Paper review expert intelligent recommendation method and system

Abstract

The invention relates to the technical field of intelligent retrieval of papers, and discloses a paper review expert intelligent recommendation method and system.According to the method, word paraphrasing is carried out on review expert text information and to-be-reviewed paper text information, semantic connotation supplementation and extension can be carried out on text information with limited information amount, and therefore the content of paper review experts is improved. Richer and more useful information is provided; multi-source information fusion is carried out on the expanded text information, the expanded rich multi-source text information can be summarized, extracted and summarized, and comprehensive text information with sufficient information content and high distinction degree is provided; and performing overall vectorization on the comprehensive text information, performing similarity calculation and sorting on the review expert text vector and the to-be-reviewed paper text vector, and screening out review expert information with relatively high similarity as output. The method can effectively improve the matching precision of the paper'small peer 'review experts, reduces the rejection rate caused by inconsistent research directions, and provides technical support for improving the review quality and shortening the review period.

Description

Technical Field

[0001] This application relates to the technical field of intelligent retrieval of papers, and particularly to a method and system for intelligent recommendation of paper review experts. Background Art

[0002] In the field of higher education, the review of dissertations is a key measure to examine the quality of dissertations and ensure the quality of student cultivation. China has formed a system of pre-defense submission and post-defense random inspection for undergraduate, master's, and doctoral dissertations. According to the requirements of relevant documents, a dissertation needs to be independently evaluated by multiple review experts, and the review results will become important bases for whether students can participate in the graduation defense on schedule and whether they meet the degree awarding conditions.

[0003] As the starting link of dissertation review, the selection of review experts aims to select "peer" experts with the same research direction for dissertations, expecting that the review experts can provide more high-quality comments and more objective scores. Traditionally, the distribution of dissertations mainly includes manual matching and feature vector matching. Manual matching is often time-consuming, and the matching results are unstable due to the limitations of the professional knowledge of the submission personnel; feature vector matching can effectively reduce the matching time, but due to technical limitations, it is difficult to deeply understand Chinese texts with the characteristics of free expression and rich semantic connotations, resulting in difficulty in ensuring the best match between the review experts and the dissertation topics. With the development of information technology, especially the technological progress in the fields of natural language processing and artificial intelligence, new possibilities have been provided for the recommendation of dissertation experts. However, most of the existing technologies focus on the matching of review experts for general academic achievements such as journal papers and conference papers, lacking in-depth optimization for the characteristics of dissertations, such as interdisciplinary review and detailed research direction.

[0004] Therefore, there is an urgent need for a method and system for recommending dissertation review experts that can comprehensively process multi-dimensional long texts and semantic information such as the titles, abstracts, keywords, subject attributes of dissertations, as well as the research directions and guiding experiences of experts. Summary of the Invention

[0005] This application provides a method and system for intelligent recommendation of paper review experts to solve the problem that most of the existing technologies focus on the matching of review experts for general academic achievements such as journal papers and conference papers, lacking in-depth optimization for the characteristics of dissertations, such as interdisciplinary review and detailed research direction.

[0006] To achieve the above object, this application is implemented through the following technical solutions: In the first aspect, this application provides a method for intelligent recommendation of paper review experts, including: S1: Obtain the personal text information of review experts, perform word interpretation and multi-source information fusion on the personal text information of review experts to construct a first database; S2: Vectorize the text information in the first database to obtain a first vector library; S3: Obtain the text information of the paper to be reviewed, perform word interpretation and multi-source information fusion on the text information of the paper to be reviewed to obtain the comprehensive text information of the paper to be reviewed; S4: Vectorize the comprehensive text information to obtain the text information vector of the paper to be reviewed; S5: Calculate the similarity between the text information vector of the paper to be reviewed and the text information vector of the first vector library; S6: Output the personal text information of review experts with similarity exceeding the threshold as the recommendation result.

[0007] In a second aspect, the present application provides an intelligent recommendation system for paper review experts, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method described in the first aspect above are implemented.

