A method and apparatus for academic misconduct document detection

Abstract

The present application belongs to the field of natural language processing, and particularly relates to a method and device for detecting academic misconduct literature, aiming at solving the problems of the prior art in recognizing and distinguishing the repetition of key and non-key content of text. The method comprises: obtaining query semantic features and layout features; performing semantic similarity retrieval on the query semantic features in a multi-dimensional semantic feature index library, and respectively weighting and counting and sorting the similar paragraph IDs and similar literature IDs in the retrieval results to obtain similar sorting results; calculating semantic feature similarity; determining layout similarity; and rearranging the similar sorting results based on the semantic feature similarity and / or the layout similarity. The present application improves the detection effect of semantic repetition after deep modification of text, distinguishes the semantic repetition of key content and non-key content, and improves the recall rate and accuracy rate by structurally extracting various features for feature retrieval and fusing the method to calculate the similarity.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, and specifically relates to a method and apparatus for detecting academic misconduct documents. Background Technology

[0002] Plagiarism is a typical manifestation of academic misconduct, taking many forms. These range from simple verbatim copying to superficial rewriting such as word substitution and sentence structure adjustments, and even deep rewriting incorporating methods like expansion, abbreviation, stylistic changes, and logical transformations. Currently, plagiarism detection in academic texts primarily relies on text content similarity retrieval. There are three main technical approaches to similarity retrieval: full-text inverted indexes based on characters, words, phrases, and N-grams; semantic hash indexes; and vector indexes based on representation learning. These three types of indexes are suitable for different business scenarios, each with its own advantages and disadvantages in terms of generation quality and retrieval performance. Due to the significant improvement in vector representation learning capabilities in recent years, vector indexes have a clear advantage in semantic-level recall. Vector indexes have gradually become the mainstream technology for content retrieval, or are being integrated with other traditional indexing methods to improve similarity retrieval performance.

[0003] Existing technologies for detecting academic misconduct, utilizing vector representation and retrieval techniques, can effectively identify plagiarism in academic papers, including plagiarism resulting from superficial semantic rewriting. However, significant shortcomings remain:

[0004] First, there is the issue of deep rewriting (abbreviation, expansion, paraphrasing, etc.) of large human or AI models (such as DeepSeek, ChatGPT, Wenxin Yiyan, etc.). Because the rewritten content changes greatly in terms of word choice, language expression, narrative style, and logical organization, even though the semantics are the same, existing detection software has difficulty identifying this type of plagiarism.

[0005] Secondly, the detection results have a certain error rate. When the language and text expression style is highly repetitive with non-critical content, but the key core objects are different, the text is easily misjudged as repetitive.

[0006] The fundamental reason is that current techniques based on vector representation or word / n-gram repetition retrieval cannot accurately express the complete semantics of text. For example, the vector representation capability decreases when the text is long, making it difficult for vector-based methods to fully express the semantics; when generating vectors by dividing sentences or fragments, the generated vectors cannot accurately represent the text semantics due to the large semantic span of the core key content. Furthermore, current methods, whether based on vector semantic representation or word / n-gram, cannot distinguish between the repetition of key and non-key content in repeated text.

[0007] In summary, existing academic misconduct detection technologies are insufficient in dealing with plagiarism caused by deep rewriting and in distinguishing between repetition of key and non-key content in texts, and urgently need improvement. Summary of the Invention

[0008] To address the aforementioned problems in existing technologies, namely, the inability of current academic misconduct detection technologies to identify deep rewriting plagiarism and to distinguish between duplicated key and non-key content in duplicated text, the first aspect of this invention proposes a method for detecting academic misconduct, the method comprising the following steps:

[0009] S100. Obtain the document to be analyzed, and use a pre-trained feature extraction model to extract features from the document to obtain query semantic features and layout features; the query semantic features include text semantic features, title keyword fusion semantic features and abstract semantic features.

[0010] S200. Based on the query semantic features, perform semantic similarity retrieval in a preset multi-dimensional semantic feature index library to obtain similar semantic features and corresponding similar paragraph IDs and similar document IDs; the multi-dimensional semantic feature index library includes various document sets and semantic feature sets, and any semantic feature in the semantic feature set is associated with the original document paragraph ID, original document ID and original text position in various document sets through the index.

[0011] S300: Perform weighted counting on the similar paragraph IDs and the similar document IDs respectively, and sort them in descending order based on the obtained count values ​​to obtain the similarity ranking result;

[0012] S400. Match the documents to be analyzed and the semantic features of the query with the similar document IDs and corresponding similar semantic features in the similarity ranking results to determine the semantic feature similarity.

[0013] S500. Based on the layout features, match them with the preset layout feature library to determine the layout similarity.

[0014] S600. The similarity ranking results are rearranged based on the semantic feature similarity and / or layout similarity to obtain the detection results.

[0015] In some preferred embodiments, a pre-trained feature extraction model is used to extract features from the document to be analyzed. The method is as follows:

[0016] The document to be analyzed is structured and extracted, including the title, abstract, main text paragraphs, figures and subfigures, tables, formulas, and references. The layout order and layout area of ​​each structural element are also recorded.

