Academic misconduct text recognition method, device, electronic device and storage medium

Abstract

The disclosed embodiments provide a method, device, electronic device and storage medium for identifying academic misconduct texts, and relate to the field of data mining technology. The method includes: parsing the text to be identified, obtaining content texts of multiple different components in the text to be identified; inputting multiple content texts into a pre-trained text detection model corresponding to each content text, respectively, and obtaining the detection results output by each text detection model, and the detection results are used to indicate the probability of predicting that the text to be identified is an academic misconduct text based on the corresponding text content; according to each detection result, determining the recognition result of the text to be identified, and the recognition result is used to indicate whether the text to be identified is an academic misconduct text. By parsing and obtaining the content text corresponding to the text to be identified and using the corresponding text detection model for detection, it is possible to obtain detection results of multiple different dimensions, and determine the recognition result by combining the detection results, which can effectively improve the accuracy of the academic misconduct text recognition result.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data mining, and in particular to a method, device, electronic device and storage medium for identifying academic misconduct text. Background Art

[0002] In the highly competitive learning and work environment, some students or researchers intend to spend less energy and time to quickly obtain invention results, thus giving rise to "paper mills". In order to meet the needs of ordering, scale and modeling, paper mills usually plagiarize, tamper with / forge, reuse charts, manipulate data or paper signatures, and may systematically manipulate the publishing process. Therefore, it is particularly important to promptly and effectively stop such academic misconduct as paper mills.

[0003] In the related technology, for academic misconduct texts such as papers generated by paper factories (hereinafter referred to as paper factory papers), the significant features of a large number of "paper projects" are usually manually mined based on statistics and the features are identified, thereby realizing academic misconduct text recognition. This method is highly dependent on expert manual analysis, and directly applied to batch automatic detection, it has problems such as poor generalization ability, low accuracy and low usability. The automatic identification tools for academic misconduct texts developed based on machine learning or deep learning technology focus on single-dimensional features, and usually can only mine surface features, making it impossible to accurately identify academic misconduct texts. It can be seen that the current academic misconduct text recognition methods all have the problem of low accuracy of recognition results.

[0004] In this case, there is an urgent need to provide an academic misconduct text recognition solution to improve the accuracy of academic misconduct text recognition results. Summary of the invention

[0005] The embodiments of the present disclosure provide a method, device, electronic device and storage medium for identifying academic misconduct texts, which are used to solve the technical problem of low accuracy of academic misconduct text recognition results in the prior art.

[0006] According to one aspect of an embodiment of the present disclosure, a method for identifying academic misconduct text is provided, comprising:

[0007] Parse the text to be recognized and obtain the content texts of multiple different components in the text to be recognized;

[0008] Inputting multiple content texts into pre-trained text detection models corresponding to the respective content texts, respectively, and obtaining detection results output by the respective text detection models, wherein the detection results are used to indicate the probability of predicting that the to-be-identified text is an academic misconduct text based on the corresponding text content;

[0009] According to each detection result, determining the recognition result of the text to be recognized, the recognition result is used to indicate whether the text to be recognized is an academic misconduct text;

[0010] Among them, the text detection model is trained based on sample content text carrying labels, and the labels are used to indicate whether the sample content text comes from academic misconduct text.

[0011] According to another aspect of the embodiment of the present disclosure, there is provided an academic misconduct text recognition device, comprising:

[0012] A text parsing module is used to parse the text to be recognized and obtain the content text of multiple different components in the text to be recognized;

[0013] A text detection module, used to input multiple content texts into pre-trained text detection models corresponding to each content text, and obtain detection results output by each text detection model, wherein the detection results are used to indicate the probability of predicting that the text to be identified is academic misconduct text based on the corresponding text content;

[0014] An abnormality recognition module is used to determine the recognition result of the text to be recognized according to each detection result, and the recognition result is used to indicate whether the text to be recognized is an academic misconduct text;

[0015] Among them, the text detection model is trained based on sample content text carrying labels, and the labels are used to indicate whether the sample content text comes from academic misconduct text.

[0016] According to another aspect of an embodiment of the present disclosure, an electronic device is provided. The electronic device includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the method provided in any of the above embodiments.

[0017] According to another aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method provided in any of the above embodiments are implemented.

[0018] According to one aspect of an embodiment of the present disclosure, a computer program product is provided, including a computer program, and the computer program is executed by a processor to perform the steps of the method provided in any of the above embodiments.

[0019] The technical solution provided by the embodiments of the present disclosure brings the following beneficial effects:

[0020] Obtain the content texts corresponding to multiple different components in the text to be identified, input each content text into the corresponding text detection model, obtain the corresponding output results, and obtain the final recognition result of the text to be identified based on each detection result. Based on the division of content text, the text to be identified is divided into multiple dimensions, and the content texts corresponding to different dimensions are detected and identified respectively, which can fully explore the mid- and deep-level features corresponding to the academic misconduct text, comprehensively determine the recognition results, and effectively improve the accuracy of the academic misconduct text recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings required for describing the embodiments of the present disclosure are briefly introduced below.

[0022] Figure 1 A flowchart of a method for identifying academic misconduct text provided by an embodiment of the present disclosure;

[0023] Figure 2 One of the structural schematic diagrams of a text detection model provided in an embodiment of the present disclosure;

[0024] Figure 3 A second structural diagram of a text detection model provided in an embodiment of the present disclosure;

[0025] Figure 4 A flowchart of a method for identifying academic misconduct papers provided by an embodiment of the present disclosure;

[0026] Figure 5 A schematic diagram of the structure of an academic misconduct text recognition device provided by an embodiment of the present disclosure;

[0027] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0028] The embodiments of the present disclosure are described below in conjunction with the drawings in the present disclosure. It should be understood that the implementation methods described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present disclosure and do not constitute a limitation on the technical solutions of the embodiments of the present disclosure.

[0029] It will be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the terms "including" and "comprising" used in the embodiments of the present disclosure refer to that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude the implementation as other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the technical field. It should be understood that when we say that an element is "connected" or "coupled" to another element, the one element may be directly connected or coupled to the other element, or it may refer to that the one element and the other element establish a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein indicates at least one of the items defined by the term, for example, "A and / or B" may be implemented as "A", or as "B", or as "A and B".

[0030] In order to make the objectives, technical solutions and advantages of the present disclosure more clear, the embodiments of the present disclosure will be further described in detail below with reference to the accompanying drawings.

[0031] The following describes several exemplary embodiments to illustrate the technical solutions of the embodiments of the present disclosure and the technical effects produced by the technical solutions of the present disclosure. It should be noted that the following embodiments can refer to, draw on or combine with each other, and the same terms, similar features and similar implementation steps in different embodiments will not be described repeatedly.

[0032] The following is a description of the related technologies involved in this application:

[0033] Related technology 1: A method for automatically detecting academic misconduct documents, creating features for the hierarchical content of the document to be detected and the table data in the document; creating features for the hierarchical content of the existing document and the table data in the document; matching the hierarchical content features of the document to be detected and the table data features in the document to be detected with the hierarchical content features of the existing document and the table features in the existing document; and determining whether the document to be detected contains academic misconduct content, misconduct table data, and the type of misconduct content.

[0034] Related technology 2: A user profiling and blockchain technology evidence collection method for academic misconduct scenarios. It includes uploading article information and user operations to the blockchain, calculating the user anomaly index through the isolation forest, using crawlers to obtain basic information about scholars, matching the identities of scholars, and labeling users as abnormal based on identity matching and anomaly index.

[0035] The above-mentioned related technologies 1 and 2 can only mine and identify the features of specific content of papers, but cannot mine the deep commonalities in papers from paper factories. Both have the problem of low accuracy of recognition results.