[0008] Beneficial effects: The intelligent recommendation method for paper review experts provided by the present application first performs word interpretation on the text information of review experts and the text information of the paper to be reviewed, which can supplement and expand the semantic connotation of the text information with limited information volume and provide richer and more useful information. Then, multi-source information fusion is performed on the expanded text information, which can summarize, refine, and summarize the rich multi-source text information after expansion and provide comprehensive text information with sufficient information content and high discrimination. Finally, the comprehensive text information is vectorized as a whole, and the similarity between the review expert text vector and the paper to be reviewed text vector is calculated and sorted, and the review expert information with higher similarity is selected as the output. Through this method, the matching accuracy of "small peer" review experts for papers can be effectively improved, the rejection rate caused by inconsistent research directions can be reduced, and technical support is provided for improving the review quality and shortening the review cycle. Description of the drawings

[0009] Figure 1 It is a flowchart of an intelligent recommendation method for paper review experts according to a preferred embodiment of the present application. Detailed implementation manners

[0010] The technical solutions of the present application will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts belong to the protection scope of the present application.

[0011] Unless otherwise defined, the technical terms or scientific terms used in this application shall have the ordinary meanings understood by those of ordinary skill in the art to which this application pertains. The terms "first", "second" and similar words used in this application do not denote any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "a" or "one" do not denote a quantity limitation, but mean that there is at least one. The terms "connected" or "coupled" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right", etc. are only used to indicate relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0012] Please refer to Figure 1 , this application provides an intelligent recommendation method for thesis review experts, including: S1: Obtain the text information of review experts, perform word interpretation and multi-source information fusion on the text information of review experts, and construct a first database.

[0013] In this step, the text information of review experts includes, but is not limited to, text information such as the research directions of review experts, the scientific fields they are engaged in, and the disciplines they belong to.

[0014] S2: Vectorize the text information of the first database to construct a first vector library.

[0015] S3: Obtain the text information of the thesis to be reviewed, perform word interpretation and multi-source information fusion on the text information of the thesis to be reviewed, and obtain the comprehensive text information of the thesis to be reviewed.

[0016] In this step, the thesis to be reviewed includes, but is not limited to, bachelor's, master's, and doctoral theses, journal papers, and conference papers. The text information of the thesis to be reviewed includes, but is not limited to, text information such as the discipline to which the thesis belongs, the title of the thesis, keywords, and abstract.

[0017] S4: Vectorize the comprehensive text information to obtain the text information vector of the thesis to be reviewed.

[0018] S5: Calculate the similarity between the text information vector of the thesis to be reviewed and the text information vectors in the first vector library.

[0019] In this step, the principles for constructing the screening function include, but are not limited to, the principle of cross-discipline avoidance and the principle of same-unit avoidance between the thesis to be reviewed and the review experts.

[0020] S6: According to the similarity results, take the information of the review experts with higher similarity as the recommended result for output.

[0021] The above intelligent recommendation method for paper review experts first performs word interpretation on the text information of review experts and the text information of papers to be reviewed, which can supplement and expand the semantic connotations of text information with limited information content, providing richer and more useful information. Then, multi-source information fusion is performed on the expanded text information, which can summarize, refine, and summarize the rich multi-source text information after expansion, providing comprehensive text information with sufficient information content and high discrimination. Finally, the comprehensive text information is vectorized as a whole, and the similarity between the text vectors of review experts and the text vectors of papers to be reviewed is calculated and sorted, and the review expert information with higher similarity is selected as the output. Through this method, the matching accuracy of "narrow-scope" review experts for papers can be effectively improved, the rejection rate caused by inconsistent research directions can be reduced, and technical support can be provided for improving the review quality and shortening the review cycle.

[0022] This embodiment is illustrated by taking a certain doctoral dissertation as an example. The title of this doctoral dissertation is "Cross-Domain Fault Diagnosis Method for Industrial Processes Based on Sub-Domain Adapted Dictionary Learning", and the text information used is shown in Table 1.

