Research achievement attribution prediction method and system based on natural language processing and social network

Abstract

The application discloses a scientific research achievement attribution prediction method and system based on natural language processing and social networks, first, based on the existing scientific research achievement recognition platform to construct a training sample library, then based on the pre-training model to train a binary classification task model, and to construct a scientific research personnel cooperation network, to construct a scientific research achievement attribute filtering module, and finally to complete the attribution prediction of the scientific research achievement by scoring the unclaimed scientific research achievements. The application can make full use of the development relationship between scientific research achievements and the cooperation information between scientific research personnel, more accurately obtain the attribution prediction result of the scientific research achievements, and greatly improve the accuracy of the achievement attribution prediction while reducing the manual participation.

Description

Technical Field

[0001] This invention belongs to the field of information technology and mainly relates to a method and system for predicting the attribution of scientific research results based on natural language processing and social networks. Background Technology

[0002] Scientific research achievements mainly include academic papers, invention patents, monographs, etc. Attribution refers to determining the authors and their order of authorship. In scientific research achievement display systems, to showcase researchers' academic achievements, researchers need to claim published research results on an achievement claiming platform. However, due to the time-sensitive nature of scientific research achievements, researchers typically do not proactively claim past achievements that are no longer included in performance evaluations. Furthermore, the sheer number and complexity of historical achievements make confirmation by research management personnel impossible. Consequently, the attribution of numerous achievements remains unconfirmed, hindering the visualization of research achievements and subsequent platform analysis. Therefore, in the field of scientific research achievement display, there is an urgent need for a method that can predict the attribution of numerous research achievements, analyze the attribution of the latest achievements, provide intelligent recommendations for researchers to claim their achievements, and predict data from past achievements for use in the display and analysis of results.

[0003] Currently, the academic community is paying increasing attention to the issue of predicting the attribution of research results, as exemplified by the published Chinese invention patent application CN114416959A, "A Method and System for Pushing and Claiming Research Results." Traditional methods for predicting attribution only consider whether various fields of the result match the author's profile. Typically, the first tag of the research result is used to match the author; if no match is found, or multiple authors are matched, the second tag is used, and so on. The shortcoming of this approach is that it only considers whether the institution, author name, journal type, and keywords of the research result (papers, patents, books, etc.) match the institution, name, and keywords of the researcher. It fails to consider the value of the researcher's act of claiming the result, the connections between research results, and the collaborative network of researchers. The problem with this approach is that...

[0004] 1. It is difficult to determine the allocation of research results by utilizing the connection between researchers' existing and potential results. Researchers' research content often evolves over time, and the results of the same researcher are usually not unrelated. Results from similar time periods are often in similar or related research fields.

[0005] 2. Difficulty in utilizing information contained in the social networks of research collaborators. Research output often relies on collaboration among researchers, and these outputs typically have multiple authors. Social networks can be built through these collaborations. In actual collaborative relationships, researchers often collaborate more than once, and these social networks exhibit community-like characteristics.

[0006] With the significant advancements in pre-trained models in natural language processing in recent years, suitable models can be built with minimal training using task corpora. This allows for the discovery of connections between research findings through textual semantic comparison in natural language processing, serving as one of the influencing factors in author allocation. Simultaneously, the collaborative networks among researchers can be used as one of the influencing factors in predicting the attribution of other authors to research findings where some authors have already been identified, better aligning with the characteristics of progressively improving and updating research findings through collaborative efforts. Summary of the Invention

[0007] This invention addresses the shortcomings of existing technologies in predicting the attribution of research achievements, which fail to consider the temporal characteristics of the achievements and the relationships between researchers, resulting in a single matching feature. It provides a method and system for predicting the attribution of research achievements based on natural language processing and social networks. First, a training sample library is constructed based on an existing research achievement claiming platform. Then, a binary classification task model is trained based on a pre-trained model, and a network of researcher collaborations is built. A research achievement attribute filtering module is also constructed. Finally, the attribution of unclaimed research achievements is predicted by assigning a score to each achievement. This invention fully utilizes the developmental relationships between research achievements and information about collaborations among researchers, resulting in more accurate attribution predictions while significantly improving the accuracy of attribution predictions and reducing human intervention.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a method for predicting the attribution of scientific research results based on natural language processing and social networks, comprising the following steps:

[0009] S1, Establishment of training sample library: Based on the existing achievement claiming platform, construct an achievement set for each researcher who has claimed scientific research achievements. In the achievement set, the same type of achievements of the same researcher are sorted by time to form a list L.

