Paper detection method and device, storage medium and terminal

Abstract

The invention discloses a paper detection method and device, a storage medium and a terminal, and belongs to the technical field of text processing. Inputting the to-be-detected paper into a pre-trained paper detection large language model; extracting multi-dimensional features of a to-be-detected paper through a paper detection large language model; generating a first detection result based on the multi-dimensional features through a first analysis module in a thesis detection large language model; and comparing the features of the to-be-detected paper with paper features in the comparison paper library through a second analysis module in the paper detection large language model to generate a second detection result. According to the paper detection method and device, the storage medium and the terminal provided by the invention, AI generation detection and plagiarism identification are integrated, and the academic integrity guarantee capability is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of text processing technology, and in particular to a paper detection method and apparatus, storage medium, and terminal. Background Technology

[0002] With the rapid development of artificial intelligence technology, especially the breakthroughs achieved by large language models represented by the GPT series in text generation, the quality of generated text is increasingly approaching or even surpassing human levels, and is widely used in academic writing, news reporting, creative writing, and other fields. However, this technological advancement has also brought new challenges to academic integrity: on the one hand, AI-generated academic papers may have potential semantic inconsistencies, style templates, and logical jumps, affecting academic quality and credibility; on the other hand, traditional plagiarism detection methods mostly rely on surface character matching or simple semantic similarity calculations, making it difficult to effectively identify new forms of academic misconduct such as AI rewriting, multi-source splicing, or cross-language plagiarism. Existing detection tools have significant shortcomings in accuracy, interpretability, and multi-dimensional detection capabilities, urgently requiring an intelligent paper detection method that integrates deep semantic analysis, style detection, and quality assessment to meet the academic integrity protection needs of the AI ​​era. Summary of the Invention

[0003] The purpose of this invention is to provide a method, apparatus, storage medium, and terminal for plagiarism detection, in order to solve the problems existing in the background art.

[0004] To achieve the above objectives, the present invention provides a paper detection method, comprising the following steps: Obtain the paper to be checked; Input the paper to be detected into a pre-trained large language model for paper detection; The paper detection big language model extracts multi-dimensional features of the paper to be detected, including deep semantic features, text statistical features and quality assessment features. Deep semantic features represent the semantic information and consistency of the paper at the word, sentence and paragraph levels, while text statistical features represent the style patterns of the paper at the lexical and syntactic levels. The first analysis module in the large language model for paper detection generates a first detection result based on multi-dimensional features. The first detection result is the judgment result that characterizes whether the paper to be detected is an artificial intelligence-generated paper. The second analysis module in the large language model for paper detection compares the features of the paper to be detected with the features of papers in the comparison paper database to generate a second detection result. The second detection result is a judgment result that characterizes whether the paper to be detected has plagiarism and / or a similarity metric value with a specific comparison paper.

[0005] Preferably, the first detection result is generated through the first analysis module, including: S101. Perform multi-level fusion processing on deep semantic features, text statistical features, and quality assessment features, including: Cross-channel feature alignment uses a learnable linear transformation layer to map deep semantic features, textual statistical features, and quality assessment features from different feature channels to a unified feature space dimension, resulting in aligned semantic feature vectors. Style feature vector and quality eigenvectors ; Dynamic attention weighting is used to construct a multi-head attention mechanism and calculate semantic feature vectors. Style feature vector and quality eigenvectors The mutual attention weight matrix between them; through the mutual attention mechanism, semantic features can dynamically pay attention to style anomaly patterns and quality anomaly indicators, while style and quality features can capture anomalous expressions at the semantic level. Gated feature fusion employs a gating mechanism to control the contribution of different features to the final fused feature. The fused feature is calculated using the following formula. : ; in, , , The gating weight vector is calculated by the sigmoid function, with a value range of [0,1], which controls the fusion ratio of semantic, style and quality features respectively; S102, Based on fusion features Identify typical AI-generated anomaly patterns in academic papers: Semantic consistency analysis extracts inter-sentence semantic consistency matrix features from deep semantic features, calculates their entropy and variance, and quantifies the degree of semantic jump; a semantic consistency threshold is set, and when the semantic consistency is lower than the preset threshold, it is marked as a suspicious feature generated by AI. Language style anomaly detection involves extracting template word usage frequency, sentence complexity distribution, and connector density from text statistical features, comparing them with a baseline model of language style in human-written papers, and calculating the style deviation. Anomaly assessment of quality indicators involves extracting text diversity and text predictability indicators from the quality assessment features. When text diversity is below the first threshold and text predictability is above the second threshold, it is marked as a quality anomaly feature. S103. The final judgment result is generated using a hierarchical classification mechanism: The primary classifier will fuse features. Input a fully connected neural network and output a primary classification probability. , which indicates the probability that the paper was generated by AI; Anomaly feature enhancement converts the semantic consistency anomaly markers, style anomaly markers, and quality anomaly markers identified in step S102 into binary feature vectors. and fusion features Concatenate to form an enhanced feature vector ; The final decision-making layer will strengthen the feature vectors. Input the decision tree classifier and combine it with the initial classification probability. The final judgment result is generated according to the following rules: like Threshold ,and If a paper contains at least two anomalous flags, it is determined to be an "AI-generated paper"; like Threshold ,and If a paper contains at least one anomalous flag, it is considered a "suspected AI-generated paper". like ≤threshold ,or If no abnormal markers are found, it is determined to be a "human-written paper"; Where the threshold > And threshold , Obtained through model training; S104. Calculate the confidence score of the judgment result: Feature consistency score assesses the degree of consistency between semantic features, style features and quality features in AI generation judgment. A high consistency score is given when all three types of features support the same judgment. Model confidence is calculated based on the purity of the leaf nodes and the sample distribution of the decision tree classifier. The final confidence score is fused by weighting the feature consistency score and the model confidence score, and the final confidence score in the range of [0,1] is output along with the first detection result. S105. Generate an interpretability report of the judgment results: Key feature extraction identifies the feature dimensions that have the greatest impact on classification decisions, including specific semantically inconsistent segments, high-frequency template words, and abnormal quality indicators; Comparative analysis compares the features of the paper to be detected with typical features in a database of human-written paper features to highlight the differences. Natural language interpretation, based on template filling technology, transforms the identified abnormal features into natural language descriptions, generating highly readable explanations of the judgment reasons.

