Paper withdrawal reason automatic classification system

Abstract

The invention relates to the technical field of text data processing, in particular to a thesis withdrawal reason automatic classification system which comprises a task division module, a model construction module, a multi-task joint module and an inference analysis module. According to the method, the content in the draft withdrawal statement is subdivided, and specific information is extracted and analyzed, so that the accuracy and depth of draft withdrawal reason classification can be remarkably improved, and through automatic text analysis and reasoning, the draft withdrawal reasons can be quickly identified and can be deeply understood, so that the accuracy and the accuracy of draft withdrawal reason classification are improved. For example, legal dispute and periodical operation problems in a draft withdrawal statement are analyzed, a data processing flow is optimized, efficient analysis of the draft withdrawal statement is achieved, manual intervention is reduced, the processing speed is increased, the transparency and processing efficiency of the draft withdrawal flow are improved, automatic reasoning analysis further enhances exploration of the relation between draft withdrawal reasons and potential laws and editing problems, and the method is suitable for popularization and application. And more accurate decision support can be provided for the publication, so that the future manuscript withdrawal risk and related disputes can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of text data processing, and in particular to an automatic classification system for paper withdrawal reasons. Background Art

[0002] The core technologies in the field of text data processing technology include natural language processing (NLP), text classification, sentiment analysis, information extraction and text mining. Natural language processing technology uses algorithms to understand and process human language, including basic tasks such as text segmentation, part-of-speech tagging, entity recognition, and dependency analysis. Text data processing also involves machine learning and deep learning technologies, using supervised learning, unsupervised learning, reinforcement learning and other methods to analyze, model and predict large-scale text data, especially in tasks such as automatic classification, topic modeling and sentiment analysis. This technology is widely used in scenarios such as public opinion analysis, social media monitoring, information filtering, automatic question-and-answer systems, and is an indispensable technical support in the information society.

[0003] Among them, the automatic classification system for reasons for paper withdrawal aims to classify and analyze the reasons for paper withdrawal in an automated way. The system uses text data processing technology, including natural language processing and machine learning algorithms, to extract relevant content from retraction notices, statements and other relevant documents, and automatically identify and classify reasons for withdrawal, such as academic misconduct, experimental data errors, peer review issues, etc. The main purpose of this system is to help the academic and publishing communities to quickly and efficiently organize and analyze the reasons for withdrawal, improve the transparency and efficiency of the withdrawal process, and provide decision-making support for researchers, publishers and regulators.

[0004] Although existing technologies cover a wide range of application scenarios, they are unable to handle complex text data, such as implicit information and reasoning analysis in retraction statements. In particular, when identifying legal disputes and journal operation issues, conventional methods cannot effectively identify the underlying reasons for retraction due to the lack of sufficient in-depth analysis and reasoning capabilities. This leads to opacity in the retraction process, inefficient processing, and misunderstanding or misclassification of the reasons for retraction. Failure to correctly identify errors in journal operations will lead to the recurrence of similar problems, affecting the reputation of the journal and academic quality assurance. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings in the prior art and propose an automatic classification system for paper withdrawal reasons.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: The automatic classification system of paper withdrawal reasons includes: The task division module identifies and extracts the reasons for retraction based on the paper retraction statement, annotates the types of academic misconduct, data errors, and journal operations, generates a classification system for retraction reasons, analyzes the information in the retraction statement content, annotates it as reasoning tasks, assigns processing rules to each task, and obtains the task input data set; The model building module, based on the task input data set and according to the characteristics of the differentiated tasks, respectively annotates the retraction statement text, extracts related features, generates a task feature extraction set, uses the task feature extraction set for training, and fuses it with the shared features to obtain a task processing optimization model; The multi-task joint module jointly processes the retraction reason classification and reasoning tasks based on the task processing optimization model, inputs the shared features into the multi-task network, performs merge optimization through each task loss function, and generates the multi-task joint processing result; Based on the results of the multi-task joint processing, the reasoning analysis module performs reasoning analysis on the legal and journal editing issues in the retraction statement, determines whether there are legal disputes and journal operation issues in the retraction statement, generates legal and journal operation reasoning classification results, and obtains the reasoning classification output results of the paper retraction reasons by analyzing the reasoning tasks and the retraction reasons.

[0007] As a further solution of the present invention, the steps for obtaining the retraction reason classification system are specifically as follows: Based on the paper retraction statement, we use text extraction to separate the key information of the retraction, filter out sentences with keywords such as retraction, error, and forgery, exclude information not related to the retraction, and obtain the extracted information set through text segmentation; The extracted information set is classified into three categories: academic misconduct, data error and journal operation error by checking the key features of the text, and an identifier is set separately for each category. The annotation process is corrected to avoid classification errors, and the consistency of the annotation results is repeatedly checked to obtain a classified annotation set; According to the classification annotation set, the data is summarized and analyzed, including analyzing the frequency of occurrence of each type of withdrawal reason, summarizing the withdrawal categories, eliminating abnormal data points in the analysis, verifying the integrity of the data set through comparative analysis, and obtaining a classification system for withdrawal reasons.

[0008] As a further solution of the present invention, the step of obtaining the task input data set is specifically: According to the retraction reason classification system, extract text fields from the retraction statement, perform text segmentation processing, identify each retraction reason in the retraction statement by matching and classifying key information, and screen feature items in combination with previously defined retraction type rules to generate a feature extraction set; Perform frequency analysis and weighted processing on the feature extraction set, identify feature items associated with the task, extract features of withdrawal time, withdrawal processing cycle, and withdrawal reason frequency in the text, sort the feature items by key points, select feature items that meet the task input conditions, and generate classification processing rules using the relationship between frequency and weight; Combined with the classification processing rules, multi-dimensional correlation analysis is performed, and a weighted value is assigned to each feature item according to the correlation between features, using the formula: ; Weight each feature item to generate the task input dataset; in, Represents the task input dataset, Representative The weight of the feature items, Representative The weight coefficient of the feature item, is the influencing factor of the feature, is the adjustment coefficient, Indicates the total number of feature items.

