corresponding verification guidance information from a preset verification rule base according to the dimension type, outputting it together. This scheme transforms the abstract outliers output by the clustering algorithm into specific verification suggestions that reviewers can directly operate, realizing a complete closed loop from anomaly identification, quantitative grading to targeted verification guidance, providing objective technical support for the auxiliary review of scientific and technological documents.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the auxiliary review method for scientific and technological documents as described in this embodiment of the invention; Figure 2 This is a DBSCAN clustering scatter plot of the document feature dimension of the auxiliary review method for scientific and technological documents described in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of the auxiliary review system for scientific and technological documents as described in this embodiment of the invention; Figure 4 This is a schematic diagram of the auxiliary review device for scientific and technological documents as described in this embodiment of the invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] It should be noted that similar reference numerals or letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] Example 1 like Figure 1 As shown in the figure, this embodiment provides an auxiliary review method for scientific and technological literature, which includes steps S1, S2 and S3.

[0022] Step S1: Obtain the feature data of the target scientific and technological literature to be analyzed, as well as the feature data and quality score of historical scientific and technological literature. The feature data includes the values ​​of multiple document feature dimensions. In this step, the feature data includes the values ​​of multiple document feature dimensions. Specifically, the document feature dimensions include, but are not limited to: (1) document self-attribute dimensions, such as the number of references, the number of pages (total number of pages in the paper), the length of the abstract (number of words in the abstract), the length of the title (number of words in the paper title), the number of keywords, the number of paragraphs, and the overlap of title-abstract keywords (overlap = number of words in the title and abstract / total number of words in the title × 100%), etc.; (2) author association dimensions, such as the number of authors, etc.; (3) journal association dimensions, such as the average number of citations of references (the arithmetic mean of the number of citations of each of the references cited by the target document), etc. Historical scientific and technological literature is a collection of published scientific and technological literature; the quality score can be assigned after being scored by experts. The following uses nine document feature dimensions as one of the nine document feature dimensions in this embodiment for explanation. The document feature dimensions in this invention are not limited to these.

[0023] Table 1. Calculation method of document feature dimensions used in Embodiment 1 of the present invention. The following example illustrates this concept using a specific target document. The target document is a scientific paper submitted to a journal in the field of computer vision. The values ​​of its nine document feature dimensions and the median of historical documents in this field are shown in Table 2.

[0024] Table 2 shows the value of the feature dimensions of the target documents and historical documents in Example 1. Step S2: Perform variable permutation operation on each document feature dimension in the feature data of the target scientific and technological literature to generate permutation feature data corresponding to each document feature dimension; input each permutation feature data into the preset literature evaluation model to obtain the permutation analysis result corresponding to each permutation feature data. The literature evaluation model is used to predict the quality score of scientific and technological literature. In this step, the variable substitution operation includes at least the following: replacing unreasonable feature dimension values ​​of historical scientific and technological literature with reasonable feature dimension values, based on the feature data and quality scores of the historical scientific and technological literature. The method for obtaining the literature evaluation model is as follows: using a multiple linear regression model as the basic framework, with the feature data of scientific and technological literature as the input variable and the quality score as the output variable. The training data consists of a large number of historical scientific and technological literatures, where each training sample includes input features and output labels, i.e., the input features are the feature data of historical scientific and technological literatures, and the output labels are the quality scores of historical scientific and technological literatures. The model training aims to minimize the mean squared error loss function, and iteratively updates the model parameters using the gradient descent method until the loss function converges to a preset threshold, thereby obtaining an optimal model with predictive ability.

[0025] In this step, the specific implementation steps for performing variable permutation operations on each document feature dimension in the feature data of the target scientific and technological documents to generate permuted feature data corresponding to each document feature dimension include steps S21-S23. Step S21: For each document feature dimension, obtain all historical values ​​of that document feature dimension from historical scientific and technological literature, and use multiple historical values ​​as a set of candidate replacement values; perform a rationality check on each candidate replacement value in the candidate replacement value set, and remove unreasonable candidate replacement values ​​according to the check results to obtain the qualified candidate replacement value corresponding to that document feature dimension. In this step, for each feature dimension of historical scientific and technological literature data, all historical values ​​that have appeared in that dimension are statistically analyzed, and these values ​​are used as a set of candidate replacement values. Simultaneously, in this step, anomaly detection can be performed on all historical values, and abnormal historical values ​​are deleted to form a set of candidate replacement values ​​before proceeding to step S211. In this embodiment, the specific anomaly detection steps include: The K-means clustering algorithm is used to cluster all historical values ​​to obtain multiple clusters. For each cluster, the average value of all historical values ​​in the cluster is calculated as the centroid of the cluster. The distance from each historical value to the centroid of the cluster is calculated, and the standard deviation of all distances is used as the intra-cluster dispersion of the cluster. The similarity threshold corresponding to the cluster is dynamically set according to the intra-cluster dispersion. The similarity threshold is positively correlated with the intra-cluster dispersion. In this step, intra-cluster dispersion reflects the density of samples within a cluster. Higher intra-cluster dispersion indicates a sparser sample distribution, requiring a higher similarity threshold to tolerate larger natural differences within the cluster. Conversely, lower intra-cluster dispersion indicates a more compact sample distribution, requiring a lower similarity threshold to sensitively identify subtle anomalies within the cluster. This achieves an adaptive match between the similarity judgment standard and the data distribution characteristics of different clusters. When setting the value, the similarity threshold can be set within the range of 0.3 to 0.95. If the threshold is below 0.3, a large number of irrelevant data points are easily misclassified as similar, introducing excessive noise; if the threshold is above 0.95, the judgment standard is too stringent, easily missing actually similar data points, leading to a significant increase in the false negative rate.