[0006] Preferably, the second detection result generated by the second analysis module is as follows: S201, Feature Enhancement and Candidate Set Construction: Processing the paper to be detected and the comparison paper database, including multi-granularity word frequency vector construction, semantic enhancement vector generation, and dynamic candidate set screening; S202, Multi-level similarity calculation and fusion; Refining the candidate set Each candidate paper The calculations include detecting sentence-level similarity and calculating paragraph-level similarity. Sentence-level similarity detection: A sliding window matching algorithm is used to compare the sentence-level augmented vector sequence of the paper to be detected with... Align the sentence-level augmented vector sequences; For each sentence pair Calculate their weighted similarity: ; in , , Adjustable weights; Identify highly similar sentence pairs ( Record its location and similarity value; Paragraph-level similarity calculation: Based on sentence-level matching results, adjacent similar sentences are clustered into matching paragraph blocks; For each matching paragraph block, calculate the similarity of its paragraph-level vectors: ; Considering paragraph length weighting, longer matching paragraphs are assigned a higher similarity contribution.

[0007] S203, Comprehensive Judgment; Plagiarism detection: If it matches a paragraph block If a consecutive match exceeds 50 words, it is marked as a suspected plagiarized segment; For each suspected plagiarized passage, calculate its confidence score: ; in The number of words in the matched segment.

[0008] Plagiarism determination: Count the total number of matching words for all suspected plagiarized segments. Total number of words in the papers to be checked Ratio: ; Set dynamic judgment threshold Its value is adjusted according to the paper type and length; like If so, it is determined that there is plagiarism. like If it is found to be plagiarism, it will be judged as suspected plagiarism and requires manual review. like If so, it is determined that there is no plagiarism. Similarity metric calculation: Overall Similarity: Calculate the combined similarity between the paper to be detected and all candidate papers. ; in The weight of the k-th matching paragraph block; Highest source similarity: Identify the single source paper with the highest similarity to the paper to be detected, and record its similarity value. ; Similarity distribution: Calculate the similarity distribution with each candidate paper to identify whether there is multi-source splicing; Finally, a detection report is generated based on the judgment results and similarity.

[0009] Preferably, in S201, the following processing is performed on the paper to be detected and the comparison database: Multi-granularity word frequency vector construction: The papers to be checked are cleaned and structured, and after removing format information, word frequency vectors are constructed at three granularities: sentence level, paragraph level, and chapter level. Sentence-level vectors are generated by counting the frequency of words in each sentence. ; Paragraph-level vectors are generated by statistically analyzing the word frequencies in each paragraph, combining this with TF-IDF weighting, highlighting the core vocabulary of each paragraph, and forming vectors. ; Chapter-level vectors are generated by statistically analyzing the word distribution of each chapter (e.g., Introduction, Methods, Results, Discussion). ; By comparing the historical papers in the paper database, word frequency vectors of the above three granularities are constructed simultaneously, and a multi-level inverted index is established, in which keywords are associated with specific sentence, paragraph and chapter positions.