[0009] As a further solution of the present invention, the step of obtaining the task feature extraction set is specifically: By analyzing the task input data set, identifying differentiated task characteristics, marking the task type as a categorical variable, performing text structured analysis, marking the core text elements in each retraction statement, and generating a differentiated task category set; Based on the differentiated task category set, extract the associated feature parameters, including the associated keywords in the task text, the contextual logical relationship and the semantic pattern of the retraction statement, filter the features using the classification judgment criteria, retain the associated features with the task goal, and use the formula: ; Calculate the total score of the associated feature set and sort the results to generate an associated feature set; in, represents the total score of the associated feature set, Indicates The weight of the keywords, Indicates the frequency of keywords. is the impact factor, is the adjustment coefficient, Indicates the total number of keywords; By using the associated feature set, the features are prioritized and cross-analyzed according to the associated features and screening rules, and a task feature extraction set is established in combination with the annotation information of the task category.

[0010] As a further solution of the present invention, the step of obtaining the task processing optimization model is specifically as follows: Based on the task feature extraction set, each feature in the data set is classified, cleaning rules are set for each type of feature according to the data type and range, data standardization is performed, and a task feature initialization set is obtained; Using the task feature initialization set, the task features and shared features are processed in layers, each layer of features is assigned a differentiated processing weight, the features are merged according to the weights, and the feature fusion technology is used to optimize the feature performance to form a fusion feature model; By using the fusion feature model, adjusting feature parameters, performing multiple iterative tests, adjusting parameter settings for each test, and establishing a task processing optimization model.

[0011] As a further solution of the present invention, the step of obtaining the multi-task joint processing result is specifically: Based on the task processing optimization model, shared features are extracted as input feature sets, the shared requirements of the retraction reason classification task and the reasoning task are analyzed, the corresponding parameters are extracted, and the multi-task shared feature set is established by combining the key ranking and frequency analysis results of the feature parameters; The multi-task shared feature set is input into the multi-task network model, the loss value of the task is calculated, and the loss value is standardized according to the key objectives of the classification task and the reasoning task, and the loss value of each task is merged, using the formula: ; Calculate the joint loss value of the task; in, is the joint loss value of the task, Indicates The weight of the task, Indicates The loss value of each task, To adjust the parameters, represents the weight normalization factor, Indicates the total number of tasks; According to the joint loss value of the task, the weight parameters and loss function in the multi-task network are readjusted, the multi-task processing model is optimized, the joint output of the classification task and the reasoning task is reallocated to the task, and the multi-task joint processing result is generated.

[0012] As a further solution of the present invention, the steps for obtaining the classification results of the legal and journal operation reasoning are specifically as follows: Extracting core features associated with the retraction statement from the multi-task joint processing results, identifying key parameters of legal disputes and journal operation issues, annotating the key content of each retraction statement through text analysis technology, performing quantitative analysis on the features, and compiling the annotated results into a problem feature set; Based on the problem feature set, reasoning analysis is performed. For each legal dispute and journal operation problem, a logical reasoning algorithm is used to determine the classification basis of the problem, weight the features and score them, and optimize the accuracy of reasoning. The formula is: ; Get the classification weight results; in, represents the classification weight, is the characteristic frequency, is the feature effectiveness score, is the feature redundancy coefficient, Tuning parameters for feature sensitivity; Using the classification weight results, the characteristics under each classification standard are analyzed, the classification results of each legal dispute and journal operation issue are evaluated, the weight and reasoning basis of each classification result are data verified and logically reviewed, and the legal and journal operation reasoning classification results are generated.

[0013] As a further solution of the present invention, the steps for obtaining the output results of the reasoning classification of the paper retraction reasons are specifically as follows: According to the classification results of the above-mentioned legal and journal operation reasoning, collect and organize related legal documents and journal editorial guidelines, classify the contents of legal documents and journal editorial guidelines, identify and record the key elements of legal terms and journal operations, annotate and code each element, and construct a set of legal and journal operation specifications; Based on the legal and journal operational norms, analyze the connection between the differentiated reasoning tasks and the retracted papers, record and classify each connection, adjust the data connection points and optimize the identification process of the retraction reasons, and form a retraction reason analysis model; The retraction reason analysis model is used to classify differentiated retraction cases, adjust the classification parameters in the model and match the cases, repeatedly verify the accuracy and consistency of the model, and obtain the reasoning classification output results of the paper retraction reasons.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, by subdividing the content of the retraction statement and extracting and analyzing specific information, the accuracy and depth of the classification of the reasons for withdrawal can be significantly improved. Through automated text analysis and reasoning, the reasons for withdrawal are not only quickly identified, but also can be more deeply understood, such as analyzing the legal disputes and journal operation problems in the retraction statement, optimizing the data processing process, achieving efficient analysis of the retraction statement, reducing manual intervention, speeding up the processing speed, and improving the transparency and processing efficiency of the retraction process. Automated reasoning analysis also strengthens the exploration of the connection between the reasons for withdrawal and potential legal and editorial issues, which helps to provide more accurate decision-making support for the publishing industry, thereby reducing future retraction risks and related disputes. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a system flow chart of the present invention; Figure 2 This is a flow chart of the classification system for withdrawal reasons in this invention; Figure 3 Flowchart for inputting data set for the task in the present invention; Figure 4 It is a flowchart of the task feature extraction set in the present invention; Figure 5 A flowchart of the task processing optimization model in the present invention; Figure 6 A flowchart of the multi-task joint processing result in the present invention; Figure 7 Flowchart of reasoning classification results for the legal and journal operations in the present invention; Figure 8 This is a flow chart of the reasoning and classification output results of the paper withdrawal reasons in the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0017] In the description of the present invention, it should be understood that the terms length, width, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc. indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific position, be constructed and operated in a specific position, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, the meaning of multiple is two or more, unless otherwise clearly and specifically defined.