[0010] For any result p in the list L, and the previous list L of that result p, ... before The post-list L of the result p after Together they constitute a positive sample (p,L) before ,L after );

[0011] Replace result p with a random result p′ from a different researcher, and compare it with the previous list L of result p. beforeList L of results p after Together they constitute the negative sample (p′, L) before ,L after );

[0012] S2, Training the Researcher Achievement Recognition Model: Defining whether any achievement p is in a researcher's achievement list L as a binary classification problem. A pre-trained model for researcher achievement recognition is constructed based on RoBERTa. The search text in the positive and negative sample sets from step S1 is converted into tokens using a tokenizer. These tokens are then converted into an acceptable format for model training input. The model is then trained to obtain the researcher achievement recognition model. The search text includes at least achievement keywords and titles.

[0013] S3: Constructing a Research Collaboration Network: Collaborators on the same research outcome are treated as the two ends of an edge, and their weights are calculated. After calculating the edge weights between all researchers, a weighted directed research collaboration network is obtained. The method for calculating the edge weights of the collaboration network is as follows: The list of authors of the research outcome p is sorted according to their level of participation as A(p) = [a1, a2, a3, ..., a...]. n ], where n represents the number of authors of output p; where author a i With a j The collaborative value in outcome p is Author a i With a j The sum of the values ​​of all collaborative outcomes can be obtained from the edges of the collaborative network. The weights are:

[0014]

[0015] S4: Construct an output attribute filter: For the output p to be assigned, exclude researchers who do not match based on basic attributes; the basic attributes include at least the researcher's name, affiliation, and publication time;

[0016] S5: Scoring Prediction: Set the minimum effective value S for model inference based on the median probability of attribution in the researchers' achievement identification model. min ;

[0017] For results that have not yet been claimed, the researcher result recognition model trained in step S2 is used to calculate the model reasoning score S1 of the candidates in the set for the result, and the collaboration attribution score S2 of the candidates in the set for the result is set to 0.

[0018] For works that have been partially claimed by collaborators, but not fully claimed by them, the claiming authors are denoted as [a1, a2, a3, ..., a...]. nThe candidate set m′ is obtained through the researcher collaboration network constructed in step S3. Using the researcher achievement recognition model trained in step S2, the inference score S1 of the candidate in the set for the achievement and the original collaboration attribution score of the candidate in the set for the achievement are calculated. Where α is a system parameter, and S1 is higher than the lowest effective value S. min The candidate's S2′ is proportionally normalized to obtain the cooperation affiliation score S2, which takes a value between 0 and 1, while ensuring that S1 does not exceed the minimum effective value S. min The candidate's cooperation affiliation score S2 is 0;

[0019] The total score S is calculated by combining S1 and S2. total When the highest S in the candidate set total >S min At that time, the results will be attributed to S. total The top candidate.

[0020] As a further improvement of the present invention, in step S2, the method for constructing the pre-trained model for identifying scientific research achievements is as follows: the RoBERTa pre-trained model is loaded using the deep learning framework PyTorch, and a Dropout layer and two fully connected layers are added after the RoBERTa model; the parameters are optimized using PyTorch's SGD optimizer, and the cross-entropy loss is used as the loss function to train the model.

[0021] As an improvement to the present invention, the result types in the result set of step S1 at least include papers L. paper Patent L patent Project L project Monograph L book The total number of positive samples in the training set is

[0022]

[0023] Here, the len() function represents the number of results in the list; m represents the number of researchers;

[0024] The ratio of positive to negative samples is 1:10.

