A plant paper detection method, system and device based on migration features

Abstract

The present application relates to the technical field of academic detection, and particularly relates to a factory paper detection method, system and device based on migration characteristics; aiming at solving the problems of high dependence on factory paper data scale and being easily affected by modification of explicit characteristics of the existing detection method; it includes respectively extracting the features of fine-grained structure, citation density, style and semantic content of the paper to be detected and the comparative paper; based on the extracted features, calculating the migration characteristic value between the paper to be detected and the comparative paper; comparing the size between the migration characteristic value and the preset threshold value, and judging whether the paper to be detected is a factory paper according to the comparison result and the attribute of the comparative paper; if the migration characteristic value is greater than the preset threshold value and the comparative paper is a normal paper, a first prompt information is generated; if the migration characteristic value is greater than the preset threshold value and the comparative paper is a factory paper, a second prompt information is generated, thereby improving the accuracy and applicability of factory paper detection.

Description

Technical Field

[0001] This invention relates to the field of academic plagiarism detection technology, and in particular to a method, system and equipment for detecting factory-made papers based on migration features. Background Technology

[0002] "Paper mills" refer to organizations that specialize in producing and selling fabricated scientific papers. Their methods are diverse, including plagiarizing others' research, falsifying experimental data, fabricating research conclusions, reusing figures and tables, improper authorship, and systematically manipulating the publication process. With the rapid development of generative artificial intelligence (AI) technology, the operating model of paper mills has been further upgraded. The cost of using large AI models to replace topics with existing legitimate papers and create counterfeit factory-generated papers has been significantly reduced. These newly generated fake factory-generated papers are more concealed, posing a significant challenge to academic misconduct detection efforts.

[0003] The factory-related papers that have been exposed so far exhibit several typical characteristics, including: high uniformity in format and writing style, templated title designs, and similar text descriptions and overall paper structures; generic descriptions of research hypotheses and experimental methods lacking uniqueness and innovation; frequent reuse of image data across different factory-related papers; anomalies in the references, such as citing irrelevant literature or using incorrect citation formats; clear patterns in the usernames, passwords, email addresses, and phone numbers used by the contributors; numerous manuscripts from different locations originating from the same IP address; and instances of unrelated collaborators and institutions sharing authorship.

[0004] Current technologies for detecting "factory-worker" papers primarily employ a combination of methods, including rule statistics and model learning. Based on identified factory-worker paper data, they learn classification features and then use combinations of these features to determine whether a paper to be detected belongs to the factory-worker category. However, these existing methods have significant limitations and are insufficient to meet the actual needs of current factory-worker paper detection.

[0005] High dependence on the scale of factory paper data: Most existing detection methods require a large amount of labeled factory paper data as training samples to build detection models by learning the features in this data. For papers produced by newly emerging paper factories, due to the lack of sufficient sample data, existing models cannot effectively learn their features, resulting in the inability to accurately identify these new factory papers.

[0006] Vulnerable to modifications made by paper factories using explicit features: Existing methods rely heavily on features that are explicit characteristics of paper factories, such as the frequency of use of special symbols or phrases, sentence length features, citation features of paired and group citations, and overall layout features. However, paper factories can easily alter these explicit features through simple manual or AI rewriting and polishing, rendering detection methods based on these features ineffective.

[0007] New AI-generated fake factory papers cannot be effectively identified: Using large AI models, new fake factory papers can be quickly generated by changing the topic and using non-factory papers as a basis. Because these newly generated papers have not appeared in the existing factory paper database, and their explicit features are not significantly different from normal papers, existing detection methods, which rely on the characteristics of already discovered factory papers, cannot identify these new types of factory papers in a timely and effective manner.

[0008] To address some of the aforementioned issues, several improved detection methods have been proposed in related fields, but shortcomings still exist:

[0009] Technical Solution 1 (Method, Device, Electronic Equipment, and Storage Medium for Recognizing Academic Misconduct Text, Patent No. 202410802140.4): This method constructs a detection model by learning the features of a collection of factory papers. The feature vectors of the text are input into a feature detection module for category prediction, essentially a binary classifier model. In feature selection, this method identifies several obvious features as traces of factory papers, including special symbols, the number of sentences containing numbers, and sentence length-related features. These features are categorized into symbol-level, sentence-level, and word-level features, and different types of features are assigned positive or negative values ​​based on their significance in normal and factory papers. However, this method still heavily relies on the existing scale of factory paper data, and the 12 selected trace features are all explicit features. Factory papers can easily alter these features through simple modifications such as character replacement, rewriting, and polishing, causing the detection method to fail.

[0010] Technical Solution 2 (Method for Automatic Detection of Paper Factory Papers Based on Heterogeneous Citation Network, Patent No. 202410911583.7): This method recognizes the lack of in-depth feature mining in citation-related research by existing identification tools. It proposes constructing a heterogeneous citation network graph from the perspective of citations and cited texts, and combining this graph with a heterogeneous attention network model for paper factory paper detection. Specific steps include acquiring raw data of the paper factory, constructing a heterogeneous citation graph network and generating multiple meta-paths, extracting textual features to obtain initial embedding vectors, obtaining final embedding feature vectors through node-level and semantic-level attention learning, and finally training a classifier for detection. However, this method relies on a large amount of citation relationship data from paper factory papers, focusing only on citation relationships such as mutual citations and self-citations between papers or journals. It does not delve into the deep features of citation behavior during the paper factory generation process, such as the positional distribution of cited content within the paper and the stylistic features of the cited content (whether it emphasizes citation concepts, method names, data facts, or viewpoints and summaries). Furthermore, the detection performance of this method depends entirely on a large amount of citation relationship data between paper factory papers and their references; without such data, the detection performance drops significantly.

[0011] Technical Solution 3 (A Method and Apparatus for Detecting Academic Misconduct Documents, Patent No. CN202510558613.5): This method addresses the highly consistent formatting of factory-produced papers by designing a formatting feature representation method centered on figures, tables, and formulas, along with a corresponding document formatting matching method. Its advantage lies in not relying on large amounts of historical data, but its limitations are also significant: it only considers the consistency of formatting features and does not address the analysis of the paper's semantic structure; for papers that contain little or no figures, tables, or formulas, the detection effect is greatly reduced due to the lack of effective formatting features for extraction and matching, failing to meet the diverse detection needs of factory-produced papers.

[0012] In summary, existing methods for detecting plagiarism purportedly created papers have certain shortcomings and deficiencies when dealing with newly generated, modified, or poorly characterized plagiarized papers. Therefore, there is an urgent need for a detection method that is less dependent on the scale of plagiarism purportedly created papers, unaffected by explicit feature modifications, and capable of effectively identifying AI-generated plagiarism and rewritten / polished plagiarism purportedly created papers. This would fill the gaps in existing technology and improve the overall level of academic misconduct detection. Summary of the Invention

[0013] To address the problems in existing technologies, this invention provides a method, system, and device for detecting factory-generated papers based on transfer features. Starting from the characteristics of factory-generated paper generation methods and processes, it achieves accurate detection of factory-generated papers by calculating various transfer features of the papers. This aims to solve the problems of existing methods, such as high dependence on the scale of factory-generated paper data, susceptibility to explicit feature modifications, and inability to effectively identify AI-imitated and rewritten factory-generated papers, thereby improving the applicability, accuracy, and robustness of factory-generated paper detection.

[0014] To achieve the aforementioned objectives, the present invention adopts the following technical solution:

[0015] In a first aspect, the present invention provides a factory paper detection method based on migration features, the method comprising:

[0016] Fine-grained feature extraction of structure, citation density and style, and semantic content is performed on the papers to be tested and the comparison papers respectively.

[0017] Based on the extracted features, the transfer feature values ​​between the paper to be detected and the comparison paper are calculated.

[0018] Compare the migration feature value with a preset threshold, and determine whether the paper to be detected is a factory paper based on the comparison result and the attributes of the paper being compared.