[0018] See also Figure 1, the automatic classification system for paper withdrawal reasons includes: The task division module identifies and extracts the reasons for retraction based on the paper retraction statement, annotates the types of academic misconduct, data errors, and journal operations, generates a classification system for retraction reasons, analyzes the information in the retraction statement content, annotates it as reasoning tasks, assigns processing rules to each task, and obtains the task input data set; The model building module is based on the task input data set. According to the characteristics of the differentiated tasks, it annotates the retraction statement text, extracts related features, generates a task feature extraction set, uses the task feature extraction set for training, and fuses it with the shared features to obtain a task processing optimization model. The multi-task joint module jointly processes the retraction reason classification and reasoning tasks based on the task processing optimization model, inputs the shared features into the multi-task network, and combines and optimizes each task loss function to generate the multi-task joint processing results; The reasoning analysis module performs reasoning analysis on the legal and journal editing issues in the retraction statement based on the results of multi-task joint processing, determines whether there are legal disputes and journal operation issues in the retraction statement, generates legal and journal operation reasoning classification results, and obtains the reasoning classification output results of the paper retraction reasons by analyzing the reasoning tasks and the reasons for retraction.

[0019] The classification system for reasons for withdrawal includes academic misconduct categories, data error categories, and journal operation categories. The task input data set includes retraction statement text samples, text processing results, and initial annotation data. The task feature extraction set includes text annotation features and correlation features. The task processing optimization model includes network training strategies and feature optimization algorithms. The results of multi-task joint processing include comprehensive classification capabilities and task loss optimization. The legal and journal operation reasoning classification results include legal dispute judgment results and editorial operation judgment results. The output results of the paper retraction reason reasoning classification include legal issue refinement and journal operation issue refinement.

[0020] See also Figure 2 , the specific steps for obtaining the retraction reason classification system are as follows: Based on the paper retraction statement, we use text extraction to separate the key information of the retraction, filter out sentences with keywords such as retraction, error, and forgery, exclude information not related to the retraction, and obtain the extracted information set through text segmentation; Extract text content from the paper retraction statement, filter out sentences containing the keywords retraction, error, and forgery. The keywords appear in the sentences involving the retraction statement, indicating the main reason for the retraction. To ensure the comprehensiveness of key information, a text matching rule can be designed according to the different formats of the retraction statement, such as looking for words with similar meanings such as retraction or withdrawal in each paragraph, extract each sentence containing keywords from the statement, and exclude content unrelated to the retraction, such as the preface of the statement or descriptions unrelated to the reason for the retraction. By reasonably segmenting the text content (for example, by sentence or paragraph), further filter out text blocks with relevant information such as retraction, error, and forgery, so as to form a preliminary information set. The steps need to ensure the integrity of the text and distinguish different types of retraction reasons according to the context. Assuming that the information set extracted from a retraction statement contains the following key information: retraction of the paper, data forgery, retraction of this paper due to experimental data errors, research method errors, and retraction, it will constitute a preliminary extracted information set, on which subsequent classification and analysis will be carried out.

[0021] The extracted information set is classified into three categories: academic misconduct, data errors, and journal operation errors by checking the key features of the text. An identifier is set for each category separately. The annotation process is corrected to avoid classification errors. The consistency of the annotation results is repeatedly checked to obtain a classified annotation set. When classifying the extracted information set, we first need to design a set of rules to distinguish different reasons for retraction based on the keywords and contextual information in the text. If the information set mentions keywords such as forgery, plagiarism, and data tampering, this type of text should be marked as academic misconduct. If the text contains statements such as data errors, inaccurate experimental results, and repeated experimental data, it should be classified as data errors. Sentences involving journal operation errors, such as editing errors and journal typesetting errors, should be classified as journal operation errors. To ensure the accuracy of classification, each category will be set with a unique identifier, such as the academic misconduct identifier is 001, the data error identifier is 002, and the journal operation error identifier is 003. During the annotation process, special attention should be paid to avoid classification errors, such as mislabeling editing errors as data errors or mislabeling data duplications as academic misconduct. In order to correct errors and ensure classification accuracy, the consistency of the annotation results can be repeatedly checked to ensure that the same type of text can be consistently classified among multiple annotators. During the annotation process, if there is a dispute over the classification of a certain piece of information, corrections can be made through collective discussion or reference to more contextual information to ensure the accuracy and consistency of the final classification.