[0025] As another improvement of the present invention, the basic achievement attributes in step S4 include at least the author's name, author's affiliation, and publication date, wherein the author's name and author's affiliation include both Chinese and English. As a further improvement of the present invention, in step S5, the total score S... total The calculation method is as follows:

[0026] S total =βS1+(1-β)S2

[0027] Where β is a system parameter.

[0028] To achieve the above objectives, the present invention also adopts the following technical solution: a scientific research achievement attribution prediction system based on natural language processing and social networks, comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) By leveraging the collaborative relationships among researchers in social networks, the accuracy of predicting other unclaimed authors for research results that have already been claimed by some authors can be improved. Information about research collaborations is very helpful in predicting the attribution of results. Researchers often form communities, and collaborations between individuals often occur more than once. Compared to the traditional method of creating individual profiles of researchers to predict the attribution of results, adding collaborative relationships in social networks to predict the attribution of results yields more reliable and interpretable results.

[0031] (2) Improving the accuracy of outcome prediction by leveraging the connection between researchers' existing and potential outcomes. Traditional models firstly consider the literal similarity between research outcome labels and researcher labels, without considering the semantic similarity between the keywords and titles of the outcomes and the semantic similarity between the researchers' research fields and content; secondly, they fail to consider the developmental trends of research outcomes over time, and how changes during the development process can better infer the likelihood of an outcome belonging to a particular researcher. This method constructs a dataset to train the model by extracting outcomes from a list sorted by time, which adds a consideration of the time dimension compared to traditional models.

[0032] (3) This invention can make full use of the information of scientific research results, and can continuously and automatically optimize the model as new claims and new results are added. This invention greatly improves the accuracy of result attribution prediction while reducing human intervention. Attached Figure Description

[0033] Figure 1 This is a flowchart of the steps of the method of the present invention; Detailed Implementation

[0034] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0035] Example 1

[0036] Methods for predicting the attribution of research findings based on natural language processing and social networks, such as Figure 1As shown, the specific steps include the following:

[0037] Step S1: Construct a training sample library based on the existing achievement claiming platform. For each researcher who has already claimed research achievements, construct an achievement set T. The achievement set is divided into multiple subsets T = {L} based on papers, patents, projects, monographs, etc. paper ,L patent ,L book ,L project}, where L is a list of a research member's research achievements sorted chronologically, for example, where L paper =[p1,p2,p3,…,p n Then construct training samples, taking p from list L. i Take the element out, in list L, in p i The elements before L constitute before =[p1,p2,p3,…,p i-1 Similarly, L can be placed in p i The elements after form L after =[p i+1 ,p i+2 ,p i+3 ,…,p n ]. Thus (p i ,L before ,L after A positive sample is constructed for training. Each element in each outcome list L can be used to construct such a record. The total number of positive samples from m research collaborators in the training set is:

[0038]

[0039] The len() function represents the number of elements in the list.

[0040] Then (p) i ,L before ,L after p in ) i A negative sample (p′, L) is formed by replacing the random result p′ of a random person who is not the researcher with the random result p′ of the random person. before ,L after A training set was constructed with positive and negative samples in a ratio of 1:10.

[0041] Step S2: Train a binary classification task model based on the pre-trained model: Define whether any result p is in a researcher's result list L as a binary classification problem. Load the RoBERTa pre-trained model using the deep learning framework PyTorch, and add a Dropout layer and two fully connected layers after the RoBERTa model; use PyTorch's SGD optimizer to optimize the parameters, and use cross-entropy loss as the loss function to train the model.

[0042] In step S1, the text in the positive and negative sample sets is converted into a token using a tokenizer (using jieba for Chinese word segmentation). Then, the tokens are converted into an acceptable permutation format for model training. The tokens are then input into the model for training to obtain the scientific research achievement recognition model.

[0043] The researcher achievement identification model can deduce the probability that any achievement belongs to a certain researcher, with the probability value between 0 and 1.