[0017] The extracted summary is divided into five parts: background, purpose, method, result, and conclusion, and each part is vectorized to obtain structured semantic features of the summary in different dimensions.

[0018] The extracted natural paragraphs of the main text are sorted and the content logic links are extracted. The content logic links include main logic link sentences and side logic link sentences. The main logic link sentences and side logic link sentences are vectorized to obtain the semantic features of the main text. The main logic link sentences, side logic link sentences and original paragraphs of the main text corresponding to the semantic features are recorded.

[0019] The extracted titles and keywords are vectorized separately, and then the semantic features of the title and keywords are calculated by weighting.

[0020] The extracted structural content is analyzed to obtain layout features centered on figures, tables, and formulas.

[0021] In some preferred embodiments, semantic similarity retrieval is performed based on query semantic features in a pre-defined multi-dimensional semantic feature index, and the method is as follows:

[0022] Based on the query semantic features, similarity retrieval is performed in the same type of semantic feature index in the multi-dimensional semantic feature index library to obtain several candidate similar semantic features for each query semantic feature; wherein, the types include main text semantic features, title keyword fusion semantic features and abstract semantic features, and similarity thresholds are set for the main logical link sentences and side logical link sentences in the main text semantic features respectively.

[0023] Select candidate similar semantic features with a similarity greater than a preset similarity threshold as similar semantic features;

[0024] Obtain the similar semantic features, the similarity of the similar semantic features, and the original document paragraph ID and original document ID pointed to by the vector index of the similar semantic features. Use the original document paragraph ID and original document ID as the corresponding similar paragraph ID and similar document ID.

[0025] In some preferred embodiments, the method for obtaining layout features centered on figures, tables, and formulas is as follows:

[0026] S110. Starting from the figure, table and formula respectively, perform a traversal of the preorder sequence and obtain the area of ​​the first preorder sequence that is the main text as the area of ​​the preorder text.

[0027] S120. Starting from the figure, table, and formula respectively, perform subsequent sequence traversal, and take the area of ​​the first subsequent sequence that is the main text as the area of ​​the subsequent text.

[0028] S130. Repeat steps S110-S120 until the layout feature sequence of all figures, tables, and formulas is obtained:

[0029] FigureID i ={ Figure , SubFigure , PreTxt , NextTxt};

[0030] TableID i ={ Table , CellNumber , PreTxt , NextTxt};

[0031] FormulaID i ={ Formula , LatexNum , PreTxt , NextTxt};

[0032] in, FigureID i For the first i The layout feature sequence of each image, Figure , SubFigure , PreTxt , NextTxt These represent the area of ​​the graph, the average area of ​​the subgraph, the area of ​​the preceding text, and the area of ​​the following text, respectively. TableID i For the first i The format feature sequence of each table, Table , CellNumber , PreTxt , NextTxt These represent the surface area, the number of table cells, the area of ​​the preceding text, and the area of ​​the following text, respectively. FormulaID i For the first i The format feature sequence of each table, Formula , LatexNum , PreTxt , NextTxt Represent the area and formula respectively. latex Number of commands, area of ​​preceding text, area of ​​following text.

[0033] In some preferred embodiments, a weighted count is performed on the similar paragraph IDs and the similar document IDs, respectively, using the following method:

[0034] Based on the type of query semantic features on which the similar semantic features are retrieved, different weights are assigned to similar paragraph IDs or similar document IDs;

[0035] The types include semantic features of the main text, semantic features of the title keywords, and semantic features of the abstract. The main logical link sentences and the side logical link sentences in the semantic features of the main text are assigned weights respectively.

[0036] The number of times each similar paragraph ID or similar document ID was recalled was calculated based on the weight.

[0037] In some preferred embodiments, the semantic feature similarity includes any one or more of the following: main logical link sentence similarity, side logical link sentence similarity, title keyword fusion semantic feature similarity, and abstract semantic feature similarity.

[0038] In some preferred embodiments, the method for determining layout similarity is as follows:

[0039] The layout features are tested using the difference ratio to determine whether the difference ratios in the four dimensions between the layout features of the document to be analyzed and the layout features of the same type of document to be compared simultaneously satisfy the preset parameter set K.

[0040] If the conditions are met, the document layout feature matching count corresponding to the current layout feature is increased by 1. However, the same figure, table, and formula features that meet the conditions are not counted cumulatively; otherwise, they are not counted.

[0041] The format similarity is obtained only after all format features of the documents to be analyzed have been tested.

[0042] In some preferred embodiments, the similarity ranking results are rearranged based on the semantic feature similarity, and the method is as follows:

[0043] The similarity ranking results are rearranged using semantic feature similarity as weight, and the repetition between paragraphs is taken into account; among them, the repetition weight of the main logical link sentence is greater than the repetition weight of the side logical link sentence, the repetition weight of the same paragraph is greater than the repetition weight of the cross paragraph, and the repetition weight of the continuous repetition is greater than the repetition weight of the scattered repetition.

[0044] In some preferred embodiments, if there is only an independent format feature library generated from the factory's collection of papers, after analyzing the papers to obtain the format features, the process jumps to step S500.