[0022] According to the classification annotation set, the data is summarized and analyzed, including analyzing the frequency of occurrence of each type of retraction reason, summarizing the retraction categories, eliminating abnormal data points in the analysis, verifying the integrity of the data set through comparative analysis, and obtaining a classification system for retraction reasons; After completing the classification and annotation of the reasons for retraction, all classified data are summarized and statistically analyzed. It is necessary to count the frequency of each reason for retraction (i.e., academic misconduct, data error, journal operation error) in the entire data set, and traverse each entry in the classification annotation set to calculate the number of each category, and then obtain the frequency of different reasons for retraction. For example, if 60 of the 100 retraction statements are due to data errors, 30 are due to academic misconduct, and 10 are due to journal operation errors, then the proportion of each category can be obtained. It is very important to eliminate abnormal data points. Thresholds can be set to identify and exclude retraction statements that are obviously deviated from the norm. For example, some retraction statements cannot be accurately classified due to unclear expression or overly vague information. A threshold can be set to eliminate data points to avoid affecting the analysis results. Through comparative analysis, it is possible to check whether there is an abnormal frequency of a certain reason for retraction, indicating a problem in the data collection or classification process. After completing data cleaning and eliminating anomalies, a classification system for reasons for retraction can be obtained, providing basic data support for further trend analysis and research. Through this analysis, we can not only understand the distribution of reasons for retraction, but also provide improvement suggestions for journals and the academic community, such as which fields have more data errors and which academic misconduct urgently needs to be strengthened. Supervision, etc.

[0023] See also Figure 3 , the specific steps for obtaining the task input data set are: According to the retraction reason classification system, text fields are extracted from the retraction statement, and text segmentation is performed. Each retraction reason in the retraction statement is identified by matching and classifying key information. Combined with the previously defined retraction type rules, feature items are screened to generate a feature extraction set. According to the retraction statement text, the retraction reason classification system data is extracted, and the text is cleaned to remove irrelevant information and noise data to ensure the accuracy and completeness of the data. Based on text analysis, the retraction reason classification system involved is identified, such as scientific misconduct, duplicate publication, ethical issues, etc. The keywords in the text are matched and classified by regular expressions to generate preliminary retraction category data. Then, the categories are screened and sorted according to the established retraction reason type rules, and the reason category of each retraction statement is marked and classified. The key fields of each retraction statement, such as retraction reason, retraction date, retraction statement, etc., are extracted, and the information is stored in a predefined data structure to further provide structured input data for subsequent analysis. Finally, a feature extraction set is generated, which contains multiple retraction reason types and their corresponding feature items as the basic data source of the task input data set.

[0024] Perform frequency analysis and weighted processing on the feature extraction set to identify the feature items associated with the task, extract the features of withdrawal time, withdrawal processing cycle, and withdrawal reason frequency in the text, sort the feature items by key points, select the feature items that meet the task input conditions, and use the relationship between frequency and weight to generate classification processing rules; According to the feature extraction set, perform text feature frequency analysis, analyze the frequency of keyword occurrence in each category, and calculate the occurrence weight of each feature item. According to the frequency of occurrence of the feature item in the retraction statement, calculate the relative importance of each feature, generate a weighted value, and apply it to the subsequent feature screening process. According to the set feature importance sorting rules, based on the correlation between the features and task requirements, further screen the most representative and task-related feature items. It is also necessary to extract features such as withdrawal time, withdrawal processing cycle, and withdrawal frequency, and use frequency weighted processing to screen the feature values ​​according to the set rules to generate classification processing rules. The rules contain key features related to the retraction statement to ensure that the input data of subsequent tasks are representative and accurate.

[0025] Combined with the classification processing rules, multi-dimensional correlation analysis is performed, and weighted values ​​are assigned to each feature item according to the correlation between features. The formula is: ; Weight each feature item to generate the task input dataset; in, Represents the task input dataset, Representative The weight of the feature items, Representative The weight coefficient of the feature item, is the influencing factor of the feature, is the adjustment coefficient, Indicates the total number of feature items; The benefit of the formula is that the introduction of weight coefficients and adjustment factors can effectively improve the importance evaluation of feature items in the task input set, taking into account the correlation between feature items and their actual contribution to the task results. This weighted processing method can ensure multi-dimensional analysis of the task input data set; In this example, assume that there are 3 weight coefficients of feature items ,in , , , weights represent the importance of feature items, and the weight coefficients of the corresponding feature items are , , , the influencing factor of the feature and adjustment coefficient ; Substitute into the formula to calculate: ; ; ; ; The result shows that the weighted value of the task input data set is 1.1232, which means that under the weighted analysis of the given retraction statement data and feature items, the task input data set forms a more comprehensive output value under the influence of relevant feature items, which will be used for the feature input analysis of subsequent tasks.

[0026] See also Figure 4 , the specific steps for obtaining the task feature extraction set are: By analyzing the task input data set, identifying differentiated task characteristics, annotating task types as categorical variables, performing text structured analysis, marking the core text elements in each retraction statement, and generating a differentiated task category set; Based on the task input data set, the differentiated task characteristics are analyzed, the parameters in the task input data set are disassembled, and the specific feature values ​​related to the task type are extracted respectively. By summarizing the characteristics of each task type, a preliminary classification variable set is formed, and a weight analysis is performed on each feature in the task input data set. The correlation between the feature and the target task is determined by the weight allocation. After removing the features with low correlation, the remaining features are sorted and screened according to specific standards. The input features of each task are compared with the classification variables, and the feature distribution differences under different task categories are calculated. At the same time, combined with the statistical results of the differences, the task characteristics are annotated one by one. During the annotation process, the correlation of the input features is dynamically evaluated to adjust the matching parameters of the classification variables, and finally a differentiated task category set is formed. On this basis, combined with the core requirements of the task, the annotated differentiated task category set is preliminarily structured to provide input data support for the subsequent extraction of task-related features and the construction of task feature extraction sets.