[0019] If the migration feature value is greater than a preset threshold and the comparison paper is a normal paper, a first prompt message is generated; if the migration feature value is greater than a preset threshold and the comparison paper is a factory paper, a second prompt message is generated.

[0020] Optionally, the fine-grained feature extraction of structure, citation density and style, and semantic content for the paper to be detected and the comparison paper respectively includes:

[0021] The sentences in the paper are classified according to the preset functional roles; adjacent functional role classification labels with the same functional role in the sentence functional role classification results are merged to obtain a merged classification sequence; using a single character of the non-classification label as a delimiter, the merged classification sequence is converted into a symbol string and defined as a fine-grained structural feature;

[0022] The system identifies the citations in the introduction, methods, results, and discussion sections of the paper, categorizes the citations according to a preset style, and concatenates the corresponding tags according to the order of their appearance in the text to obtain citation density and style characteristics.

[0023] Semantic content represented by entity relation triples is extracted from the paper, and the triples are grouped according to the functional role classification results of the sentences to obtain a set of triples grouped by functional role; each set of triples corresponds to a functional role classification label of a sentence and the weight coefficient corresponding to the label, thus obtaining semantic content features.

[0024] Optionally, classifying the paper sentences according to preset functional roles includes: segmenting the entire paper based on predefined fine-grained structural classification rules, determining the start and end positions of each part of the IMRD; labeling each sentence in the segmented paper to obtain functional role classification labels corresponding to each sentence, forming a sentence classification label sequence;

[0025] The fine-grained structural classification rules include: the IMRD classification table and the sentence functional role classification table;

[0026] The IMRD classification table is used to divide a paper into four parts: introduction, methods, results, and discussion. Each part corresponds to an IMRD classification label, namely introduction (I), methods (M), results (R), and discussion (D).

[0027] The sentence functional role classification table is used to classify sentences into multiple functional roles based on their function in the paper. Each functional role corresponds to a functional role classification label and a weight coefficient. The functional role classification labels are represented by symbols of equal length. The prefix letters of the functional role classification labels are consistent with the IMRD classification labels.

[0028] Optionally, the step of identifying the cited content in the introduction, methods, results, and discussion sections of the paper, classifying the cited content according to a preset style, and concatenating the tags corresponding to the classification results according to the order of their appearance in the text includes:

[0029] Based on the location information of each part of the paper's IMRD, sentences containing citations are extracted from the four parts of the paper: introduction, methods, results, and discussion, forming a set of citations for each part;

[0030] Based on a predefined citation content classification table, each citation in the citation content set of each part is classified, and the corresponding classification label is determined to form a classification label sequence for the citation content of each part.

[0031] According to the IMRD structure, the IMRD category tags of each part are merged with the corresponding reference content category tags, and then the merged category tags are concatenated.

[0032] Optionally, the step of grouping the triples according to the sentence functional role classification results to obtain a set of triples grouped by functional role includes: extracting the scientific entities contained in each sentence of the paper and the relationships between the entities, and constructing a set of entity-relation triples for each sentence; each triple consists of a first entity, a relation, and a last entity;

[0033] Based on the functional role classification labels of sentences, the entity relation triples corresponding to sentences with the same functional role classification labels are grouped together to form a set of triples under each functional role classification.

[0034] Optionally, the step of calculating the transfer feature value between the paper to be detected and the comparison paper based on the extracted features includes: calculating the fine-grained structural transfer degree, citation density and style transfer degree, and semantic content transfer degree between the paper to be detected and the comparison paper based on the fine-grained structural features, citation density and style features, and semantic content features, and summing the transfer degrees to obtain the transfer feature value.

[0035] Optionally, the step of calculating the fine-grained structural transfer, citation density and style transfer, and semantic content transfer between the paper to be detected and the comparison paper based on their fine-grained structural features, citation density and style features, and semantic content features includes: preprocessing the fine-grained structural features of the paper to be detected and the comparison paper by replacing common substrings with lengths greater than a preset threshold in the character sequences of the two fine-grained structural features with single characters that do not repeat existing characters; using the Levenstein distance calculation method to calculate the edit distance between the two fine-grained structural features after preprocessing, where the edit distance is the minimum number of editing operations required to convert the fine-grained structural features of the paper to be detected into the fine-grained structural features of the comparison paper; and calculating the fine-grained structural transfer based on the edit distance and the lengths of the two fine-grained structural features.

[0036] A method for calculating fine-grained structural mobility is used to obtain citation density and style mobility.

[0037] For two sets of triples under the same sentence functional role classification in the paper to be tested and the comparison paper, the transferability relationship between the triples in the set is determined, and the semantic transferability under the same functional role classification is calculated.

[0038] Optionally, the step of determining the transferability relationship between the triples in the two sets of triples under the same sentence functional role classification in the paper to be detected and the comparison paper, and calculating the semantic transferability under the same functional role classification, includes:

[0039] Determine whether the similarity between the triple G1 group of the paper to be detected and the triple G2 group of the comparison paper is greater than a preset threshold, and whether the first entity of G1 and the first entity of G2, and the last entity of G1 and the last entity of G2 belong to the same type of entity or have the same superordinate class; wherein, the relationship similarity is obtained by querying a preset relationship similarity table or semantic calculation, and the entity of the same type or superordinate class relationship is obtained by verifying through knowledge graph and semantic dictionary;

[0040] The number of transferable relationships between each triplet in the set of triplets in the papers to be detected and the set of triplets in the comparison papers is counted; where a single triplet in the paper to be detected corresponds to one or more triplets in the comparison papers, it is counted as 1 match.

[0041] Record the number of matching triples between the paper to be detected and the comparison paper, and the number of matching triples between the comparison paper and the paper to be detected;

[0042] Based on the number of matching triples between the paper to be detected and the comparison paper and the number of triples in the triple set, the semantic transfer degree under the functional role classification is calculated.

[0043] Based on the semantic transferability under the functional role classification and the weight coefficient corresponding to the functional role classification, the semantic content transferability between the paper to be detected and the comparison paper is calculated.

[0044] Secondly, the present invention provides a factory paper detection system based on migration features, the system comprising:

[0045] The feature extraction module is used to extract fine-grained features of structure, citation density and style, and semantic content from the paper to be detected and the comparison paper, respectively.

[0046] The calculation module is used to calculate the transfer feature values ​​between the paper to be detected and the comparison paper based on the extracted features.

[0047] The determination module is used to compare the migration feature value with a preset threshold, and determine whether the paper to be detected is a factory paper based on the comparison result and the attributes of the paper to be compared.

[0048] The result prompting module is used to generate a first prompt message if the migration feature value is greater than a preset threshold and the comparison paper is a normal paper; and to generate a second prompt message if the migration feature value is greater than the preset threshold and the comparison paper is a factory paper.

[0049] Thirdly, the present invention provides an electronic device, the electronic device comprising:

[0050] At least one processor; and

[0051] A memory communicatively connected to the at least one processor; wherein,

[0052] The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method described in any one of the first aspects.

[0053] Compared with the closest existing technology, the present invention has the following advantages:

[0054] This invention, based on the characteristics of factory paper generation methods and processes, provides a factory paper detection method, system, and device based on transfer features. It achieves factory paper detection by calculating the transfer features between the paper to be detected and the comparison papers (factory papers and already published papers). Considering three core transfer features in the factory paper generation process—fine-grained structural transfer features, citation density and style transfer features, and semantic content transfer features—it constructs detection indicators from three dimensions: paper structure, citation features, and semantic content. By integrating and calculating the fine-grained structural features, citation density and style features, and semantic content features of the paper, it provides a new solution for factory paper detection, effectively improving the overall level of academic misconduct detection.