[0026] For each cluster, perform the following steps in sequence: Calculate the covariance matrix of the cluster, perform eigenvalue decomposition on the covariance matrix to obtain the maximum and minimum eigenvalues, and use the ratio of the maximum eigenvalue to the minimum eigenvalue as the morphological coefficient of the cluster. Based on the morphological coefficient, preset the adjustment coefficient corresponding to the cluster. The adjustment coefficient is positively correlated with the morphological coefficient. In this step, the morphological coefficient characterizes the uniformity of cluster distribution in all directions within the feature space. A larger value indicates a more pronounced banded feature of the cluster, requiring a larger adjustment coefficient in subsequent judgments to reduce the risk of edge samples being misclassified as anomalies. When setting the value, the adjustment coefficient can be set to an integer not less than 1, and typically not exceeding 5. If the preset value is lower than 1 (i.e., 0 or a negative number), the second cardinality will be greater than or equal to the total number of historical values ​​within the cluster, resulting in an overly large second value. This means that the vast majority of clusters cannot be labeled as candidate anomalies, leading to severely insufficient sensitivity in anomaly detection. Conversely, if the adjustment coefficient is higher than 5, the penalty for cluster size is too strong, causing the second cardinality to shrink excessively, easily misclassifying normal clusters as anomalies, resulting in a significantly increased false alarm rate.

[0027] For each historical value in the cluster, calculate the distance from that historical value to the cluster center. Divide this distance by the maximum distance from all historical values ​​within the cluster to the cluster center to obtain the outlier contribution of that historical value. Pairwise, combine every two historical values ​​to form a data pair. Calculate the first logarithm of each data pair and the similarity value between the two historical values ​​in each pair. Multiply this similarity value by the product of the outlier contributions of each historical value in the data pair to obtain a weighted similarity value. Data pairs that do not exceed the similarity threshold corresponding to the cluster are considered dissimilar data pairs. The second logarithm of the dissimilar data pairs is calculated. The first logarithm is added to the total number of historical values ​​in the cluster to obtain the first cardinality. The second logarithm is divided by the first cardinality and recorded as the first value. The adjustment coefficient is subtracted from the total number of historical values ​​in the cluster to obtain the second cardinality. The second cardinality is divided by the first cardinality and recorded as the second value. If the first value is greater than the second value, the cluster is marked as a candidate anomalous cluster; otherwise, the cluster is marked as a normal cluster. In this step, the outlier contribution reflects the degree to which each historical value deviates from the cluster center. It is used as a weight to weight the similarity value, so that edge samples far from the cluster center have a greater influence on the determination of dissimilar data pairs. By comparing the actual proportion of dissimilar data pairs within the cluster with the retention benchmark set based on the cluster size, it is finally determined whether the cluster is a candidate outlier cluster.

[0028] In addition, for each cluster, every two historical values ​​in the cluster are paired to form data pairs. The first logarithm of the data pairs is counted, and the similarity between the two data points in each pair is calculated. In this step, assuming that the cluster contains N data points, the number of data pairs is... When calculating similarity, x and y are two historical values. When x+y>0, the similarity is 1-|xy| / (x+y). When x+y=0, the two historical values ​​are considered to be completely identical, and the similarity is directly set to 1.

[0029] For a candidate anomalous cluster, calculate the silhouette coefficient of the candidate anomalous cluster, and at the same time calculate the average silhouette coefficient of all non-candidate anomalous clusters. If the silhouette coefficient of the candidate anomalous cluster is less than the product of the average silhouette coefficient of all non-candidate anomalous clusters and the preset proportion threshold, then the candidate anomalous cluster is finally determined as an anomalous cluster and deleted; otherwise, the candidate anomalous cluster is reclassified as a normal cluster. In this step, the preset proportion threshold can be 80%. By comparing the silhouette coefficient of the candidate anomalous cluster with the average silhouette coefficient of all non-candidate anomalous clusters, if the silhouette coefficient of the candidate anomalous cluster is significantly lower than the average level of the normal cluster, it indicates that the internal structure of the cluster is loose and the differentiation from the normal cluster is insufficient, thus meeting the removal condition of outlier clusters; otherwise, it indicates that although there are certain differences in the cluster, it still has a reasonable clustering structure and should be retained as a normal cluster.

[0030] Collect all historical values ​​contained in the normal clusters to form a set of candidate replacement values.

[0031] In addition to the steps above, all historical values ​​can be grouped and selected to form a candidate replacement value set. Then, proceed to step S211. The specific filtering steps include: Clustering algorithms are used to cluster all historical values, resulting in multiple clusters and their corresponding cluster centers. The number of historical values ​​contained in each cluster is counted, and the cluster with the most historical values ​​is determined as the primary cluster. If the number of historical values ​​in the primary cluster is lower than a preset threshold, the entire set of historical values ​​is directly used to form a candidate replacement value set. Otherwise, the remaining clusters outside the primary cluster are recorded as secondary clusters. Pairwise merging is performed between all secondary clusters: any two secondary clusters whose inter-cluster distance is less than a preset merging threshold are merged into a new cluster. The inter-cluster distance is the Euclidean distance between the cluster centers, and the new cluster center is the midpoint of the line connecting the two old cluster centers. This merging operation is repeated until the inter-cluster distance between any two secondary clusters is not less than the preset merging threshold. In this step, the preset scale threshold can be 1% to 5% of the total number of historical values; After merging, the inter-cluster distances between each secondary cluster and the main cluster are recalculated to obtain a set of all inter-cluster distances. The minimum, maximum, mean, and standard deviation of this set are calculated. The mean minus twice the standard deviation is compared with the minimum value, and the larger of the two is taken as the lower boundary value. The mean plus twice the standard deviation is compared with the maximum value, and the smaller of the two is taken as the upper boundary value. The mean of the lower and upper boundary values ​​is taken as the inter-cluster assignment threshold. Secondary clusters with inter-cluster distances greater than the inter-cluster assignment threshold are designated as distant clusters, and secondary clusters with inter-cluster distances less than or equal to the inter-cluster assignment threshold are designated as close clusters. The historical values ​​of the main cluster and all close clusters are summarized to form the first dataset. The historical values ​​of the distant clusters are used as candidate supplementary data, and a portion of the historical values ​​are randomly selected from them according to a preset selection ratio and added to the second dataset. The first and second datasets are combined to form a candidate replacement value set.

[0032] The preset selection ratio is a value between 1% and 30%, for example, 5%. Cluster analysis is used to divide all historical values ​​into three levels according to their group affiliation: the main cluster, the nearest cluster, and the far-distance cluster. Based on this, historical values ​​in the main group and the nearest cluster are fully retained, while historical values ​​in the far-distance cluster are only randomly selected according to a preset ratio and added to the candidate set. Through this hierarchical division and differentiated retention strategy, while ensuring sufficient data diversity in the candidate replacement value set, obviously outlier historical values ​​are filtered out at the source.