[0044] Step S3: Construct a collaborative network of researchers. The list of n authors of research output p, ordered by their level of participation, is A(p) = [a1, a2, a3, ..., a...]. n Generally, the higher the level of involvement between two authors, the closer their collaboration is considered to be, and the greater the value of their collaboration. For result p, author a... i With a j The value of cooperation is Author a i With a j In all the results of the collaboration, author a i Higher levels of participation lead to greater accumulated value in collaborations. As a side in the cooperative network The weights of the edges between all authors are calculated to obtain a weighted directed network of researchers.

[0045] Step S4: Construct the achievement attribute filtering module. Before predicting the attribution of unclaimed achievements, researchers whose information does not match the author's name, affiliation, publication date, etc., need to be removed to reduce computational load and improve the efficiency of system prediction. In this step, the researcher information will be extracted from already claimed works. The researcher's name includes Chinese name, English name, and English name abbreviation, and the affiliation name includes the affiliation's Chinese name, English name, Chinese abbreviation, and English abbreviation.

[0046] Step S5: Assignment scoring for research results. The researchers selected through the filtering module will be comprehensively scored, and the researcher with the highest score will be chosen as the assignee of the research result. Author a of result p. i In the case of unclaimed claims, there is a set of filtered candidate owners {m1, m2, ..., m}.s}. Compare the results p and m i L t Construct an input (p,L) t,before ,L t,after The input is given to the model constructed in step S2 to obtain the model inference score S1, and p and m are related. i The collaboration attribution score S2 = 0; for some author a of outcome p j Or a few authors {a1,a2,…,a j If the collaboration is claimed, the original collaboration attribution score will also be obtained through the collaborative social network of researchers constructed in step S3. Here, α is a system parameter, which is adjusted according to the actual situation, and is usually greater than 0.5. S1 is set higher than the set minimum effective value S. min Normalize S2′ of the candidate affiliation to obtain S2, and calculate the total score S. total =βS1 + (1-β)S2. When the highest S in the candidate set... total >S min At that time, result p will be attributed to S. total The top candidate.

[0047] Example 2

[0048] Consider the following dataset and perform outcome attribution prediction. The dataset content is only used to illustrate the outcome attribution prediction process:

[0049] Claimed results:

[0050] 619_2548_001,paper,2011,[D,E]

[0051] 619_2548_002,paper,2011,[E,D,C]

[0052] 619_2548_003,paper,2012,[D,B]

[0053] 619_2548_004,paper,2013,[D,C]

[0054] 619_2548_005,paper,2013,[E,A,D]

[0055] 619_2548_006,paper,2013,[D,F,C]

[0056] 619_2548_007,paper,2012,[A,B,C]

[0057] 619_2548_008,paper,2013,[A,C,B,D]

[0058] 619_2548_009,paper,2013,[B,A,E,C]

[0059] The data is separated into columns by commas and records by line breaks. The first column is the unique ID of the paper, the second column is the type of achievement, the third column is the publication time, and the fourth column is the list of authors.

[0060] Researcher Information:

[0061] Researcher A, 10001, Dong Fang

[0062] Researcher B, 10001, Dong Fang

[0063] Researcher C, 10001, Gu Xin

[0064] Researcher D,10002, Gu Xin

[0065] Researcher E, 10002, Dong Fan

[0066] The data is separated into columns by commas and into records by line breaks. The first column is the personnel code, the second column is the organization to which the personnel belong, and the third column is the personnel name.

[0067] Expected outcome:

[0068] 613_2444_010,paper,2012,["Gu X,Dong F"],10001

[0069] The data is separated into columns by commas and records by line breaks. The first column is the unique ID of the paper, the second column is the type of work, the third column is the publication date, the fourth column is the list of authors, and the fifth column is the affiliation of the authors. Researcher C has already claimed the first authorship of this paper, and the second author, DongF, is a candidate to be predicted.

[0070] Step S1: Establishing the training sample library: Taking researcher D's achievements as an example, extract several achievement IDs in ascending order of time to form a positive sample [619_2548_001, 619_2548_002, 619_2548_003, 619_2548_004, 619_2548_005], with a label of 0. Take another researcher's paper, its record is as follows:

[0071] 619_2548_007,paper,2012,[A,B,C]

[0072] Replace any one of the positive samples with the ID of this paper record to obtain negative samples [619_2548_001,619_2548_007,619_2548_003,619_2548_004,619_2548_005], with a label of 1.