[0055] This invention proposes a method, system, and device for detecting factory papers based on transfer features. It is not only applicable to the detection of traditional factory papers but also effectively addresses new types of factory papers generated and rewritten by AI. It can effectively identify fake factory papers generated by AI models that mimic normal papers with different topics, or factory papers whose language style has been altered through AI rewriting and polishing. Furthermore, it has low dependence on the scale of factory paper data and is not limited by the explicit features of factory papers. Even with only one factory paper or only a copy of the original paper, factory paper detection can be achieved. It enables detection with small samples or even single samples, significantly improving the applicability of the method.

[0056] This invention's method can adjust parameters such as the sentence functional role classification table, the cited content classification table, and the migration feature threshold according to the characteristics of different disciplines, making it applicable to the detection of factory-style papers in multiple disciplines, including natural sciences, social sciences, and engineering technology. Furthermore, this method can be used in conjunction with other existing detection methods, further improving detection accuracy through multi-method fusion, and exhibits good scalability and compatibility. Attached Figure Description

[0057] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0058] Figure 1 This is a flowchart of a factory paper detection method based on migration features provided by the present invention;

[0059] Figure 2 This is a flowchart illustrating the working process of the factory paper detection method based on migration features provided by the present invention.

[0060] Figure 3 This is a schematic diagram of the structural modules of the factory paper detection system based on migration features provided by the present invention;

[0061] Figure 4 This is an internal structural diagram of the electronic device provided by the present invention. Detailed Implementation

[0062] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0063] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0064] This invention provides a method, system, and device for detecting factory-generated papers based on transfer features. It can be widely applied to factory-generated paper detection, AI-generated factory-generated paper detection, and academic misconduct paper detection scenarios. The aim is to solve the problems of existing detection methods, such as high dependence on the scale of factory-generated paper data and susceptibility to explicit feature modifications, thereby improving the accuracy and applicability of factory-generated paper detection. Embodiments of this invention are described below with reference to the accompanying drawings.

[0065] Example 1: Please refer to Figure 1 , Figure 1 Embodiment 1 of the present invention provides a method for detecting factory-style papers based on transfer features. This method detects factory-style papers by calculating transferable features between the paper to be detected and a comparison paper (the comparison paper is either a published normal paper or a factory-style paper). The transferable features include fine-grained structural transfer features, citation density and style transfer features, and semantic content transfer features. The method specifically includes the following steps:

[0066] S101 performs fine-grained feature extraction on the paper to be detected and the comparison paper, respectively, focusing on structural, citation density and style, and semantic content features.

[0067] S102 calculates the migration feature values ​​between the paper to be detected and the comparison paper based on the extracted features;

[0068] S103 compares the migration feature value with the preset threshold, and determines whether the paper to be detected is a factory paper based on the comparison result and the attributes of the paper to be compared.

[0069] S104 If the migration feature value is greater than the preset threshold and the comparison paper is a normal paper, then a first prompt message is generated; if the migration feature value is greater than the preset threshold and the comparison paper is a factory paper, then a second prompt message is generated.

[0070] In step S101 above, fine-grained structural features, citation density and style features, and semantic content features are extracted from the paper to be detected and the comparison paper, respectively. Through the feature extraction in step S101, fine-grained structural features, citation density and style features, and semantic content features are obtained, respectively.

[0071] Specifically, the feature extraction process for fine-grained structures includes:

[0072] The sentences in the paper are classified according to the preset functional roles; adjacent functional role classification labels with the same functional role in the sentence functional role classification results are merged to obtain a merged classification sequence; using a single character of the non-classification label as a delimiter, the merged classification sequence is converted into a symbol string and defined as a fine-grained structural feature;

[0073] In the above embodiments, classifying the paper sentences according to preset functional roles includes: segmenting the entire paper based on predefined fine-grained structural classification rules, determining the start and end positions of each part of the IMRD; labeling each sentence in the segmented paper to obtain functional role classification labels corresponding to each sentence, forming a sentence classification label sequence;

[0074] The fine-grained structural classification rules include: the IMRD classification table and the sentence functional role classification table;

[0075] The IMRD classification table is used to divide a paper into four parts: introduction, methods, results, and discussion. Each part corresponds to an IMRD classification label, namely introduction (I), methods (M), results (R), and discussion (D).

[0076] The sentence functional role classification table is used to classify sentences into multiple functional roles based on their function in the paper. Each functional role corresponds to a functional role classification label and a weight coefficient. The functional role classification labels are represented by symbols of equal length. The prefix letters of the functional role classification labels are consistent with the IMRD classification labels.

[0077] The IMRD classification table is defined as follows:

[0078] Following the internationally accepted IMRD structure, the classification table 1 is set as follows:

[0079] Table 1

[0080]

[0081] In this classification table, the four sections I, M, R, and D are presented in the order of the paper's writing, with no repetition in the content of each section. The introduction mainly introduces the research background, motivation, and purpose; the research methods section describes the experimental design, materials, and procedures; the research results section presents the experimental data and analysis results; and the discussion section interprets the results and provides an outlook.

[0082] The sentence function role classification table includes the following settings:

[0083] The sentences in the paper are categorized into multiple functional roles, as shown in Table 2:

[0084] Table 2

[0085]

[0086] In this classification table, the prefix letters "I", "M", "R" and "D" correspond to the four parts of IMRD, indicating the more likely position of the label (e.g., I1-I4 are more likely to appear in the introduction, M1-M6 are more likely to appear in the research methods section, etc.). The weight coefficients are set according to the importance of each functional role in the paper, and the total weight is 1. They can be fine-tuned according to actual detection needs.

[0087] In the above embodiments, the feature extraction process for citation density and style includes:

[0088] The system identifies citations in the introduction, methods, results, and discussion sections of a paper, categorizes these citations according to a preset style, and then concatenates the corresponding tags according to their order of appearance in the text to obtain citation density and style characteristics.

[0089] Specifically, the system identifies citations in the introduction, methods, results, and discussion sections of the paper, categorizes these citations according to a predefined style, and then concatenates the corresponding tags based on their order of appearance in the text, including:

[0090] Based on the location information of each part of the paper's IMRD, sentences containing citations are extracted from the four parts of the paper: introduction, methods, results, and discussion, forming a set of citations for each part;

[0091] Based on a predefined citation content classification table, each citation in the citation content set of each part is classified, and the corresponding classification label is determined to form a classification label sequence for the citation content of each part.

[0092] According to the IMRD structure, the IMRD category tags of each part are merged with the corresponding reference content category tags, and then the merged category tags are concatenated.

[0093] In the above embodiments, extracting the paper's introduction specifically includes: using a citation extraction model, based on the annotation results of the four parts of the full-text IMRD, extracting all sentences containing cited content in each part, including not only direct citations but also indirect citations:

[0094] IC = {R1, R2, ...}; represents the content cited in the introduction section of the paper.

[0095] MC = {R1, R2, ...}; represents the cited content in the methods section of the paper.

[0096] RC = {R1, R2, ...}; represents the content cited in the results section of the paper.

[0097] DC = {R1, R2, ...} represents the content cited in the discussion section of the paper.

[0098] R i This represents the i-th referenced content within each section.

[0099] In the above embodiments, classifying each reference in the reference content set of each part based on a predefined reference content classification table includes: classifying the extracted IC, MC, RC, and DC reference content using a classification model, with classification labels shown in Table 3:

[0100] Table 3

[0101]

[0102] The table above does not limit the number of categories. Different numbers of categories can be set based on data from different disciplines.

[0103] IC = {C1, C2, C3, C1…}; Categorization tags for citations in the introduction section of the paper;

[0104] MC = {C1, C2, C1, C1, ...}; Categorization tags for the cited content in the methods section of the paper.

[0105] RC = {C6, C2, C5, ...}; Categorization tags for citations in the results section of the paper.

[0106] DC = {C5, C1, C1, ...}; categorization tags for citations in the discussion section of the paper.

[0107] Constructing citation density and style features: Based on the IMRD structure, construct citation density and citation style representation symbols in the form of characters, merge the IMRD category labels with the citation content category labels, and then concatenate all category labels.