[0036] The following describes several exemplary embodiments to illustrate the technical solutions of the embodiments of the present disclosure and the technical effects produced by the technical solutions of the present disclosure. It should be noted that the following embodiments can refer to, draw on or combine with each other, and the same terms, similar features and similar implementation steps in different embodiments will not be described repeatedly.

[0037] It can be understood that in the academic misconduct text identification method provided by the embodiment of the present disclosure, any method step can be executed by the terminal and / or the server, and all the steps in the method can be executed independently by the terminal or the server, or jointly by the terminal and the server.

[0038] The server may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services. The terminal may be a smart phone, tablet computer, laptop computer, desktop computer, intelligent voice interaction device (such as smart speaker), wearable electronic device (such as smart watch), vehicle-mounted terminal, smart home appliance (such as smart TV), AR / VR device, etc., but is not limited thereto.

[0039] The following describes the embodiments of the present disclosure using a server as an execution subject; however, this does not constitute a limitation to the embodiments of the present disclosure.

[0040] Figure 1 A flowchart of a method for identifying academic misconduct text provided by an embodiment of the present disclosure is shown in FIG. Figure 1 As shown, the technical solution provided by the embodiment of the present disclosure includes the following steps:

[0041] Step S101, parsing the text to be recognized, and obtaining content texts of multiple different components in the text to be recognized.

[0042] The text to be identified may be text used to convey information in different fields, such as paper text, project fund text, book text or patent text.

[0043] Specifically, the embodiment of the present disclosure provides a technical solution for identifying whether a text to be identified is an academic misconduct text. Corresponding to a normal text, the academic misconduct text in the embodiment of the present disclosure can also be referred to as an abnormal text. Taking the specific steps of identifying any text to be identified as an example, the method provided by the embodiment of the present disclosure is described in detail.

[0044] In step S101, the text to be recognized is parsed, the components corresponding to the text to be recognized are recognized and the corresponding content text is extracted, and the content texts of multiple different components in the text to be recognized are obtained.

[0045] It is understandable that, according to the division and definition of the structure and content corresponding to the text to be recognized, the text to be recognized can be divided into several components, and the type of content text corresponding to each component is different. In the embodiment of the present disclosure, the type of text to be recognized, the division method of the components of the text to be recognized, the type of content text preset, and the specific number of content texts obtained can be determined according to actual needs.

[0046] For example, when the text to be identified is a paper text, the types of its corresponding content text may include: title text, abstract text, keyword text, main text and conclusion text, etc.; when the text to be identified is a patent text, the types of its corresponding content text may include: patent name text, specification abstract text, claim text, specification text and specific implementation method text, etc.

[0047] In addition, in the embodiments of the present disclosure, the file format corresponding to the text to be recognized may be .txt, .doc, .docx, .pdf, HTML, XML, etc., that is, when parsing the text to be recognized, it may be plain text content or formatted text content that is parsed, and the components corresponding to the text to be recognized are determined by methods including but not limited to recognizing information such as keywords, special characters or numbers in the text, and recognizing information such as fonts, font sizes and paragraph marks in the text, and the content text corresponding to each component is plain text content.

[0048] Step S102, input multiple content texts into the pre-trained text detection models corresponding to each content text, and obtain the detection results output by each text detection model. The detection results are used to indicate the probability of predicting that the text to be identified is an academic misconduct text based on the corresponding text content.

[0049] Among them, the text detection model is trained based on sample content text carrying labels, and the labels are used to indicate whether the sample content text comes from academic misconduct text.

[0050] Specifically, since different types of content texts are quite different, the features that can be mined from the content texts and the methods of detecting whether the corresponding content texts are derived from academic misconduct texts based on the mined features are also significantly different. Therefore, in order to make full use of the feature differences between different types of content texts and ensure the corresponding detection effects, the embodiments of the present disclosure construct dedicated text detection models for each type of content text and pre-train them.

[0051] The text detection model corresponding to each type of content text is pre-trained based on sample content text with labels. The labels are used to indicate whether the sample content text originates from academic misconduct text. The sample content text is consistent with the type of content text input when the model is applied. The algorithm adopted by the model, the specific structure of the model and the model training method can be determined according to actual needs.

[0052] For example, when the text to be identified is a paper text and the type of the content text is a title text, the sample content text of the text detection model includes paper titles derived from normal text and academic misconduct text, wherein the label is used to indicate that the paper title comes from normal text or academic misconduct text, normal text refers to a paper text with academic norms, and academic misconduct text refers to a paper text with academic misconduct.

[0053] In step S102, for any content text, the content text is input into a text detection model corresponding to the type of the text content, and the detection result output by the text detection model is obtained. Each content text is processed in the above manner to obtain the detection result corresponding to each content text.

[0054] The detection result is used to indicate the probability of predicting that the text to be identified is an academic misconduct text based on the corresponding text content. The specific form of the detection result can be the probability percentage of predicting that the text to be identified is an academic misconduct text, or the ratio of the two results of predicting whether the text to be identified is an academic misconduct text, etc. The specific form depends on the output type of the corresponding text detection model and can be determined according to actual needs.

[0055] Step S103, based on each detection result, determining the recognition result of the text to be recognized, the recognition result is used to indicate whether the text to be recognized is an academic misconduct text. Specifically, in step S103, the recognition result of the text to be recognized is determined based on each detection result, the recognition result is used to indicate whether the text to be recognized is an academic misconduct text.

[0056] In order to obtain an identification result that can indicate whether the text to be identified is an academic misconduct text, it is necessary to integrate the various detection results. The integration method can be a weighted method, a logical rule judgment method, etc. The specific integration method can be determined according to actual needs.

[0057] Among them, the weighted method is to assign a corresponding weight to each detection result, and perform weighted calculation or weighted average calculation on each detection result to obtain the recognition result. Logical rule judgment refers to determining the recognition result based on each detection result and its preset logical relationship. For example, when there are three detection results, when the probability that the detection result indicates that the corresponding content text is derived from the academic misconduct text is greater than that from the normal text, it is determined that the detection result indicates that the corresponding content text is derived from the academic misconduct text. When at least two detection results indicate that the corresponding content text is derived from the academic misconduct text, the recognition result is finally determined to be that the text to be recognized is an academic misconduct text.

[0058] In addition, it should be noted that academic misconduct refers to acts that violate academic norms and academic ethics, including but not limited to fabrication, falsification and plagiarism. Depending on the type of text to be identified, the corresponding classification rules for normal text and academic misconduct text may also be different, and the specific classification rules can be determined according to actual needs.

[0059] For example, when the text to be identified is a paper text, the corresponding academic misconduct text may refer to a paper factory paper; when the text to be identified is a project fund text, the corresponding academic misconduct text may refer to a fund project application with fabricated data problems; when the text to be identified is a book text or a patent text, the corresponding academic misconduct text may refer to a patent application document with plagiarism problems, etc.

[0060] The technical solution provided by the embodiment of the present disclosure obtains the content texts corresponding to multiple different components in the text to be identified, inputs each content text into the corresponding text detection model, obtains the corresponding output result, and obtains the final recognition result of the text to be identified based on each detection result. Based on the division of the content text, the text to be identified is divided into multiple dimensions, and the content texts corresponding to different dimensions are detected and identified respectively, which can fully explore the middle and deep features corresponding to the academic misconduct text, comprehensively determine the recognition results, and effectively improve the accuracy of the academic misconduct text recognition results.

[0061] In a possible implementation, the types of content text include: title text, abstract text and body text.

[0062] Specifically, in order to comprehensively mine the features corresponding to each dimension of the text to be identified, so as to improve the accuracy of the identification result indicating whether the text to be identified is academic misconduct text, the embodiment of the present disclosure performs detection from three dimensions: title, abstract and text. Correspondingly, the content text includes: title text, abstract text and text.