[0027] Based on the differentiated task category set, the associated feature parameters are extracted, including the associated keywords in the task text, the contextual logical relationship and the semantic pattern of the retraction statement. The features are filtered using the classification judgment criteria, and the associated features with the task goal are retained. The formula is used: ; Calculate the total score of the associated feature set and sort the results to generate an associated feature set; in, represents the total score of the associated feature set, Indicates The weight of the keywords, Indicates the frequency of keywords. is the impact factor, is the adjustment coefficient, Indicates the total number of keywords; The benefit of the formula is that it uses the weight value of the keyword 、Influence Factor and adjustment coefficient , the frequency and contribution of each feature item can be dynamically weighed in multiple dimensions, avoiding the deviation caused by relying solely on feature frequency and optimizing the importance assessment of task features; Assume that the number of associated features of a task is 3, which are The weight values ​​are , the corresponding characteristic frequency is , impact factor , adjustment coefficient , substitute the value into the formula to calculate: ; ; ; ; The result shows that the total score of the associated feature set is 2.588. The value reflects the comprehensive weight and contribution of each feature, which provides an important basis for the screening of the task feature extraction set and guides the final construction of the subsequent feature set.

[0028] Using the associated feature set, according to the associated features and screening rules, the features are prioritized and cross-analyzed, and the task feature extraction set is established in combination with the annotation information of the task category; Combined with the associated features in the differentiated task category set, the importance relationship between features is deeply analyzed, and each feature in the feature set is prioritized according to the feature importance score. Features with higher importance are screened, and features are cross-compared according to feature priority. The degree of correlation between each feature and the other features is extracted, and a feature correlation matrix is ​​constructed based on the degree of correlation. In the correlation matrix, the similarities and differences between features are measured, and then redundant feature values ​​are eliminated. Based on the screened feature set, the applicability of the features in the task category is calculated item by item. Combined with the results of the applicability and correlation matrix, the distribution between features is dynamically adjusted. The feature set with the highest applicability and the strongest correlation is selected through cross-analysis. Finally, the screened high-priority associated feature set is integrated into the task feature extraction set to form a corresponding feature labeling model with the differentiated task category set.

[0029] See also Figure 5,The specific steps for obtaining the task processing optimization model are: Based on the task feature extraction set, each feature in the data set is classified, and cleaning rules are set for each type of feature according to the data type and range, and data standardization is performed to obtain the task feature initialization set; Perform range checking and missing value filling on numerical data. For example, for numerical features, set the upper and lower limits to a specific range (such as interval values ​​set based on business rules), filter out abnormal values ​​that exceed the range and replace them with the mean or median. For categorical features, process abnormal categories through frequency analysis, merge low-frequency categories into other categories, and delete blank or invalid categories. For text features, perform length screening to remove illogical feature content. For example, set the lower limit of the length of text features to 10 bytes to ensure its rationality. After completing the cleaning rules, perform standardization on numerical features, and use linear scaling methods to uniformly map data to an interval (such as between 0 and 1) to eliminate the impact of feature dimension differences. Categorical features are converted through one-hot encoding so that they can be quantified in subsequent modeling to obtain a unified structure of task feature initialization sets.

[0030] Using the task feature initialization set, the task features and shared features are processed in layers, each layer of features is given a differentiated processing weight, the features are merged according to the weights, and the feature fusion technology is used to optimize the feature performance to form a fusion feature model; First, the importance of each type of feature is clarified. For example, core features and auxiliary features are distinguished in task features, and core features are given higher weights. The importance of different features is quantified by calculating the correlation coefficient of the features to the target variables. The features are divided into three layers: high, medium, and low. Different processing weights are given to each layer of features. In the process of calculating the weights, comprehensive values ​​are assigned based on the distribution, importance ranking, and interaction information of the features. The features of each layer are weighted and merged according to the weights to construct preliminary fusion features. In the process of feature merging, multi-dimensional features are de-redundanted, such as removing redundant dimensions and reducing the impact of highly correlated features, thereby enhancing the representation ability of fusion features. The performance is optimized through the feature fusion mechanism, and the accuracy, relevance, and prediction performance of features are improved in multiple dimensions. Finally, a fusion feature model is formed as input for subsequent modeling.

[0031] By fusing feature models, adjusting feature parameters, conducting multiple iterative tests, and adjusting parameter settings for each test, a task processing optimization model is established; First, specific settings are made for each iteration, such as dynamically adjusting feature weights to enhance the model's sensitivity to high-importance features. In each test, the parameter settings are optimized step by step, such as gradually adjusting the weight values ​​of feature combinations to remove features that contribute less to the output results, thereby improving model performance. After each test, output indicators such as model accuracy, error, and recall are recorded, and the results are compared with the target performance. If the model performance does not meet expectations, the feature parameters are further corrected, such as modifying the layering rules of certain features, re-dividing the weights, or adjusting the fusion method until the model is stable. Through multiple rounds of iterations, a task processing optimization model is finally established to ensure efficient and reliable performance in complex task scenarios.

[0032] See also Figure 6 , the specific steps for obtaining the results of multi-task joint processing are: Based on the task processing optimization model, shared features are extracted as the input feature set, the shared requirements of the retraction reason classification task and the reasoning task are analyzed, the corresponding parameters are extracted, and the multi-task shared feature set is established by combining the key ranking and frequency analysis results of the feature parameters; By analyzing the sharing requirements of the retraction reason classification task and the reasoning task, we first counted the distribution of each feature parameter in the task input data set, and calculated the matching degree of these features for the classification task and the reasoning task according to the task objectives, extracted the feature parameters that showed significant correlation in both tasks, eliminated the features with correlation below a certain threshold, and retained the shared feature set. Then, we conducted a multi-dimensional analysis on the screened shared feature set, focusing on calculating the synergy between different features. We further refined the parameters in the shared feature set through frequency analysis and importance ranking methods. In the feature set, we evaluated the multi-task adaptability of the shared features in combination with the classification variables of the tasks, further eliminated redundant feature values ​​using parameter distribution analysis, and dynamically adjusted the weight distribution in the feature set using classification variables. We input the adjusted shared feature set into the multi-task network as input parameters to generate a multi-task shared feature set, providing standardized input data for subsequent multi-task joint optimization.