[0108] Example: The citation density and citation style notation obtained by concatenating the citation content classification results with IMRD classification labels is: IMRDC=IC1-IC2-IC3-IC1...MC1-MC2-MC1-MC1...RC6-RC2-RC5...DC5-DC1-DC1…

[0109] Citation density represents the total number of citations and their distribution within each part of the IMRDC, while citation style represents the number and distribution of category labels for the citation content. As shown in IMRDC, the converted character sequence is used to characterize citation density and citation style.

[0110] Specifically, the feature extraction process for semantic content includes:

[0111] Semantic content represented by entity relation triples is extracted from the paper, and the triples are grouped according to the functional role classification results of the sentences to obtain a set of triples grouped by functional role; each set of triples corresponds to a functional role classification label of a sentence and the weight coefficient corresponding to the label, thus obtaining semantic content features.

[0112] In the above embodiments, the step of grouping the triples according to the sentence functional role classification results to obtain a set of triples grouped by functional role includes: extracting the scientific research entities contained in each sentence of the paper and the relationships between the entities, and constructing a set of entity-relation triples for each sentence; each triple consists of a first entity, a relation, and a last entity;

[0113] Based on the functional role classification labels of sentences, the entity relation triples corresponding to sentences with the same functional role classification labels are grouped together to form a set of triples under each functional role classification.

[0114] In one embodiment, the specific technical solution for representing and extracting fine-grained structural features of a paper is as follows:

[0115] First, a fine-grained structural classification system for academic papers is established, consisting of an IMRD (Introduction, Methods, Results, Discussion) classification table and a sentence functional role classification table. The IMRD classification table divides the paper into four parts: Introduction (I), Methods (M), Results (R), and Discussion (D), presented sequentially without repetition. The sentence functional role classification table categorizes sentences according to their function within the paper, classifying them into various functional roles such as Background (I1), Motivation (I2), Purpose (I3), Significance (I4), Research Subject (M1), Research Topic (M2), Research Hypothesis (M3), Research Method (M4), Model (M5), Algorithm (M6), Experiment (R1), Data (R2), Results (R3), Conclusion (D1), Discussion (D2), and Outlook (D3). Each functional role corresponds to a classification label, and these labels use symbols of equal length. A common prefix letter indicates the IMRD position where the label is more likely to appear (e.g., labels prefixed with "I" are more likely to appear in the Introduction). In addition, the sentence function role classification table can be refined and adjusted according to the characteristics of different disciplines, and classification labels can be added or modified.

[0116] Next, fine-grained structural features were extracted from the paper:

[0117] The entire paper is segmented using sequence labeling or classification models to determine the start and end positions of each of the four parts of the IMRD.

[0118] Similarly, sequence labeling models or classification models are used to label each sentence in the full text of the paper, determine the functional role classification label corresponding to each sentence, and form a sentence classification label sequence;

[0119] The obtained sentence classification label sequences are merged according to the rule of "adjacent and identical classification labels" to obtain the merged classification sequence;

[0120] The merged classification sequence is converted into a symbolic representation string, and any single non-classification label character is used as a delimiter to represent the fine-grained structural features of the paper. This string is the fine-grained structural feature of the paper.

[0121] In one embodiment, the representation and extraction of citation density and style features of a paper are as follows:

[0122] First, a categorization table for cited content is set up. Based on the type of cited content, it is divided into several categories such as cited concept name (C1), cited data (C2), cited formula (C3), cited image (C4), cited viewpoint (C5), and cited rule (C6). Each category corresponds to a category label, and the number and type of category labels can be adjusted according to the characteristics of data from different disciplines.

[0123] Then, in the extraction of citation density and style features of the paper, the citation extraction model is used, combined with the determined position information of each part of the paper's IMRD, to extract all sentences containing citations in the four parts of the introduction (I), method (M), results (R), and discussion (D), including direct and indirect citations, and to form the citation content set of each part respectively.

[0124] The classification model is used to classify each piece of referenced content in the collection of referenced content in each part, determine its corresponding referenced content classification label, and obtain the classification label sequence of referenced content in each part;

[0125] Following the order of the IMRD structure, the IMRD classification labels of each part are merged with the corresponding citation content classification labels. Then, all classification labels are concatenated in sequence to form a symbolic representation string. This string represents the citation density and style characteristics of the paper. The structure and length of the string reflect the total number of citations and their distribution in different parts of the IMRD (citation density), while the type and distribution of citation content classification labels reflect the style characteristics of the citations.

[0126] In one embodiment, the representation and extraction of semantic content features of a paper are as follows:

[0127] First, the research entities and their relationships are extracted from the sentences in the full text of the paper. Using the entity recognition and relationship extraction model, the research entities (such as research objects, concepts, methods, data, etc.) contained in each sentence and the relationships between these entities are extracted. The entity relationship triple set of each sentence is constructed. Each triple consists of a first entity, a relation, and a last entity (i.e., the form of "first entity-relation-last entity").

[0128] Then, based on the functional role classification labels of sentences in the paper, the extracted entity relation triples are grouped: the entity relation triples corresponding to sentences with the same functional role classification labels are grouped together to form a set of triples under each functional role classification. Each set of triples corresponds to a sentence functional role classification label, and each classification label corresponds to a preset weight coefficient. This weight coefficient is set according to the importance of different functional roles in the paper and can be adjusted according to the subject area and detection requirements.

[0129] Further, in step S102, the calculation of the transfer feature value between the paper to be detected and the comparison paper based on the extracted features includes: calculating the fine-grained structural transfer degree, citation density and style transfer degree, and semantic content transfer degree between the paper to be detected and the comparison paper based on the fine-grained structural features, citation density and style features, and semantic content features, and summing the transfer degrees to obtain the transfer feature value.

[0130] In the above embodiments, the calculation of the fine-grained structural transfer, citation density and style transfer, and semantic content transfer between the paper to be detected and the comparison paper based on their fine-grained structural features, citation density and style features, and semantic content features includes: preprocessing the fine-grained structural features of the paper to be detected and the comparison paper by replacing common substrings with lengths greater than a preset threshold in the character sequences of the two fine-grained structural features with single characters that do not repeat existing characters; using the Levenstein distance calculation method to calculate the edit distance between the two fine-grained structural features after preprocessing, where the edit distance is the minimum number of editing operations required to convert the fine-grained structural features of the paper to be detected into the fine-grained structural features of the comparison paper; and calculating the fine-grained structural transfer based on the edit distance and the lengths of the two fine-grained structural features.

[0131] A method for calculating fine-grained structural mobility is used to obtain citation density and style mobility.

[0132] For two sets of triples under the same sentence functional role classification in the paper to be tested and the comparison paper, the transferability relationship between the triples in the set is determined, and the semantic transferability under the same functional role classification is calculated.

[0133] In the above embodiments, determining the transferability relationship between the triples in the two sets of triples under the same sentence functional role classification in the paper to be detected and the comparison paper, and calculating the semantic transferability under the same functional role classification includes:

[0134] Determine whether the similarity between the triple G1 group of the paper to be detected and the triple G2 group of the comparison paper is greater than a preset threshold, and whether the first entity of G1 and the first entity of G2, and the last entity of G1 and the last entity of G2 belong to the same type of entity or have the same superordinate class; wherein, the relationship similarity is obtained by querying a preset relationship similarity table or semantic calculation, and the entity of the same type or superordinate class relationship is obtained by verifying through knowledge graph and semantic dictionary;

[0135] The number of transferable relationships between each triplet in the set of triplets in the papers to be detected and the set of triplets in the comparison papers is counted; where a single triplet in the paper to be detected corresponds to at least one triplet in the comparison papers, it is counted as 1 match.

[0136] Record the number of matching triples between the paper to be detected and the comparison paper, and the number of matching triples between the comparison paper and the paper to be detected;

[0137] Based on the number of matching triples between the paper to be detected and the comparison paper and the number of triples in the triple set, the semantic transfer degree under the functional role classification is calculated.

