A content originality verification method and system based on editing behavior history

Abstract

The application discloses a content originality verification method and system based on editing behavior history, and the verification method is applied to a user end and a server, and the method comprises the following steps: collecting editing behavior data generated in the process of editing a document by a user at the user end, forming static granularity history records and dynamic granularity history records; transmitting the granularity history records to the server; analyzing the granularity history records at the server, forming identification results of multiple content generation modes; and statistically analyzing the identification results of the multiple content generation modes at the server, comparing the statistical quantity with a related preset threshold, comprehensively judging the originality of the content, and generating an originality verification report. By constructing a soft controllable creation environment at the front end to collect two kinds of history records with different granularities, and then transmitting the history records to the back end for originality analysis, the detection capability is not dependent on the text content itself, and the problem that the prior art is easily evaded or invalid due to AI iteration is fundamentally solved.

Description

Technical Field

[0001] This invention relates to a method and system for verifying the originality of content based on editing behavior history. Background Technology

[0002] With the rapid development of AI-generated content (AIGC) technology, content originality verification faces unprecedented challenges. Existing detection technologies can be mainly categorized as follows:

[0003] I. Finished Article Detection Technologies: These include methods such as text similarity comparison, perplexity detection, linguistic feature analysis, and machine learning classification. Representative products include CNKI AIGC Detection, Turnitin AI Writing Detection, GPTZero, and PaperYY. The fundamental flaw of this type of technology is that it only analyzes the static features of the final text, and its effectiveness gradually diminishes with the iteration of AI generation technology. Numerous "AI rewriting" tools already exist on the market; users can easily bypass detection by simply rewriting the AI-generated content once. II. Plagiarism Detection Systems: These detect plagiarism through text search and comparison. Representative products include CNKI Plagiarism Detection, VIP, and MOSS. This type of technology can only compare existing publicly available text records and cannot identify entirely new AI-generated content. III. Examination Proctoring Systems: These monitor users through screen recording, webcams, and restricting page switching. Representative products include iFlytek and Xuexitong. This type of technology offers a poor user experience, is unsuitable for long-duration tasks such as essay writing, and cannot detect cross-screen cheating. IV. Large Model Watermarking Technology: This technology adds invisible watermarks to the output of large models. Representative products include Google SynthID and Alibaba Cloud Watermarking. These technologies can only detect content generated by specific models. Watermarks become almost entirely ineffective after simple rewriting, and watermarks can be avoided by using open-source models. Fifth, version-level writing process analysis technology: This technology detects anomalies by periodically saving document version snapshots and analyzing differences between versions. While this technology focuses on the writing process, its granularity is too coarse, only detecting the simplest copy-paste behavior. Its fundamental limitation is its inability to detect cross-screen copying—users can completely circumvent detection by simply displaying AI-generated content on their phone or another screen and then manually typing it word by word. Cross-screen copying is extremely low-cost, but current technology has virtually no ability to detect it.

[0004] The common problems with the aforementioned existing technologies are: either they only analyze the static features of the final product (finished product detection, plagiarism checking, watermarking), making them easily circumvented or ineffective; or their data collection granularity is insufficient (version-level process analysis), failing to identify sophisticated circumvention methods; or they excessively intrude on user privacy (proctoring systems), making them unsuitable for typical content creation scenarios. Therefore, a new technological framework is urgently needed to fundamentally solve the problem of content originality verification and effectively prevent widespread cheating behaviors, including direct pasting, cross-screen copying, and AI-assisted rewriting. Summary of the Invention

[0005] This invention provides a method and system for verifying the originality of content based on editing behavior history, in order to solve the problem that existing technologies cannot fundamentally solve the problem of content originality verification.

[0006] The technical solution of this invention is implemented as follows:

[0007] The first objective of this invention is to provide a content originality verification method based on editing behavior history, applicable to both the user end and the server end. The method comprises: Step A, collecting editing behavior data generated during the user's document editing process on the user end, forming static granularity historical records and dynamic granularity historical records; Step B, transmitting each static granularity historical record and dynamic granularity historical record to the server end; Step C, analyzing each static granularity historical record and dynamic granularity historical record on the server end to form identification results for multiple content generation methods; Step D, statistically analyzing the identification results for multiple content generation methods on the server end, comparing the statistical results with relevant preset thresholds to comprehensively determine the originality of the content, and generating an originality verification report.

[0008] Step C above, analyzing the historical records of each static granularity and dynamic granularity on the server side to generate identification results for multiple content generation methods, specifically includes:

[0009] Step C1: Receive static and dynamic granular historical records on the server. Step C2: Preprocess and extract features from the static and dynamic granular historical records to form multiple feature information. Step C3: Input each feature information into a multi-layer analysis architecture for analysis, classify the deletion events in the feature information, and form identification results for multiple content generation methods. The identification results for the content generation methods are one or more of "original input", "external pasting", "internal copying" and "copying input".

[0010] Step C3 above, which involves inputting various feature information into a multi-layer analysis architecture for analysis, classifying deletion events in the feature information, and forming identification results for multiple content generation methods, specifically includes:

[0011] Step C31: Input each feature information into the statistical feature analysis layer, perform statistical feature analysis on each feature information, extract the statistical features of the feature information, and filter out obviously normal feature information and obviously abnormal feature information based on preset rules and thresholds. The remaining feature information is listed as intermediate information. Step C32: Convert each intermediate information into a feature vector sequence, input the feature vector sequence into the temporal behavior machine learning model analysis layer, and obtain high-risk copying information and low-risk copying information. Step C33: Input each low-risk copying information into the large language model semantic understanding analysis layer, perform semantic understanding and classification of the editing intent of low-confidence information, classify the deletion events in the feature information in combination with the context, and determine whether the low-confidence information is a semantic revision or a mechanical correction. Step C34: Collect the analysis results of each layer and perform statistics to obtain the recognition results of multiple content generation methods.