[0023] Table 1 Text Information of Doctoral Dissertation

[0024] Optionally, the S1 includes: S11: Obtain the text information of review experts. Let the set of review expert text information be: ; Among them, represents the set of review expert text information obtained, represents the th subset of text information of the th review expert, represents the th type of text information of the th review expert; , ; represents the total number of review experts obtained, represents the total number of types of review expert text information.

[0025] Specifically, the expert database used in this embodiment covers 80 subject professional fields and contains relevant information of more than 18,000 review experts. Therefore, the total number of review experts , and the total number of types of review expert text information . The text information of the first review expert is shown in Table 2.

[0026] Table 2 Text Information of the First Review Expert

[0027]

[0028] S12: Interpret the text information of the review experts to obtain the extended text information of the review experts. Let the word interpretation model for the text information of the review experts be: ; In the formula, represents the set of all extended text information of the review experts, represents the word interpretation function for the text information of the review experts, represents the th set of extended text information of the review expert, as follows: ; represents the th type of extended text information of the th review expert. Specifically, in this embodiment, the word interpretation function for the text information of the review experts is implemented using electronic tools such as encyclopedias (e.g., Baidu Encyclopedia, etc.), word dictionaries (e.g., Xinhua Dictionary, etc.), large language models (e.g., Wenyan Yixin, etc.) to interpret the subject nouns, research directions, and scientific fields in the text information of the review experts; no extension is made to the personal academic qualification information such as the review expert's title and supervisor type in the text information of the review experts. Among them, the extended text information of the first review expert is shown in Table 3.

[0029] Table 3 Extended Text Information of the First Review Expert

[0030]

[0031] S13: Fuse the extended text information of the review experts from multiple sources to obtain the comprehensive text information of the review experts and construct the first database. Let the multi-source information fusion model for the extended text information of the review experts be: ; In the formula, represents the first database composed of the set of all comprehensive text information of the review experts, represents the multi-source information fusion function for the extended text information of the review experts, represents the th comprehensive text information of the review expert.

[0032] Specifically, in this embodiment, the multi-source information fusion function Implemented using large language models (e.g., Wenyan Yixin, Kimi, ChatGPT, etc.), the extended text information of the review experts is input into the large language model as background knowledge, and multi-source information fusion of the extended text information of the review experts is performed under standardized prompts. Among them, the comprehensive text information of the first review expert is shown in Table 4.

[0033] Table 4 Comprehensive Text Information of the First Review Expert

[0034]

[0035] Optionally, S2 includes vectorizing the text information of the first database to obtain the text vector of the comprehensive text information of the review experts, and constructing the first vector library. Let the text vectorization model for the comprehensive text information of the review experts be: ; In the formula, represents the first vector library composed of the set of text vectors of all the comprehensive text information of the review experts, represents the text vectorization function for the comprehensive text information of the review experts, represents the th text vector of the comprehensive text information of the review experts.

[0036] Specifically, in this embodiment, the text vectorization function for the comprehensive text information of the review experts is implemented using a Bi-Encoder model. Among them, the text vector of the comprehensive information of the first review expert is shown in Table 5.

[0037] Table 5 Text Vector of the Comprehensive Information of the First Review Expert

[0038] Optionally, S3 includes: S31: Obtain the text information of the paper to be reviewed, denoted as: ; Among them, represents the set of text information of the paper to be reviewed, represents the th type of text information of the paper to be reviewed, ; represents the total number of types of text information of the paper to be reviewed.

[0039] Specifically, in this embodiment, . The text information of the paper to be reviewed is shown in Table 6.

[0040] Table 6 Text Information of the Thesis

[0041] S32: Interpret the text information of the paper to be reviewed to obtain the extended text information of the paper to be reviewed. Let the word interpretation model for the text information of the paper to be reviewed be: ; In the formula, represents the extended text information of the paper to be reviewed, represents the word interpretation function for the text information of the paper to be reviewed, represents the th type of extended text information of the paper to be reviewed.