[0045] A second aspect of the present invention provides an apparatus for detecting academic misconduct documents, the apparatus comprising:

[0046] The index building module is configured to build a multi-dimensional semantic feature index library using a trained feature extraction model. The multi-dimensional semantic feature index library includes various document sets and semantic feature sets. Any semantic feature in the semantic feature set is associated with the original document paragraph ID, the original document ID, and the original text position in various document sets through the index.

[0047] The feature extraction module is configured to acquire the documents to be analyzed and use a trained feature extraction model to extract features from the documents to obtain query semantic features and layout features.

[0048] The semantic similarity retrieval module is configured to perform semantic similarity retrieval in a preset multi-dimensional semantic feature index based on the query semantic features, and obtain similar semantic features as well as the corresponding similar paragraph IDs and similar document IDs;

[0049] The semantic similarity calculation module is configured to match the semantic features of the document to be analyzed and the query with the similar document IDs and corresponding similar semantic features in the similarity ranking results to determine the semantic feature similarity.

[0050] The layout similarity calculation module is configured to match the layout features with a preset layout feature library to determine the layout similarity.

[0051] The result sorting module is configured to perform weighted counting on the similar paragraph IDs and the similar document IDs respectively, and sort them in descending order based on the obtained count values ​​to obtain similarity sorting results; it is also configured to rearrange the similarity sorting results based on the semantic feature similarity and / or layout similarity to obtain detection results.

[0052] The beneficial effects of this invention are:

[0053] (1) The present method extracts semantic features of the main text, semantic features of title keywords, semantic features of the abstract and layout features through structured extraction. Based on the retrieval results of different semantic features, it adopts multiple fusion methods to calculate similarity and perform semantic feature matching to obtain the original text content and perform comparison calculation together. It can improve the effect of semantic repetition detection after deep text modification by multiple methods such as abbreviation, expansion, fusion of narrative style and narrative logic rewriting, improve the recall rate, improve the accuracy of detection results, and distinguish between semantic repetition of key content and non-key content.

[0054] (2) Based on the existing literature text content, the text is decomposed by designing main logical links and side logical links. The main logical links and side logical links are transformed into their respective logical link sentences and then vectors are generated as semantic feature vectors. The main logical link sentences are used as inputs for extracting key semantic features, and the side logical link sentences are used as inputs for non-key semantic features. This can ensure the integrity of the semantic segmentation of the text content, and at the same time, it can also distinguish the key and non-key parts in the main text content.

[0055] (3) Based on the structural characteristics of academic literature and the importance of different parts, a weight-based mean vectorization representation method was designed for non-textual content according to the structural characteristics and importance of different parts of the literature, and the title keyword weights were integrated as independent semantic features; after the abstract was structured, semantic features of “background”, “purpose”, “method”, “result” and “conclusion” were established respectively to improve the semantic feature representation ability and refine the semantic feature representation granularity.

[0056] (4) A method for representing the layout structure features with figures, tables and formulas as the core content and a method for matching the layout structure of documents based on the method are proposed, which is a general auxiliary method for comparing the layout of various documents; and it can realize the rapid matching of factory papers. Even with only a small amount of factory paper data, the layout feature comparison of factory papers can still be realized. Attached Figure Description

[0057] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0058] Figure 1 This is a flowchart of a method for detecting academic misconduct documents in an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the structure of the semantic feature set in the multi-dimensional semantic feature index library in this embodiment of the invention;

[0060] Figure 3 This is a flowchart illustrating the extraction of features and the parsing of each structural element in an embodiment of the present invention. Detailed Implementation

[0061] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0062] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0063] This invention addresses the issue of inability to distinguish between key and non-key content in repeated text. It extracts logical chains from the main text, performs structured transformation on the abstract and title, and then performs vector transformation on each. It uses a pre-built multi-dimensional semantic feature index library for similarity retrieval and merges the matching results to achieve in-depth retrieval of the input documents and can determine whether the repetition is in the main logical chain or in the side logical chain.

[0064] To more clearly explain the method of this invention for improving the detection effect of academic misconduct documents, the following is combined with... Figure 1 The steps in the embodiments of the present invention will be described in detail below.

[0065] A method for detecting academic misconduct documents according to a first embodiment of the present invention includes steps S100-S600, each step of which is described in detail below:

[0066] S100. Obtain the document to be analyzed, and use a pre-trained feature extraction model to extract features from the document to obtain query semantic features and layout features; the query semantic features include text semantic features, title keyword fusion semantic features and abstract semantic features.

[0067] Preferably, a pre-trained feature extraction model is used to extract features from the document to be analyzed. The method is as follows:

[0068] The document to be analyzed is structured and extracted, including the title, abstract, main text paragraphs, figures and subfigures, tables, formulas, and references. The layout order and layout area of ​​each structural element are also recorded.

[0069] The extracted summary is divided into five parts: background, purpose, method, result, and conclusion, and each part is vectorized to obtain structured semantic features of the summary in different dimensions.