[0012] Step C33 above, which involves inputting each low-risk copying information into the semantic understanding and analysis layer of the large language model, performing semantic understanding and classification of the editing intent of the low-confidence information, and classifying the deletion events in the feature information in conjunction with the context to determine whether the low-confidence information is a semantic revision or a mechanical correction, specifically includes:

[0013] Step C331: Identify deletion type events in each low-risk copied information, and merge consecutive deletion operations with an interval less than a preset threshold into a deletion group; for each deletion group, extract the context before and after deletion, the inserted content after deletion, and the time interval between deletion and insertion; output a structured list of deletion event evidence; Step C332: Construct a structured natural language prompt from the evidence information of each deletion event; call the large language model to classify each event into one of the following categories: "semantic revision", "input method correction", "spelling error", "pure deletion" and "uncertain", and output the classification result of each deletion event and a statistical summary of each type; Step C333: Divide the editing process into multiple time slices at fixed time intervals, calculate the text difference for each slice, display the content changes within the time period, then associate the deletion events that occur in the slice and their classification results with the corresponding slice, and finally output a formatted time slice description; Step C334: Construct a complete judgment prompt from the time slice information and the final text, call the large language model for comprehensive judgment, and output a structured judgment result.

[0014] The aforementioned static granular history record saves a complete snapshot of the document at a preset time point and records the differences between adjacent versions; the dynamic granular history record records the operation sequence formed by each keyboard input and mouse operation of the user, using the smallest indivisible editing unit as the unit. The operation sequence includes the time information field when the operation occurs, the operation type field, the location information field, the content change field, and the auxiliary information field. The collected field data is structured and stored according to a preset format to form an editing behavior log.

[0015] Step D above, which involves statistically analyzing the identification results of multiple content generation methods on the server side, comparing the statistical results with relevant preset thresholds, comprehensively determining the originality of the content, and generating an originality verification report, specifically includes:

[0016] The system statistically analyzes the identification results of multiple content generation methods to obtain multiple writing features and statistical features. After comparing the above statistical features with relevant preset thresholds, the originality of the content is comprehensively determined, and an originality verification report including an originality score, judgment criteria, and risk warnings is generated. An API interface is also provided for external systems to call. The writing features include input speed distribution, deletion frequency, modification density, and the proportion of revision behavior. The statistical features include input speed, pasting ratio, deletion rate, and editing location distribution.

[0017] Step B above, transmitting the static and dynamic granular historical records to the server, specifically involves the user end employing a batch asynchronous upload strategy and a multi-channel redundant recording mechanism. The final upload is triggered when the page is closed or the session ends, and the temporarily stored editing behavior logs are periodically uploaded to the server.

[0018] A second objective of this invention is to provide a content originality system based on editing behavior history, characterized by comprising:

[0019] The system comprises the following modules: an editing behavior collection module (deployed on the user's end) and a data transmission module (deployed on the user's end) to collect editing behavior data during document editing, generating static and dynamic granular history records. A data receiving module (deployed on the server's end) receives the static and dynamic granular history records from the data transmission module. A behavior analysis module (deployed on the server's end) analyzes the static and dynamic granular history records to identify multiple content generation methods. A statistical comparison module (deployed on the server's end) statistically analyzes these results and compares them with preset thresholds to generate multiple comparison results. An originality verification module (deployed on the server's end) comprehensively determines the originality of the content based on these comparison results and generates an originality verification report.

[0020] The aforementioned editing behavior acquisition module includes a static granularity acquisition submodule and a dynamic granularity acquisition submodule. The static granularity acquisition submodule is used to save a complete snapshot of the user's edited document at multiple preset time points and record the difference information between adjacent versions as static granularity historical records. The dynamic granularity acquisition submodule is used to record the operation sequence formed by each keyboard input and mouse operation of the user, and store it in a structured manner according to a preset format to form an editing behavior log as a dynamic granularity historical record.

[0021] The aforementioned behavior analysis module includes a historical record preprocessing submodule, used to preprocess and extract features from historical records at various static and dynamic granularities to form multiple feature information; and a multi-layer analysis architecture analysis submodule, used to analyze multiple feature information to form recognition results for multiple content generation methods.

[0022] The aforementioned multi-layered analysis architecture includes the following sub-modules: a statistical feature analysis unit, a temporal behavior machine learning model analysis unit, and a large language model semantic understanding analysis unit.

[0023] The aforementioned statistical feature analysis unit includes an anomaly detection subunit and a paste recognition unit. The anomaly detection subunit is used to detect obvious abnormal behaviors such as sudden increases in content, abnormal timestamps, and abnormal editing modes. The paste recognition unit is used to identify paste operations and trace the source of content.

[0024] The aforementioned large language model semantic understanding and analysis unit includes a copying recognition subunit, which is used to distinguish between original input and copied input by classifying and analyzing deletion events.

[0025] The aforementioned content originality system based on editing behavior history also includes auxiliary function modules deployed on the server side. These auxiliary function modules include:

[0026] The document state reconstruction unit is used to reconstruct the state of a document at any historical moment based on dynamic granularity historical records; and the editing process visualization replay unit is used to provide video-style editing process replay, converting the history of editing behavior into a visual animation demonstration.