[0063] Among them, the title text refers to a short and eye-catching sentence or phrase used to describe the content or theme of the entire text to be recognized. Generally speaking, the title text exists at the beginning of the text to be recognized. In the embodiment of the present disclosure, in addition to the title text at the beginning of the text to be recognized, the title text can also include the titles of each chapter, etc.

[0064] Abstract text refers to a brief overview of the content of the text to be recognized, usually used to summarize the main content and theme of the text to be recognized.

[0065] It is understandable that since different categories of text to be identified do not all have components corresponding to the summary, in the embodiment of the present disclosure, the summary text, in addition to the content corresponding to the summary part of the text to be identified, may also be other parts whose substantive content is equivalent to the summary part, such as a review, conclusion, etc.

[0066] The body text is the subject part of the text to be identified, including detailed information, argumentation data and specific content. In the embodiment of the present disclosure, the scope of the text component to be identified corresponding to the body text can be determined according to the actual content, such as all the body content of the text to be identified, or the body content corresponding to a specific chapter in the text to be identified.

[0067] It is understandable that the corresponding title text, abstract text and main text of different categories of text to be recognized have different components corresponding to the text to be recognized. The specific method of determining the title text, abstract text and main text from the text to be recognized can be determined according to actual needs.

[0068] For example, when the text to be identified is a paper, the title text corresponds to the paper title and chapter titles, the abstract text corresponds to the paper abstract, and the body text corresponds to the chapter contents of the paper except the research background and acknowledgments. When the text to be identified is a patent, the title text corresponds to the patent name, the abstract text corresponds to the specification abstract, and the body text corresponds to the specific implementation method of the specification.

[0069] The title and abstract of the technical solution provided by the embodiment of the present disclosure can provide the theme and key information of the text, while the main text contains detailed content and arguments. More comprehensive and multi-angle information is obtained from the three dimensions of title, abstract and main text, and the recognition result of the text to be recognized is comprehensively determined to determine whether the text to be recognized is an academic misconduct text. Through multi-dimensional detection, the corresponding mid- and deep-level features of the academic misconduct text can be fully excavated, and a comprehensive judgment is made based on the features of the three dimensions to enhance the reliability of the recognition results, and further effectively improve the accuracy of academic misconduct text recognition.

[0070] In a possible implementation, the text detection model corresponding to the body text includes: a text processing module, a random walk module, a feature extraction module, a feature fusion module and a feature detection module;

[0071] Input multiple content texts into the pre-trained text detection models corresponding to the content texts, respectively, and obtain the detection results output by each text detection model, including:

[0072] Input the main text into the corresponding text detection model, perform word segmentation on the main text through the text processing module, and output the word segmentation sequence;

[0073] The word segmentation sequence is input into the random walk module, and multiple target word segmentation groups are determined and output from the word segmentation sequence by the random walk module, each target word segmentation group includes multiple consecutive word segmentations;

[0074] Input each target word group into the feature extraction module, and use the self-attention mechanism to extract features from each target word group, and output the feature vector of each target word group;

[0075] The feature vectors of each target word group are input into the feature fusion module, and the feature vectors of each target word group are fused by the feature fusion module to output the feature vector of the main text;

[0076] The feature vector of the main text is input into the feature detection module, and the feature detection module performs category prediction on the feature vector of the main text and outputs the detection result.

[0077] Specifically, Figure 2 This is one of the structural diagrams of a text detection model provided by an embodiment of the present disclosure, such as Figure 2 As shown, the text detection model corresponding to the main text includes: a text processing module, a random walk module, a feature extraction module, a feature fusion module and a feature detection module.

[0078] Correspondingly, the specific steps of inputting the body text into the corresponding text detection model and obtaining the detection result output by the text detection model include:

[0079] The main text is input into the corresponding text detection model, and the main text is segmented through the text processing module to output the segmentation sequence. By segmenting the main text and converting it into a segmentation sequence, the original main text can be split into meaningful vocabulary units and converted into a machine-readable representation to prepare for subsequent processing.

[0080] For example, the body text of the input text detection model is processed by the Tokenizer, which converts the input body text into an index (token) sequence and maps each word or subword to the corresponding number in the vocabulary of the text detection model. The vocabulary of the text detection model is a predefined vocabulary set that contains all the words and special tags used in model training, which determines the scope of vocabulary that the model can understand and process.

[0081] It is understandable that the above Tokenizer-based word segmentation method is only a specific example provided in the embodiment of the present disclosure. The word segmentation method in the embodiment of the present disclosure can be selected according to actual needs, including but not limited to any word segmentation method corresponding to the word segmentation method based on string matching, the word segmentation method based on understanding, and the word segmentation method based on statistics. Considering that the main text may be too long, directly performing feature extraction and category prediction on the word segmentation sequence may lead to negative effects such as increased computing resource consumption, reduced performance, difficulty in processing, and poor accuracy. Therefore, the embodiment of the present disclosure adds a random walk module before performing feature extraction on the word segmentation sequence.

[0082] The word segmentation sequence is input into the random walk module, and through the random walk strategy, the random walk module determines and outputs multiple target word segmentation groups from the word segmentation sequence, each target word segmentation group includes multiple consecutive word segments, which can effectively overcome the challenge of long text processing while retaining the context of the main text. It can be understood that the random walk strategy is used to randomly select multiple target word segmentation groups from the word segmentation sequence, and its specific method, the number of target word segmentation groups, and the number of words in the target word segmentation groups can be determined according to actual needs.

[0083] For example, the random walk strategy can be to determine a fixed walk window (including a window with a preset number of segmented words), randomly determine the window position from the segmented word sequence, and use the segmented words corresponding to the window as the target segmented word group. It can also be to preset the number of segmented words in the segmented word group, divide the segmented word sequence into multiple segmented word groups, randomly sample each segmented word group, and select a fixed number of target segmented word groups.

[0084] It is understandable that when training a text detection model, the sample body text can also be processed using the same random walk strategy mentioned above. While retaining text of sufficient length for training, multiple training samples can be randomly sampled to expand the number of samples so that the model can grasp the features of different parts of the input body text, thereby increasing the generalization ability of the model and enabling the model to learn the common features used for detection in different parts of the body text.

[0085] Each target word group is input into the feature extraction module, and the feature extraction module uses the self-attention mechanism to extract features of each target word group and output the feature vector of each target word group.

[0086] The feature extraction module in the disclosed embodiment uses the self-attention mechanism to capture the associated information in the text. The self-attention mechanism allows the model to dynamically weight different positions in the input sequence during the encoding process, thereby better capturing the semantic relationship and contextual information between words. Its expression is as follows:

[0087]

[0088] in, , and Represents the query, key, and value, respectively. is the dimension of the self-attention mechanism. softmax The function normalizes the attention weights and then multiplies them by the value to get the final attention output. In this way, through the use of Tokenizer processing and self-attention mechanism, the key information in the text can be extracted, providing strong support for subsequent processing and analysis.

[0089] By applying the self-attention mechanism, the feature extraction module is able to extract feature vectors corresponding to rich text features from each target word segment. These features can cover semantic information at the word, sentence, and paragraph levels, including but not limited to word associations, sentence structures, and paragraph topics. During the feature extraction process, the self-attention mechanism encodes each position in the target word segment and captures semantic representations of different granularities at different levels.

[0090] It can be understood that the specific architecture corresponding to the self-attention mechanism used by the feature extraction model in the embodiment of the present disclosure can be determined according to actual needs, and the architecture includes but is not limited to the transformer architecture, the bidirectional encoder representation transformer (Bidirectional Encoder Representations from Transformers, BERT) and the generative pre-trained transformer (Generative Pre-trained Transformer, GPT), etc.