[0070] The extracted natural paragraphs of the main text are sorted and the content logic links are extracted. The content logic links include main logic link sentences and side logic link sentences. The main logic link sentences and side logic link sentences are vectorized to obtain the semantic features of the main text. The main logic link sentences, side logic link sentences and original paragraphs of the main text corresponding to the semantic features are recorded.

[0071] The extracted titles and keywords are vectorized separately, and then the semantic features of the title and keywords are calculated by weighting.

[0072] The extracted structural content is analyzed to obtain layout features centered on figures, tables, and formulas.

[0073] In this embodiment, if the extracted summary is not a structured summary, it is first converted into a structured summary, and then extracted into five parts: "background", "purpose", "method", "results" and "conclusion".

[0074] Based on the structural characteristics of academic literature and the importance of different parts, different vectorization representation methods are designed. The weighted mean representation method of combining literature titles and keywords improves the accuracy of semantic feature representation. A method that first structures the abstract and then treats each part as an independent semantic feature is designed to refine the granularity of semantic feature representation.

[0075] In this embodiment, dependency parsing or a large AI model is used to extract the content logic links, but those skilled in the art can also use other methods to extract the logic links. The content logic links are various content elements such as entities, concepts, terms, and actions, and the chain relationships between them, which is a formal semantic representation of text content.

[0076] The content logic links mentioned in this application include main logic link sentences and side logic link sentences. There is only one main logic link sentence, which reflects the key elements of the core content of the text and the key semantics of the text content. There may be multiple side logic link sentences, which serve as supplements to the content on the main link.

[0077] By designing main logical link sentences and side logical link sentences to decompose the text, the main logical link sentences serve as input for extracting key semantic features, while the side logical link sentences serve as input for non-key semantic features. Compared with existing text segmentation methods based on complete text, sentence segmentation, fixed length, sliding window segmentation, etc., the logical link approach can ensure the integrity of text content semantic segmentation, while also distinguishing between key and non-key content in the main text of academic documents.

[0078] In this embodiment, the extraction of logical links is performed on a paragraph in a paper based on the method of the present invention, specifically as follows:

[0079] Natural paragraphs in the text:

[0080] Biodiversity is the comprehensive ecological complex formed by organisms and their environment, along with the various ecological processes associated with it. It includes animals, plants, microorganisms, their genes, and the complex ecosystems they form with their environments. It typically comprises three components: genetic diversity, species diversity, and ecosystem diversity. Human survival is inseparable from other organisms; the rich diversity of organisms and ecosystems provides the material and environmental basis for human survival. In recent years, with the continuous growth of the world's population and the increasing scope and intensity of human activities, how to protect and rationally utilize biodiversity has become a challenge facing countries worldwide.

[0081] The logical chain of the extracted content is as follows:

[0082] {Main link sentence:}

[0083] Biodiversity → Ecological Complexes and Ecological Processes → Human Dependence

[0084] Side link sentence:

[0085] 1. Biodiversity → Genetic diversity, species diversity, ecosystem diversity

[0086] 2. Biodiversity → Impact of Human Activities → Challenges of Conservation and Utilization

[0087] More preferably, the extracted titles and keywords are vectorized separately, and then the weighted semantic features of the title and keywords are calculated. The method is as follows:

[0088] 1) Vectorize the obtained title using any method to obtain TV, which represents the semantic features of the title;

[0089] 2) Vectorize one or more keywords obtained using the same method as in step 1), to obtain {KV} i} represents the semantic features of keywords. i Indicates the keyword sequence number;

[0090] 3) Calculate the semantic features of title keywords. The calculation method is as follows:

[0091] ;

[0092] Wherein, TKV represents the semantic feature vector fused with title keywords. Indicates the weight of the title. n W represents the number of keywords. i Indicates the first i The weight of each keyword.

[0093] In this embodiment, the weights of the n keywords can be the same, decrease according to the keyword sequence number, or be partially decreased and partially the same.

[0094] A method for representing layout structure features with figures, tables, and formulas as core content, and a document layout structure matching method based on this method, which can serve as a general auxiliary method for comparing the layouts of various documents.

[0095] More preferably, the method for obtaining layout features centered on figures, tables, and formulas is as follows:

[0096] S110. Starting from the figure, table and formula respectively, perform a traversal of the preorder sequence and obtain the area of ​​the first preorder sequence that is the main text as the area of ​​the preorder text.

[0097] S120. Starting from the figure, table, and formula respectively, perform subsequent sequence traversal, and take the area of ​​the first subsequent sequence that is the main text as the area of ​​the subsequent text.

[0098] S130. Repeat steps S110-S120 until the layout feature sequence of all figures, tables, and formulas is obtained:

[0099] FigureID i ={ Figure , SubFigure , PreTxt , NextTxt};

[0100] TableID i ={ Table , CellNumber , PreTxt , NextTxt};

[0101] FormulaID i ={ Formula , LatexNum , PreTxt , NextTxt};

[0102] in, FigureID i For the first i The layout feature sequence of each image, Figure , SubFigure , PreTxt , NextTxt These represent the area of ​​the graph, the average area of ​​the subgraph, the area of ​​the preceding text, and the area of ​​the following text, respectively. TableID i For the first i The format feature sequence of each table, Table , CellNumber , PreTxt , NextTxt These represent the surface area, the number of table cells, the area of ​​the preceding text, and the area of ​​the following text, respectively. FormulaID i For the first i The format feature sequence of each table, Formula , LatexNum , PreTxt , NextTxt Represent the area and formula respectively. latex Number of commands, area of ​​preceding text, area of ​​following text.