[0027] The aforementioned auxiliary function modules also include: a content source tracing unit, used to trace the source type of each character in the document, use different colors to mark content from different sources to form a visual display, count the number and proportion of characters from each source type, and generate a source distribution report; and a traceability rate calculation unit, used to calculate the ratio of the number of characters with a clearly traceable source to the total number of characters in the document as the traceability rate.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] 1. The content originality verification method based on editing behavior history, applied to both the user end and the server end, is characterized by comprising: Step A, collecting editing behavior data generated during the user's document editing process on the user end to form static granular history records and dynamic granular history records; Step B, transmitting each static granular history record and dynamic granular history record to the server end; Step C, analyzing each static granular history record and dynamic granular history record on the server end to form identification results for multiple content generation methods; Step D, statistically analyzing the identification results for multiple content generation methods on the server end, comparing the statistical quantity with relevant preset thresholds to comprehensively determine the originality of the content, and generating an originality verification report. The front end constructs a softly controllable creation environment. In addition to collecting traditional static granular history records, it also collects complete user creation process data to form dynamic granular history records, which are then transmitted to the back end for originality analysis. This makes the detection capability independent of the text content itself, fundamentally solving the problem that existing technologies are easily circumvented or become ineffective with AI iteration.

[0030] 2. Other advantages of the present invention are described in detail in the Embodiments section. Attached Figure Description

[0031] Figure 1 This is a flowchart of the content originality verification method based on editing behavior history provided in Embodiment 1 of the present invention;

[0032] Figure 2 This is a flowchart of step C in the content originality verification method based on editing behavior history;

[0033] Figure 3 This is a flowchart of step C3 in the content originality verification method based on editing behavior history;

[0034] Figure 4 This is a flowchart of step C33 in the content originality verification method based on editing behavior history;

[0035] Figure 5 This is a block diagram of a content originality system based on editing behavior history provided in Embodiment 2 of the present invention;

[0036] Figure 6 A schematic diagram of the behavior analysis module in a content originality system based on editing behavior history;

[0037] Figure 7 This is a block diagram of an auxiliary function module in a content originality system based on editing behavior history. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1:

[0040] like Figures 1 to 4 As shown, this embodiment provides a content originality verification method based on editing behavior history, applied to both the user end and the server end, characterized by including:

[0041] Step A: Collect editing behavior data generated by users during the document editing process on the user's end, forming static granular history records and dynamic granular history records;

[0042] Step B: Transmit the historical records of each static granularity and the historical records of each dynamic granularity to the server.

[0043] Step C: Analyze the historical records of each static granularity and the historical records of each dynamic granularity on the server side to generate identification results of multiple content generation methods;

[0044] Step D: On the server side, the identification results of multiple content generation methods are statistically analyzed. The statistical results are compared with relevant preset thresholds to comprehensively determine the originality of the content and generate an originality verification report.

[0045] The content originality verification method based on editing behavior history builds a soft and controllable creation environment on the front end. In addition to collecting traditional static granular historical records, it also collects complete user creation process data to create dynamic granular historical records, which are then transmitted to the back end for originality analysis. This makes the detection capability independent of the text content itself, fundamentally solving the problem that existing technologies are easily circumvented or become ineffective with AI iteration.

[0046] Step C above, analyzing the historical records of each static granularity and dynamic granularity on the server side to generate identification results for multiple content generation methods, specifically includes:

[0047] Step C1: Receive historical records of each static granularity and the historical records of each dynamic granularity on the server side;

[0048] Step C2: Preprocess and extract features from the historical records of each static granularity and the historical records of each dynamic granularity to form multiple feature information;

[0049] Step C3: Input each feature information into the multi-layer analysis architecture for analysis, classify the deletion events in the feature information, and form identification results of multiple content generation methods; wherein the identification results of the content generation methods are one or more of "original input", "external pasting", "internal copying" and "copy input".

[0050] User-generated input refers to content created and entered independently by the user; external pasting refers to content copied and pasted from outside the document; internal copying refers to content copied and pasted within the document; and transcribed input refers to content manually entered by the user referring to external content. Transcribed input specifically refers to the act of manually entering content by the user referring to an external content source, which may include a mobile phone screen, printed materials, or another monitor. The characteristic of transcribed input is that the user does not create content but simply inputs predetermined content word by word.

[0051] The deletion characteristic of original writing is that there is semantic-level correction after deletion, reflecting the author's thought and consideration in word choice; the deletion characteristic of copying is only mechanical correction, with the content before and after deletion being semantically identical. Therefore, by classifying deletion events in conjunction with context (including semantic revision, input method correction, spelling errors, and pure deletion, etc.), we can distinguish between "original input" and "copying input." Consideration represents the user's semantic-level modification of the input content during the creation process, reflecting the user's thought and creative process. This includes semantic revision (deleting existing content and inserting new content with different semantics), structural adjustment (moving or reorganizing document content), and expression optimization (improving the wording or expression of existing content).

[0052] Non-reflective behavior refers to mechanical corrections made by users during the input process, which do not reflect the creative thought process. It includes input method correction (correction made after selecting the wrong candidate word when using the Pinyin input method), spelling correction (correction caused by Pinyin input errors), and typing error correction (correction caused by pressing the wrong key).

[0053] Step C3 above, which involves inputting various feature information into a multi-layer analysis architecture for analysis, classifying deletion events in the feature information, and forming identification results for multiple content generation methods, specifically includes:

[0054] Step C31: Input each feature information into the statistical feature analysis layer, perform statistical feature analysis on each feature information, extract the statistical features of the feature information, and filter out obviously normal feature information and obviously abnormal feature information based on preset rules and thresholds. The remaining feature information is listed as intermediate information.