[0010] Semantic augmentation vector generation: Based on the word-level encoded vector sequence in deep semantic features, a context-sensitive word vector is calculated for each word. The word vectors are weighted by position to generate sentence-level semantic vectors. ; Aggregate sentence vectors to generate paragraph-level semantic vectors. and chapter-level semantic vectors ; The word frequency vector and the semantic vector are concatenated to form an enhanced feature vector. ; Dynamic candidate set filtering: Calculate the chapter-level enhancement vector of the paper to be detected The cosine similarity between the vectors of corresponding chapters of each paper in the comparison paper database and the vectors of the corresponding chapters. Set the first filter threshold Retain those with similarity greater than The paper chapters were used to construct an initial candidate set. ; For candidate set The similarity of paragraph-level augmented vectors in the papers is further calculated, and a second screening threshold is set. To form a refined candidate set .

[0011] The preferred content extracted using a large language model for paper detection is as follows: Text hierarchical parsing and initialization establish a five-level hierarchical structure of "document → chapter → paragraph → sentence → word" for structured input paper text. Word-level context encoding is performed through a pre-trained language model encoder, and word representation is enhanced by combining part-of-speech tagging information to form a word-level encoding vector with both syntactic and semantic information.

[0012] Intra-sentence semantic modeling and aggregation are performed using graph attention networks to model syntactic dependencies within sentences and capture structural semantic relationships between lexical units. A hierarchical attention mechanism is employed to calculate the contribution weight of each lexical unit to the sentence semantics, and a weighted aggregation is used to generate a sentence vector that integrates lexical information and syntactic structural features.

[0013] Inter-sentence semantic consistency modeling uses a sequence encoder to contextually encode sentence vectors within a paragraph, constructing an inter-sentence semantic similarity matrix. By calculating the statistical features of the matrix (average similarity, variance, diagonal offset), the semantic coherence within the paragraph is quantified, and anomalous patterns such as semantic jumps are detected.

[0014] The document-level semantic aggregation and feature output employs a hierarchical attention network to generate overall document vectors and chapter vectors from sentence vectors, analyzing cross-chapter semantic relationships and logical flows. The final output includes a complete semantic feature system at the word, sentence, paragraph, and document levels, along with metadata such as semantic consistency anomaly markers.

[0015] Optimization and interpretability enhancement are achieved by integrating semantic representations from different levels through multi-scale feature fusion technology to form a unified deep semantic feature vector. Interpretable functions such as key semantic unit identification, semantic anomaly localization, and semantic flow visualization are provided, while a feature quality assessment mechanism is established to ensure the reliability and effectiveness of extracted features.

[0016] Preferably, text statistical features are extracted using a large language model for paper detection, including: The usage patterns of the papers to be checked at the lexical level are statistically analyzed to obtain lexical statistical features, including the frequency of template word usage, distribution of high-frequency words, density of conjunction words, proportion of stop words, and number of unique professional terms. The structural patterns of the papers to be tested at the sentence level are statistically analyzed to obtain sentence statistical features, including average sentence length and variance, sentence complexity distribution, and template sentence usage.

[0017] A plagiarism detection device, comprising: The acquisition module is used to acquire the paper to be checked; The large language model processing module integrates a pre-trained large language model for paper detection, which is used to receive papers to be detected and extract multi-dimensional features of the papers to be detected. The multi-dimensional features include at least deep semantic features and text statistical features. The AI-generated detection module, as the first analysis module in the large language model processing module, is used to generate the first detection result based on multi-dimensional features. The first detection result is the judgment result that characterizes whether the paper to be detected is an AI-generated paper. The plagiarism detection module, as the second analysis module in the large language model processing module, is used to compare the features of the paper to be detected with the comparison paper database to generate a second detection result. The second detection result is the judgment result and / or similarity quantification value characterizing whether the paper to be detected has plagiarism behavior. The output module is used to output the first and second detection results.

[0018] A storage medium storing at least one executable instruction that causes a processor to perform an operation corresponding to the above-described paper detection method.

[0019] A terminal includes: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform operations corresponding to the above-described paper detection method.