[0103] In this embodiment, the areas of preceding and subsequent text are further explained:

[0104] In step S100, a pre-trained feature extraction model is used to extract features from the document to be analyzed, including the title, abstract, main text paragraphs, figures and subfigures, tables, formulas, and references. The layout order and area of ​​each structural element are recorded. The extraction order is assumed to be:

[0105] Title, Abstract, [Main Text Paragraph P1, Main Text Paragraph P2, Main Text Paragraph P3], Figure 1 , Figure 2 Formula 1, [Main Text Paragraph P4, Main Text Paragraph P5], Formula 2 Figure 3 Figure 4, [Main Text Paragraphs P6 and P7], Table 1 [Main Text Paragraphs P8 and P9], [References];

[0106] but, Figure 3 The corresponding preceding text area is the area of ​​[Main Text Paragraph P4, Main Text Paragraph P5], and the following text area is the area of ​​[Main Text Paragraph P6, Main Text Paragraph P7].

[0107] Alternatively, in this embodiment, the text area can also be replaced by the text length.

[0108] In this embodiment, the structure is extracted using a visual model. Alternatively, other commonly used structure methods in the art can also be used.

[0109] S200. Based on each query semantic feature, perform semantic similarity retrieval one by one in the preset multi-dimensional semantic feature index library to obtain a retrieval result set for each semantic feature. The retrieval result set includes similar semantic features and corresponding similar paragraph IDs and similar document IDs. The multi-dimensional semantic feature index library includes various document sets and semantic feature sets. Any semantic feature in the semantic feature set is associated with the original document paragraph ID, original document ID and original text position in various document sets through the index.

[0110] Preferably, the multi-dimensional semantic feature index library is obtained through various types of documents. By using a pre-trained feature extraction model to extract features from various document sets, semantic features corresponding to various documents are obtained, including semantic features of the main text, semantic features of title keywords, and semantic features of the abstract. At the same time, layout features are obtained. Based on the semantic features of various documents, indexes are constructed separately. Each semantic feature is linked to an identifier pointing to the original text content, including the document ID and the original text location.

[0111] Preferably, the various document collections include academic document collections and factory paper collections, and the format feature library includes a factory paper format feature library and an academic document collection format feature library. The factory paper format feature library and the academic document collection format feature library are extracted based on the academic document collection and the factory paper collection, respectively.

[0112] Preferably, if there is only an independent factory paper format feature library generated from the factory paper literature collection, after analyzing the literature to obtain the format features, the process jumps to step S500.

[0113] More preferably, the index construction method can adopt various existing vector indexing methods depending on the scale of the data; semantic features generated using the same vector generation method can also be merged into the same index when constructing the index.

[0114] Specifically, in this embodiment, the structured content obtained in step S100 can be stored in blocks or stored as a whole. The original text location (collections of various documents) can be either the location of the whole storage or the location of the block storage, without limitation.

[0115] Preferably, the query semantic features are searched for similarity in the same type of semantic feature index in the multi-dimensional semantic feature index library to obtain a set of search results for each semantic feature. The method is as follows:

[0116] First, based on the query semantic features, a similarity retrieval of the same type of semantic features is performed in a multi-dimensional semantic feature index library to obtain several candidate similar semantic features for each query semantic feature, resulting in a title + keyword retrieval result set T={ t i}, Structured summary search results {A={ a i}, B={ b i}, C={ c i}, D={ d i}, E={ e i}}、Main link retrieval results { M1={ m 1i}, M2={ m 2i},…} and side link retrieval results { S1={ s 1i},S2={ s 2i},…};

[0117] Wherein, A, B, C, D, and E represent the search result sets of "background", "purpose", "method", "results", and "conclusion" respectively; the types include semantic features of the main text, semantic features of the title keyword fusion and semantic features of the abstract, and similarity thresholds are set for the main logical link sentences and the side logical link sentences in the semantic features of the main text respectively;

[0118] Then, candidate similar semantic features with a similarity greater than a preset similarity threshold are selected as similar semantic features;

[0119] Finally, obtain the similar semantic features, the similarity of the similar semantic features, and the original document paragraph ID and original document ID pointed to by the vector index of the similar semantic features. Use the original document paragraph ID and original document ID as the corresponding similar paragraph ID and similar document ID.

[0120] In this embodiment, the number of elements in each set is flexibly controlled or a fixed number is set according to the similarity threshold.