[0073] This process is repeated to construct a positive and negative sample set with a ratio of 1:10.

[0074] Step S2: Training the research results recognition model: Taking the positive samples constructed in step S1 as an example, train the binary classification model:

[0075] [619_2548_001,619_2548_002,619_2548_003,619_2548_004,619_2548_005],0

[0076] First, retrieve the paper's title from the database based on the paper ID, and then combine these titles to obtain the input:

[0077] 0,Evaluation of Effectiveness and Cost-Effectiveness of AsphaltPavement Rehabilitations Utilizing LTPP Data; Validating MEPDG with TennesseePavement Performance Data; Cost-Effectiveness Analyses of MaintenanceTreatments for Low-and Moderate-Traffic Asphalt Pavements in Tennessee; Analyzing Influence Factors of Transverse Cracking on LTPP Resurfaced AsphaltPavements through NB and ZINB Models; Performance comparison of laboratory and field produced previous concrete mixtures;

[0078] The RoBERTa pre-trained model was loaded using the PyTorch deep learning framework. A Dropout layer and two fully connected layers were added after the RoBERTa model. The parameters were optimized using PyTorch's SGD optimizer, and cross-entropy loss was used as the loss function to train the model. The sample set described in step S1 was converted to the above format and input into the model for training, resulting in a research achievement recognition model. The median probability of attribution obtained from the sample set in the recognition pattern, 0.6, was set as the minimum effective value S for model inference. min .

[0079] Step S3: Constructing a Research Collaboration Network: Taking the weight calculation between researcher A and researcher B as an example, first, query the database for collaborative research results already claimed by researchers A and B. The results are as follows:

[0080] 619_2548_007,paper,2012,[A,B,C]

[0081] 619_2548_008,paper,2013,[A,C,B,D]

[0082] 619_2548_009,paper,2013,[B,A,E,C]

[0083] The collaborative value between researchers A and B regarding the paper with ID 619_2548_007 is: Calculate the collaborative value between authors A and B based on the value of the paper. As an edge The weights, As an edge The weights are then used to construct the cooperative network.

[0084] Step S4: Result Attribute Filtering: For the results to be predicted

[0085] 613_2444_010,paper,2012,["Gu X,Dong F"],10001

[0086] Filtering by author name and affiliation, the following research projects might correspond to the second author, Dong F:

[0087] Researcher A, 10001, Dong Fang

[0088] Researcher B, 10001, Dong Fang

[0089] Step S5: Taking the paper and possible authors from Step S4 as an example, this step illustrates how to determine their attribution:

[0090] Samples of researchers A and B’s paper outputs and the paper with ID 613_2444_010 are constructed using the method in step S1.

[0091] Researcher A Sample:

[0092] [619_2548_005,619_2548_007,619_2548_008,619_2548_009,619_2548_010]

[0093] Researcher B sample:

[0094] [619_2548_003,619_2548_007,619_2548_008,619_2548_009,619_2548_010]

[0095] Input the sample into the model trained in step S2 to obtain the score S. 1A =0.887 and S 1B =0.632, all of which are greater than the lowest effective value of 0.6 for model inference.

[0096] Since the paper with ID 613_2444_010 has already been claimed by researcher C, it is still necessary to calculate the collaboration attribution score: calculate the edges in the collaboration graph according to the method described in step S3. Calculate S 2A ′=0.5×3.583+0.5×0=1.792, S 2B = 0.5 × 0.833 + 0.5 × 2.083 = 1.458. Normalization yields S. 2A =1,S 2A =0.814.

[0097] Calculate the total score S total,A =0.944,S total,B =0.723, therefore the second author of the paper with ID 613_2444_16963 is assigned to researcher A. In summary, this invention can fully utilize the developmental relationships between research results and information on collaboration among researchers to more accurately predict the attribution of research results, reducing human intervention while significantly improving the accuracy of result attribution prediction.