[0020] Therefore, the present invention, employing the above-mentioned paper detection method, apparatus, storage medium, and terminal, has the following beneficial effects: (1) By integrating deep semantic features, text statistical features and quality assessment features, and combining cross-channel alignment, dynamic attention mechanism and gating fusion strategy, the system can comprehensively capture the semantic inconsistencies, style anomalies and quality defects of AI-generated text, and significantly improve the accuracy and robustness of detection. (2) The hierarchical classification mechanism combined with abnormal feature reinforcement is adopted. While outputting the judgment result, confidence score and interpretability report are provided to help users understand the judgment basis, improve the transparency and credibility of the system, and are suitable for scenarios with high requirements for the interpretability of results, such as education and publishing. (3) By calculating similarity at the sentence and paragraph levels and combining it with a dynamic threshold adjustment mechanism, the system can effectively detect complex plagiarism behaviors such as local plagiarism and multi-source splicing, while reducing the false alarm rate and adapting to the detection needs of papers of different disciplines and lengths. (4) Integrating AI-generated detection and plagiarism recognition into the same large language model framework enables feature sharing and process collaboration, greatly improving detection efficiency; it supports storage media and terminal deployment, has good system compatibility and scalability, and is suitable for various scenarios such as cloud services and localized applications.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of a paper detection method according to the present invention; Figure 2 This is a schematic diagram of the process by which the first analysis module generates the first detection result in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process by which the second analysis module generates the second detection result in an embodiment of the present invention; Figure 4 This is a block diagram of the paper detection device according to an embodiment of the present invention. Detailed Implementation

[0023] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] Please see Figures 1-4 A paper plagiarism detection method includes the following steps: Obtain the paper to be checked.

[0025] The paper to be detected is input into a pre-trained large language model for paper detection.

[0026] The paper detection big language model extracts multi-dimensional features of the paper to be detected, including deep semantic features, text statistical features and quality assessment features. Deep semantic features represent the semantic information and consistency of the paper at the word, sentence and paragraph levels, while text statistical features represent the style patterns of the paper at the lexical and syntactic levels.

[0027] The following content was extracted from the deep semantic features using a large language model for paper detection: Text hierarchical parsing and initialization establish a five-level hierarchical structure of "document → chapter → paragraph → sentence → word" for structured input paper text. Word-level context encoding is performed through a pre-trained language model encoder, and word representation is enhanced by combining part-of-speech tagging information to form a word-level encoding vector with both syntactic and semantic information.

[0028] Intra-sentence semantic modeling and aggregation are performed using graph attention networks to model syntactic dependencies within sentences and capture structural semantic relationships between lexical units. A hierarchical attention mechanism is employed to calculate the contribution weight of each lexical unit to the sentence semantics, and a weighted aggregation is used to generate a sentence vector that integrates lexical information and syntactic structural features.

[0029] Inter-sentence semantic consistency modeling uses a sequence encoder to contextually encode sentence vectors within a paragraph, constructing an inter-sentence semantic similarity matrix. By calculating the statistical features of the matrix (average similarity, variance, diagonal offset), the semantic coherence within the paragraph is quantified, and anomalous patterns such as semantic jumps are detected.

[0030] The document-level semantic aggregation and feature output employs a hierarchical attention network to generate overall document vectors and chapter vectors from sentence vectors, analyzing cross-chapter semantic relationships and logical flows. The final output includes a complete semantic feature system at the word, sentence, paragraph, and document levels, along with metadata such as semantic consistency anomaly markers.

[0031] Optimization and interpretability enhancement are achieved by integrating semantic representations from different levels through multi-scale feature fusion technology to form a unified deep semantic feature vector. Interpretable functions such as key semantic unit identification, semantic anomaly localization, and semantic flow visualization are provided, while a feature quality assessment mechanism is established to ensure the reliability and effectiveness of extracted features.

[0032] Textual statistical features are extracted using a large language model for paper detection, including: The usage patterns of the papers to be checked at the lexical level are statistically analyzed to obtain lexical statistical features, including the frequency of template word usage, distribution of high-frequency words, density of conjunction words, proportion of stop words, and number of unique professional terms. The structural patterns of the papers to be tested at the sentence level are statistically analyzed to obtain sentence statistical features, including average sentence length and variance, sentence complexity distribution, and template sentence usage.

[0033] The first analysis module in the large language model for paper detection generates a first detection result based on multi-dimensional features. The first detection result is the judgment result that characterizes whether the paper to be detected is an artificial intelligence-generated paper.