[0055] Step C32: Convert each intermediate information into a feature vector sequence, and input the feature vector sequence into the analysis layer of the time-series behavioral machine learning model to obtain high-risk and low-risk information for copying.

[0056] Step C33: Input each low-risk copying information into the semantic understanding and analysis layer of the large language model, perform semantic understanding and classification of the editing intent of the low-confidence information, classify the deletion events in the feature information in combination with the context, and determine whether the low-confidence information is a semantic revision or a mechanical correction.

[0057] Step C34: Collect the analysis results of each layer and perform statistics to obtain the identification results of multiple content generation methods.

[0058] The obvious anomalies include content surge anomalies, timestamp anomalies, and editing mode anomalies. The temporal behavior machine learning model analysis layer is a Transformer-based temporal behavior model that uses Multi-Instance Learning (MIL) for deep feature extraction and behavior pattern recognition. The large language model semantic understanding analysis layer employs the Agentic Workflow approach, utilizing a large language model to perform semantic-level intent analysis on editing behaviors along the timeline. This semantic-level intent analysis includes: extracting the editing operations to be analyzed and their context from the editing behavior history; classifying the editing operations into semantic-level intents; constructing a temporal representation of the editing process based on the classification results; making a comprehensive originality determination based on the temporal representation; and making semantic-level modifications to the input content.

[0059] Through cascading processing of a multi-layered analysis architecture, not only can pasting behavior be quickly detected in the first analysis layer, but also more subtle cross-screen copying behavior can be detected based on the "deliberation behavior" feature through the depth of the second and third analysis layers.

[0060] The statistical feature analysis layer extracts statistical features of editing behavior, including input speed, pasting ratio, deletion rate, and editing location distribution; it performs rapid screening based on preset rules and thresholds. The core indicators extracted by the statistical feature analysis layer include: traceability rate (the ratio of characters with a clearly traceable source to the total number of characters in the document); pasting ratio (the ratio of externally pasted characters to the total number of characters in the document); manual input ratio; and deletion rate (the ratio of the number of deletion operations to the total number of operations). If the pasting ratio exceeds a preset threshold (e.g., 50%), it is directly determined that there is a large amount of external pasting. The statistical feature analysis layer is characterized by low computational complexity and low resource consumption, making it suitable for initial screening of all submissions.

[0061] The temporal behavior machine learning model analysis layer converts the editing behavior sequence into a feature vector sequence, which is then input into the pre-trained temporal behavior machine learning model. The temporal behavior machine learning model adopts a Transformer-based multi-instance learning (MIL) architecture, which can handle variable-length operation sequences. The model outputs the plagiarism risk probability P∈[0,1], where a higher probability indicates a greater suspicion of plagiarism.

[0062] The Transformer-MIL architecture comprises an input layer, a Transformer encoding layer, a pooling layer, and a classification layer. The input layer divides the edit operation sequence into multiple instances with a fixed window size and stride, extracting a feature vector for each instance. The Transformer encoding layer employs a multi-head self-attention mechanism to encode the instance sequence, capturing the temporal dependencies between operations. The pooling layer uses average pooling to aggregate the variable-length instance sequence into a fixed-dimensional representation. The classification layer outputs binary classification probabilities.

[0063] In this embodiment, the input features of the temporal behavioral machine learning model include the following dimensions: basic operation features, including insertion length, deletion length, and paste flag; time features, including operation interval Δt and long pause flag; position features, including cursor movement distance Δcursor; behavioral semantic features, including undo operation flag and continuous paste detection flag; and input method features, used to identify the Chinese Pinyin input process.

[0064] The training strategy for the time-series behavioral machine learning model includes: using regularization techniques to alleviate overfitting; dividing the training and test sets into length-based hierarchical partitions to ensure that the model performs well on documents of different lengths; and setting performance target constraints, requiring a low false positive rate and a high true positive rate.

[0065] Step C33 above, which involves inputting each low-risk copying information into the semantic understanding and analysis layer of the large language model, performing semantic understanding and classification of the editing intent of the low-confidence information, and classifying the deletion events in the feature information in conjunction with the context to determine whether the low-confidence information is a semantic revision or a mechanical correction, specifically includes:

[0066] Step C331: Identify deletion type events in each low-risk information copying, and merge consecutive deletion operations with an interval less than a preset threshold into a deletion group; for each deletion group, extract the context before and after deletion, the inserted content after deletion, and the time interval between deletion and insertion; output a structured list of deletion event evidence.

[0067] Step C332: Construct the evidence information for each deletion event into a structured natural language prompt; call the large language model to classify each event into one of the following categories: "semantic revision", "input method correction", "spelling error", "pure deletion" and "uncertain", and output the classification results for each deletion event and the statistical summary of each type;

[0068] Step C333: Divide the editing process into multiple time slices at fixed time intervals, calculate the text difference for each slice, display the content changes within the time period, associate the deletion events that occur in the slice and their classification results with the corresponding slice, and finally output a formatted time slice description.

[0069] Step C334: Construct a complete judgment prompt from the time slice information and the final text, call the large language model to make a comprehensive judgment and output a structured judgment result.

[0070] The judgment results include originality score, judgment confidence level, final judgment, list of evidence supporting the judgment, list of evidence contradicting the judgment, detailed analysis, and comprehensive analysis summary; the characteristics of the semantic understanding analysis layer of the large language model are strong understanding ability and accurate judgment, but high computational cost, and it is usually used for in-depth analysis of suspicious cases.