[0033] In addition, in this step, the rationality of each candidate replacement value in the candidate replacement value set is checked, and unreasonable candidate replacement values ​​are removed according to the check results. The specific implementation steps for obtaining the qualified candidate replacement values ​​corresponding to each document feature dimension include steps S211-S212. Step S211: Calculate the upper quartile and lower quartile of all candidate replacement values ​​in the candidate replacement value set; calculate the difference between the upper quartile and the lower quartile to obtain the interquartile range; In this step, the upper quartile (Q3, i.e., the 75th percentile) and lower quartile (Q1, i.e., the 25th percentile) of all candidate replacement values ​​in the candidate replacement value set are calculated, and the interquartile range IQR = Q3 - Q1 is calculated.

[0034] The quartiles are calculated as follows: The candidate replacement value set contains n data points. Sort the data in the candidate replacement value set in ascending order. The position of the lower quartile Q1 is p1 = (n+1) × 25%; The position of the upper quartile Q3 is p3 = (n+1) × 75%; When the position is an integer, the data value corresponding to that position is directly taken as the quartile; when the position is not an integer, let the integer part of the position be i and the decimal part be d, then the quartile values ​​are: Q=x(i)+d×(x(i+1)−x(i)) Where x(i) and x(i+1) are the i-th and i+1-th data after sorting, respectively.

[0035] In this step, taking the number of references as an example, after sorting the 18 candidate replacement values, n=18. The lower quartile position p1 = (18 + 1) × 25% = 4.75. Q1 = 30 + 0.75 × (33 - 30) = 32; The upper quartile position p3 = (18 + 1) × 75% = 14.25. Q3 = 65 + 0.25 × (72 - 65) = 67; Interquartile Range (IQR) = 67 - 32 = 35. The calculation results are shown in Table 3.

[0036] Table 3. Results of Iquartile Calculation for the Number of References Step S212: Obtain a set of candidate adjustment coefficients, which includes multiple candidate adjustment coefficients that increase sequentially. Calculate the product of each candidate adjustment coefficient and its interquartile range to obtain multiple weighted interquartile ranges. For each candidate adjustment coefficient, determine a temporary normal value interval based on the lower quartile, upper quartile, and the weighted interquartile range corresponding to the candidate adjustment coefficient. Count the number of candidate replacement values ​​outside this temporary normal value interval in the candidate replacement value set, and calculate the ratio of this number to the total number of candidate replacement values ​​in the candidate replacement value set. Use this ratio as the outlier removal ratio corresponding to the candidate adjustment coefficient. Obtain qualified candidate replacement values ​​based on all outlier removal ratios.

[0037] In this step, a candidate adjustment coefficient set is first obtained. This set includes multiple candidate adjustment coefficients that increase in increments of a preset step size from a first preset value to a second preset value. For example, the candidate adjustment coefficient set could consist of multiple coefficients ranging from 1.0 to 3.0 with a step size of 0.1 or 0.2. For instance, the candidate adjustment coefficient set K = {1.0, 1.1, 1.2, ..., 3.0}. For each candidate adjustment coefficient k, the weighted interquartile range k × IQR is calculated to determine the temporary normal value interval [Q1 - k × IQR, Q3 + k × IQR]. The number of candidate replacement values ​​outside this interval is counted, and the ratio of this number to the total number of candidate replacement values ​​is calculated to obtain the outlier removal ratio.

[0038] In this step, the specific implementation steps for obtaining qualified candidate replacement values ​​based on all outlier removal ratios include steps S2121-S2122; Step S2121: According to the candidate adjustment coefficients in the candidate adjustment coefficient set in ascending order, compare the outlier removal ratio corresponding to each candidate adjustment coefficient with the preset ratio range. The candidate adjustment coefficient whose outlier removal ratio falls into the preset ratio range for the first time is determined as the target adjustment coefficient. Determine the normal value range based on the lower quartile, upper quartile, and the weighted interquartile range corresponding to the target adjustment coefficient. If the candidate replacement value is outside the normal value range, then the candidate replacement value is determined as the first type of unreasonable candidate replacement value and is removed to obtain the remaining candidate replacement values. In this step, the outlier removal ratios are compared with a preset ratio range (e.g., [5%, 15%]) in ascending order of candidate adjustment coefficients. The candidate adjustment coefficients that first fall into the preset ratio range are determined as the target adjustment coefficients. Then, the final normal value range is determined according to [Q1-k×IQR, Q3+k×IQR]. Candidate replacement values ​​outside this range are removed to obtain the remaining candidate replacement values.

[0039] In this step, taking the number of references as an example, the candidate replacement value set contains 18 values ​​(18, 22, 25, 28, 30, 33, 36, 40, 42, 45, 48, 52, 56, 60, 65, 72, 85, 110). The candidate adjustment coefficients are scanned from smallest to largest. When k=1.0, the normal range is [-3, 102]. Only value 110 is outside this range, resulting in an outlier elimination rate of 1 / 18 ≈ 5.6%. Since the value falls within the preset range [5%, 15%] for the first time, the target adjustment coefficient k=1.0, and value 110 is eliminated. The scanning results are shown in Table 4.

[0040] Table 4. Results of Candidate Adjustment Coefficient Scanning and First-Layer IQR Removal Step S2122: For each remaining candidate replacement value, search for historical scientific and technological literature in which the value of the literature feature dimension is equal to the candidate replacement value, and obtain its corresponding quality score value; calculate the dispersion between all quality score values. If the dispersion exceeds the preset stability threshold, the candidate replacement value is identified as the second type of unreasonable candidate replacement value and is removed, and finally qualified candidate replacement values ​​are obtained.