[0034] The first detection result is generated through the first analysis module, including: S101. Perform multi-level fusion processing on deep semantic features, text statistical features, and quality assessment features, including: Cross-channel feature alignment uses a learnable linear transformation layer to map deep semantic features, textual statistical features, and quality assessment features from different feature channels to a unified feature space dimension, resulting in aligned semantic feature vectors. Style feature vector and quality eigenvectors ; Dynamic attention weighting is used to construct a multi-head attention mechanism and calculate semantic feature vectors. Style feature vector and quality eigenvectors The mutual attention weight matrix between them; through the mutual attention mechanism, semantic features can dynamically pay attention to style anomaly patterns and quality anomaly indicators, while style and quality features can capture anomalous expressions at the semantic level. Gated feature fusion employs a gating mechanism to control the contribution of different features to the final fused feature. The fused feature is calculated using the following formula. : ; in, , , The gating weight vector is calculated by the sigmoid function, with a value range of [0,1], which controls the fusion ratio of semantic, style and quality features respectively; S102, Based on fusion features Identify typical AI-generated anomaly patterns in academic papers: Semantic consistency analysis extracts inter-sentence semantic consistency matrix features from deep semantic features, calculates their entropy and variance, and quantifies the degree of semantic jump; a semantic consistency threshold is set, and when the semantic consistency is lower than the preset threshold, it is marked as a suspicious feature generated by AI. Language style anomaly detection involves extracting template word usage frequency, sentence complexity distribution, and connector density from text statistical features, comparing them with a baseline model of language style in human-written papers, and calculating the style deviation. Anomaly assessment of quality indicators involves extracting text diversity and text predictability indicators from the quality assessment features. When text diversity is below the first threshold and text predictability is above the second threshold, it is marked as a quality anomaly feature. S103. The final judgment result is generated using a hierarchical classification mechanism: The primary classifier will fuse features. Input a fully connected neural network and output a primary classification probability. , which indicates the probability that the paper was generated by AI; Anomaly feature enhancement converts the semantic consistency anomaly markers, style anomaly markers, and quality anomaly markers identified in step S102 into binary feature vectors. and fusion features Concatenate to form an enhanced feature vector ; The final decision-making layer will strengthen the feature vectors. Input the decision tree classifier and combine it with the initial classification probability. The final judgment result is generated according to the following rules: like Threshold ,and If a paper contains at least two anomalous flags, it is determined to be an "AI-generated paper"; like Threshold ,and If a paper contains at least one anomalous flag, it is considered a "suspected AI-generated paper". like ≤threshold ,or If no abnormal markers are found, it is determined to be a "human-written paper"; Where the threshold > And threshold , Obtained through model training; S104. Calculate the confidence score of the judgment result: Feature consistency score assesses the degree of consistency between semantic features, style features and quality features in AI generation judgment. A high consistency score is given when all three types of features support the same judgment. Model confidence is calculated based on the purity of the leaf nodes and the sample distribution of the decision tree classifier. The final confidence score is fused by weighting the feature consistency score and the model confidence score, and the final confidence score in the range of [0,1] is output along with the first detection result. S105. Generate an interpretability report of the judgment results: Key feature extraction identifies the feature dimensions that have the greatest impact on classification decisions, including specific semantically inconsistent segments, high-frequency template words, and abnormal quality indicators; Comparative analysis compares the features of the paper to be detected with typical features in a database of human-written paper features to highlight the differences. Natural language interpretation, based on template filling technology, transforms the identified abnormal features into natural language descriptions, generating highly readable explanations of the judgment reasons.

[0035] The second analysis module in the large language model for paper detection compares the features of the paper to be detected with the features of papers in the comparison paper database to generate a second detection result. The second detection result is a judgment result that characterizes whether the paper to be detected has plagiarism and / or a similarity metric value with a specific comparison paper.

[0036] The second analysis module generates the following second detection result: S201, Feature Enhancement and Candidate Set Construction: Processing the paper to be detected and the comparison paper database, including multi-granularity word frequency vector construction, semantic enhancement vector generation, and dynamic candidate set screening.

[0037] Multi-granularity word frequency vector construction: The papers to be checked are cleaned and structured, and after removing format information, word frequency vectors are constructed at three granularities: sentence level, paragraph level, and chapter level. Sentence-level vectors are generated by counting the frequency of words in each sentence. ; Paragraph-level vectors are generated by statistically analyzing the word frequencies in each paragraph, combining this with TF-IDF weighting, highlighting the core vocabulary of each paragraph, and forming vectors. ; Chapter-level vectors are generated by statistically analyzing the word distribution of each chapter (e.g., Introduction, Methods, Results, Discussion). ; By comparing the historical papers in the paper database, word frequency vectors of the above three granularities are constructed simultaneously, and a multi-level inverted index is established, in which keywords are associated with specific sentence, paragraph and chapter positions.

[0038] Semantic augmentation vector generation: Based on the word-level encoded vector sequence in deep semantic features, a context-sensitive word vector is calculated for each word. The word vectors are weighted by position to generate sentence-level semantic vectors. ; Aggregate sentence vectors to generate paragraph-level semantic vectors. and chapter-level semantic vectors ; The word frequency vector and the semantic vector are concatenated to form an enhanced feature vector. ; Dynamic candidate set filtering: Calculate the chapter-level enhancement vector of the paper to be detected The cosine similarity between the vectors of corresponding chapters of each paper in the comparison paper database and the vectors of the corresponding chapters. Set the first filter threshold Retain those with similarity greater than The paper chapters were used to construct an initial candidate set. ; For candidate set The similarity of paragraph-level augmented vectors in the papers is further calculated, and a second screening threshold is set. To form a refined candidate set .