[0121] S300. Weighted counts are performed on the similar paragraph IDs and the similar document IDs respectively, and the results are sorted in descending order based on the obtained count values ​​to obtain the similarity ranking result. The method is as follows:

[0122] Based on the type of query semantic features on which the similar semantic features are retrieved, different weights are assigned to similar paragraph IDs or similar document IDs;

[0123] The types include semantic features of the main text, semantic features of the title keywords, and semantic features of the abstract, but the main logical link sentences and the side logical link sentences in the semantic features of the main text are assigned weights respectively;

[0124] The number of times each similar paragraph ID or similar document ID was recalled was calculated based on the weights.

[0125] Based on the above counting results, the IDs of similar paragraphs or similar documents are sorted in descending order.

[0126] Preferably, in this embodiment, the retrieval results of the semantic features of the main text are the main factor in group counting, and the similar document ID count value recalled by the main link retrieval can be set to have the same or greater weight than that of the side links.

[0127] More preferably, in this embodiment, a weighted count is performed based on the above method, and the result is:

[0128] If a similar document ID is retrieved 5 times, twice through the main logical link sentence and 3 times through the secondary logical link sentence, then the weighted count value corresponding to this similar document ID is 3*MW+2*SW;

[0129] Suppose that in the 5 retrievals of the similar document ID, a total of 2 similar paragraph IDs were hit. One natural paragraph ID was recalled through 2 main logical link sentences and 1 side logical link sentence, and the weighted count value of this similar paragraph ID is 2*MW+1*SW; the other similar paragraph ID was recalled through 1 main logical link sentence and 1 side logical link sentence, and the weighted count value of this similar paragraph ID is 1*MW+1*SW.

[0130] Where MW is the weight for retrieval and recall of the main logical link sentence, and SW is the weight for retrieval and recall of the side logical link sentence.

[0131] More preferably, in this embodiment, when performing weighted counting, the search results of title + keywords and structured summary features can be considered, or they can be ignored.

[0132] More preferably, when the number of IDs of various types is extremely high and some results need to be filtered, only IDs with high count values ​​can be selected as needed, or the similarity threshold can be further increased to continue filtering IDs, thereby reducing the subsequent comparison process and improving the detection speed.

[0133] S400. Match the documents to be analyzed and the semantic features of the query with the similar document IDs and corresponding similar semantic features in the similarity ranking results to determine the semantic feature similarity.

[0134] The semantic feature similarity includes any one or more of the following: main logical link sentence similarity, side logical link sentence similarity, title keyword fusion semantic feature similarity, and abstract semantic feature similarity.

[0135] Preferably, when calculating similarity, vector similarity calculation methods are used, including but not limited to vector distance, dot product and other methods.

[0136] S500. Based on the layout features, match them with a preset layout feature library to determine the layout similarity. The method is as follows:

[0137] The layout features are tested using the difference ratio to determine whether the difference ratios in the four dimensions between the layout features of the document to be analyzed and the layout features of the same type of document to be compared simultaneously satisfy the preset parameter set K.

[0138] The test formula is: ,in i =1,2,3,4;

[0139] Furthermore, the three layout types—figures, tables, and formulas—can use different sets of preset parameters. k i , aThis indicates the formatting features of any figure, table, or formula in the document entered by the user. b This indicates a formatting feature of a figure, table, or formula of the same type in the documents to be compared.

[0140] If the conditions are met, the document layout feature matching degree count corresponding to the current layout feature is increased by 1, and the same figure, table, and formula features that meet the conditions are not counted cumulatively; otherwise, no count is added.

[0141] The format similarity is obtained only after all format features of the documents to be analyzed have been tested.

[0142] Preferably, if the document to be analyzed is a factory paper, after obtaining the layout features, the process jumps to step S500, where the layout features of the user-input paper are directly matched with the layout features of the factory papers in the current document set. Papers with high matching degree can be fused and calculated with other factory paper feature detection methods to indicate the potential risk of factory papers.

[0143] S600. The similarity ranking results are rearranged based on the semantic feature similarity and / or layout similarity to obtain the detection results.

[0144] Preferably, different sorting rules are selected based on user input. In this embodiment, the semantic feature similarity of the main text semantic features is used as the primary criterion:

[0145] The similarity ranking results are rearranged using semantic feature similarity as weight. The repetition ratio of main logical link sentences and side logical link sentences in the natural paragraphs of the user input document and the natural paragraphs of the result document is calculated, and the repetition between paragraphs is also considered. Among them, the repetition weight of main logical link sentences is greater than that of side logical link sentences, the repetition weight of the same paragraph is greater than that of the cross-paragraph repetition weight, and the repetition weight of continuous repetition is greater than that of scattered repetition.

[0146] More preferably, when calculating the repetition ratio, one or more methods can be used, such as text ratio, number of text numbers, link repetition ratio, and number of link repetitions.

[0147] Preferably, the various semantic and format features constructed by the above methods can be applied independently or, based on the retrieval results of different semantic features, can be matched using multiple fusion methods. Furthermore, the identified duplicate content is encapsulated together with the calculation results and integrated into existing academic misconduct detection software as a detection result indicating potential plagiarism such as rewriting, abbreviation, expansion, or paraphrasing. The detection results indicate whether the duplicate content belongs to the key content of the main logical link, the non-key content of the side logical link, or both, thereby improving the detection effectiveness of existing academic misconduct detection software.