[0071] The aforementioned static granular history record saves a complete snapshot of the document at a preset time point and records the differences between adjacent versions; the dynamic granular history record records the operation sequence formed by each keyboard input and mouse operation of the user, using the smallest indivisible editing unit as the unit. The operation sequence includes the time information field when the operation occurs, the operation type field, the location information field, the content change field, and the auxiliary information field. The collected field data is structured and stored according to a preset format to form an editing behavior log.

[0072] The data structure of the version snapshot includes: a version identifier, a timestamp of version saving, the complete content of the document, the length of the document content L_n, and the difference information from the previous version. Designers can save version snapshots of the document at preset time intervals T (e.g., T=30s or T=60s).

[0073] The operation type field includes: event type, with values ​​including input, key press, paste, cut, and copy; input type, with values ​​including insert text, delete backward, delete forward, insert from paste, and insert from drag and drop. The content change field includes the text content inserted during the operation, the text content deleted during the operation, the total length of the document before the operation L_before, the total length of the document after the operation L_after, and a flag indicating whether this operation is only cursor movement.

[0074] The location information field includes: the cursor position and selection range before the operation, used to locate the location where the editing occurred; and the cursor position and selection range after the operation, used to calculate the cursor movement distance Δcursor.

[0075] The content change fields include: the text content inserted in this operation; the text content deleted in this operation; the total length of the document before the operation L_before; the total length of the document after the operation L_after; and a flag indicating whether this operation is only a cursor movement.

[0076] The auxiliary information fields include clipboard metadata, source information of clipboard operations, key metadata, key information and modifier key status, input method metadata, and combination status and combination content of input methods.

[0077] Dynamic granular history records can reconstruct all intermediate operations between any two points in time without information loss; they can directly identify operations such as pasting and cutting, rather than inferring them from content differences; they can analyze user editing behavior patterns, including input rhythm, modification habits, and refinement characteristics; and they can trace the source type and input time of each character.

[0078] Step B above, transmitting the static and dynamic granular historical records to the server, specifically involves the user end employing a batch asynchronous upload strategy and a multi-channel redundant recording mechanism. The final upload is triggered when the page is closed or the session ends, and the temporarily stored editing behavior logs are periodically uploaded to the server.

[0079] A batch asynchronous upload strategy is adopted to reduce the number of network requests and client-side performance overhead, implementing a breakpoint resume mechanism to automatically retry in case of network interruption or upload failure, ensuring data integrity. A final upload is triggered when the page is closed or the session ends, ensuring all data has been transmitted to the server. A multi-channel redundant recording mechanism simultaneously monitors multiple event sources, including beforeinput, input, keydown, paste, and cut events. For situations where the same operation may be captured by multiple events, the system ensures recording accuracy through deduplication and verification. The data transmission module also stores editing behavior data first in the client's local storage to protect against data loss due to unexpected events such as network interruptions and page refreshes.

[0080] Step D above, which involves statistically analyzing the identification results of multiple content generation methods on the server side, comparing the statistical results with relevant preset thresholds, comprehensively determining the originality of the content, and generating an originality verification report, specifically includes:

[0081] The system statistically analyzes the identification results of multiple content generation methods to obtain multiple writing features and statistical features. After comparing the above statistical features with relevant preset thresholds, the originality of the content is comprehensively determined, and an originality verification report including an originality score, judgment criteria, and risk warnings is generated. An API interface is also provided for external systems to call. The writing features include input speed distribution, deletion frequency, modification density, and the proportion of revision behavior. The statistical features include input speed, pasting ratio, deletion rate, and editing location distribution.

[0082] The percentage of deliberation behavior is calculated by classifying deletion events during the editing process and then statistically analyzing each category.

[0083] In this embodiment, the preset threshold mainly includes two independent indicators: the existence of revision behavior and the density of revision behavior. The existence of revision behavior is a Boolean indicator, representing whether at least one semantic revision event exists within a specified text segment or time window. A value of 1 indicates the presence of revision behavior, and a value of 0 indicates its absence. The density of revision behavior is a numerical indicator, calculated as: Revision Density D = N_semantic / L × 100, where N_semantic is the number of semantic revision events, L is the text length (in characters), and the unit is times per hundred characters. This indicator quantifies the frequency of revision behavior per unit of text. Based on statistical analysis of a large number of real writing samples, this invention determines the discrimination thresholds and reference ranges for each indicator:

[0084] (1) Refinement density: The normal range of original writing is 1.5 to 4.0 times / 100 words, and the typical characteristic of copying behavior is less than 0.5 times / 100 words. The judgment threshold is set at 0.8 times / 100 words. If it is higher than this threshold, it is judged as an original tendency.

[0085] (2) Semantic revision ratio: defined as the ratio of the number of semantic revision events to the total number of all revision events (including semantic revision, typographical error correction, and pinyin error correction). The normal range for original writing is 30% to 60%, and the typical characteristic of copying behavior is less than 10%. The discrimination threshold is set at 15%.

[0086] (3) Evenness of the distribution of scrutiny behavior: Measured by the coverage of time slices, original writing usually covers more than 60% of time slices, while the scrutiny events of copying behavior are concentrated or missing. The judgment threshold is set to cover more than 40% of time slices.

[0087] (4) Paste ratio: Original writing is usually less than 10%, exceeding 30% triggers a warning, and exceeding 50% is judged as non-original;

[0088] (5) Deletion rate: The normal range for original writing is 8% to 20%. A deletion rate below 3% indicates plagiarism, and a deletion rate above 40% indicates data anomalies.