[0091] The feature vectors of each target word group are input into the feature fusion module, and the feature vectors of each target word group are fused by the feature fusion module, and multiple feature vectors are synthesized into a single feature vector of the main text and output. It can be understood that the specific method of feature fusion includes but is not limited to linking, pooling, weighted summation and feature transformation.

[0092] The feature vector of the main text is input into the feature detection module, and the feature detection module predicts the category of the feature vector of the main text and outputs the detection result. It can be understood that the feature detection module handles the classification task corresponding to normal text and academic misconduct text, that is, the feature detection module is actually equivalent to a binary classifier. To construct a binary classifier, algorithms such as logistic regression, decision tree, random forest, and support vector machine (SVM) can be used.

[0093] The technical solution provided by the embodiment of the present disclosure determines and outputs multiple target word segments from the word segmentation sequence corresponding to the main text, and extracts features from each target word segmentation group to obtain a feature vector of each word segmentation group, further fuses each feature vector to obtain a feature vector of the main text, and performs category prediction on the feature vector of the main text, and outputs the detection result. This method can effectively reduce the amount of data that the model needs to process, effectively overcome the challenge of long text processing, and at the same time improve the training effect and generalization ability of the model.

[0094] In a possible implementation, a random walk module is used to determine and output multiple target word groups from a word segmentation sequence, including:

[0095] The word segmentation sequence is divided into multiple word segmentation groups according to the preset grouping rules through the random walk module;

[0096] A preset number of target word groups are randomly obtained from each word group, and the target word groups are output.

[0097] Specifically, considering that the current model based on the self-attention mechanism architecture has a maximum input limit when processing text, and generally speaking, the data corresponding to the main text far exceeds the input limit of the general model. For example, when the text to be recognized is a paper, the length of the main text data is generally between 4,000 and 10,000 tokens, which far exceeds the input limit of the general transformer model.

[0098] In order to solve the above problems, the embodiment of the present disclosure sets up a random walk module, adopts a random walk strategy for long texts, effectively processes long text data while maintaining the richness of the model input data, so as to solve the limitations when processing long texts and make full use of the capabilities of the model corresponding to the self-attention mechanism architecture for training and feature extraction. The random walk strategy specifically includes two steps: text segmentation and random selection.

[0099] Text segmentation refers to dividing a word segmentation sequence into multiple word segmentation groups according to preset grouping rules.

[0100] For example, the preset grouping rule may be to divide the word sequence into groups with no preset fixed number of word segments to obtain multiple word groups, or to divide the word sequence into groups with a preset fixed number of word segments, etc. The specific preset grouping rule may be determined according to actual needs.

[0101] Random selection refers to extracting word groups from each word group by random sampling, and randomly obtaining a preset number of word groups as target word groups and outputting them.

[0102] For example, for the transformer-based pre-trained language representation models BERT, DistilBERT (ADistilled version of BERT), and RoBERTa (A Robustly Optimized BERTPretraining Approach), the maximum number of tokens that can be processed by the tokenizer used when processing text is 510 (after removing the two special tokens [CLS] and [SEP]). Among them, the [CLS] token is used to represent the main idea or classification information of the entire sequence, and the [SEP] token is used to separate different parts of the sequence.

[0103] The word segmentation sequence corresponding to the main text is divided into groups of 510 words each to obtain the corresponding word segmentation group tokens.

[0104] Randomly select 10 word groups as target word groups select_groups , each target word group contains 510 words, .

[0105] The formula can be expressed as:

[0106]

[0107] It is understandable that during the training process, the sample body text can also be processed using the same random walk strategy as above. During each model training, 10 target grouping tokens are randomly selected and simultaneously input into the model for training, so that multiple training samples can be obtained for each sample body text. This random selection method can retain sufficient length text for training, and because of multiple random sampling, it can focus on the features of different parts of the input paper, which can enhance the generalization of the model and enable the model to learn the common features used for detection in different parts of the body text.

[0108] The random walk strategy provided by the technical solution provided by the embodiment of the present disclosure effectively overcomes the challenge of long text processing, while improving the training effect and generalization ability of the model, and has the following advantages:

[0109] Take full advantage of the model's input limitations: By splitting long texts into processable segments, the input limitations of the model corresponding to the self-attention mechanism architecture are fully utilized, avoiding the limitation of the model's input length.

[0110] Increase the diversity of training samples: Random sampling increases the diversity of training samples, enabling the model to learn more comprehensive and representative text features. By randomly selecting different grouping tokens for training, the model can be exposed to different parts of the text, thereby learning richer feature representations.

[0111] Learning contextual relationships: The grouping tokens selected during each training are random, which means that the model can learn the contextual relationships between different groups. This helps the model better understand the overall semantics of the text, not just the semantics of the local fragment.

[0112] In a possible implementation, the feature fusion module performs feature fusion on the feature vectors of each target word group and outputs the feature vector of the main text, including:

[0113] The feature fusion module performs average pooling on the feature vectors of each target word group and outputs the feature vector of the main text.

[0114] Specifically, the embodiment of the present disclosure performs feature fusion by average pooling operation, that is, after the feature vectors of the target word groups are input into the feature fusion module, the feature fusion module performs average pooling operation on the feature vectors of each target word group to output the feature vector of the main text.

[0115] For example, select 10 target word groups token As input, it is encoded through the BERT model to obtain the corresponding feature vector These feature vectors capture the semantic information of each segment and help understand the local meaning and context of the text.

[0116] The feature vectors of each target word group are integrated through the feature fusion module (average pooling layer) to obtain a comprehensive representation of the entire text. The specific steps are as follows:

[0117] The average pooling layer applies an average pooling operation (AvgPool2d) to these 10 feature vectors Integrate and merge the features of multiple fragments into a single feature vector .

[0118] The average pooling operation averages the corresponding positions of each feature vector to obtain a comprehensive representation of the entire body text. It contains the semantic information of the entire text, providing strong support for subsequent model training and text processing.

[0119]

[0120] The feature vector of the integrated text Input into the feature detection module to obtain the final detection result. The feature detection module further processes and transforms the features and outputs the detection result [P 1 ,P 2 ]. In this test result, P 1 and P 2 They respectively represent the probability distribution of predicting the text to be identified as two different categories, normal papers and paper factory papers, based on the main text. According to the probability distribution, we judge which category the input text belongs to and make corresponding classification decisions.

[0121] The technical solution provided by the embodiment of the present disclosure uses an average pooling operation through a feature fusion module to merge the feature vectors of each target word group of the main text into a single feature vector to obtain the feature vector of the main text. The average pooling operation can retain contextual information, so that the acquired feature vector of the entire main text can better reflect the semantic information of the entire main text, and more accurately mine the features of the main text of the text to be identified, thereby improving the accuracy of academic misconduct text recognition.

[0122] In a possible implementation, the text detection model corresponding to the title text or the summary text includes: a text processing module, a feature extraction module and a feature detection module;

[0123] Input multiple content texts into the pre-trained text detection models corresponding to the content texts, respectively, and obtain the detection results output by each text detection model, including:

[0124] Input the content text into the corresponding text detection model, perform word segmentation on the content text through the text processing module, and output the word segmentation sequence;

[0125] The word segmentation sequence is input into the feature extraction module, and the feature extraction module uses the self-attention mechanism to extract features of the word segmentation sequence and outputs the feature vector of the word segmentation sequence;

[0126] The text feature vector is input into the feature detection module, and the feature detection module performs category prediction on the feature vector of the word segmentation sequence and outputs the detection result.