[0098] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the attribution of research results based on natural language processing and social networks, characterized in that, Includes the following steps: S1, Establishment of the training sample library: Based on the existing achievement claiming platform, an achievement set is constructed for each researcher who has claimed research achievements. In the achievement set, achievements of the same type by the same researcher are sorted by time to form a list. ; The list any of the results With this achievement The previous list before This achievement The list after after Together they constitute a positive sample ; Random results from different researchers Replacement results , and results The previous list before Results The list after after Together they constitute negative samples ; S2, Training a binary classification task model based on a pre-trained model: arbitrarily selected results. Is it in a researcher's list of achievements? Defined as a binary classification problem, a pre-trained model for identifying scientific research achievements is built based on RoBERTa. The search text in the positive and negative sample sets in step S1 is converted into tokens through a tokenizer. Then, the tokens are converted into an acceptable arrangement format for model training input. The tokens are then input into the model for training to obtain the scientific research achievement identification model. The search text includes achievement keywords and titles. S3: Constructing a Research Collaboration Network: Collaborators on the same research outcome are treated as the two ends of an edge, and their weights are calculated. After calculating the edge weights between all researchers, a weighted directed research collaboration network is obtained. The method for calculating the edge weights of the collaboration network is as follows: Research outcome... The author list, sorted by level of involvement, is as follows: ,in, Indicate results The number of authors; among them, the authors and In results The value of cooperation in China is , the author and The sum of the values ​​of all collaborative outcomes can be obtained from the edges of the collaborative network. The weights are: ； S4: Construct outcome attribute filtering: For outcomes to be assigned Based on basic attributes, researchers who do not match are excluded; these basic attributes include the researcher's name, affiliation, and publication date. S5: Scoring Prediction: Set the minimum effective value for model inference based on the median probability of attribution in the researchers' achievement identification model. ; For unclaimed achievements, based on the filtering method described in step S4, a set of candidate attribution m is constructed. The researcher achievement identification model trained in step S2 is then used to calculate the model inference scores of the candidates in the set for that achievement. And assign a score to each candidate in the set for their collaboration on the achievement. =0; For works that have been partially claimed by collaborators, but not fully claimed by them, the authors who claimed them are denoted as follows: The candidate set is obtained through the research collaboration network constructed in step S3. Using the researcher achievement recognition model trained in step S2, calculate the inference score of the candidate in the set for that achievement model. The score attribution of the original collaboration to the achievement by the candidates in the set. ,in For system parameters, Higher than the lowest effective value Candidates The cooperative affiliation score is obtained by proportional normalization. Let its value be between 0 and 1. Not higher than the minimum effective value Candidate's cooperation attribution score =0; total and Calculate the total score When the highest in the candidate set At that time, the author will be attributed to The top candidate.

2. The method for predicting the attribution of scientific research results based on natural language processing and social networks as described in claim 1, characterized in that: The types of results in the result set of step S1 include papers. paper ,patent patent ,project project ,monograph book The total number of positive samples in the training set is ； in, The function represents the number of results in the list; Indicates the number of researchers; The ratio of positive to negative samples is 1:

10.

3. The method for predicting the attribution of scientific research results based on natural language processing and social networks as described in claim 1, characterized in that: In step S2, the method for constructing the pre-trained model for identifying scientific research achievements is as follows: the RoBERTa pre-trained model is loaded using the deep learning framework PyTorch, and a Dropout layer and two fully connected layers are added after the RoBERTa model; the parameters are optimized using PyTorch's SGD optimizer, and the cross-entropy loss is used as the loss function to train the model.

4. The method for predicting the attribution of scientific research results based on natural language processing and social networks as described in claim 1, characterized in that: The basic attributes of the results in step S4 include the author's name, author's affiliation, and publication date, wherein the author's name and author's affiliation can be in Chinese or English.

5. The method for predicting the attribution of scientific research results based on natural language processing and social networks as described in claim 1, characterized in that: In step S5, the total score The calculation method is as follows: ； in, These are system parameters.

6. A research outcome attribution prediction system based on natural language processing and social networks, comprising a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-5 above.

Images