[0089] (6) Linear input ratio: Original writing is usually below 70%, exceeding 85% triggers a plagiarism warning, and exceeding 90% can be judged as plagiarism behavior in combination with other indicators.

[0090] The content originality verification method based on editing behavior history described in this embodiment employs an adaptive adjustment mechanism for the discrimination threshold. The aforementioned thresholds are default baseline values, which the system dynamically adjusts based on document metadata: for short documents (text length less than 300 characters), the lower limit requirement for refinement density is appropriately relaxed; for long documents (text length greater than 2000 characters), the threshold for linear input ratio is tightened; for documents completed quickly (writing time less than 3 minutes), all thresholds are tightened by approximately 20% to improve detection sensitivity; for documents written over a long period (writing time greater than 30 minutes), the thresholds are appropriately relaxed to reduce the false alarm rate. Furthermore, the system adjusts the weight coefficient of pinyin error correction events based on the input method type (pinyin input method, Wubi input method, handwriting input, etc.) and adopts differentiated baseline distribution models for different text types (argumentative essays, expository texts, narrative texts, etc.).

[0091] In this embodiment, the specific method for detecting copying behavior includes the following steps:

[0092] Step 1: Delete event extraction. Extract all deletion events from the editing history. These events include single backspace deletion, consecutive backspace deletion, deletion after selection, and cut deletion. Step 2: Contextual evidence extraction. For each deletion event group, extract the deleted text content and its length, the local context before and after deletion, the inserted content after deletion, the time interval between deletion and insertion, and the relative position of the deletion location in the document. Step 3: Delete event classification. Classify each deletion event semantically. Step 4: Originality determination. Determine the originality based on the classification results of the deletion events.

[0093] The deletion event categories include: semantic revision, where the content before and after deletion has different semantics, reflecting the author's deliberation process; typographic error correction, where corrections are made due to incorrect key presses, and the deleted content has no actual semantic meaning; pinyin error correction, where corrections are made due to incorrect word selection in the pinyin input method, and the content before and after deletion has the same semantics but similar pronunciation; and deletion only, where there is no subsequent insertion after deletion.

[0094] The criteria for determining originality include: original writing signals, namely, the existence of a sufficient number and density of semantic revision events, and these events are distributed in various positions of the document, reflecting a continuous thinking process; and copying signals, namely, very few or no semantic revision events, deletion events mainly consisting of corrections of typing errors and pinyin errors, and editing positions mainly concentrated at the end of the document in a linear input pattern.

[0095] The following three specific examples illustrate the analysis results of the content originality verification method based on editing behavior history of this invention.

[0096] Example 1: During the process of a user completing an 800-word argumentative essay on the user's end, the system recorded 2347 editing operations. The server performed three analyses on all editing operations. The first analysis layer (statistical feature analysis layer) detected no pasting behavior (pasting rate less than 10%, within the normal range), a traceability rate of 94% (traceability rate not less than 90%, within the normal range requirement), a deletion rate of 12.3% (deletion rate between 8% and 20%, within the normal range), and an average input speed of 3.2 words / second (speed between 2 and 6 words / second, within the normal range). The second analysis layer (temporal behavior machine learning model analysis layer) output a plagiarism risk probability P=0.08, classifying it as low risk. The analysis results of the third analysis layer (large language model semantic understanding analysis layer) are as follows: In the evidence extraction stage, 67 deletion events were extracted; in the deletion event classification stage, 23 semantic revisions, 12 typing error corrections, and 28 pinyin error corrections were identified, with only 4 being deleted; based on this, the deliberation density is calculated to be 23 ÷ 8 = 2.88 times / 100 characters, exceeding the originality threshold (0.8 times / 100 characters); the semantic revision ratio is 23 ÷ (23 + 12 + 28) = 36.5%, exceeding the originality threshold (threshold is 15%); in the time slice construction stage, 15 valid time slices were generated, with semantic revision events distributed in 12 of them, and the deliberation behavior distribution coverage is 12 ÷ 15 = 80%, exceeding the originality threshold (threshold is 40%). The overall judgment result is: originality score of 92 points, judgment result is "original", confidence level 95%.

[0097] Example 2: A user submitted an 800-word document, and the system recorded 87 editing operations. The first analysis layer (statistical feature analysis layer) detected that the proportion of pasted characters reached 77.9%, exceeding the preset threshold (threshold is 50%), directly triggering a non-original warning; the traceability rate was 98%, within the normal range; the deletion rate was 2.1%, lower than the lower limit of the abnormal threshold (threshold is 3%), showing abnormal characteristics; the input speed fluctuated abnormally, with the characteristic of a large number of characters being inserted instantaneously. The second analysis layer (temporal behavior machine learning model analysis layer) output a copying risk probability P=0.94, judging it as high risk. The analysis results of the third analysis layer (large language model semantic understanding analysis layer) are as follows: only 3 deletion events were extracted in the evidence extraction stage, and 0 semantic revisions, 2 typing error corrections, and 1 pinyin error correction were identified in the deletion event classification stage; based on this, the deliberation density was calculated to be 0 times / 100 characters, lower than the copying judgment threshold (threshold is 0.5 times / 100 characters); the semantic revision proportion was 0%, lower than the copying judgment threshold (threshold is 10%). The overall assessment result is: originality score of 25 points, judgment result of "non-original", confidence level of 98%. The main judgment criteria include: the proportion of pasted characters is 77.9%, far exceeding the 50% threshold, pasting operations are concentrated within 2 minutes after the start of editing, and the scrutiny density is 0.