[0127] Specifically, considering that the title text and abstract text data usually do not exceed the input limit of the general model, the title text or abstract text can be directly segmented and then feature extracted. Figure 3 The second structural diagram of a text detection model provided in the embodiment of the present disclosure is as follows: Figure 3 As shown, the text detection model corresponding to the title text or summary text provided by the embodiment of the present disclosure includes: a text processing module, a feature extraction module and a feature detection module.

[0128] It can be understood that there is a one-to-one correspondence between the title text or summary text and the text detection model, that is, the title text corresponds to the first text detection model, and the summary text corresponds to the second detection model, where "first" and "second" are used to distinguish the two text detection models and have no practical meaning.

[0129] The content text (title text or summary text) is input into the corresponding text detection model, and the content text is segmented through the text processing module to output the segmentation sequence.

[0130] The word segmentation sequence is input into the feature extraction module, and the feature extraction module uses the self-attention mechanism to extract features of the word segmentation sequence and outputs the feature vector of the word segmentation sequence.

[0131] The text feature vector is input into the feature detection module, and the feature detection module performs category prediction on the feature vector of the word segmentation sequence and outputs the detection result.

[0132] It should be noted that compared with the text detection model corresponding to the main text, the detection model corresponding to the title text or summary text does not require the random walk module and the feature fusion module. The remaining modules are consistent with the text detection model corresponding to the main text. The detailed description of each module is given above and will not be repeated here.

[0133] The technical solution provided by the embodiments of the present disclosure extracts corresponding text feature vectors from the two dimensions of title and abstract corresponding to the text to be identified, and performs probability prediction through the text feature vectors, so as to obtain detection results corresponding to the two dimensions. The detection results corresponding to the main text dimension can be further combined to achieve comprehensive judgment based on the detection results corresponding to different dimensions, so as to enhance the reliability of the recognition results and further effectively improve the accuracy of academic misconduct text recognition.

[0134] It should be noted that the self-attention mechanism architecture adopted by the detection models corresponding to the title text, abstract text and body text can be the same or different, which can be determined according to actual needs.

[0135] For example, the embodiment of the present disclosure uses three different self-attention architectures. By comparing the three, the model that best suits each type of content text is selected, thereby achieving better performance and results in the text feature extraction process.

[0136] 1. BERT is a Transformer-based pre-trained language model that learns a large amount of contextual information and semantic information of text through large-scale unsupervised training. BERT uses a bidirectional encoding method and uses a masked language model and next sentence prediction tasks for pre-training, so that the model can better understand language context and semantics.

[0137] 2. DistilBERT is a lightweight pre-trained language model based on BERT, which is obtained by distilling BERT. DistilBERT significantly reduces the model size and computing resource requirements while maintaining BERT's high performance, making it suitable for natural language processing tasks with limited computing resources.

[0138] 3. RoBERTa is a pre-trained language model based on BERT. It optimizes the BERT pre-training method by using larger data sets and longer training time, further improving the performance and robustness of the language model.

[0139] Taking the text to be identified as a paper text as an example, and the academic misconduct text as a paper factory paper, in order to verify the difficulty of the paper factory paper detection task and analyze the performance of different detection technologies for different paper structures, the disclosed embodiment uses the constructed paper factory paper and normal paper data sets of the same topic to train a group of machine learning and deep learning methods, and performs classification training on the title, abstract, and text of the paper, respectively, so as to find the characteristic performance of different methods for different data. The disclosed embodiment finds a method that is most suitable for paper factory paper detection through these detection technologies and comparisons.

[0140] Use precision, recall, and f1 to measure the training effects of different methods on the data set:

[0141] Precision: The proportion of true positive examples among samples predicted to be positive examples.

[0142] Recall: The proportion of samples predicted as positive among the true positive samples.

[0143] F1: A model evaluation indicator that takes into account both precision and recall.

[0144] Since the construction of the data set fully ensures the balance of sample categories, it is reasonable and effective to use precision, recall and F1 value as evaluation criteria. Since the three text structures of title, abstract and body are different in length, writing structure and information density, the embodiments of this disclosure use different detection technologies for comparative exploration.

[0145] Table 1

[0146]

[0147] Table 1 shows the comprehensive results of the paper detection experiment in Paper Factory. Based on the analysis of the data shown in Table 1, it can be observed that the BERT-based methods have achieved the best detection performance for the three text structures of title, abstract and text. For the title structure, the BERT model (BERT) with a shallower model layer achieved the best performance of F1=0.767, while for the abstract and text structures, the BERT model (RoBERTa) with a deeper model layer achieved the best performance of F1=0.842 and F1=0.847.

[0148] For the three structures, the Transformer-based model achieved better results than the traditional machine learning-based model, which reveals the powerful performance of the Transformer model on text tasks. In addition, by comparing the best BERT models among the three text structures, it can be found that for the title structure, the shallower BERT model achieves the best result, while for the abstract and text structure, the deeper BERT model achieves the best result. This may be because the length of the title is generally short, and a model with a very deep number of layers can be detected; while the length of the abstract and the text is long, and a deeper model is required to detect them. The experimental results show that the BERT model can be used to detect paper mill papers very well, and for the longer text structure, a more complex model structure is required to achieve better results.

[0149] In a possible implementation, the method further includes:

[0150] Recognize preset symbols, preset words and preset sentences from the text to be recognized;

[0151] Analyze the positions and quantities of preset symbols, preset words and preset sentences in the text to be recognized, and obtain symbol-level features, word-level features and sentence-level features, wherein the symbol-level features are used to indicate the distribution characteristics of the preset symbols in the text to be recognized, the word-level features are used to indicate the distribution characteristics of the preset words in the text to be recognized, and the sentence-level features are used to indicate the distribution characteristics of the preset sentences in the text to be recognized;

[0152] Obtain feature recognition results based on symbol-level features, word-level features, and sentence-level features;

[0153] According to each detection result, the recognition result of the text to be recognized is obtained, including:

[0154] According to the detection results and feature recognition results, the recognition result of the text to be recognized is obtained.

[0155] Specifically, in order to further improve the accuracy of the recognition result of whether the text to be recognized is an academic misconduct text, in addition to obtaining the detection results from the three dimensions of title, abstract, and text in a machine learning manner, the embodiment of the present disclosure also analyzes the distribution characteristics of symbols, words, and sentences in the text to be recognized, and recognizes whether the text to be recognized is an academic misconduct text from the distribution characteristics of symbols, words, and sentences in the entire text to be recognized, and obtains a feature recognition result. The specific steps are as follows:

[0156] Identify preset symbols, preset words and preset sentences from the text to be recognized, wherein the preset symbols, preset words and preset sentences, from the perspective of symbols, words and sentences, can refer to all symbols, words and sentences in the text to be recognized, or can be special symbols, words and sentences in the text to be recognized. The specific recognition method and content can be determined according to actual needs.

[0157] For example, the preset symbols are multiple different types of special symbols in the text to be recognized, the preset words include but are not limited to special words, words of special parts of speech, and all words, and the preset sentences include but are not limited to sentences containing special content, sentences with special sentence patterns, etc.

[0158] After identifying the word preset symbols, preset words and preset sentences in the text to be recognized, the positions and quantities of the preset symbols, preset words and preset sentences in the text to be recognized are analyzed to obtain symbol-level features, word-level features and sentence-level features.

[0159] Among them, symbol-level features, word-level features and sentence-level features are obtained. The symbol-level features are used to indicate the distribution characteristics of preset symbols in the text to be recognized, the word-level features are used to indicate the distribution characteristics of preset words in the text to be recognized, and the sentence-level features are used to indicate the distribution characteristics of preset sentences in the text to be recognized.

[0160] It can be understood that the distribution characteristics are used to reflect the number, position and distribution pattern of preset symbols, preset words and preset sentences in the text to be recognized. Its specific forms of expression include but are not limited to the total number in the text, the average number in each paragraph, the average number in each sentence (i.e., sentence), the difference in number between adjacent paragraphs, the difference in number in adjacent sentences, etc.