[0138] Based on the semantic transferability under the functional role classification and the weight coefficient corresponding to the functional role classification, the semantic content transferability between the paper to be detected and the comparison paper is calculated.

[0139] Furthermore, based on the above content in step S102, the migration feature calculation process is explained in detail:

[0140] Fine-grained structural transfer feature calculation: The edit distance formula is used to calculate the edit distance between the fine-grained structural features of the paper to be detected and the comparison paper (normal paper or factory paper). This edit distance represents the minimum number of editing operations (including deletion, insertion, and replacement) required to convert the fine-grained structural features of the paper to be detected into the fine-grained structural features of the comparison paper. To solve the problem of large edit distance caused by local structural interchange, before calculating the edit distance, common substrings with a length greater than a preset threshold in the two features can be replaced with a single character that does not repeat the existing characters in the feature. Then, based on the edit distance and the length of the two features, the fine-grained structural transfer degree is calculated. The core logic of the calculation formula is: Fine-grained structural transfer degree = 1 - (edit distance / (length of fine-grained structural feature of paper to be detected + length of fine-grained structural feature of comparison paper)). The value range of the transfer degree is [0,1]. The closer the transfer degree is to 1, the more similar the fine-grained structure of the two papers is, and the stronger the fine-grained structural transfer feature is.

[0141] Citation density and style transfer feature calculation: Using the same method as the fine-grained structural transfer feature calculation, firstly, the edit distance between the citation density and style features of the paper to be detected and the comparison paper is calculated (common substring replacement preprocessing can be performed in the same way). Then, based on the edit distance and the length of the two features, the citation density and style transfer degree are calculated. The calculation logic is consistent with the fine-grained structural transfer degree. The closer the transfer degree is to 1, the more similar the citation density and style of the two papers are, and the stronger the citation density and style transfer features are.

[0142] Semantic content transfer feature calculation: First, for the sets of triples under the same sentence functional role classification in the paper to be detected and the comparison paper, the transferability relationship between the triples in the two sets is calculated. The criteria for determining whether there is a transferability relationship between two triples (triple G1 of the paper to be detected and triple G2 of the comparison paper) are: the similarity of the relationship between the two triples is greater than a preset threshold α, and the first entity of G1 and the first entity of G2 belong to the same type of entity or have the same superordinate class, while the last entity of G1 and the last entity of G2 belong to the same type of entity or have the same superordinate class. Among them, the relationship similarity can be obtained by querying a preset relationship similarity table, or by calculating the semantic similarity between the vector representations of the two relationships; whether the entities are of the same type or have the same superordinate class can be verified by large model inference, or by using existing knowledge graphs, WordNet semantic dictionaries, etc.

[0143] Next, the semantic transferability under the same functional role classification is calculated: the number of triples in the set of triples of the paper to be detected that have a transferable relationship with the set of triples of the comparison papers is counted (when a triple of the paper to be detected corresponds to multiple triples of the comparison papers, it is only counted as 1 match), and the sum is used to obtain the number of matching triples from the paper to the comparison papers; similarly, the number of matching triples from the set of triples of the comparison papers to the paper to be detected is counted; then, the sum of the number of matching triples in the two directions is divided by the sum of the total number of triples in the two sets of triples to obtain the semantic transferability under the functional role classification.

[0144] Finally, the overall semantic content transfer degree is calculated: the semantic transfer degree under each functional role category is multiplied by the weight coefficient corresponding to that category, and then all results are summed to obtain the semantic content transfer degree between the paper to be detected and the comparison paper. The closer the transfer degree is to 1, the more similar the semantic content of the two papers is, and the stronger the semantic content transfer feature is.

[0145] In step S103 above, the factory paper detection judgment includes: adding the fine-grained structure transfer degree, citation density and style transfer degree, and semantic content transfer degree to obtain the transfer feature value between the paper to be detected and the comparison paper. A transfer feature threshold θ is set (this threshold can be calibrated based on a large amount of experimental data, and different thresholds can be set for different disciplines). If the transfer feature value is greater than the threshold θ, a judgment is made based on the type of the comparison paper: if the comparison paper is a normal paper, it is suggested that the paper to be detected may be a factory paper generated by imitating the normal paper; if the comparison paper is a factory paper, it is suggested that the paper to be detected may belong to the same type of factory paper produced by the same paper factory as the factory paper.

[0146] For example, 1. The user input paper A is processed through step 101 to obtain fine-grained structural features, citation density and style features, and semantic features:

[0147] For example, the paper's fine-grained structural features:

[0148] A_ SCF="I1-I2-I4-I5-M1-I1-M3-M2-...-S n "

[0149] Citation density and style characteristics:

[0150] A_IMRDC "IC1-IC2-IC3-IC1...MC1-MC2-MC1-MC1...RC6-RC2-RC5...DC5-DC1-DC1..."

[0151] Semantic features:

[0152] A_LG={LG1, LG2, LG3, ... LG n};

[0153] In the formula, LG n LG represents all triple combinations under the nth category label. n ={G1,G2,…G m};G m Represents a single triple: G m ={h, r, t}; h and t represent the first and last research entities extracted, respectively, and r represents the relationship between the two research entities.

[0154] After processing through step 101, the fine-grained structural features, citation density and style features, and semantic features of the comparative paper B are obtained respectively:

[0155] For example, the paper's fine-grained structural features:

[0156] B_ SCF=“I1-I2-I4-I5-M1-I1-M3-M2-...-S n "

[0157] Citation density and style characteristics:

[0158] B_IMRDC="IC1-IC2-IC3-IC1...MC1-MC2-MC1-MC1...RC6-RC2-RC5...DC5-DC1-DC1..."

[0159] Semantic features:

[0160] B_LG={LG1, LG2, LG3, ... LG n};

[0161] In the formula, LG n LG represents all triple combinations under the nth category label. n={G1,G2,…G m};G m Represents a single triple: G m ={h, r, t}; h and t represent the first and last research entities extracted, respectively, and r represents the relationship between the two research entities.

[0162] 2. Calculate the structural transfer features of the paper:

[0163] Calculate the edit distance between A_SCF and B_SCF, which is the minimum number of editing operations (deletion, insertion, replacement) required to transform the structure A_SCF of paper A into the structure B_SCF of paper B, and use it as the paper structure transferability feature.

[0164] Among them, the smaller the edit distance, the more similar the fine-grained structure of the papers, and the stronger the paper structure transfer characteristics.

[0165] The structural transfer features of a paper are calculated using the following formula:

[0166]

[0167] In the formula, len(A_SCF) and len(B_SCF) represent the lengths of A_SCF and B_SCF, respectively. EditDistance(A_SCF, B_SCF) represents the edit distance, which is calculated directly using the Levenshtein distance method, resulting in the following recursive formula:

[0168] ;

[0169] Indicator functions: ;

[0170] In the formula, a i , b j These represent the edit distances of the first i and the first j characters, respectively.

[0171] Using the Levenstein distance calculation formula above, the number of edits (insertion, deletion, replacement) required between A_SCF and B_SCF is obtained as the edit distance EditDistance(A_SCF, B_SCF), which in turn yields... .

[0172] Optionally, before calculating the edit distance, common substrings with a length greater than the threshold in the two strings A_SCF and B_SCF can be replaced with other single characters (not repeating the characters contained in A_SCF and B_SCF) to resolve the impact of large edit distance caused by local structure swapping.

[0173] 3. Calculate citation density and style transfer features:

[0174] Using the same method as the paper structure transfer, the transfer degree between A_IMRDC and B_IMRDC was calculated, and the results were obtained.

[0175] StrTrans(A_IMRDC, B_IMRDC) serves as a feature for citation density and citation style transfer.

[0176] 4. Calculate semantic transfer degree:

[0177] The transferability of semantic features A_LG and B_LG is calculated as the semantic transfer features between paper A and paper B.