[0098] Example 3: A user submitted an 800-word document. The system recorded 1892 editing operations, all of which were manual input with no pasting. The first analysis layer (statistical feature analysis layer) detected a 0% pasting rate, which is within the normal range; a traceability rate of 96%, which is also within the normal range; a deletion rate of 2.8%, close to the lower limit of the anomaly threshold (3%), showing marginal anomaly characteristics; and a 100% manual input rate, which is within the normal range; however, the linear input rate was as high as 92%, exceeding the anomaly threshold (85%), showing significant anomaly characteristics. The second analysis layer (temporal behavior machine learning model analysis layer) output a copying risk probability P=0.72, classifying it as medium to high risk, and automatically triggering the third analysis layer for in-depth analysis. The analysis results of the third analysis layer (the semantic understanding analysis layer of the large language model) are as follows: In the evidence extraction stage, a total of 53 deletion events were extracted; in the deletion event classification stage, only 2 semantic revisions, 4 typing error corrections, and 47 pinyin error corrections were identified; based on this, the deliberation density is calculated to be 2÷8=0.25 times / 100 characters, which is lower than the copying judgment threshold (0.5 times / 100 characters); the semantic revision ratio is 2÷(2+4+47)=3.8%, which is lower than the copying judgment threshold (10%); in the time slice construction stage, 18 effective time slices were generated, and semantic revision events were only distributed in 2 of them, with a deliberation behavior distribution coverage of 2÷18=11%, which is lower than the copying judgment threshold (20%). The comprehensive judgment result is: originality score of 30 points, judgment result of "non-original", confidence level of 92%. The main criteria for judgment include: the revision density of 0.25 times / 100 characters is far lower than the original threshold of 0.8 times / 100 characters; the semantic revision rate is only 3.8%; 92% of the editing operations are linear appending; and the proportion of pinyin error correction is as high as 89%, which is consistent with the behavior of only correcting input method errors when copying.

[0099] The content originality verification method based on editing behavior history described in this embodiment can effectively distinguish between "original input" and "copy input", and has a high accuracy rate in originality verification.

[0100] Example 2:

[0101] like Figures 5 to 7 As shown, this embodiment provides a content originality system based on editing behavior history, characterized by including:

[0102] The editing behavior collection module 11 is deployed on the user terminal 1. It is used to collect a number of editing behavior data during the user's document editing process and form static granular history records and dynamic granular history records.

[0103] The data transmission module 12, deployed on the user terminal 1, is used to transmit static granular history records and dynamic granular history records to the server.

[0104] The data receiving module 21, deployed on the server 2, is used to receive static granular history records and dynamic granular history records from the data transmission module.

[0105] The behavior analysis module 22, deployed on server 2, is used to analyze static granular history records and dynamic granular history records to generate identification results for multiple content generation methods.

[0106] The statistical comparison module 23, deployed on the server 2, is used to perform statistics on multiple recognition results and compare the statistical quantity with relevant preset thresholds to form multiple comparison results;

[0107] The originality verification module 24, deployed on server 2, is used to comprehensively determine the originality of content based on multiple comparison results and generate an originality verification report.

[0108] The originality system for the history of editing behavior can collect various types of editing behavior data, has a simple structure, and has a high accuracy rate in originality verification.

[0109] The aforementioned editing behavior acquisition module 11 includes a static granularity acquisition submodule 111 and a dynamic granularity acquisition submodule 112. The static granularity acquisition submodule 111 is used to save a complete snapshot of the user's edited document at multiple preset time points and record the difference information between adjacent versions as static granularity historical records. The dynamic granularity acquisition submodule 112 is used to record the operation sequence formed by each keyboard input and mouse operation of the user, and store it in a structured manner according to a preset format to form an editing behavior log as a dynamic granularity historical record.

[0110] The behavior analysis module 22 described above includes a historical record preprocessing submodule 221, which is used to preprocess and extract features from historical records of various static granularities and historical records of dynamic granularities to form multiple feature information.

[0111] And a multi-layer analysis architecture analysis submodule 222, which is used to analyze multiple feature information to form recognition results of multiple content generation methods.

[0112] The aforementioned multi-layer analysis architecture analysis submodule 222 includes: a statistical feature analysis unit 26, a temporal behavior machine learning model analysis unit 27, and a large language model semantic understanding analysis unit 28.

[0113] The statistical feature analysis unit 26 described above includes an anomaly detection subunit 261 and a paste recognition unit 262. The anomaly detection subunit 261 is used to detect obvious abnormal behaviors such as sudden increases in content, abnormal timestamps, and abnormal editing modes. The paste recognition unit 262 is used to identify paste operations and trace the source of content.

[0114] The large language model semantic understanding and analysis unit 28 described above includes a copying recognition subunit 281, which is used to distinguish between original input and copied input by classifying and analyzing deletion events.

[0115] The aforementioned content originality system based on editing behavior history also includes an auxiliary function module 25 deployed on server 2. The auxiliary function module 25 includes:

[0116] Document state reconstruction unit 251 is used to reconstruct the state of a document at any historical moment based on dynamic granularity historical records;

[0117] The editing process visualization playback unit 252 is used to provide video-style editing process playback, converting the history of editing behavior into a visual animation demonstration.