[0161] The feature recognition results are obtained based on symbol-level features, word-level features and sentence-level features.

[0162] Furthermore, the detection results and feature recognition results can be integrated to obtain the recognition result of the text to be recognized.

[0163] It can be understood that the steps of obtaining feature recognition results based on symbol-level features, word-level features and sentence-level features, and obtaining recognition results of the text to be recognized by integrating the detection results and feature recognition results, are essentially to integrate multiple results. The integration method may be a weighted method, a logical rule judgment method and the like. The specific method is similar to the method of determining the recognition result by integrating multiple detection results, and will not be elaborated here.

[0164] For example, taking the text to be identified as a paper text as an example, the embodiment of the present disclosure also starts from the features that may contain paper factory paper clues, such as the use of paper punctuation marks, sentence length, tone or conjunctions, and constructs 33 paper factory paper features using preset conditions as shown in Table 2 below, where the feature categories can be divided into 3 categories:

[0165] ① Text expression characteristics: In the process of producing papers in the same batch of paper factories, the same text generation tools may be used, so the generated texts have highly similar expressions. The expression characteristics studied in this paper include the use of punctuation marks and conjunctions.

[0166] ②Title sentence characteristics: According to existing research, the titles of papers from paper mills have similar grammatical structures.

[0167] ③ Paragraph structure characteristics: In order to speed up the paper production process, paper mill paper makers often use templates for batch production. This template has similarities in the structure of the entire text, and has commonalities in sentence length and paragraph length.

[0168] Table 2

[0169]

[0170] The Shapley Additive exPlanations (SHAP) tool is used to perform an explanatory analysis of the process of model detection of paper factory papers. SHAP assigns a SHAP value to each feature or attribute, and simulates the order of adding different features to measure its contribution to the model output. The SHAP values ​​of the model training are extracted for feature importance analysis, and the results show that 12 features are significant. According to the schematic diagram of the importance analysis of the structural features of the paper factory papers obtained in the embodiment of the present disclosure, it can be seen that the y-axis represents each feature, and these features are sorted according to importance; the x-axis represents the SHAP value, the negative direction is a normal paper, and the positive direction is a paper factory paper; each point represents a sample in the data set, and the color of the point represents the size of the feature value, red is high and blue is low; overlapping points jitter in the y-axis direction. By observing whether the red sample points on each feature are concentrated in the positive direction or the negative direction, it can be obtained whether the feature is more significant in the paper factory paper or the normal paper.

[0171] Table 3 shows the significance of the above features for normal papers or paper mill papers, summarized by feature category. Features 1 to 5 count the number of times five special symbols appear in each paper; Feature 6 counts the number of sentences containing numbers in each paper; Features 7 to 11 count the distribution of the number of words in each sentence, long sentences are defined as sentences with more than 34 words, and short sentences are defined as sentences with less than 11 words. The average and standard deviation of the number of words in each sentence are calculated to obtain features 9 and 10, and the average value of the difference in the number of words between two adjacent sentences is calculated to obtain feature 11; Feature 12 counts the number of times the word "et" appears in each paper, which is equivalent to "etal." used to mark references.

[0172] Paper mill papers use more single quotes "'" and semicolons ";", while normal papers use more brackets ")", hyphens "'" and colons ":". This indicates that paper mill papers may use single quotes and semicolons more frequently.

[0173] In terms of sentence distribution, normal papers have more sentences containing numbers, reflecting that formal research results cannot be separated from the support of experimental data. Paper mill papers contain more short sentences, and the standard deviation of word count and the difference between the number of words in two adjacent sentences are larger. Normal papers contain more long sentences and the average number of words is larger, indicating that the writing structure of paper mill papers is disorganized and long and short sentences are mixed.

[0174] It was also found that normal papers contain more words "et", that is, formal papers (without academic misconduct) will contain more "et al." to mark references.

[0175] Table 3

[0176]

[0177] Therefore, the disclosed embodiment uses these 12 obvious features as trace features of paper factory papers. Specifically, for a paper data to be detected, first count the original values ​​of the trace features c1, c2, c3…, c12 (such as the number of special symbols “)”, the number of sentences containing numbers, the average value of the difference in the number of words in two adjacent sentences, etc.), and then classify them into three categories, namely 1~5 and 12 are symbol-level features, 6~8 are sentence-level features, and 9~11 are word-level features. In each category, the original value of the more significant features in the “paper factory papers” is positive, and the original value of the more significant features in the “normal papers” is set to negative.

[0178] It should be noted that the word "et" is equivalent to "et al." used to mark references. In the paper, "et" itself does not have strong semantic information, but rather reflects the writing habits of the authors of the paper. Its nature is similar to the habit of using symbols. Therefore, in the present disclosed embodiment, it is classified as a symbol-level feature.

[0179] Sum all raw values ​​in each category and normalize Sigmoid Operation to scale it to a feature value between 0 and 1, that is:

[0180]

[0181]

[0182]

[0183] The disclosed embodiment calculates the title probability corresponding to the detection results of the title text, abstract text and body text. p1 , Summary Probability p2 , text probability p3 , symbol-level features obtained from trace features f1 , sentence-level features f2 , word-level features f3 , the six scores are summed up to get the final score, which is used to decide whether a paper is a paper mill paper. The larger the score, the higher the possibility.

[0184] Final Score=p1+p2+p3+f1+f2+f3

[0185] The technical solution provided by the embodiment of the present disclosure adds a statistically-based academic misconduct text recognition method on the basis of detecting by extracting text features according to a model, obtains corresponding features according to the distribution features corresponding to preset symbols, preset words and preset sentences in the text to be recognized, and comprehensively obtains feature recognition results, and obtains the final recognition result of the text to be recognized by combining the feature recognition results and each detection result, so as to improve the accuracy of the recognition result.

[0186] The following is a description of the academic misconduct text recognition method provided by the embodiment of the present disclosure in conjunction with a specific application example:

[0187] Figure 4 A flowchart of a method for identifying academic misconduct papers provided by the present disclosure is shown in FIG. Figure 4 As shown, the type of text to be identified is a paper text, the corresponding file format of the paper text is the PDF file format, and the academic misconduct text is a paper factory paper.

[0188] After obtaining the paper text in any PDF format, the PDF file is automatically parsed. The specific steps are as follows:

[0189] ① Byte stream serialization: The original PDF data of the paper is serialized into a character sequence through byte stream serialization.

[0190] ② Patterned search results: Perform a patterned search on the character sequence to obtain the font regular features of the title, abstract, and text.

[0191] ③ Extract text based on font: Combine character sequence and font rules to extract title, abstract and main text.

[0192] ④ Stability processing: Perform stability processing on the extracted title, abstract, and body text, including detecting spaces and font mixing to ensure the accuracy of the extracted text.

[0193] ⑤Output results: Finally, the extraction results of the paper title, abstract, and main text are obtained.

[0194] After obtaining the title, abstract, and body text extraction results, use the corresponding text detection models to perform automatic detection of papers in the paper factory. The specific steps are as follows:

[0195] ① Word vectorization: Through the text processing module, the title, abstract, and body text of the paper are word vectorized respectively to obtain the title, abstract, and body tokens.

[0196] ② Random walk: Since the main text is too long, it needs to be divided into multiple groups. Through the random walk module, a random walk strategy is adopted for the main text token to sample a fixed number of main text group tokens.

[0197] ③ Feature extraction: Input the title, abstract, and text grouping tokens into the feature extraction module to obtain the title, abstract, and text grouping text features.