[0033] Input the multi-task shared feature set into the multi-task network model, calculate the loss value of the task, standardize the loss value according to the key goals of the classification task and reasoning task, merge the loss value of each task, and use the formula: ; Calculate the joint loss value of the task; in, is the joint loss value of the task, Indicates The weight of the task, Indicates The loss value of each task, To adjust the parameters, represents the weight normalization factor, Indicates the total number of tasks; The benefit of the formula is that by dynamically introducing the weight parameter Normalization factor and adjustment coefficient , which realizes flexible adjustment and balanced optimization of multi-task loss values, and effectively solves the impact of different task complexity and priority on the overall loss value; There are three tasks, and their joint loss values ​​are The weights of the tasks are , adjustment parameters , weight normalization factor ; Substituting the parameters into the formula: ; ; ; ; The result shows that the multi-task joint loss value is 3.198, which reflects the comprehensive balance of the weights of each task and its loss value, provides an important guiding parameter for the subsequent optimization adjustment of the multi-task network, and lays the foundation for the generation of multi-task results.

[0034] According to the joint loss value of the task, the weight parameters and loss function in the multi-task network are readjusted to optimize the multi-task processing model, and the joint output of the classification task and the reasoning task is reallocated to the task to generate the multi-task joint processing result; The weight of the joint output of the classification task and the reasoning task is adjusted, and the task priority is reallocated. The task output features are standardized by normalizing the feature parameters in the output of each task, and the importance of the task output features is calculated in turn to generate the importance ranking of the output features. Based on the loss optimization results of the multi-task network, the cross-adaptability between the features of each task is analyzed, and the synergy between the output features is extracted through correlation matrix analysis. The redundant features are eliminated and the optimal output features are retained. After feature elimination and reconstruction, the output features are matched for multi-task adaptability, and the comprehensive contribution of each feature in the joint output is calculated in turn. The output of each task is reallocated and dynamically adjusted to obtain the multi-task joint processing results, and the results are standardized into a unified task output matrix to provide reliable basic output data support for subsequent result application scenarios.

[0035] See also Figure 7,The steps to obtain the classification results of legal and journal operations are as follows: Extract the core features associated with the retraction statement from the multi-task joint processing results, identify the key parameters of legal disputes and journal operation issues, annotate the key content of each retraction statement through text analysis technology, perform quantitative analysis on the features, and compile the annotated results into a problem feature set; Identify text fragments involving legal disputes and journal operation issues in the retraction statement, mark the key content in the retraction statement based on text parsing technology, screen out controversial or problematic sentences or keywords by extracting key paragraphs and analyzing the contextual logical relationship, quantify each marked feature item, calculate its feature strength under the relevant task category, and construct a feature strength matrix to screen out redundant information or repeated features. Use the optimized and screened feature matrix as the problem feature set, and perform preliminary relevance sorting. At the same time, establish the corresponding relationship between features according to different task classification requirements to provide initial feature input for the next step of analysis and classification.

[0036] Based on the problem feature set, we perform reasoning analysis. For each legal dispute and journal operation problem, we use the logical reasoning algorithm to determine the classification basis of the problem, assign weights and score the features, and optimize the accuracy of reasoning. The formula is: ; Get the classification weight results; in, represents the classification weight, is the characteristic frequency, is the feature effectiveness score, is the feature redundancy coefficient, Tuning parameters for feature sensitivity; The benefit of the formula is that by introducing the characteristic frequency , feature effectiveness score , feature redundancy coefficient , feature sensitivity adjustment parameters Key indicators such as , can comprehensively measure the influence of each feature in classification and enhance the rationality and flexibility of classification results; : The final classification weight of the reasoning analysis is used to indicate the classification importance and urgency of the problem; : The frequency of occurrence of relevant features in the text directly affects the basic data of classification judgment; : Feature effectiveness score, obtained through expert evaluation or automated tool analysis; : Feature redundancy coefficient, used to adjust the influence of features that appear repeatedly in multiple categories; : Feature sensitivity adjustment parameters are adjusted according to the current analysis environment and the sensitivity of the specific case; The following characteristics of legal controversy and journal issues are considered in a particular retraction statement: Feature 1 Frequency , feature validity score , feature redundancy coefficient , sensitivity adjustment parameter ; Feature 2 Frequency , feature validity score , feature redundancy coefficient , sensitivity adjustment parameter ; For feature 1: ; ; For feature 2: ; ; This result shows that the classification weight of feature 1 is , the classification weight of feature 2 is ,The weight values ​​can be further used for classification decisions, to determine the priority and importance of legal disputes and journal operation issues, and support the generation of the final reasoning classification results.

[0037] Using the classification weight results, analyze the characteristics under each classification standard, evaluate the classification results of each legal dispute and journal operation issue, conduct data verification and logical review on the weight and reasoning basis of each classification result, and generate the classification results of legal and journal operation reasoning; Combine the feature weights with the logic of the classification task, conduct a comprehensive analysis of each type of problem, evaluate the adaptability of the legal dispute and journal problem features in different categories in turn, screen out the most representative features, and construct a feature adaptability matrix for the screened features. The credibility of each classification feature is evaluated through the correlation between the eigenvalues ​​in the matrix. At the same time, each eigenvalue is classified and labeled according to the classification standard, and a preliminary matching relationship between the classification basis and the reasoning model is generated. The classification result with the highest final weight is reallocated to the legal dispute problem and journal operation problem category. After verification and confirmation, the legal and journal operation reasoning classification results are obtained to form the final output result of the task.