[0178] Calculate whether there is a transferable relationship between the triples of A_LG and B_LG in each class:

[0179] The method for calculating the migration relationship of triplet is as follows:

[0180] ;

[0181] In the formula, T(t) 1 ，t 2 ） This is used to calculate whether a transfer relationship exists between two triples. 1 indicates yes, and 0 indicates no. Triples are considered to have a transferable relationship when the similarity threshold is greater than α and the first and last entities belong to the same class or have the same superclass.

[0182] Sim(r 1 ，r 2 ) This indicates the similarity or consistency between relations r1 and r2 contained in two triples. This can be achieved by looking up a predefined relation similarity table, or by calculating the semantic similarity represented by the vectors r1 and r2. No limitation is imposed here. Same_or_hyper(h1, h2) indicates whether two entities or their corresponding upper and lower-order entities are the same. This can be verified through large-scale model inference, or based on existing knowledge graphs or WordNet semantic dictionaries. No limitation is imposed here.

[0183] The semantic transferability feature calculation method for the same category is to divide the number of mutually transferable features by the sum of their respective total numbers. The formula is:

[0184]

[0185] In the formula, G i G jThese represent the triplets at their respective i-th and j-th positions within the same category of two papers (grouped according to the sentence function roles in the sentence function role classification table).

[0186] Calculate the semantic transfer features between paper A and paper B:

[0187]

[0188] Each category's transfer feature is multiplied by its respective category weight coefficient (Table 2 shows the weights corresponding to the sentence functional role classification in the paper).

[0189] Calculate paper transfer features: Trans(A,B) = SemTrans(A,B) + StrTrans(A_SCF,B_SCF) + StrTrans(A_IMRDC, B_IMRDC)

[0190] If Trans(A,B)>θ, then:

[0191] If B is a normal paper, it suggests that A may be a factory paper that plagiarizes B.

[0192] If B is a factory-related paper, it suggests that A may be a similar factory-related paper to B.

[0193] Step S104 above further describes the method for determining normal papers and factory papers, and can also issue corresponding reminders to users based on the output determination results. Specifically, if the migration feature value is greater than a preset threshold and the comparison paper is a normal paper, a first prompt message is generated; if the migration feature value is greater than the preset threshold and the comparison paper is a factory paper, a second prompt message is generated.

[0194] Furthermore, this method can be applied independently to detect factory-related papers, or it can be combined with other existing factory-related paper detection methods (such as rule-based statistical methods and model-based classification methods) to further improve the accuracy of factory-related paper detection through multi-method collaborative detection.

[0195] Example 2: To make the technical solution of the present invention clearer and easier to understand, the following detailed description of the factory paper detection method based on migration features of the present invention is provided in conjunction with Example 2. This example takes paper detection in a certain medical field as an example, and the specific steps are as follows:

[0196] I. Based on the characteristics of medical papers, the internationally accepted IMRD structure is adopted. A fine-grained structural classification table is established based on Table 1, and a sentence functional role classification table is used in Table 2 to categorize sentences in medical papers into multiple functional roles. In this fine-grained structural classification table, the four parts (I, M, R, D) are presented sequentially according to the writing order of the paper, with no repetition in the content of each part. The introduction mainly introduces the research background, motivation, and purpose; the research methods section describes the experimental design, materials, and procedures; the research results section presents experimental data and analysis results; and the discussion section interprets the results and provides future prospects.

[0197] In the sentence functional role classification table, the prefix letters "I", "M", "R" and "D" correspond to the four parts of IMRD, indicating the position where the tag is more likely to appear (e.g., I1-I4 are more likely to appear in the introduction, M1-M6 are more likely to appear in the research methods section, etc.). The weight coefficients are set according to the importance of each functional role in the medical paper, and the total weight is 1. They can be fine-tuned according to actual detection needs.

[0198] II. Feature Extraction of the Paper to be Detected and the Comparison Paper

[0199] Paper A (suspected factory paper) and paper B (normal paper in the medical field) were selected as the detection objects. Fine-grained structural features, semantic features, citation density and style features were extracted from the two papers respectively.

[0200] (1) Fine-grained structural feature extraction

[0201] IMRD Segmentation: Using a sequence labeling model based on the Transformer architecture, the full texts of Paper A and Paper B are segmented to determine the location of each part of the IMRD. For example, the introduction of Paper A begins with the first sentence of paragraph 1 and ends with the fourth sentence of paragraph 3; the methods section begins with the first sentence of paragraph 4 and ends with the third sentence of paragraph 7; the results section begins with the first sentence of paragraph 8 and ends with the second sentence of paragraph 12; and the discussion section begins with the first sentence of paragraph 13 and ends with the fifth sentence of paragraph 16. The IMRD segmentation results for Paper B are similar and will not be repeated here.

[0202] Sentence Functional Role Labeling: A sequence labeling model based on the Transformer architecture was used to label the functional roles of each sentence in the full text of Paper A and Paper B. Taking the sentences in the introduction of Paper A as an example: "Angiostrongylus cantonensis infection is an important zoonotic parasitic disease with an increasing incidence rate worldwide," labeled as I1 (disease background); "Current diagnostic methods for this disease have low sensitivity, and new detection technologies are urgently needed," labeled as I2 (research motivation); "This study aims to establish a highly sensitive diagnostic method based on molecular biology," labeled as I3 (research objective). Through labeling, the sentence classification label sequence of Paper A is obtained as S_A={I1,I1,I2,I3,I4,M1,M1,M2,M3,M4,R1,R2,R2,R3,D1,D2,D3,...}, and the sentence classification label sequence of Paper B is similar.

[0203] Label sequence merging: S_A and S_B are merged according to the rule of "adjacent and with the same classification label". For example, two consecutive I1s in S_A are merged into one I1, and two consecutive M1s are merged into one M1. After merging, the classification sequence of paper A is SC_A={I1,I2,I3,I4,M1,M2,M3,M4,R1,R2,R3,D1,D2,D3,...}, and the merged classification sequence of paper B is similar.

[0204] Feature transformation: Using "-" as the delimiter, the merged classification sequences are converted into symbol strings. The fine-grained structural feature A_SCF of paper A is "I1-I2-I3-I4-M1-M2-M3-M4-R1-R2-R3-D1-D2-D3-...", and the fine-grained structural feature B_SCF of paper B is similar.

[0205] (2) Semantic feature extraction

[0206] Entity-Relation Triple Extraction: Using a pre-trained entity recognition and relation extraction model in the medical field (such as BioBERT + relation classification model), scientific entities and relations are extracted from each sentence of Paper A and Paper B. Taking the sentence "This study used SPF-grade Kunming mice as experimental subjects and employed real-time quantitative PCR to detect the egg load of Angiostrongylus cantonensis" in Paper A as an example, the extracted entity-relation triples include: G1={"SPF-grade Kunming mice", "as", "experimental subjects"}, G2={"real-time quantitative PCR method", "used to detect", "egg load of Angiostrongylus cantonensis"}. Through extraction, the entity-relation triple set SG_A={SG1,SG2,SG3,...} of Paper A is obtained (SGn is the triple set of the nth sentence), and the triple set SG_B of Paper B is similar.

[0207] Triple grouping: Based on the sentence functional role classification labels obtained in step 2, the triples in SG_A and SG_B are grouped. For example, in paper A, the triples corresponding to sentences labeled M1 (experimental subjects) are grouped into one group LG_A1, and the triples corresponding to sentences labeled M4 (detection methods) are grouped into one group LG_A4. Finally, the grouped triple set of paper A is A_LG={LG_A1,LG_A2,LG_A3,...}; LG_An is the set of triples under the nth functional role classification of paper A. The grouped triple set of paper B is B_LG={LG_B1,LG_B2,LG_B3,...}.