[0148] Although the steps in the above embodiments are described in the above order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not need to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple variations are all within the protection scope of this invention.

[0149] A second embodiment of the present invention provides an apparatus for improving the detection effect of academic misconduct documents, the apparatus comprising:

[0150] The index building module is configured to build a multi-dimensional semantic feature index library using a trained feature extraction model. The multi-dimensional semantic feature index library includes various document sets and semantic feature sets. Any semantic feature in the semantic feature set is associated with the original document paragraph ID, the original document ID, and the original text position in various document sets through the index.

[0151] The feature extraction module is configured to acquire the documents to be analyzed and use a trained feature extraction model to extract features from the documents to obtain query semantic features and layout features.

[0152] The semantic similarity retrieval module is configured to perform semantic similarity retrieval in a preset multi-dimensional semantic feature index based on the query semantic features, and obtain similar semantic features as well as the corresponding similar paragraph IDs and similar document IDs;

[0153] The semantic similarity calculation module is configured to match the semantic features of the document to be analyzed and the query with the similar document IDs and corresponding similar semantic features in the similarity ranking results to determine the semantic feature similarity.

[0154] The layout similarity calculation module is configured to match the layout features with a preset layout feature library to determine the layout similarity.

[0155] The result sorting module is configured to perform weighted counting on the similar paragraph IDs and the similar document IDs respectively, and sort them in descending order based on the obtained count values ​​to obtain similarity sorting results; it is also configured to rearrange the similarity sorting results based on the semantic feature similarity and / or layout similarity to obtain detection results.

[0156] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the system described above can be found in the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0157] It should be noted that the system for detecting academic misconduct provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0158] An electronic device according to a third embodiment of the present invention includes:

[0159] At least one processor; and

[0160] A memory communicatively connected to at least one of the processors; wherein,

[0161] The memory stores instructions that can be executed by the processor to implement the aforementioned method for detecting academic misconduct documents.

[0162] A fourth embodiment of the present invention provides a computer-readable storage medium storing computer instructions, which are executed by the computer to implement the above-described method for detecting academic misconduct documents.

[0163] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the electronic device and computer-readable storage medium described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0164] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0165] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0166] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0167] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0168] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0169] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for detecting academic misconduct documents, characterized in that, Includes the following steps: S100. Obtain the document to be analyzed, and use a pre-trained feature extraction model to extract features from the document to obtain query semantic features and layout features; the query semantic features include text semantic features, title keyword fusion semantic features, and abstract semantic features; the layout features are centered on figures, tables, and formulas, and are obtained using the following method: S110. Starting from the figure, table and formula respectively, perform a traversal of the preorder sequence and obtain the area of ​​the first preorder sequence that is the main text as the area of ​​the preorder text. S120. Starting from the figure, table, and formula respectively, perform subsequent sequence traversal, and take the area of ​​the first subsequent sequence that is the main text as the area of ​​the subsequent text. S130. Repeat steps S110-S120 until the layout feature sequence of all figures, tables, and formulas is obtained: FigureID i ={Figure,SubFigure,PreTxt,NextTxt}; TableID i ={Table,CellNum,PreTxt,NextTxt}; FormulaID i ={Formula,LatexNum,PreTxt,NextTxt}; Among them, FigureID i Let TableID be the layout feature sequence of the i-th figure, where Figure, SubFigure, PreTxt, and NextTxt represent the figure area, the average area of ​​the subfigure, the area of ​​the preceding text, and the area of ​​the following text, respectively; i Let FormulaID be the layout feature sequence of the i-th table, where Table, CellNum, PreTxt, and NextTxt represent the surface area, number of table cells, area of ​​preceding text, and area of ​​following text, respectively; i Let be the layout feature sequence of the i-th table, where Formula, LatexNum, PreTxt, and NextTxt represent the formula area, the number of Latex commands in the formula, the area of ​​the preceding text, and the area of ​​the following text, respectively. S200. Based on the query semantic features, perform semantic similarity retrieval in a preset multi-dimensional semantic feature index library to obtain similar semantic features and corresponding similar paragraph IDs and similar document IDs; the multi-dimensional semantic feature index library includes a document set and a semantic feature set, and any semantic feature in the semantic feature set is associated with the original document paragraph ID, the original document ID and the original text position in the document set through an index. S300: Perform weighted counting on the similar paragraph IDs and the similar document IDs respectively, and sort them in descending order based on the obtained count values ​​to obtain the similarity ranking result; S400. Match the documents to be analyzed and the semantic features of the query with the similar document IDs and corresponding similar semantic features in the similarity ranking results to determine the semantic feature similarity. S500. Based on the layout features, match them with the preset layout feature library to determine the layout similarity. S600. The similarity ranking results are rearranged based on the semantic feature similarity and / or layout similarity to obtain the detection results.