[0041] In this step, for the remaining candidate replacement values ​​after the first layer of verification, a second layer of verification is performed. For each remaining candidate replacement value v, all documents in historical scientific literature with a feature dimension value equal to v are searched, and the corresponding quality scores for these documents are obtained. The mean μ and standard deviation σ of these quality scores are calculated, and the coefficient of variation CV = σ / μ is obtained by dividing the standard deviation by the mean. If the CV exceeds a preset stability threshold (e.g., 0.8), the candidate replacement value is judged as a second type of unreasonable candidate replacement value and is removed. The quality scores of historical scientific literature are obtained by averaging the scores given independently by three domain experts on a scale of 1 to 10, and are used as training labels for the literature evaluation model; where a higher score indicates better literature quality, a score below 4 is considered literature with obvious quality problems.

[0042] In this step, for the 17 candidate replacement values ​​remaining after the first-level validation, historical references with the same number of references for each value are searched, and the average μ and standard deviation σ of their quality scores are calculated. The coefficient of variation (CV) is calculated as CV = σ / μ. If CV > 0.8, the candidate replacement value is removed. After validation, candidate values ​​18 (CV = 0.91) and 85 (CV = 0.82) were removed, and the remaining 15 qualified candidate replacement values ​​are shown in Table 5.

[0043] Table 5 Stability verification of coefficient of variation and qualified candidate replacement values Step S22: Replace the original value of each qualified candidate replacement value with the original value of the document feature dimension, while keeping the values ​​of other document feature dimensions unchanged, construct preliminary replacement feature data, input each preliminary replacement feature data into the document evaluation model for prediction, and obtain the prediction result; In this step, for example, for the document feature dimension of the number of references, the original value is 95, and the qualified candidate replacement values ​​are 22 and 25. Then, 95 in the feature data is replaced with 22, while other dimension values ​​remain unchanged, generating one preliminary replacement feature data; 95 in the feature data is replaced with 25, while other dimension values ​​remain unchanged, generating another preliminary replacement feature data. This yields the preliminary replacement feature data corresponding to the document feature dimension of the number of references. In this step, the original number of references for the target literature was 95. The predicted quality score output by the literature evaluation model (i.e., the original analysis result, with a maximum score of 10) was 4.2, corresponding to the "poor quality" range in the expert scoring. The 15 qualified candidate replacement values ​​were replaced with 95, while the values ​​of the other 8 dimensions remained unchanged. These were then input into the model sequentially to obtain the predicted results. The absolute value of the difference between the predicted and original analysis results was calculated, as shown in Table 6.

[0044] Table 6. Prediction results of reference quantity dimension permutation Step S23: Calculate the absolute value of the difference between the original analysis results and each prediction result. Sort the qualified replacement values ​​in descending order of the absolute value of the difference. Select the top-ranked replacement values ​​as the final replacement values. Replace the original value of the document feature dimension with each final replacement value, while keeping the values ​​of other document feature dimensions unchanged. Construct the replacement feature data corresponding to the document feature dimension.

[0045] In this step, the characteristic data of the target scientific and technological literature are input into the literature evaluation model to obtain the raw analysis results. After sorting the qualified replacement values ​​according to the absolute value of the difference from largest to smallest, the top 3 replacement values ​​are selected as the final replacement values. In this step, the 15 qualified candidate replacement values ​​are sorted from largest to smallest based on the absolute value of the difference, and the top 3 are selected as the final replacement values, namely 22, 25, and 28. The sorting results are shown in Table 7.

[0046] Table 7 shows the results of sorting candidate replacement values ​​by absolute value of difference. Step S3: Input the feature data of the target scientific and technological literature into the literature evaluation model to obtain the original analysis results; based on the degree of difference between the original analysis results and the replacement analysis results, determine the feature dimensions of the target literature from multiple literature feature dimensions, generate comprehensive information on the feature dimensions of the target literature, and use the comprehensive information to assist in the review of the target scientific and technological literature.

[0047] In this step, the specific implementation steps for determining the target document feature dimension from multiple document feature dimensions based on the degree of difference between the original analysis results and the replacement analysis results include steps S31-S32. Step S31: For each document feature dimension, calculate the average of all permutation analysis results corresponding to that document feature dimension, and use it as the comprehensive permutation analysis result corresponding to that document feature dimension; calculate the absolute difference between the original analysis result and the comprehensive permutation analysis result, and use it as the first difference corresponding to that document feature dimension; and calculate the standard deviation of all permutation analysis results corresponding to that document feature dimension. In this step, the average value of all permutation analysis results for each of the nine document feature dimensions was calculated. The absolute difference between this average and the original analysis results was used as the first difference, and the standard deviation of the permutation analysis results was also calculated. The statistical results of the first difference and standard deviation for each dimension are shown in Table 8.

[0048] Table 8. First Difference and Standard Deviation of Each Document Feature Dimension Step S32: Treat each document feature dimension as a sample point. The feature vector of each sample point includes the first difference and standard deviation corresponding to the document feature dimension. All sample points constitute the first dataset. Based on the preset first density neighborhood radius and first density point number threshold, use a density-based clustering algorithm to cluster the sample points in the first dataset to obtain at least one first cluster. The first cluster containing the most sample points is determined as the main cluster. Sample points in the first dataset that are not assigned to the main cluster are regarded as outlier sample points. Determine the target document feature dimension based on the outlier sample points.

[0049] In this step, the specific implementation steps for determining the feature dimensions of the target document based on outlier sample points include steps S321-S322; Step S321: The second dataset is composed of outlier sample points. Based on the preset second density neighborhood radius and second density point number threshold, the sample points in the second dataset are clustered using a density-based clustering algorithm to obtain at least one second cluster. The first density neighborhood radius is greater than the second density neighborhood radius, and the first density point number threshold is greater than the second density point number threshold. Step S322: Calculate the average of the first differences corresponding to each sample point in the main cluster, as the main cluster average difference; calculate the average of the first differences corresponding to each sample point in each second cluster, as the secondary cluster average difference corresponding to that second cluster; obtain the first difference of isolated sample points that are not assigned to any second cluster, as the isolated point difference; compare the main cluster average difference, the secondary cluster average difference, and the isolated point difference, and determine the main cluster, secondary cluster, or isolated sample point corresponding to the largest value among the three as the target cluster; if the target cluster is an isolated sample point, then determine the document feature dimension corresponding to that isolated sample point as the target document feature dimension; if the target cluster is a main cluster or a secondary cluster, then select the document feature dimension with the largest first difference from all the document feature dimensions contained in the target cluster, and determine it as the target document feature dimension.