[0208] (3) Extraction of citation density and style features

[0209] Citation Extraction: A rule-based and deep learning-based citation extraction model was used, combined with the IMRD segmentation results of Paper A and Paper B, to extract citations from each part. For example, the sentence quoted in the introduction of Paper A, "Smith et al. (2022) pointed out that the prevalence of Angiostrongylus cantonensis disease in the Pacific region exceeds 10%", was extracted as citation content R1; the sentence quoted in the methods section, "The primer sequences used in this study were designed according to the sequence reported by Wang et al. (2021)", was extracted as citation content R2. Through extraction, the citation content sets of each part of Paper A were obtained: IC_A (introduction), MC_A (methods), RC_A (results), and DC_A (discussion). The citation content sets of Paper B, IC_B, MC_B, RC_B, and DC_B, were similar.

[0210] Citation content classification: A classification table for medical citation content was set up (Table 3), and the extracted citation content was classified using a classification model.

[0211] For example, the citation R1 in the introduction of paper A ("Smith et al. (2022) pointed out that the prevalence of Angiostrongylus cantonensis infection in the Pacific region exceeds 10%) is classified as C2 (cited data); the citation R2 in the methods section ("The primer sequences used in this study refer to the sequence design reported by Wang et al. (2021)") is classified as C3 (cited experimental methods). Through classification, the classification label sequences of the citations in each part of paper A are obtained: IC_A={C2,C5,C1,...}, MC_A={C3,C3,C2,...}, RC_A={C4,C2,C5,...}, DC_A={C5,C1,C3,...}, and the classification label sequences IC_B, MC_B, RC_B, and DC_B of paper B are similar.

[0212] Feature transformation: Following the IMRD structure order, the IMRD category labels of each part are merged and concatenated with the citation content category labels, using "-" as the separator. Citation density and style features of paper A: A_IMRDC = "IC2-IC5-IC1-...-MC3-MC3-MC2-...-RC4-RC2-RC5-...-DC5-DC1-DC3-...". The feature B_IMRDC of paper B is similar.

[0213] III. Calculation of Transfer Features

[0214] (1) Calculation of fine-grained structure migration characteristics

[0215] Common substring replacement preprocessing: A common substring length threshold of 3 is set. Both A_SCF and B_SCF are analyzed separately to find common substrings with a length greater than 3. For example, common substrings “I1-I2-I3-I4” and “M1-M2-M3-M4” are found and replaced with the single, non-repeating characters “X” and “Y”. After the replacement, A_SCF becomes “X-M5-M6-Y-R1-R2-R3-…”, and B_SCF becomes “Y-M5-M6-X-R1-R2-R3-…”.

[0216] Edit distance calculation: The Levenstein distance calculation method is used to calculate the edit distance between A_SCF and B_SCF after replacement. Assuming that the length of A_SCF after replacement is 20 and the length of B_SCF is 20, the calculated edit distance between the two is 2 (only 2 replacement operations are needed to make them completely identical).

[0217] Fine-grained structure mobility calculation: According to the formula (fine-grained structure mobility = 1 - (edit distance / (A_SCF length + B_SCF length))), substituting the data, we can get: fine-grained structure mobility = 1 - (2 / (20 + 20)) = 1 - 0.05 = 0.95.

[0218] (2) Calculation of citation density and style transfer features

[0219] Using the same method as for calculating fine-grained structural transfer features, A_IMRDC and B_IMRDC are first preprocessed with common substring replacement (with the threshold also set to 3), and then the edit distance is calculated. Assuming that the lengths of A_IMRDC and B_IMRDC are both 30 after replacement, and the edit distance is 3, then the citation density and style transfer degree are equal to 1 - (3 / (30+30)) = 1 - 0.05 = 0.95.

[0220] (3) Semantic content transfer feature calculation

[0221] Determining the transferability of triplet relationships under the same functional role classification: Setting a similarity threshold α=0.8, taking LG_A1 (M1 classification, experimental subjects) and LG_B1 (M1 classification, experimental subjects) of papers A and B as examples, we selected the triplet G_A1={“SPF-grade Kunming mice”, “as”, “experimental subjects”} in LG_A1, and the triplet G_B1={“SPF-grade ICR mice”, “as”, “experimental subjects”} in LG_B1. The similarity between the relationships “as” and “as” was calculated to be 1.0 (greater than the threshold of 0.8). Through medical knowledge graph verification, both “SPF-grade Kunming mice” and “SPF-grade ICR mice” belong to “SPF-grade experimental mice”, which are the same type of entity. “Experimental subjects” and “experimental subjects” are also the same entity. Therefore, we determined that there is a transferable relationship between G_A1 and G_B1 (T=1).

[0222] Semantic transferability calculation within the same category: Count the number of transferable relationships between each triple in LG_A1 and LG_B1. Assuming LG_A1 contains 5 triples, and each triple can be transferred to another triple in LG_B1, then MatchTriples(A_LG1,B_LG1)=5; similarly, LG_B1 contains 5 triples, and each triple can be transferred to another triple in LG_A1, so MatchTriples(B_LG1,A_LG1)=5. The total number of triples in LG_A1 and LG_B1 is 5+5=10. Therefore, the semantic transferability under this category = (5+5) / 10 = 1.0. Using the same method, the semantic transferability under other functional role categories was calculated. For example, the semantic transferability between LG_A4 (M4 classification, detection method) and LG_B4 was 0.9, and the semantic transferability between LG_A10 (R1 classification, experimental design) and LG_B10 was 0.92.

[0223] Overall semantic transferability calculation: Multiply the semantic transferability of each functional role category by the corresponding weight coefficient, and then sum them. For example, the semantic transferability of category M1 is 1.0 × weight 0.10 = 0.10, the semantic transferability of category M4 is 0.9 × weight 0.10 = 0.09, and the semantic transferability of category R1 is 0.92 × weight 0.08 = 0.0736. Other categories are calculated in the same way and then summed. Finally, the overall semantic transferability of papers A and B is 0.93.

[0224] IV. Factory-based paper detection and judgment

[0225] Adding the fine-grained structure transfer (0.95), citation density and style transfer (0.95), and semantic transfer (0.93), we get the transfer feature value = 0.95 + 0.95 + 0.93 = 2.83. Based on experimental calibration in the medical field, a threshold θ = 2.5 is set. Since the transfer feature value 2.83 > 2.5, and the comparison paper B is a normal paper, it is suggested that the paper to be detected, A, may be a factory paper generated by imitating paper B.

[0226] As can be seen from the above Example 2, the method of the present invention can clearly and accurately realize the detection of factory-made papers, with strong operability of the steps and reliable detection results, fully demonstrating the advanced nature and practicality of the method.

[0227] Based on the same inventive concept, this application also provides a migration feature-based factory paper detection system for implementing the above-described migration feature-based factory paper detection method. The solution provided by this system is similar to the implementation scheme described in the above-described embodiments. Therefore, the specific limitations of one or more migration feature-based factory paper detection system embodiments provided below can be found in the limitations of the migration feature-based factory paper detection method described above, and will not be repeated here.

[0228] In one embodiment, the present invention provides a factory paper detection system based on migration features, such as... Figure 3 As shown, it includes: a feature extraction module 210, a calculation module 220, a decision module 230, and a result display module 240, wherein:

[0229] The feature extraction module 210 is used to extract fine-grained features of structure, citation density and style, and semantic content from the paper to be detected and the comparison paper, respectively.

[0230] Calculation module 220 is used to calculate the migration feature values ​​between the paper to be detected and the comparison paper based on the extracted features;

[0231] The determination module 230 is used to compare the migration feature value with the preset threshold, and determine whether the paper to be detected is a factory paper based on the comparison result and the attributes of the paper to be compared.

[0232] The result prompt module 240 is used to generate a first prompt message if the migration feature value is greater than a preset threshold and the comparison paper is a normal paper; and to generate a second prompt message if the migration feature value is greater than the preset threshold and the comparison paper is a factory paper.