[0198] ④Pooling average: The feature fusion module adopts a pooling average strategy for the main text grouping text, and synthesizes multiple grouping features of the main text grouping text into a single main text feature.

[0199] ⑤ Judgment results of the paper factory: Input the title, abstract, and text features into the corresponding feature detection modules, namely the title classifier, abstract classifier, and text classifier, respectively, to obtain the automatic detection results of the paper factory papers in three dimensions. The automatic detection results are expressed in the form of probability. p1 , p2 , p3 They respectively represent the probability scores of the title text, abstract text, and body text being papers from paper factories, and are all values ​​between 0 and 1. Further, the output probability scores are used to determine which category the input text belongs to.

[0200] It is understandable that compared with academic misconduct such as plagiarism which is easy to detect, there is currently an industrial chain for paper factories to produce papers. The papers generated by paper factories are more concealed and difficult to identify, which has brought serious negative impacts on science and education. However, the current existing technologies for detecting papers are not effective in identifying paper factory papers. Based on this, the present disclosed embodiment proposes an academic misconduct text recognition method that can be effectively applied to the identification of paper factory papers.

[0201] The disclosed embodiment mainly serves the automatic detection product of academic misconduct of "paper factory papers", and can support end-to-end detection process. By automatically converting PDF files into structured text, a paper factory paper detection model based on text random walk strategy and text attention mechanism is used. Compared with the previous automatic detection tools that can only detect simple abstract content, the method provided by the disclosed embodiment realizes the automatic detection of academic misconduct such as paper factory papers at the full text level.

[0202] A paper factory paper detection model based on a text-based random walk strategy and self-attention mechanism is proposed. It can adaptively detect paper texts of any length, and automatically detect whether a paper is a paper factory paper from multiple feature points such as text, grammar, and semantics through the three dimensions of title, abstract, and text. The content feature extraction of academic misconduct of paper factory papers is realized at the structured article content level. For academic misconduct such as "paper factory papers", an end-to-end model automation tool is proposed for the first time. As long as an original paper PDF is input, the article content can be automatically extracted for paper factory paper detection. This key technical point should be protected.

[0203] It should be understood that the embodiments of the present disclosure are only used as a specific example to explain the present invention in detail, and the described embodiments are only intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0204] Figure 5 A schematic diagram of a device for identifying academic misconduct text provided by an embodiment of the present disclosure is shown in FIG. Figure 5 As shown, the academic misconduct text recognition device 500 includes:

[0205] The text parsing module 501 is used to parse the text to be recognized and obtain the content texts of multiple different components in the text to be recognized;

[0206] The text detection module 502 is used to input multiple content texts into the pre-trained text detection models corresponding to the content texts, and obtain the detection results output by the text detection models, where the detection results are used to indicate the probability of predicting that the text to be identified is an academic misconduct text based on the corresponding text content;

[0207] The abnormality identification module 503 is used to determine the recognition result of the text to be identified according to each detection result, and the recognition result is used to indicate whether the text to be identified is an academic misconduct text;

[0208] Among them, the text detection model is trained based on sample content text carrying labels, and the labels are used to indicate whether the sample content text comes from academic misconduct text.

[0209] In a possible implementation, the types of content text include: title text, abstract text and body text.

[0210] In a possible implementation, the text detection model corresponding to the body text includes: a text processing module, a random walk module, a feature extraction module, a feature fusion module and a feature detection module;

[0211] Text detection module, used to:

[0212] Input the main text into the corresponding text detection model, perform word segmentation on the main text through the text processing module, and output the word segmentation sequence;

[0213] The word segmentation sequence is input into the random walk module, and multiple target word segmentation groups are determined and output from the word segmentation sequence by the random walk module, each target word segmentation group includes multiple consecutive word segmentations;

[0214] Input each target word group into the feature extraction module, and use the self-attention mechanism to extract features from each target word group, and output the feature vector of each target word group;

[0215] The feature vectors of each target word group are input into the feature fusion module, and the feature vectors of each target word group are fused by the feature fusion module to output the feature vector of the main text;

[0216] The feature vector of the main text is input into the feature detection module, and the feature detection module performs category prediction on the feature vector of the main text and outputs the detection result.

[0217] In a possible implementation, a random walk module is used to determine and output multiple target word groups from a word segmentation sequence, including:

[0218] The word segmentation sequence is divided into multiple word segmentation groups according to the preset grouping rules through the random walk module;

[0219] A preset number of target word groups are randomly obtained from each word group, and the target word groups are output.

[0220] In a possible implementation, the feature fusion module performs feature fusion on the feature vectors of each target word group and outputs the feature vector of the main text, including:

[0221] The feature fusion module performs average pooling on the feature vectors of each target word group and outputs the feature vector of the main text.

[0222] In a possible implementation, the text detection model corresponding to the title text or the summary text includes: a text processing module, a feature extraction module and a feature detection module;

[0223] Text detection module, used to:

[0224] Input the content text into the corresponding text detection model, perform word segmentation on the content text through the text processing module, and output the word segmentation sequence;

[0225] The word segmentation sequence is input into the feature extraction module, and the feature extraction module uses the self-attention mechanism to extract features of the word segmentation sequence and outputs the feature vector of the word segmentation sequence;

[0226] The text feature vector is input into the feature detection module, and the feature detection module performs category prediction on the feature vector of the word segmentation sequence and outputs the detection result.

[0227] In a possible implementation, the apparatus for identifying academic misconduct texts further includes: a distribution detection module;

[0228] The distribution detection module is used to identify preset symbols, preset words and preset sentences from the text to be recognized;

[0229] Analyze the positions and quantities of preset symbols, preset words and preset sentences in the text to be recognized, and obtain symbol-level features, word-level features and sentence-level features, wherein the symbol-level features are used to indicate the distribution characteristics of the preset symbols in the text to be recognized, the word-level features are used to indicate the distribution characteristics of the preset words in the text to be recognized, and the sentence-level features are used to indicate the distribution characteristics of the preset sentences in the text to be recognized;

[0230] Obtain feature recognition results based on symbol-level features, word-level features, and sentence-level features;

[0231] The anomaly recognition module is used to obtain the recognition result of the text to be recognized based on the detection results and feature recognition results.

[0232] The device of the embodiments of the present disclosure can execute the method provided by the embodiments of the present disclosure, and the implementation principles are similar. The actions performed by each module in the device of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed functional description of each module of the device, please refer to the description in the corresponding method shown in the previous text, which will not be repeated here.

[0233] In addition, in the embodiments of the present disclosure, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories), or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0234] An embodiment of the present disclosure provides an electronic device (computer device / equipment / system), including a memory, a processor, and a computer program stored in the memory, and the processor executes the above-mentioned computer program to implement the steps of the method provided in any optional embodiment of the present disclosure and achieve corresponding technical effects.

[0235] In an optional embodiment, an electronic device is provided. Figure 6 A schematic diagram of the structure of an electronic device provided by an embodiment of the present disclosure is shown in FIG. Figure 6 As shown, the electronic device 600 includes: a processor 601 and a memory 603. The processor 601 and the memory 603 are connected, such as through a bus 602. Optionally, the electronic device 600 may also include a transceiver 604, which may be used for data interaction between the electronic device and other electronic devices, such as data transmission and / or data reception. It should be noted that in actual applications, the transceiver 604 is not limited to one, and the structure of the electronic device 600 does not constitute a limitation on the embodiments of the present disclosure.

[0236] Processor 601 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of the present invention. Processor 601 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0237] The bus 602 may include a path to transmit information between the above components. The bus 602 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 602 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0238] The memory 603 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to carry or store computer programs and can be read by a computer, without limitation herein.

[0239] The memory 603 is used to store the computer program for executing the embodiment of the present disclosure, and the execution is controlled by the processor 601. The processor 601 is used to execute the computer program stored in the memory 603 to implement the steps shown in the above method embodiment.