2. The method for detecting academic misconduct documents according to claim 1, characterized in that, The method for extracting features from the document to be analyzed using a pre-trained feature extraction model is as follows: The document to be analyzed is structured and extracted, including the title, abstract, main text paragraphs, figures and subfigures, tables, formulas, and references. The layout order and layout area of ​​each structural element are also recorded. The extracted summary is divided into five parts: background, purpose, method, result, and conclusion, and each part is vectorized to obtain structured semantic features of the summary in different dimensions. The extracted natural paragraphs of the main text are sorted and the content logic links are extracted. The content logic links include main logic link sentences and side logic link sentences. The main logic link sentences and side logic link sentences are vectorized to obtain the semantic features of the main text. The main logic link sentences, side logic link sentences and original paragraphs of the main text corresponding to the semantic features are recorded. The extracted titles and keywords are vectorized separately, and then the semantic features of the title and keywords are calculated by weighting. The extracted structural content is analyzed to obtain layout features centered on figures, tables, and formulas.

3. The method for detecting academic misconduct documents according to claim 2, characterized in that, Based on the query semantic features, semantic similarity retrieval is performed in a pre-defined multi-dimensional semantic feature index. The method is as follows: Based on the query semantic features, similarity retrieval is performed in the same type of semantic feature index in the multi-dimensional semantic feature index library to obtain several candidate similar semantic features for each query semantic feature; wherein, the types include main text semantic features, title keyword fusion semantic features and abstract semantic features, and similarity thresholds are set for the main logical link sentences and side logical link sentences in the main text semantic features respectively. Select candidate similar semantic features with a similarity greater than a preset similarity threshold as similar semantic features; Obtain the similar semantic features, the similarity of the similar semantic features, and the original document paragraph ID and original document ID pointed to by the vector index of the similar semantic features. Use the original document paragraph ID and original document ID as the corresponding similar paragraph ID and similar document ID.

4. The method for detecting academic misconduct documents according to claim 2, characterized in that, The similar paragraph IDs and similar document IDs are weighted and counted separately, and the method is as follows: Based on the type of query semantic features on which the similar semantic features are retrieved, different weights are assigned to similar paragraph IDs or similar document IDs; The types include semantic features of the main text, semantic features of the title keywords, and semantic features of the abstract. The main logical link sentences and the side logical link sentences in the semantic features of the main text are assigned weights respectively. The number of times each similar paragraph ID or similar document ID was recalled was calculated based on the aforementioned weights.

5. The method for detecting academic misconduct documents according to claim 4, characterized in that, The semantic feature similarity includes one or more of the following: main logical link sentence similarity, side logical link sentence similarity, title keyword fusion semantic feature similarity, and abstract semantic feature similarity.

6. The method for detecting academic misconduct documents according to claim 2, characterized in that, The method for determining layout similarity is as follows: The layout features are tested using the difference ratio to determine whether the difference ratios in the four dimensions between the layout features of the document to be analyzed and the layout features of the same type of document to be compared simultaneously satisfy the preset parameter set. If the conditions are met, the document layout feature matching count corresponding to the current layout feature is increased by 1. However, the same figure, table, and formula features that meet the conditions are not counted cumulatively; otherwise, they are not counted. The format similarity is obtained after all format features of the documents to be analyzed have been tested.

7. A method for detecting academic misconduct documents according to claim 5 or 6, characterized in that, The method for rearranging the similarity ranking results based on the semantic feature similarity is as follows: The similarity ranking results are rearranged using semantic feature similarity as weight, and the repetition between paragraphs is taken into account; among them, the repetition weight of the main logical link sentence is greater than the repetition weight of the side logical link sentence, the repetition weight of the same paragraph is greater than the repetition weight of the cross paragraph, and the repetition weight of the continuous repetition is greater than the repetition weight of the scattered repetition.

8. The method for detecting academic misconduct documents according to claim 1, characterized in that, If only an independent format feature library generated from the factory's paper document collection exists, after analyzing the documents to obtain the format features, proceed to step S500.

9. An apparatus for detecting academic misconduct documents, and a method for detecting academic misconduct documents according to any one of claims 1-8, characterized in that, The device includes: The index building module is configured to build a multi-dimensional semantic feature index library using a trained feature extraction model. The multi-dimensional semantic feature index library includes various document sets and semantic feature sets. Any semantic feature in the semantic feature set is associated with the original document paragraph ID, the original document ID, and the original text position in various document sets through the index. The feature extraction module is configured to acquire the documents to be analyzed and use a trained feature extraction model to extract features from the documents to obtain query semantic features and layout features. The semantic similarity retrieval module is configured to perform semantic similarity retrieval in a preset multi-dimensional semantic feature index based on the query semantic features, and obtain similar semantic features as well as the corresponding similar paragraph IDs and similar document IDs; The semantic similarity calculation module is configured to match the semantic features of the document to be analyzed and the query with the similar document IDs and corresponding similar semantic features in the similarity ranking results to determine the semantic feature similarity. The layout similarity calculation module is configured to match the layout features with a preset layout feature library to determine the layout similarity. The result sorting module is configured to perform weighted counting on the similar paragraph IDs and the similar document IDs respectively, and sort them in descending order based on the obtained count values ​​to obtain similarity sorting results; it is also configured to rearrange the similarity sorting results based on semantic feature similarity and / or layout similarity to obtain detection results.

Images