[0038] See also Figure 8 ,The specific steps for obtaining the output results of the reasoning classification of paper retraction are: According to the classification results of legal and journal operation reasoning, collect and organize related legal documents and journal editorial guidelines, classify the contents of legal documents and journal editorial guidelines, identify and record the key elements of legal terms and journal operations, annotate and code each element, and construct a set of legal and journal operation specifications; Screen relevant laws and regulations from the legal database, classify documents according to field, promulgation time and legal effect, for example, organize data privacy protection-related laws and academic ethics regulations separately, and in the compilation of journal editorial guidelines, focus on submission requirements, review process and editorial decision-making rules, classify them according to journal impact factors or subject areas, parse the content of each document paragraph by paragraph, identify the legal terms and key operational elements, such as extracting terms such as data minimization and informed consent in the field of privacy protection, and extracting key operational points such as peer review and academic integrity in journal editorial guidelines, annotate each element in detail, including related legal terms, corresponding editorial rules and specific operational suggestions, store them using unique codes to ensure their traceability and accuracy in subsequent analysis, and build a set of legal and journal operational specifications to lay the foundation for subsequent analysis and model establishment.

[0039] Based on the legal and journal operational norms, we analyze the connection between the differentiated reasoning tasks and the retracted papers, record and classify each connection, adjust the data connection points and optimize the identification process of the retraction reasons to form a retraction reason analysis model; First, we extract detailed descriptions of the reasons for retraction from the retracted paper database, such as data falsification, ethical issues, or peer review issues, and compare them one by one with the regulatory points in laws, regulations, and editorial guidelines. For each connection, such as the correspondence between the failure to obtain subject consent mentioned in the law and the data collection violations described in the retracted paper, we record it in detail and classify it as an ethical issue connection. According to the frequency and severity of the connection, we optimize the data connection points, such as adding clear indexes for high-frequency issues to facilitate subsequent queries and reasoning. When optimizing the process of identifying the reasons for retraction, we use the items in the laws and operational specifications as the rule basis, check the descriptions of the retracted papers one by one, and ensure the accuracy and clarity of the connection, so as to form a retraction reason analysis model to improve the model's ability to interpret the reasons for retraction.

[0040] Use the retraction reason analysis model to classify differentiated retraction cases, adjust the classification parameters in the model and match the cases, repeatedly verify the accuracy and consistency of the model, and obtain the reasoning classification output results of the paper retraction reasons; First, extract the retraction instructions involved in the case, parse the content sentence by sentence, extract the core elements, such as non-compliance with legal terms, unreliable experimental data, etc., match them to the classification rules in the analysis model, and adjust the classification parameters in the model according to the classification results. For example, adjust the model's distinction weights for data falsification and ethical issues according to the actual situation of the case to ensure that the model can adapt to a variety of retraction scenarios more carefully. In each verification, test the classification results of the model by randomly sampling retraction cases, record the classification consistency of the model output, and manually check the accuracy of the proofreading results. In case of deviations, adjust the model classification rules in the opposite direction, repeat the verification until the expected effect is achieved, and finally form the classification output results of the reasoning for paper retraction, providing accurate decision-making support for academic journal management.

[0041] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. An automatic classification system for paper withdrawal reasons, characterized by: The system comprises: The task division module identifies and extracts the reasons for retraction based on the paper retraction statement, annotates the types of academic misconduct, data errors, and journal operations, generates a classification system for retraction reasons, analyzes the information in the retraction statement content, annotates it as reasoning tasks, assigns processing rules to each task, and obtains the task input data set; The model building module, based on the task input data set and according to the characteristics of the differentiated tasks, respectively annotates the retraction statement text, extracts related features, generates a task feature extraction set, uses the task feature extraction set for training, and fuses it with the shared features to obtain a task processing optimization model; The multi-task joint module jointly processes the retraction reason classification and reasoning tasks based on the task processing optimization model, inputs the shared features into the multi-task network, performs merge optimization through each task loss function, and generates the multi-task joint processing result; Based on the results of the multi-task joint processing, the reasoning analysis module performs reasoning analysis on the legal and journal editing issues in the retraction statement, determines whether there are legal disputes and journal operation issues in the retraction statement, generates legal and journal operation reasoning classification results, and obtains the reasoning classification output results of the paper retraction reasons by analyzing the reasoning tasks and the retraction reasons.

2. The automatic classification system for paper withdrawal reasons according to claim 1 is characterized in that: The specific steps for obtaining the retraction reason classification system are as follows: Based on the paper retraction statement, we use text extraction to separate the key information of the retraction, filter out sentences with keywords such as retraction, error, and forgery, exclude information not related to the retraction, and obtain the extracted information set through text segmentation; The extracted information set is classified into three categories: academic misconduct, data error and journal operation error by checking the key features of the text, and an identifier is set separately for each category. The annotation process is corrected to avoid classification errors, and the consistency of the annotation results is repeatedly checked to obtain a classified annotation set; According to the classification annotation set, the data is summarized and analyzed, including analyzing the frequency of occurrence of each type of withdrawal reason, summarizing the withdrawal categories, eliminating abnormal data points in the analysis, verifying the integrity of the data set through comparative analysis, and obtaining a classification system for withdrawal reasons.