[0240] The electronic devices in the embodiments of the present disclosure may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), wearable devices, etc., and fixed terminals such as digital TVs, desktop computers, etc.

[0241] An embodiment of the present disclosure provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0242] The embodiments of the present disclosure also provide a computer program product, including a computer program, which can implement the steps and corresponding contents of the aforementioned method embodiments when executed by a processor.

[0243] It should be noted that the computer-readable storage medium disclosed above may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0244] In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus, or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0245] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages ​​or a combination thereof, including, but not limited to, object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate 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 a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0246] The terms "first", "second", "third", "fourth", "1", "2", etc. (if any) in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than that shown or described in the text.

[0247] It should be understood that, although the flowchart of the embodiment of the present disclosure indicates each operation step by arrows, the implementation order of these steps is not limited to the order indicated by the arrows. Unless clearly stated herein, in some implementation scenarios of the embodiment of the present disclosure, the implementation steps in each flowchart can be executed in other orders according to demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages based on the actual implementation scenario. Some or all of these sub-steps or stages can be executed at the same time, and each sub-step or stage in these sub-steps or stages can also be executed at different times. In scenarios with different execution times, the execution order of these sub-steps or stages can be flexibly configured according to demand, and the embodiment of the present disclosure does not limit this.

[0248] The above is only an optional implementation method for some implementation scenarios of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present disclosure, other similar implementation methods based on the technical ideas of the present disclosure are also within the protection scope of the embodiments of the present disclosure.

Claims

1. A method for identifying academic misconduct texts, characterized in that: include: Parsing the text to be recognized, obtaining content texts of multiple different components in the text to be recognized, wherein the types of the content texts include: title text, abstract text and body text; Inputting multiple content texts into pre-trained text detection models corresponding to the respective content texts, respectively, and obtaining detection results output by the respective text detection models, wherein the detection results are used to indicate the probability of predicting that the to-be-identified text is an academic misconduct text based on the corresponding text content; Determine the recognition result of the text to be recognized according to each detection result, wherein the recognition result is used to indicate whether the text to be recognized is an academic misconduct text; The text detection model is trained based on sample content texts with labels, and the labels are used to indicate whether the sample content texts are derived from academic misconduct texts; The text detection model corresponding to the body text includes: a text processing module, a random walk module, a feature extraction module, a feature fusion module and a feature detection module; accordingly, the plurality of content texts are respectively input into the pre-trained text detection models corresponding to the respective content texts, and the detection results output by the respective text detection models are obtained, including: Input the main text into the corresponding text detection model, perform word segmentation processing on the main text through the text processing module, and output a word segmentation sequence; Inputting the segmented word sequence into a random walk module, determining and outputting a plurality of target segmented word groups from the segmented word sequence through the random walk module, wherein each target segmented word group includes a plurality of continuous segmented words; Input each target word group into the feature extraction module, use the self-attention mechanism to extract features from each target word group, and output a feature vector for each target word group; Inputting the feature vectors of each target word group into the feature fusion module, performing feature fusion on the feature vectors of each target word group through the feature fusion module, and outputting the feature vectors of the main text; Inputting the feature vector of the body text into a feature detection module, performing category prediction on the feature vector of the body text through the feature detection module, and outputting a detection result; The text detection model corresponding to the title text or the summary text includes: a text processing module, a feature extraction module and a feature detection module; accordingly, the plurality of content texts are respectively input into the pre-trained text detection models corresponding to the respective content texts, and the detection results output by the respective text detection models are obtained, including: Input the content text into the corresponding text detection model, perform word segmentation processing on the content text through the text processing module, and output a word segmentation sequence; Inputting the word segmentation sequence into the feature extraction module, performing feature extraction on the word segmentation sequence by using a self-attention mechanism through the feature extraction module, and outputting a feature vector of the word segmentation sequence; The text feature vector is input into a feature detection module, and the feature detection module performs category prediction on the feature vector of the word segmentation sequence and outputs a detection result.

2. The method for identifying academic misconduct text according to claim 1, characterized in that: The step of determining and outputting a plurality of target word segmentation groups from the word segmentation sequence by the random walk module includes: Dividing the word segmentation sequence into a plurality of word segmentation groups according to a preset grouping rule by the random walk module; A preset number of target word groups are randomly obtained from each word group, and the target word groups are output.

3. The method for identifying academic misconduct text according to claim 1, characterized in that: The feature fusion module is used to perform feature fusion on the feature vectors of each target word group and output the feature vector of the main text, including: The feature fusion module performs an average pooling operation on the feature vectors of each target word group, and outputs the feature vector of the main text.

4. The method for identifying academic misconduct text according to any one of claims 1 to 3, characterized in that: The method further comprises: Recognize preset symbols, preset words and preset sentences from the text to be recognized; Analyze the positions and quantities of the preset symbols, preset words and preset sentences in the text to be recognized to obtain symbol-level features, word-level features and sentence-level features, wherein the symbol-level features are used to indicate the distribution characteristics of the preset symbols in the text to be recognized, the word-level features are used to indicate the distribution characteristics of the preset words in the text to be recognized, and the sentence-level features are used to indicate the distribution characteristics of the preset sentences in the text to be recognized; Acquire feature recognition results according to the symbol-level features, the word-level features, and the sentence-level features; The step of obtaining the recognition result of the text to be recognized according to each detection result includes: According to the detection results and the feature recognition results, a recognition result of the text to be recognized is obtained.

5. An academic misconduct text recognition device, characterized in that: include: A text parsing module, used for parsing the text to be recognized, and obtaining content texts of multiple different components in the text to be recognized, wherein the types of the content texts include: title text, abstract text and body text; A text detection module, used to input multiple content texts into pre-trained text detection models corresponding to each content text, and obtain detection results output by each text detection model, wherein the detection results are used to indicate the probability of predicting that the text to be identified is an academic misconduct text based on the corresponding text content; An abnormality identification module, used for determining the identification result of the text to be identified according to each detection result, wherein the identification result is used for indicating whether the text to be identified is an academic misconduct text; The text detection model is trained based on sample content texts with labels, and the labels are used to indicate whether the sample content texts are derived from academic misconduct texts; The text detection model corresponding to the body text includes: a text processing module, a random walk module, a feature extraction module, a feature fusion module and a feature detection module; accordingly, the text detection module is used to: Input the main text into the corresponding text detection model, perform word segmentation processing on the main text through the text processing module, and output a word segmentation sequence; Inputting the segmented word sequence into a random walk module, determining and outputting a plurality of target segmented word groups from the segmented word sequence through the random walk module, wherein each target segmented word group includes a plurality of continuous segmented words; Input each target word group into the feature extraction module, use the self-attention mechanism to extract features from each target word group, and output a feature vector for each target word group; Inputting the feature vectors of each target word group into the feature fusion module, performing feature fusion on the feature vectors of each target word group through the feature fusion module, and outputting the feature vectors of the main text; Inputting the feature vector of the body text into a feature detection module, performing category prediction on the feature vector of the body text through the feature detection module, and outputting a detection result; The text detection model corresponding to the title text or the summary text includes: a text processing module, a feature extraction module and a feature detection module; accordingly, the text detection module is used to: Input the content text into the corresponding text detection model, perform word segmentation processing on the content text through the text processing module, and output a word segmentation sequence; Inputting the word segmentation sequence into the feature extraction module, performing feature extraction on the word segmentation sequence by using a self-attention mechanism through the feature extraction module, and outputting a feature vector of the word segmentation sequence; The text feature vector is input into a feature detection module, and the feature detection module performs category prediction on the feature vector of the word segmentation sequence and outputs a detection result.

6. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images