[0050] In step S3, comprehensive information on the feature dimensions of the target document is generated. The specific implementation steps for assisting the review of the target scientific and technological document based on the comprehensive information include steps S33 and S34. Figure 2 This is a DBSCAN clustering scatter plot of the document feature dimensions of the auxiliary review method for scientific and technological literature described in Embodiment 1 of this invention. (Reference) Figure 2 In this step, nine document feature dimensions are used as sample points, and the feature vector of each sample point is (first difference, standard deviation). DBSCAN density clustering is used. The first differences of eight dimensions are all less than 1.0, forming a single main cluster. The first difference of the "number of references" dimension is 3.8, and the standard deviation is 0.7, far from the main cluster, and it is identified as an outlier (isolated point). Its corresponding dimension is the target document feature dimension. The specific clustering results are shown in Table 9.

[0051] Table 9. Clustering results based on document feature dimensions and density. Step S33: Calculate the product of the first difference and the standard deviation corresponding to the feature dimension of the target document, and sum the product with the first difference to obtain the comprehensive anomaly index; generate corresponding graded warning labels based on the comparison results of the preset anomaly index threshold and the comprehensive anomaly index; obtain the name, original value, and statistical distribution information of the feature dimension of the target document in historical scientific and technological literature. The statistical distribution information includes the average, median, upper quartile Q3, and lower quartile Q1 of the historical values; compare the original value with the statistical distribution information to calculate the deviation direction and deviation degree: if the original value is greater than Q3, it is judged as positively high; if the original value is less than Q1, it is judged as negatively low; if the original value is between Q1 and Q3, it is judged as being in the normal range; deviation degree = (original value − median) / median × 100%, the deviation degree is positive when the deviation direction is positively high, and negative when the deviation direction is negatively low; Step S34: Generate comprehensive information on the target document's feature dimensions. The comprehensive information should include at least the following: the name of the target document's feature dimension; the original value of the target scientific and technological document on the target document's feature dimension; the statistical distribution information of historical scientific and technological documents on this dimension; the direction and degree of deviation of the original value relative to the historical statistical distribution; and the graded warning label. Based on the type of the target document's feature dimension, match the corresponding verification guidance information from the preset verification rule base, and output the comprehensive information and verification guidance information together to guide reviewers in conducting targeted verification of the target scientific and technological document.

[0052] In this step, each target document feature dimension in the verification rule base has corresponding verification guidance information; The target document feature dimension calculated by this invention represents the feature dimension that has the greatest impact on the overall quality evaluation prediction value of the document, that is, the key anomaly driving dimension of the document. It is neither simply the dimension that deviates the most from the historical average, nor the subjectively determined important dimension, but the dimension that contributes the most to the deviation of the evaluation result of the document from the normal expectation through permutation analysis.

[0053] Based on this, steps S33 and S34 further quantify this dimension in depth: Assuming the target scientific document is a paper submitted to a computer vision journal with 95 references, the system calculates the first difference of 3.8 and the standard deviation of 0.7 for the "number of references" dimension through the aforementioned steps. Therefore, the comprehensive anomaly index = 3.8 + 3.8 × 0.7 = 6.46. If the preset high-risk threshold is 5.0, the comprehensive anomaly index of this document exceeds the high-risk threshold, and the system automatically generates a "high-risk warning" label. Subsequently, the system obtains the original 95 values ​​for the "number of references" dimension of this document and statistically obtains the reference number distribution information of similar documents in this field from historical scientific literature: the lower quartile Q1 is 32 references, the median is 42 references, and the upper quartile Q3 is 67 references. Since the original value of 95 references is greater than Q3=67 references, the system determines the deviation direction to be positively high, with a deviation degree of (95-42) / 42×100%=126.2%. The system generates comprehensive information, including: dimension name "Number of References", original value "95 references", historical statistical distribution "Q1=32 references, median=42 references, Q3=67 references", deviation direction "positively high", deviation degree "+126.2%", and graded warning label "high-risk warning". Based on the dimension type of "Number of References", the system matches the corresponding verification guidance information from the preset verification rule base: "It is recommended to focus on verifying the reference list of this manuscript to check for suspected academic misconduct characteristics such as a large number of irrelevant references, abnormal self-citations, chaotic citation formats, or citations of outdated references." The core logic of this process lies in first identifying the truly noteworthy dimensions, and then conducting precise deviation analysis and verification recommendations for those dimensions. This is because a single document typically has dozens of characteristic dimensions; comparing each dimension with historical data without impact assessment will generate numerous invalid alerts. Conversely, providing verification recommendations without deviation analysis lacks quantitative basis and therefore loses persuasiveness. Only by first identifying the most influential target dimension through substitution analysis, and then performing a progressive process of deviation analysis, tiered warnings, and matching verification guidelines, can we ensure that review resources are focused on the most critical risk points, achieving a complete closed loop of review support from anomaly identification to quantitative grading and then to actionable recommendations.

[0054] The auxiliary review method for scientific and technological literature provided in this embodiment has the following significant technical advantages compared to existing technologies: By performing variable substitution on each feature dimension of the target literature and combining it with density clustering analysis, the system can automatically identify the target dimension that has an abnormal impact on the evaluation results from multiple literature feature dimensions, locating the problem without manual page-by-page inspection. Furthermore, the system quantifies the direction and degree of deviation of the target dimension relative to the distribution of historical literature, and generates graded warning labels according to preset thresholds. Simultaneously, it automatically matches corresponding verification guidance information based on the dimension type, enabling reviewers not only to know "which indicator is abnormal," but also to clarify "the degree of abnormality" and "what content should be focused on for verification." Thus, this method transforms traditional literature quality review, which relies on human experience, into data-driven automated screening and precise location, significantly improving the efficiency and objectivity of auxiliary review of scientific and technological literature, and reducing the workload and risk of missed detections in manual review.

[0055] Example 2 like Figure 3 As shown in the figure, this embodiment provides an auxiliary review system for scientific and technological documents. The system includes an acquisition module 1, a replacement module 2, and a generation module 3.