3. The automatic classification system for paper withdrawal reasons according to claim 2 is characterized in that: The steps for obtaining the task input data set are specifically as follows: According to the retraction reason classification system, extract text fields from the retraction statement, perform text segmentation processing, identify each retraction reason in the retraction statement by matching and classifying key information, and screen feature items in combination with previously defined retraction type rules to generate a feature extraction set; Perform frequency analysis and weighted processing on the feature extraction set, identify feature items associated with the task, extract features of withdrawal time, withdrawal processing cycle, and withdrawal reason frequency in the text, sort the feature items by key points, select feature items that meet the task input conditions, and generate classification processing rules using the relationship between frequency and weight; Combined with the classification processing rules, multi-dimensional correlation analysis is performed, and a weighted value is assigned to each feature item according to the correlation between features, using the formula: ； Weight each feature item to generate the task input data set; in, Represents the task input dataset, Representative The weight of the feature items, Representative The weight coefficient of the feature item, is the influencing factor of the feature, is the adjustment coefficient, Indicates the total number of feature items.

4. The automatic classification system for paper withdrawal reasons according to claim 3 is characterized in that: The steps for obtaining the task feature extraction set are specifically as follows: By analyzing the task input data set, identifying differentiated task characteristics, marking the task type as a categorical variable, performing text structured analysis, marking the core text elements in each retraction statement, and generating a differentiated task category set; Based on the differentiated task category set, extract the associated feature parameters, including the associated keywords in the task text, the contextual logical relationship and the semantic pattern of the retraction statement, filter the features using the classification judgment criteria, retain the associated features with the task goal, and use the formula: ； Calculate the total score of the associated feature set and sort the results to generate an associated feature set; in, represents the total score of the associated feature set, Indicates The weight of the keywords, Indicates the frequency of keywords. is the impact factor, is the adjustment coefficient, Indicates the total number of keywords; By using the associated feature set, the features are prioritized and cross-analyzed according to the associated features and screening rules, and a task feature extraction set is established in combination with the annotation information of the task category.

5. The automatic classification system for paper withdrawal reasons according to claim 4 is characterized in that: The steps for obtaining the task processing optimization model are specifically as follows: Based on the task feature extraction set, each feature in the data set is classified, cleaning rules are set for each type of feature according to the data type and range, data standardization is performed, and a task feature initialization set is obtained; Using the task feature initialization set, the task features and shared features are processed in layers, each layer of features is assigned a differentiated processing weight, the features are merged according to the weights, and the feature fusion technology is used to optimize the feature performance to form a fusion feature model; By using the fusion feature model, adjusting feature parameters, performing multiple iterative tests, adjusting parameter settings for each test, and establishing a task processing optimization model.

6. The automatic classification system for paper withdrawal reasons according to claim 5 is characterized in that: The steps for obtaining the multi-task joint processing result are specifically as follows: Based on the task processing optimization model, shared features are extracted as input feature sets, the shared requirements of the retraction reason classification task and the reasoning task are analyzed, the corresponding parameters are extracted, and the multi-task shared feature set is established by combining the key ranking and frequency analysis results of the feature parameters; The multi-task shared feature set is input into the multi-task network model, the loss value of the task is calculated, and the loss value is standardized according to the key objectives of the classification task and the reasoning task, and the loss value of each task is merged, using the formula: ； Calculate the joint loss value of the task; in, is the joint loss value of the task, Indicates The weight of the task, Indicates The loss value of each task, To adjust the parameters, represents the weight normalization factor, Indicates the total number of tasks; According to the joint loss value of the task, the weight parameters and loss function in the multi-task network are readjusted, the multi-task processing model is optimized, the joint output of the classification task and the reasoning task is reallocated to the task, and the multi-task joint processing result is generated.

7. The automatic classification system for paper withdrawal reasons according to claim 6 is characterized in that: The steps for obtaining the classification results of the legal and journal operation reasoning are as follows: Extracting core features associated with the retraction statement from the multi-task joint processing results, identifying key parameters of legal disputes and journal operation issues, annotating the key content of each retraction statement through text analysis technology, performing quantitative analysis on the features, and compiling the annotated results into a problem feature set; Based on the problem feature set, reasoning analysis is performed. For each legal dispute and journal operation problem, a logical reasoning algorithm is used to determine the classification basis of the problem, weight the features and score them, and optimize the accuracy of reasoning. The formula is: ； Get the classification weight results; in, represents the classification weight, is the characteristic frequency, is the feature effectiveness score, is the feature redundancy coefficient, Tuning parameters for feature sensitivity; Using the classification weight results, the characteristics under each classification standard are analyzed, the classification results of each legal dispute and journal operation issue are evaluated, the weight and reasoning basis of each classification result are data verified and logically reviewed, and the legal and journal operation reasoning classification results are generated.

8. The automatic classification system for paper withdrawal reasons according to claim 7 is characterized in that: The specific steps for obtaining the output results of the reasoning classification of the paper retraction reasons are as follows: According to the classification results of the above-mentioned legal and journal operation reasoning, collect and organize related legal documents and journal editorial guidelines, classify the contents of legal documents and journal editorial guidelines, identify and record the key elements of legal terms and journal operations, annotate and code each element, and construct a set of legal and journal operation specifications; Based on the legal and journal operational norms, analyze the connection between the differentiated reasoning tasks and the retracted papers, record and classify each connection, adjust the data connection points and optimize the identification process of the retraction reasons, and form a retraction reason analysis model; The retraction reason analysis model is used to classify differentiated retraction cases, adjust the classification parameters in the model and match the cases, repeatedly verify the accuracy and consistency of the model, and obtain the reasoning classification output results of the paper retraction reasons.

Images