[0039] S202, Multi-level similarity calculation and fusion; Refining the candidate set Each candidate paper The calculations include detecting sentence-level similarity and calculating paragraph-level similarity. Sentence-level similarity detection: A sliding window matching algorithm is used to compare the sentence-level augmented vector sequence of the paper to be detected with... Align the sentence-level augmented vector sequences; For each sentence pair Calculate their weighted similarity: ; in , , Adjustable weights; Identify highly similar sentence pairs ( Record its location and similarity value; Paragraph-level similarity calculation: Based on sentence-level matching results, adjacent similar sentences are clustered into matching paragraph blocks; For each matching paragraph block, calculate the similarity of its paragraph-level vectors: ; Considering paragraph length weighting, longer matching paragraphs are assigned a higher similarity contribution.

[0040] S203, Comprehensive Judgment; Plagiarism detection: If it matches a paragraph block If a consecutive match exceeds 50 words, it is marked as a suspected plagiarized segment; For each suspected plagiarized passage, calculate its confidence score: ; in The number of words in the matched segment.

[0041] Plagiarism determination: Count the total number of matching words for all suspected plagiarized segments. Total number of words in the papers to be checked Ratio: ; Set dynamic judgment threshold Its value is adjusted according to the paper type and length; like If so, it is determined that there is plagiarism. like If it is found to be plagiarism, it will be judged as suspected plagiarism and requires manual review. like If so, it is determined that there is no plagiarism. Similarity metric calculation: Overall Similarity: Calculate the combined similarity between the paper to be detected and all candidate papers. ; in The weight of the k-th matching paragraph block; Highest source similarity: Identify the single source paper with the highest similarity to the paper to be detected, and record its similarity value. ; Similarity distribution: Calculate the similarity distribution with each candidate paper to identify whether there is multi-source splicing; Finally, a detection report is generated based on the judgment results and similarity.

[0042] A plagiarism detection device, comprising: The acquisition module is used to acquire the paper to be checked; The large language model processing module integrates a pre-trained large language model for paper detection, which is used to receive papers to be detected and extract multi-dimensional features of the papers to be detected. The multi-dimensional features include at least deep semantic features and text statistical features. The AI-generated detection module, as the first analysis module in the large language model processing module, is used to generate the first detection result based on multi-dimensional features. The first detection result is the judgment result that characterizes whether the paper to be detected is an AI-generated paper. The plagiarism detection module, as the second analysis module in the large language model processing module, is used to compare the features of the paper to be detected with the comparison paper database to generate a second detection result. The second detection result is the judgment result and / or similarity quantification value characterizing whether the paper to be detected has plagiarism behavior. The output module is used to output the first and second detection results.

[0043] Therefore, the present invention adopts the above-mentioned paper detection method, device, storage medium and terminal to provide a multi-dimensional, interpretable and highly efficient intelligent paper detection solution that integrates AI generation detection and plagiarism identification, significantly improving the ability to protect academic integrity.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting plagiarism, characterized in that, Includes the following steps: Obtain the paper to be checked; Input the paper to be detected into a pre-trained large language model for paper detection; The paper detection big language model extracts multi-dimensional features of the paper to be detected, including deep semantic features, text statistical features and quality assessment features. Deep semantic features represent the semantic information and consistency of the paper at the word, sentence and paragraph levels, while text statistical features represent the style patterns of the paper at the lexical and syntactic levels. The first analysis module in the large language model for paper detection generates a first detection result based on multi-dimensional features. The first detection result is the judgment result that characterizes whether the paper to be detected is an artificial intelligence-generated paper. The second analysis module in the large language model for paper detection compares the features of the paper to be detected with the features of papers in the comparison paper database to generate a second detection result. The second detection result is a judgment result that characterizes whether the paper to be detected has plagiarism and / or a similarity metric value with a specific comparison paper.