[0056] Module 1 is used to acquire the feature data of the target scientific and technological literature to be analyzed, as well as the feature data and quality score of historical scientific and technological literature. The feature data includes the values ​​of multiple feature dimensions of the literature. The permutation module 2 is used to perform variable permutation operations on each feature dimension of the target scientific and technological literature, generating permutation feature data corresponding to each feature dimension; each permutation feature data is input into a preset literature evaluation model to obtain permutation analysis results corresponding to each permutation feature data, and the literature evaluation model is used to predict the quality score of the scientific and technological literature. The generation module 3 is used to input the feature data of the target scientific and technological literature into the literature evaluation model to obtain the original analysis results; based on the degree of difference between the original analysis results and the replacement analysis results, the target literature feature dimensions are determined from multiple literature feature dimensions, and comprehensive information of the target literature feature dimensions is generated to assist in the review of the target scientific and technological literature based on the comprehensive information.

[0057] In one specific embodiment of this disclosure, the replacement module 2 further includes a verification unit 21, a construction unit 22, and a sorting unit 23.

[0058] The verification unit 21 is used to obtain all historical values ​​of each document feature dimension in historical scientific and technological literature, and use multiple historical values ​​as a set of candidate replacement values; to perform a rationality verification on each candidate replacement value in the candidate replacement value set, and to remove unreasonable candidate replacement values ​​according to the verification results, so as to obtain the qualified candidate replacement value corresponding to the document feature dimension. Construction unit 22 is used to replace the original value of the document feature dimension with each qualified candidate replacement value, while keeping the values ​​of other document feature dimensions unchanged, to construct preliminary replacement feature data. Each preliminary replacement feature data is then input into the document evaluation model for prediction to obtain the prediction result. The sorting unit 23 is used to calculate the absolute value of the difference between the original analysis results and each prediction result, sort the qualified replacement values ​​in descending order of the absolute value of the difference, select the top-ranked replacement values ​​as the final replacement values, replace the original value of the document feature dimension with each final replacement value, and leave the values ​​of other document feature dimensions unchanged, thereby constructing the replacement feature data corresponding to the document feature dimension.

[0059] In one specific embodiment of this disclosure, the verification unit 21 further includes a first calculation unit 211 and an acquisition unit 212.

[0060] The first calculation unit 211 is used to calculate the upper quartile and lower quartile of all candidate replacement values ​​in the candidate replacement value set; and to calculate the difference between the upper quartile and the lower quartile to obtain the interquartile range. The acquisition unit 212 is used to acquire a set of candidate adjustment coefficients, which includes multiple candidate adjustment coefficients that increase sequentially. It calculates the product of each candidate adjustment coefficient and its interquartile range to obtain multiple weighted interquartile ranges. For each candidate adjustment coefficient, a temporary normal value interval is determined based on the lower quartile, the upper quartile, and the corresponding weighted interquartile range. The number of candidate replacement values ​​outside this temporary normal value interval is counted, and the ratio of this number to the total number of candidate replacement values ​​in the candidate replacement value set is calculated as the outlier removal ratio corresponding to that candidate adjustment coefficient. Qualified candidate replacement values ​​are obtained based on all outlier removal ratios.

[0061] In one specific embodiment of this disclosure, the acquisition unit 212 further includes a rejection unit 2121 and a search unit 2122.

[0062] The elimination unit 2121 is used to compare the outlier elimination ratio corresponding to each candidate adjustment coefficient with a preset ratio range in ascending order of candidate adjustment coefficients in the candidate adjustment coefficient set, and determine the candidate adjustment coefficient whose outlier elimination ratio falls into the preset ratio range for the first time as the target adjustment coefficient; determine the normal value range based on the lower quartile, upper quartile and the weighted interquartile range corresponding to the target adjustment coefficient; if the candidate replacement value is outside the normal value range, then the candidate replacement value is determined as the first type of unreasonable candidate replacement value and eliminated to obtain the remaining candidate replacement values; The search unit 2122 is used to search for historical scientific and technological documents in which the value of the document feature dimension is equal to the candidate replacement value for each remaining candidate replacement value, and obtain the corresponding quality score value; calculate the dispersion between all quality score values, and if the dispersion exceeds the preset stability threshold, then the candidate replacement value is identified as the second type of unreasonable candidate replacement value and is removed, and finally qualified candidate replacement value is obtained.

[0063] In one specific embodiment of this disclosure, the generation module 3 further includes a second calculation unit 31 and a clustering unit 32.

[0064] The second calculation unit 31 is used to calculate the average value of all permutation analysis results corresponding to each document feature dimension, as the comprehensive permutation analysis result corresponding to the document feature dimension; calculate the absolute difference between the original analysis result and the comprehensive permutation analysis result, as the first difference corresponding to the document feature dimension; and calculate the standard deviation of all permutation analysis results corresponding to the document feature dimension. Clustering unit 32 is used to treat each document feature dimension as a sample point. The feature vector of each sample point includes the first difference and standard deviation corresponding to the document feature dimension. All sample points constitute the first dataset. Based on the preset first density neighborhood radius and first density point number threshold, a density-based clustering algorithm is used to cluster the sample points in the first dataset to obtain at least one first cluster. The first cluster containing the most sample points is determined as the main cluster. Sample points in the first dataset that are not assigned to the main cluster are regarded as outlier sample points. The target document feature dimension is determined based on the outlier sample points.

[0065] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0066] Example 3 Corresponding to the above method embodiments, this disclosure also provides an auxiliary review device for scientific and technological documents. The auxiliary review device for scientific and technological documents described below and the auxiliary review method for scientific and technological documents described above can be referred to in correspondence with each other.

[0067] Figure 4 This is a block diagram illustrating an auxiliary review device 300 for scientific and technological documents according to an exemplary embodiment. Figure 4 As shown, the auxiliary review device 300 for scientific and technological documents may include a processor 301 and a memory 302. The auxiliary review device 300 for scientific and technological documents may also include one or more of a multimedia component 303, an I / O interface 304, and a communication component 305.