[0042] Specifically, in this embodiment, the word interpretation function for the text information of the paper to be reviewed is implemented using electronic tools such as encyclopedias (such as Baidu Encyclopedia, etc.), word dictionaries (such as Xinhua Dictionary, etc.), large language models (such as Wenyan Yixin, etc.), etc. to interpret the words in the paper title, subject nouns, keywords, and paper direction in the text information of the degree thesis to be reviewed; no extension is made to the information such as abstract and training level in the text information of the degree thesis to be reviewed. Among them, the extended text information of the degree thesis to be reviewed is shown in Table 7.

[0043] Table 7 Extended text information of the degree thesis

[0044] S33: Fuse the extended text information of the paper to be reviewed with multi-source information to obtain the comprehensive text information of the paper to be reviewed. Let the multi-source information fusion model for the extended text information of the paper to be reviewed be: ; In the formula, represents the comprehensive text information of the paper to be reviewed, represents the multi-source information fusion function for the extended text information of the paper to be reviewed.

[0045] Specifically, in this embodiment, the multi-source information fusion function for the extended text information of the paper to be reviewed is implemented using a large language model (such as Wenyan Yixin, etc.). The extended text information of the paper to be reviewed is input into the large language model as background knowledge, and multi-source information fusion of the extended text information of the paper to be reviewed is performed under normalized prompt words. Among them, the comprehensive text information of the degree thesis to be reviewed is shown in Table 8.

[0046] Table 8 Comprehensive Text Information of Dissertation

[0047] Optionally, S4 includes vectorizing the comprehensive text information of the dissertation to be reviewed. Let the vectorization model for the comprehensive text information of the dissertation to be reviewed be ; In the formula, represents the text vector of the comprehensive text information of the dissertation to be reviewed, represents the text vectorization function for the comprehensive text information of the dissertation to be reviewed.

[0048] Specifically, in this embodiment, the text vectorization function for the comprehensive text information of the dissertation to be reviewed is implemented using a Bi-Encoder model. The text vector of the comprehensive information of the dissertation to be reviewed is shown in Table 9.

[0049] Table 9 Text Vectors of Comprehensive Information of Dissertation

[0050] Optionally, S5 includes: S51: According to the text information of the dissertation to be reviewed, screen the text vectors of review experts that meet the requirements from the first vector library, and construct a subset of text vectors of review expert information for the dissertation to be reviewed. Let the screening function be: ; In the formula, represents the subset of text vectors of expert information for the dissertation to be reviewed in the first vector library, represents the screening function for constructing the subset of text vectors of expert information, represents the th text vector of expert information in the constructed subset of text vectors of expert information, ; represents the total number of review experts in the constructed subset of text vectors of review expert information.

[0051] Specifically, in this embodiment, the basis for constructing the screening function includes the cross-discipline avoidance principle and the same-unit avoidance principle between the dissertation to be reviewed and the review experts. Among them, the cross-discipline avoidance principle requires that the discipline to which the dissertation belongs and the discipline to which the review expert belongs should avoid crossing the first-level discipline; the same-unit avoidance principle requires that the training unit of the dissertation and the work unit of the review expert should avoid being the same. The total number of review experts in the subset of text vectors of expert information constructed based on these two principles .

[0052] S52: Calculate the similarity between the text vector of the paper to be reviewed and the subset of text vectors of reviewer information screened from the first vector library. Assume the similarity model is: ; In the formula, represents the set of the similarity calculation results; represents the similarity function, represents the similarity calculation result between the text vector of the paper to be reviewed and the th text vector of reviewer information in the subset of text vectors of reviewer information.

[0053] Specifically, in this embodiment, the similarity function uses the cosine function to calculate the distance between two text vectors and takes it as the similarity calculation result.

[0054] Optionally, S6 includes: S61: Sort the similarity results. Assume the first sorting model of the similarity results is: ; In the formula, represents the result sorted with respect to the similarity , is the first sorting function.

[0055] Specifically, in this embodiment, the first sorting function uses an efficient similarity search tool (Facebook AI Similarity Search, FAISS) launched by Meta. Using the traversal optimization algorithm provided by FAISS, through performing a nearest neighbor search, the returned results are usually sorted from small to large according to the cosine distance.