2. The paper detection method according to claim 1, characterized in that, The first detection result is generated through the first analysis module, including: S101. Perform multi-level fusion processing on deep semantic features, text statistical features, and quality assessment features, including: Cross-channel feature alignment uses a learnable linear transformation layer to map deep semantic features, textual statistical features, and quality assessment features from different feature channels to a unified feature space dimension, resulting in aligned semantic feature vectors. Style feature vector and quality eigenvectors ; Dynamic attention weighting is used to construct a multi-head attention mechanism and calculate semantic feature vectors. Style feature vector and quality eigenvectors Mutual attention weight matrix between them; Gated feature fusion employs a gating mechanism to control the contribution of different features to the final fused feature. The fused feature is calculated using the following formula. : ； in, , , The gating weight vector is calculated by the sigmoid function, with a value range of [0,1], which controls the fusion ratio of semantic, style and quality features respectively; S102, Based on fusion features Identify typical AI-generated anomaly patterns in academic papers: Semantic consistency analysis extracts inter-sentence semantic consistency matrix features from deep semantic features, calculates their entropy and variance, and quantifies the degree of semantic jump; a semantic consistency threshold is set, and when the semantic consistency is lower than the preset threshold, it is marked as a suspicious feature generated by AI. Language style anomaly detection involves extracting template word usage frequency, sentence complexity distribution, and connector density from text statistical features, comparing them with the language style baseline model of the written paper, and calculating the style deviation. Anomaly assessment of quality indicators involves extracting text diversity and text predictability indicators from the quality assessment features. When text diversity is below the first threshold and text predictability is above the second threshold, it is marked as a quality anomaly feature. S103. The final judgment result is generated using a hierarchical classification mechanism: The primary classifier will fuse features. Input a fully connected neural network and output a primary classification probability. , which indicates the probability that the paper was generated by AI; Anomaly feature enhancement converts the semantic consistency anomaly markers, style anomaly markers, and quality anomaly markers identified in step S102 into binary feature vectors. and fusion features Concatenate to form an enhanced feature vector ; The final decision-making layer will strengthen the feature vectors. Input the decision tree classifier and combine it with the initial classification probability. The final judgment result is generated according to the following rules: like Threshold ,and If a paper contains at least two anomalous flags, it is determined to be an AI-generated paper. like Threshold ,and If a paper contains at least one anomalous flag, it is considered a suspected AI-generated paper. like ≤threshold ,or If no abnormal markers are found, the paper is determined to be manually written. Where the threshold > And threshold , Obtained through model training; S104. Calculate the confidence score of the judgment result: Feature consistency score assesses the degree of consistency between semantic features, style features and quality features in AI generation judgment. A high consistency score is given when all three types of features support the same judgment. Model confidence is calculated based on the purity of the leaf nodes and the sample distribution of the decision tree classifier. The final confidence score is fused by weighting the feature consistency score and the model confidence score, and the final confidence score in the range of [0,1] is output along with the first detection result. S105. Generate an interpretability report of the judgment results: Key feature extraction identifies the feature dimensions that have the greatest impact on classification decisions, including specific semantically inconsistent segments, high-frequency template words, and abnormal quality indicators; Comparative analysis compares the features of the paper to be detected with typical features in a database of human-written paper features to highlight the differences. Natural language interpretation, based on template filling technology, transforms the identified abnormal features into natural language descriptions, generating highly readable explanations of the judgment reasons.

3. The paper detection method according to claim 2, characterized in that, The second detection result generated by the second analysis module is as follows: S201, Feature Enhancement and Candidate Set Construction: Processing the paper to be detected and the comparison paper database, including multi-granularity word frequency vector construction, semantic enhancement vector generation, and dynamic candidate set screening; S202, Multi-level similarity calculation and fusion; Refining the candidate set Each candidate paper Calculate sentence-level similarity and paragraph-level similarity; Sentence-level similarity detection: A sliding window matching algorithm is used to compare the sentence-level augmented vector sequence of the paper to be detected with... Align the sentence-level augmented vector sequences; For each sentence pair Calculate their weighted similarity: ； in , , Adjustable weights; Identify highly similar sentence pairs ( Record its location and similarity value; Paragraph-level similarity calculation: Based on sentence-level matching results, adjacent similar sentences are clustered into matching paragraph blocks; For each matching paragraph block, calculate the similarity of its paragraph-level vectors: ； Considering paragraph length weighting, longer matching paragraphs are given a higher similarity contribution; S203, Comprehensive Judgment; Plagiarism detection: If it matches a paragraph block If a consecutive match exceeds 50 words, it is marked as a suspected plagiarized segment; For each suspected plagiarized passage, calculate its confidence score: ； in The number of words in the matched segment; Plagiarism determination: Count the total number of matching words for all suspected plagiarized segments. Total number of words in the papers to be checked Ratio: ； Set dynamic judgment threshold Its value is adjusted according to the paper type and length; like If so, it is determined that there is plagiarism. like If it is found to be plagiarism, it will be judged as suspected plagiarism and requires manual review. like If so, it is determined that there is no plagiarism. Similarity metric calculation: Overall Similarity: Calculate the combined similarity between the paper to be detected and all candidate papers. ； in The weight of the k-th matching paragraph block; Highest source similarity: Identify the single source paper with the highest similarity to the paper to be detected, and record its similarity value. ; Similarity distribution: Calculate the similarity distribution with each candidate paper to identify whether there is multi-source splicing; Finally, a detection report is generated based on the judgment results and similarity.