[0068] The processor 301 controls the overall operation of the auxiliary review device 300 for scientific and technological documents to complete all or part of the steps in the aforementioned auxiliary review method for scientific and technological documents. The memory 302 stores various types of data to support the operation of the auxiliary review device 300 for scientific and technological documents. This data may include, for example, instructions for any application or method operating on the auxiliary review device 300 for scientific and technological documents, as well as application-related data, such as contact data, sent and received messages, images, audio, video, etc. The memory 302 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 303 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 302 or transmitted via the communication component 305. The audio component also includes at least one speaker for outputting audio signals. I / O interface 304 provides an interface between processor 301 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 305 is used for wired or wireless communication between the auxiliary review device 300 for scientific and technical documents and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 305 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0069] In an exemplary embodiment, the auxiliary review device 300 for scientific and technological documents may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the aforementioned auxiliary review method for scientific and technological documents.

[0070] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the above-described auxiliary review method for scientific and technological documents. For example, the computer-readable storage medium may be the memory 302 including program instructions, which may be executed by the processor 301 of the auxiliary review device 300 for scientific and technological documents to complete the above-described auxiliary review method for scientific and technological documents.

[0071] Example 4 Corresponding to the above method embodiments, this disclosure also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the above-described auxiliary review method for scientific and technological documents.

[0072] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the auxiliary review method for scientific and technological documents described in the above method embodiments.

[0073] Specifically, the readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0074] The above description is merely a preferred 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 principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assisting in the review of scientific and technological literature, characterized in that, include: The process involves acquiring feature data of the target scientific and technological literature to be analyzed, as well as feature data and quality scores of historical scientific and technological literature. The feature data includes values ​​for multiple document feature dimensions. These document feature dimensions include, but are not limited to, document self-attribute dimensions, author association dimensions, and journal association dimensions. A variable substitution operation is performed on each feature dimension of the target scientific and technological literature to generate substitution feature data corresponding to each feature dimension. Each substitution feature data is then input into a preset literature evaluation model to obtain the substitution analysis result corresponding to each substitution feature data. The literature evaluation model is used to predict the quality score of the scientific and technological literature. The variable substitution operation includes at least the following: replacing unreasonable values ​​of the literature feature dimensions with reasonable values ​​of the literature feature dimensions based on the feature data and quality score of historical scientific and technological literature. The characteristic data of the target scientific and technological literature is input into the literature evaluation model to obtain the original analysis results. Based on the degree of difference between the original analysis results and the replacement analysis results, the characteristic dimensions of the target literature are determined from multiple literature characteristic dimensions, and comprehensive information on the characteristic dimensions of the target literature is generated. The comprehensive information is used to assist in the review of the target scientific and technological literature.

2. The auxiliary review method for scientific and technological documents according to claim 1, characterized in that, Perform variable permutation operations on each feature dimension of the target scientific literature's feature data to generate permuted feature data corresponding to each feature dimension, including: For each document feature dimension, all historical values ​​of that document feature dimension are obtained from historical scientific and technological documents, and multiple historical values ​​are used as a set of candidate replacement values. The rationality of each candidate replacement value in the candidate replacement value set is verified, and unreasonable candidate replacement values ​​are eliminated based on the verification results to obtain the qualified candidate replacement value corresponding to that document feature dimension. Replace the original value of each qualified candidate replacement value with the original value of the document feature dimension, while keeping the values ​​of other document feature dimensions unchanged, to construct preliminary replacement feature data. Input each preliminary replacement feature data into the document evaluation model for prediction to obtain the prediction results. Calculate the absolute value of the difference between the original analysis results and each prediction result. Sort the qualified replacement values ​​in descending order of the absolute value of the difference. Select the top-ranked replacement values ​​as the final replacement values. Replace the original value of the document feature dimension with each final replacement value, while keeping the values ​​of other document feature dimensions unchanged. Construct the replacement feature data corresponding to the document feature dimension.

3. The auxiliary review method for scientific and technological documents according to claim 2, characterized in that, Each candidate replacement value in the candidate replacement value set is validated for reasonableness. Based on the validation results, unreasonable candidate replacement values ​​are removed, resulting in qualified candidate replacement values ​​corresponding to the feature dimension of the document, including: Calculate the upper and lower quartiles of all candidate replacement values ​​in the candidate replacement value set; calculate the difference between the upper and lower quartiles to obtain the interquartile range; Obtain a set of candidate adjustment coefficients, which includes multiple candidate adjustment coefficients in ascending order. Calculate the product of each candidate adjustment coefficient and its interquartile range to obtain multiple weighted interquartile ranges. For each candidate adjustment coefficient, determine a temporary normal value interval based on the lower quartile, upper quartile, and the corresponding weighted interquartile range. Count the number of candidate replacement values ​​outside this temporary normal value interval in the candidate replacement value set, and calculate the ratio of this number to the total number of candidate replacement values ​​in the candidate replacement value set. Use this ratio as the outlier removal ratio for that candidate adjustment coefficient. Obtain qualified candidate replacement values ​​based on all outlier removal ratios.

4. The auxiliary review method for scientific and technological documents according to claim 3, characterized in that, Based on all outlier removal ratios, qualified candidate replacement values ​​were obtained, including: According to the candidate adjustment coefficient set in ascending order, the outlier removal ratio corresponding to each candidate adjustment coefficient is compared with the preset ratio range. The candidate adjustment coefficient whose outlier removal ratio falls into the preset ratio range for the first time is determined as the target adjustment coefficient. The normal value range is determined according to the lower quartile, upper quartile and the weighted interquartile range corresponding to the target adjustment coefficient. If the candidate replacement value is outside the normal value range, the candidate replacement value is determined as the first type of unreasonable candidate replacement value and is removed to obtain the remaining candidate replacement values. For each remaining candidate replacement value, search for historical scientific and technological literature in which the value of the literature feature dimension is equal to the candidate replacement value, and obtain its corresponding quality score value; calculate the dispersion among all quality score values. If the dispersion exceeds the preset stability threshold, the candidate replacement value is identified as the second type of unreasonable candidate replacement value and is removed, and finally qualified candidate replacement values ​​are obtained.