[0056] S62: According to the first sorting result of the similarity, take the top review candidate experts with higher similarity calculation results for re - sorting, and use the list of the top candidate experts in the sorting result of the second stage as the recommended reviewers of the paper. Assume the second sorting model of the similarity is ; In the formula, represents the result after re - sorting with respect to the initial sorting result of the similarity , is the second sorting function, represents taking the top reviewers arranged in descending order from the initial sorting result of the similarity .

[0057] Specifically, in this embodiment, 。The second sorting model uses a Cross-Encode model. The comprehensive text information of the paper to be reviewed and the comprehensive text information of the previous reviewers are jointly input into this model for processing, so as to obtain the scores of the similarity of reviewers in the second sorting stage. Then, the calculation results of the similarity in the second sorting stage are sorted in descending order to obtain the sorting results of the second similarity model.

[0058] Specifically, in this embodiment, 。According to the results of the second sorting model, the information of 5 reviewers with higher similarity is output as the recommended reviewers for the paper to be reviewed, as shown in Table 10.

[0059] Table 10 Information of 5 Recommended Reviewers for Dissertation ID_LW_1

[0060] In summary, this embodiment can recommend reviewers for the dissertation to be reviewed, and recommend reviewers with a high degree of matching of the dissertation theme and following the principles of interdisciplinary avoidance and same-unit avoidance, and recommends "peer" experts that meet the requirements for the dissertation.

[0061] This application also provides an intelligent recommendation system for paper reviewers, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented. This intelligent recommendation system for paper reviewers can implement various embodiments of the above intelligent recommendation method for paper reviewers and can achieve the same beneficial effects, which will not be elaborated here.

[0062] The preferred specific embodiments of this application have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations according to the concept of this application without creative work. Therefore, all technical solutions that can be obtained by those skilled in the art in this technical field based on the concept of this application through logical analysis, reasoning, or limited experiments on the basis of the prior art should be within the protection scope determined by the claims.

Claims

1. An intelligent recommendation method for paper review experts, characterized in that, Including: S1: Obtain the personal text information of review experts, perform word interpretation and multi-source information fusion on the personal text information of review experts to construct a first database; S2: Vectorize the text information in the first database to obtain a first vector library; S3: Obtain the text information of the paper to be reviewed, perform word interpretation and multi-source information fusion on the text information of the paper to be reviewed to obtain the comprehensive text information of the paper to be reviewed; S4: Vectorize the comprehensive text information to obtain the text information vector of the paper to be reviewed; S5: Calculate the similarity between the text information vector of the paper to be reviewed and the text information vectors in the first vector library; S6: Output the personal text information of review experts with similarity exceeding the threshold as the recommendation result.

2. The intelligent recommendation method for paper review experts according to claim 1, wherein The types of papers corresponding to the text information of the papers to be reviewed include bachelor's theses, master's theses, doctoral dissertations, journal papers, and conference papers.

3. The intelligent recommendation method for paper review experts according to claim 1, wherein The S1 includes: S11: Obtain the personal text information of review experts. Let the set of personal text information of review experts satisfy the following relational expression: ； Among them, represents the set of personal text information of the obtained review experts, represents the sub - set of personal text information of the th review expert, represents the th review expert's type of personal text information; , ; represents the total number of obtained review experts, represents the total number of types of personal text information of review experts; S12: Perform word interpretation on the personal text information of review experts to obtain the extended personal text information of review experts. Let the word interpretation model for the text information of review experts be as follows: ； In the formula, represents the set of all reviewers' extended personal text information, represents the word interpretation function for reviewers' personal text information, represents the th set of a reviewer's extended personal text information , represents the th reviewer's th type of extended personal text information; S13: Fuse multi-source information with the extended personal text information of review experts to obtain the comprehensive text information of review experts, and construct a first database. Let the multi-source information fusion model for the extended personal text information of review experts be as follows: ； In the formula, represents the first database composed of the comprehensive text information sets of all review experts, represents the multi-source information fusion function for the extended personal text information of review experts, represents the th comprehensive text information of a review expert.