4. The paper detection method according to claim 3, characterized in that, In S201, the following processing is performed on the paper to be detected and the comparison database: Multi-granularity word frequency vector construction: The papers to be checked are cleaned and structured, and after removing format information, word frequency vectors are constructed at three granularities: sentence level, paragraph level, and chapter level. Sentence-level vectors are generated by counting the frequency of words in each sentence. ; Paragraph-level vectors are generated by statistically analyzing the word frequencies in each paragraph, combining this with TF-IDF weighting, highlighting the core vocabulary of each paragraph, and forming vectors. ; Chapter-level vectors are generated by statistically analyzing the word distribution of each chapter. ; By comparing the historical papers in the paper database, word frequency vectors of the above three granularities are constructed simultaneously, and a multi-level inverted index is established, in which keywords are associated with specific sentence, paragraph and chapter positions; Semantic augmentation vector generation: Based on the word-level encoded vector sequence in deep semantic features, a context-sensitive word vector is calculated for each word. The word vectors are weighted by position to generate sentence-level semantic vectors. ; Aggregate sentence vectors to generate paragraph-level semantic vectors. and chapter-level semantic vectors ; The word frequency vector and the semantic vector are concatenated to form an enhanced feature vector. ; Dynamic candidate set filtering: Calculate the chapter-level enhancement vector of the paper to be detected The cosine similarity between the vectors of corresponding chapters of each paper in the comparison paper database and the vectors of the corresponding chapters. Set the first filter threshold Retain those with similarity greater than The paper chapters were used to construct an initial candidate set. ; For candidate set The similarity of paragraph-level augmented vectors in the papers is further calculated, and a second screening threshold is set. To form a refined candidate set .

5. The paper detection method according to claim 4, characterized in that, The following content was extracted from the deep semantic features using a large language model for paper detection: Text hierarchical parsing and initialization establish a five-level hierarchical structure for structured input paper text: document → chapter → paragraph → sentence → word. Word-level context encoding is performed through a pre-trained language model encoder, and word representation is enhanced by combining part-of-speech tagging information to form a word-level encoding vector with both syntactic and semantic information. Intra-sentence semantic modeling and aggregation: Based on graph attention network, we model the syntactic dependency relations within sentences and capture the structural semantic relationships between lexical units; A hierarchical attention mechanism is used to calculate the contribution weight of each word to the semantics of the sentence, and a sentence vector is generated by weighted aggregation. The sentence vector integrates lexical information and syntactic structure features. Inter-sentence semantic consistency modeling uses a sequence encoder to contextually encode sentence vectors within a paragraph, constructing an inter-sentence semantic similarity matrix; by calculating the statistical features of the matrix, the semantic coherence within the paragraph is quantified, and abnormal patterns are detected. The document-level semantic aggregation and feature output uses a hierarchical attention network to generate overall document vectors and chapter vectors from sentence vectors, and analyzes the semantic relationships and logical flow across chapters; the final output includes a complete semantic feature system at the word, sentence, paragraph, and document levels, along with metadata. Optimization and interpretability enhancement: By integrating semantic representations at different levels through multi-scale feature fusion technology, a unified deep semantic feature vector is formed; It provides functions for identifying key semantic units, locating semantic anomalies, and visualizing and interpreting semantic flows, while also establishing a feature quality assessment mechanism.

6. The paper detection method according to claim 5, characterized in that, Textual statistical features are extracted using a large language model for paper detection, including: The usage patterns of the papers to be checked at the lexical level are statistically analyzed to obtain lexical statistical features, including the frequency of template word usage, distribution of high-frequency words, density of conjunction words, proportion of stop words, and number of unique professional terms. The structural patterns of the papers to be tested at the sentence level are statistically analyzed to obtain sentence statistical features, including average sentence length and variance, sentence complexity distribution, and template sentence usage.

7. A plagiarism detection device, characterized in that, include: The acquisition module is used to acquire the paper to be checked; The large language model processing module integrates a pre-trained large language model for paper detection, which is used to receive papers to be detected and extract multi-dimensional features of the papers to be detected. The multi-dimensional features include at least deep semantic features and text statistical features. The AI-generated detection module, as the first analysis module in the large language model processing module, is used to generate the first detection result based on multi-dimensional features. The first detection result is the judgment result that characterizes whether the paper to be detected is an AI-generated paper. The plagiarism detection module, as the second analysis module in the large language model processing module, is used to compare the features of the paper to be detected with the comparison paper database to generate a second detection result. The second detection result is the judgment result and / or similarity quantification value characterizing whether the paper to be detected has plagiarism behavior. The output module is used to output the first and second detection results.

8. A storage medium storing at least one executable instruction that causes a processor to perform an operation corresponding to the paper detection method as described in any one of claims 1-6.

9. A terminal, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the paper detection method as described in any one of claims 1-6.

Images