5. The auxiliary review method for scientific and technological documents according to claim 1, characterized in that, Based on the degree of difference between the original analysis results and the permutation analysis results, the target document feature dimensions are determined from multiple document feature dimensions, including: For each document feature dimension, the average of all permutation analysis results corresponding to that document feature dimension is calculated as the comprehensive permutation analysis result corresponding to that document feature dimension; the absolute difference between the original analysis result and the comprehensive permutation analysis result is calculated as the first difference corresponding to that document feature dimension, and the standard deviation of all permutation analysis results corresponding to that document feature dimension is calculated. Each document feature dimension is treated as a sample point. The feature vector of each sample point includes the first difference and standard deviation corresponding to that document feature dimension. All sample points constitute the first dataset. Based on the preset first density neighborhood radius and first density point number threshold, a density-based clustering algorithm is used to cluster the sample points in the first dataset to obtain at least one first cluster. The first cluster containing the most sample points is determined as the main cluster. Sample points in the first dataset that are not assigned to the main cluster are regarded as outlier sample points. The target document feature dimension is determined based on the outlier sample points.

6. An auxiliary review system for scientific and technological documents, characterized in that, include: The acquisition module is used to acquire the feature data of the target scientific and technological literature to be analyzed, as well as the feature data and quality score of historical scientific and technological literature. The feature data includes the values ​​of multiple feature dimensions of the literature. The permutation module is used to perform variable permutation operations on each feature dimension of the target scientific and technological literature, generating permutation feature data corresponding to each feature dimension; each permutation feature data is input into a preset literature evaluation model to obtain permutation analysis results corresponding to each permutation feature data, and the literature evaluation model is used to predict the quality score of scientific and technological literature. The generation module is used to input the feature data of the target scientific and technological literature into the literature evaluation model to obtain the original analysis results; based on the degree of difference between the original analysis results and the replacement analysis results, the target literature feature dimensions are determined from multiple literature feature dimensions, and comprehensive information on the target literature feature dimensions is generated to assist in the review of the target scientific and technological literature.

7. The auxiliary review system for scientific and technological documents according to claim 6, characterized in that, The replacement module includes: The verification unit is used to obtain all historical values ​​of each document feature dimension from historical scientific and technological literature, and use multiple historical values ​​as a set of candidate replacement values; to perform a rationality verification on each candidate replacement value in the candidate replacement value set, and to remove unreasonable candidate replacement values ​​based on the verification results, so as to obtain the qualified candidate replacement value corresponding to the document feature dimension. The construction unit is used to replace the original value of each qualified candidate replacement value of the document feature dimension, while keeping the values ​​of other document feature dimensions unchanged, to construct preliminary replacement feature data. Each preliminary replacement feature data is then input into the document evaluation model for prediction to obtain the prediction result. The sorting unit is used to calculate the absolute value of the difference between the original analysis results and each prediction result. The qualified replacement values ​​are sorted in descending order of the absolute value of the difference, and the top-ranked replacement values ​​are selected as the final replacement values. Each final replacement value replaces the original value of the document feature dimension, while the values ​​of other document feature dimensions remain unchanged, thus constructing the replacement feature data corresponding to the document feature dimension.

8. The auxiliary review system for scientific and technological documents according to claim 7, characterized in that, The verification unit includes: The first calculation unit is used to calculate the upper quartile and lower quartile of all candidate replacement values ​​in the candidate replacement value set; and to calculate the difference between the upper quartile and the lower quartile to obtain the interquartile range. The acquisition unit is used to acquire a set of candidate adjustment coefficients, which includes multiple candidate adjustment coefficients that increase sequentially. It calculates the product of each candidate adjustment coefficient and its interquartile range to obtain multiple weighted interquartile ranges. For each candidate adjustment coefficient, a temporary normal value interval is determined based on the lower quartile, upper quartile, and the corresponding weighted interquartile range. The number of candidate replacement values ​​outside this temporary normal value interval is counted, and the ratio of this number to the total number of candidate replacement values ​​in the candidate replacement value set is calculated as the outlier removal ratio corresponding to that candidate adjustment coefficient. Qualified candidate replacement values ​​are obtained based on all outlier removal ratios.

9. The auxiliary review system for scientific and technological documents according to claim 8, characterized in that, The acquisition unit includes: The elimination unit is used to compare the outlier elimination ratio of each candidate adjustment coefficient with a preset ratio range in ascending order of candidate adjustment coefficients in the candidate adjustment coefficient set. The candidate adjustment coefficient whose outlier elimination ratio falls into the preset ratio range for the first time is determined as the target adjustment coefficient. The normal value range is determined based on the lower quartile, upper quartile, and the weighted interquartile range corresponding to the target adjustment coefficient. If the candidate replacement value is outside the normal value range, the candidate replacement value is determined as the first type of unreasonable candidate replacement value and eliminated to obtain the remaining candidate replacement values. The search unit is used to find historical scientific and technological documents from historical scientific and technological documents for each remaining candidate replacement value, whose document feature dimension value is equal to the candidate replacement value, and obtain its corresponding quality score value; calculate the dispersion among all quality score values, and if the dispersion exceeds the preset stability threshold, the candidate replacement value is identified as the second type of unreasonable candidate replacement value and is removed, and finally qualified candidate replacement values ​​are obtained.

10. The auxiliary review system for scientific and technological documents according to claim 6, characterized in that, The generation module includes: The second calculation unit is used to calculate the average value of all permutation analysis results corresponding to each document feature dimension, as the comprehensive permutation analysis result corresponding to that document feature dimension; calculate the absolute difference between the original analysis result and the comprehensive permutation analysis result, as the first difference corresponding to that document feature dimension; and calculate the standard deviation of all permutation analysis results corresponding to that document feature dimension. A clustering unit is used to treat each document feature dimension as a sample point. The feature vector of each sample point includes the first difference and standard deviation corresponding to the document feature dimension. All sample points constitute the first dataset. Based on a preset first density neighborhood radius and a first density point number threshold, a density-based clustering algorithm is used to cluster the sample points in the first dataset to obtain at least one first cluster. The first cluster containing the most sample points is determined as the main cluster. Sample points in the first dataset that are not assigned to the main cluster are considered outliers. The target document feature dimension is determined based on the outliers.

Citation Information

Patent Citations

  • A substation knowledge graph construction and optimization method based on multi-view learning

    CN117251582B

  • A power grid substation engineering knowledge graph construction and retrieval method and system

    CN121561116B