[0233] In one embodiment, an electronic device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown. The electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements the migration-feature-based factory paper detection method described in any one of steps S101 to S104. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0234] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0235] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0236] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0237] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0238] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0239] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A factory paper detection method based on transfer features, characterized in that, The method includes: Fine-grained feature extraction of structure, citation density and style, and semantic content is performed on the papers to be tested and the comparison papers respectively. Based on the extracted features, the transfer feature values ​​between the paper to be detected and the comparison paper are calculated. Compare the migration feature value with a preset threshold, and determine whether the paper to be detected is a factory paper based on the comparison result and the attributes of the paper being compared. If the migration feature value is greater than a preset threshold and the comparison paper is a normal paper, a first prompt message is generated; if the migration feature value is greater than a preset threshold and the comparison paper is a factory paper, a second prompt message is generated. The fine-grained feature extraction of structure, citation density and style, and semantic content for the paper to be detected and the comparison paper includes: The sentences in the paper are classified according to the preset functional roles; adjacent functional role classification labels with the same functional role in the sentence functional role classification results are merged to obtain a merged classification sequence; using a single character of the non-classification label as a delimiter, the merged classification sequence is converted into a symbol string and defined as a fine-grained structural feature; The system identifies the citations in the introduction, methods, results, and discussion sections of the paper, categorizes the citations according to a preset style, and concatenates the corresponding tags according to the order of their appearance in the text to obtain citation density and style characteristics. Semantic content represented by entity relation triples is extracted from the paper, and the triples are grouped according to the functional role classification results of the sentences to obtain a set of triples grouped by functional role; each set of triples corresponds to a functional role classification label of a sentence and the weight coefficient corresponding to the label, thus obtaining semantic content features.

2. The method according to claim 1, characterized in that, The process of classifying the paper sentences according to preset functional roles includes: segmenting the entire paper based on predefined fine-grained structural classification rules, determining the start and end positions of each part of the IMRD; labeling each sentence in the segmented paper to obtain the functional role classification labels corresponding to each sentence, forming a sentence classification label sequence; The fine-grained structural classification rules include: the IMRD classification table and the sentence functional role classification table; The IMRD classification table is used to divide a paper into four parts: introduction, methods, results, and discussion. Each part corresponds to an IMRD classification label, namely introduction (I), methods (M), results (R), and discussion (D). The sentence functional role classification table is used to classify sentences into multiple functional roles based on their function in the paper. Each functional role corresponds to a functional role classification label and a weight coefficient. The functional role classification labels are represented by symbols of equal length. The prefix letters of the functional role classification labels are consistent with the IMRD classification labels.

3. The method according to claim 1, characterized in that, The process involves identifying citations in the introduction, methods, results, and discussion sections of the paper, categorizing the citations according to a preset style, and then concatenating the corresponding tags based on their order of appearance in the text. Based on the location information of each part of the paper's IMRD, sentences containing citations are extracted from the four parts of the paper: introduction, methods, results, and discussion, forming a set of citations for each part; Based on a predefined citation content classification table, each citation in the citation content set of each part is classified, and the corresponding classification label is determined to form a classification label sequence for the citation content of each part. According to the IMRD structure, the IMRD category tags of each part are merged with the corresponding reference content category tags, and then the merged category tags are concatenated.

4. The method according to claim 1, characterized in that, The process of grouping triples based on sentence function role classification results to obtain a set of triples grouped by function role includes: extracting the research entities contained in each sentence of the paper and the relationships between the entities, and constructing a set of entity-relation triples for each sentence; each triple consists of a first entity, a relation, and a last entity; Based on the functional role classification labels of sentences, the entity relation triples corresponding to sentences with the same functional role classification labels are grouped together to form a set of triples under each functional role classification.

5. The method according to claim 1, characterized in that, The calculation of the transfer feature value between the paper to be detected and the comparison paper based on the extracted features includes: calculating the fine-grained structural transfer degree, citation density and style transfer degree, and semantic content transfer degree between the paper to be detected and the comparison paper based on the fine-grained structural features, citation density and style features, and semantic content features, and then adding the transfer degrees together to obtain the transfer feature value.

6. The method according to claim 5, characterized in that, The calculation of fine-grained structural transfer, citation density and style transfer, and semantic content transfer between the paper to be detected and the comparison paper, based on their fine-grained structural features, citation density and style transfer, and semantic content transfer, includes: preprocessing the fine-grained structural features of the paper to be detected and the comparison paper by replacing common substrings with lengths greater than a preset threshold in the character sequences of the two fine-grained structural features with single characters that do not repeat existing characters; using the Levenstein distance calculation method to calculate the edit distance between the two fine-grained structural features after preprocessing, which is the minimum number of editing operations required to convert the fine-grained structural features of the paper to be detected into the fine-grained structural features of the comparison paper; and calculating the fine-grained structural transfer based on the edit distance and the lengths of the two fine-grained structural features. A method for calculating fine-grained structural mobility is used to obtain citation density and style mobility. For two sets of triples under the same sentence functional role classification in the paper to be tested and the comparison paper, the transferability relationship between the triples in the set is determined, and the semantic transferability under the same functional role classification is calculated.

7. The method according to claim 6, characterized in that, The process of determining the transferability relationship between the triples in the same sentence functional role classification in the test paper and the comparison paper, and calculating the semantic transferability under the same functional role classification, includes: Determine whether the similarity between the triple G1 group of the paper to be detected and the triple G2 group of the comparison paper is greater than a preset threshold, and whether the first entity of G1 and the first entity of G2, and the last entity of G1 and the last entity of G2 belong to the same type of entity or have the same superordinate class; wherein, the relationship similarity is obtained by querying a preset relationship similarity table or semantic calculation, and the entity of the same type or superordinate class relationship is obtained by verifying through knowledge graph and semantic dictionary; The number of transferable relationships between each triplet in the set of triplets in the papers to be detected and the set of triplets in the comparison papers is counted; where a single triplet in the paper to be detected corresponds to at least one triplet in the comparison papers, it is counted as 1 match. Record the number of matching triples between the paper to be detected and the comparison paper, and the number of matching triples between the comparison paper and the paper to be detected; Based on the number of matching triples between the paper to be detected and the comparison paper and the number of triples in the triple set, the semantic transfer degree under the functional role classification is calculated. Based on the semantic transferability under the functional role classification and the weight coefficient corresponding to the functional role classification, the semantic content transferability between the paper to be detected and the comparison paper is calculated.

8. A factory paper detection system based on migration features, characterized in that, The system includes: The feature extraction module is used to extract fine-grained features of structure, citation density and style, and semantic content from the paper to be detected and the comparison paper, respectively. The calculation module is used to calculate the transfer feature values ​​between the paper to be detected and the comparison paper based on the extracted features. The determination module is used to compare the migration feature value with a preset threshold, and determine whether the paper to be detected is a factory paper based on the comparison result and the attributes of the paper to be compared. The result prompting module is used to generate a first prompt message if the migration feature value is greater than a preset threshold and the comparison paper is a normal paper; and to generate a second prompt message if the migration feature value is greater than the preset threshold and the comparison paper is a factory paper. The fine-grained feature extraction of structure, citation density and style, and semantic content for the paper to be detected and the comparison paper includes: The sentences in the paper are classified according to the preset functional roles; adjacent functional role classification labels with the same functional role in the sentence functional role classification results are merged to obtain a merged classification sequence; using a single character of the non-classification label as a delimiter, the merged classification sequence is converted into a symbol string and defined as a fine-grained structural feature; The system identifies the citations in the introduction, methods, results, and discussion sections of the paper, categorizes the citations according to a preset style, and concatenates the corresponding tags according to the order of their appearance in the text to obtain citation density and style characteristics. Semantic content represented by entity relation triples is extracted from the paper, and the triples are grouped according to the functional role classification results of the sentences to obtain a set of triples grouped by functional role; each set of triples corresponds to a functional role classification label of a sentence and the weight coefficient corresponding to the label, thus obtaining semantic content features.

Images