4. The intelligent recommendation method for paper review experts according to claim 3, wherein The text vectorization model for the comprehensive text information of review experts is as follows: ； In the formula, represents the first vector library composed of a set of text vectors of all review experts' comprehensive text information, represents the text vectorization function for review experts' comprehensive text information, represents the text vector of the comprehensive text information of the 5. The intelligent recommendation method for paper review experts according to claim 1, wherein The S3 includes: S31: Obtain the text information of the paper to be reviewed as follows: ； Among them, represents the set of text information of the papers to be reviewed, represents the type of text information of the -th paper to be reviewed; represents the total number of types of text information of the papers to be reviewed; S32: Perform word interpretation on the text information of the paper to be reviewed to obtain the extended text information of the paper to be reviewed. Let the word interpretation model for the text information of the paper to be reviewed be as follows: ； In the formula, represents the extended text information of the paper to be reviewed, represents the word interpretation function for the text information of the paper to be reviewed, represents the type of extended text information of the paper to be reviewed; S33: Fuse multi-source information with the extended text information of the paper to be reviewed to obtain the comprehensive text information of the paper to be reviewed. Among them, the multi-source information fusion model for the extended text information of the paper to be reviewed is as follows: ； In the formula, represents the comprehensive text information of the paper to be reviewed, represents the multi-source information fusion function for the extended text information of the paper to be reviewed.

6. The intelligent recommendation method for paper review experts according to claim 5, wherein The vectorization model for the comprehensive text information of the paper to be reviewed is as follows: ； Wherein, represents the text vector of the comprehensive text information of the paper to be reviewed, represents the text vectorization function for the comprehensive text information of the paper to be reviewed.

7. The intelligent recommendation method for paper review experts according to claim 1, wherein The S5 includes: S51: Based on the text information of the paper to be reviewed, screen the review expert text vectors that meet the requirements from the first vector library according to the screening function, and construct a subset of review expert information text vectors for the paper to be reviewed. Among them, the screening function is as follows: ； Wherein, represents a subset of expert information text vectors regarding the paper to be reviewed in the first vector library, represents a screening function for constructing the subset of expert information text vectors, represents the th expert information text vector in the constructed subset of expert information text vectors, ; represents the total number of review experts in the constructed subset of review expert information text vectors, represents the extended text information of the paper to be reviewed, represents the first vector library composed of a set of text vectors of comprehensive text information of all review experts; S52: Calculate the similarity between the text vector of the paper to be reviewed and the subset of review expert information text vectors screened from the first vector library based on the similarity model. The similarity model is as follows: ； In the formula, represents the set of the similarity calculation results; represents the similarity function, represents the similarity calculation result between the text vector of the paper to be reviewed and the th text vector of the expert information text vector subset of the expert information, represents the text vector of the comprehensive text information of the paper to be reviewed.

8. The intelligent recommendation method for paper review experts according to claim 7, wherein The screening function satisfies the preset principles, and the preset principles include: the cross-discipline avoidance principle and the same-unit avoidance principle between the paper to be reviewed and the review experts.

9. The intelligent recommendation method for paper review experts according to claim 1, characterized in that The S6 includes: S61: Sort the similarity based on the first similarity sorting model. The first similarity sorting model is as follows: ； In the formula, represents the result sorted with respect to the similarity , is the first sorting function; S62: According to the first sorting result of similarity, select the top review candidate experts whose similarity calculation results are greater than the threshold for re - sorting based on the second similarity sorting model, and use the list of the top candidate experts in the second sorting result as the recommended reviewers for the said paper. The second similarity sorting model is as follows: ； In the formula, represents the initial sorting result regarding similarity The result after double sorting, is the second sorting function, represents taking the first from the first sorting result of similarity review experts in descending order.

10. An intelligent recommendation system for paper review experts, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 9 above.

Images