[0118] Content source tracing unit 253 is used to trace the source type of each character in the document, mark the content from different sources with different colors to form a visual display, count the number and proportion of characters from each source type, and generate a source distribution report;

[0119] The traceability calculation unit 254 is used to calculate the ratio of the number of characters with a clearly traceable source to the total number of characters in the document as the traceability rate.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A content originality verification method based on editing behavior history, applied to both the user end and the server end, characterized in that, include: Step A: Collect editing behavior data generated by users during the document editing process on the user's end, forming static granular history records and dynamic granular history records; Step B: Transmit the historical records of each static granularity and the historical records of each dynamic granularity to the server. Step C: Analyze the historical records of each static granularity and the historical records of each dynamic granularity on the server side to generate identification results of multiple content generation methods; Step D: On the server side, the identification results of multiple content generation methods are statistically analyzed. The statistical results are compared with relevant preset thresholds to comprehensively determine the originality of the content and generate an originality verification report.

2. The content originality verification method based on editing behavior history according to claim 1, characterized in that: Step C, analyzing the historical records of each static granularity and dynamic granularity on the server side to generate identification results for multiple content generation methods, specifically includes: Step C1: Receive historical records of each static granularity and the historical records of each dynamic granularity on the server side; Step C2: Preprocess and extract features from the historical records of each static granularity and the historical records of each dynamic granularity to form multiple feature information; Step C3: Input each feature information into the multi-layer analysis architecture for analysis, classify the deletion events in the feature information, and form identification results of multiple content generation methods; wherein the identification results of the content generation methods are one or more of "original input", "external pasting", "internal copying" and "copy input".

3. The content originality verification method based on editing behavior history according to claim 2, characterized in that: Step C3, inputting each feature information into a multi-layer analysis architecture for analysis, classifying deletion events in the feature information, and forming identification results for multiple content generation methods, specifically includes: Step C31: Input each feature information into the statistical feature analysis layer, perform statistical feature analysis on each feature information, extract the statistical features of the feature information, and filter out obviously normal feature information and obviously abnormal feature information based on preset rules and thresholds. The remaining feature information is listed as intermediate information. Step C32: Convert each intermediate information into a feature vector sequence, and input the feature vector sequence into the analysis layer of the time-series behavioral machine learning model to obtain high-risk and low-risk information for copying. Step C33: Input each low-risk copying information into the semantic understanding and analysis layer of the large language model, perform semantic understanding and classification of the editing intent of the low-confidence information, classify the deletion events in the feature information in combination with the context, and determine whether the low-confidence information is a semantic revision or a mechanical correction. Step C34: Collect the analysis results of each layer and perform statistics to obtain the identification results of multiple content generation methods.

4. The content originality verification method based on editing behavior history according to claim 3, characterized in that: Step C33, inputting each low-risk copying information into the semantic understanding and analysis layer of the large language model, performing semantic understanding and classification of the editing intent of the low-confidence information, classifying deletion events in the feature information in conjunction with the context, and determining whether the low-confidence information is a semantic revision or a mechanical correction, specifically includes: Step C331: Identify deletion type events in each low-risk information copying, and merge consecutive deletion operations with an interval less than a preset threshold into a deletion group; for each deletion group, extract the context before and after deletion, the inserted content after deletion, and the time interval between deletion and insertion; output a structured list of deletion event evidence. Step C332: Construct the evidence information for each deletion event into a structured natural language prompt; call the large language model to classify each event into one of the following categories: "semantic revision", "input method correction", "spelling error", "pure deletion" and "uncertain", and output the classification result for each deletion event and the statistical summary of each type; Step C333: Divide the editing process into multiple time slices at fixed time intervals, calculate the text difference for each slice, display the content changes within the time period, associate the deletion events that occur in the slice and their classification results with the corresponding slice, and finally output a formatted time slice description. Step C334: Construct a complete judgment prompt from the time slice information and the final text, call the large language model to make a comprehensive judgment and output a structured judgment result.

5. A method for verifying content originality based on editing behavior history according to any one of claims 1 to 4, characterized in that: The static granular history record saves a complete snapshot of the document at a preset time point and records the differences between adjacent versions; the dynamic granular history record records the operation sequence formed by each keyboard input and mouse operation of the user, using the smallest indivisible editing unit as the unit. The operation sequence includes the time information field when the operation occurs, the operation type field, the location information field, the content change field, and the auxiliary information field. The collected field data is structured and stored according to a preset format to form an editing behavior log.

6. The content originality verification method based on editing behavior history according to claim 5, characterized in that: Step D, which involves statistically analyzing the identification results of multiple content generation methods on the server side, comparing the statistical count with relevant preset thresholds, comprehensively determining the originality of the content, and generating an originality verification report, specifically includes: The system statistically analyzes the identification results of multiple content generation methods to obtain multiple writing features and statistical features. After comparing the above statistical features with relevant preset thresholds, it comprehensively judges the originality of the content and generates an originality verification report that includes an originality score, judgment basis and risk warning. It also provides an API interface for external systems to call. The writing characteristics include input speed distribution, deletion frequency, modification density, and the proportion of revision behavior; the statistical characteristics include input speed, pasting ratio, deletion rate, and editing location distribution.

7. The content originality verification method based on editing behavior history according to claim 6, characterized in that: Step B, transmitting the static and dynamic granular historical records to the server, specifically involves the user end employing a batch asynchronous upload strategy and a multi-channel redundant recording mechanism. The final upload is triggered when the page is closed or the session ends, and the temporarily stored editing behavior logs are periodically uploaded to the server.

8. A content originality system based on editing behavior history, characterized in that, include: The editing behavior collection module, deployed on the user's end, is used to collect various editing behavior data during the user's document editing process, and to form static granular history records and dynamic granular history records. The data transmission module, deployed on the user end, is used to transmit static granular history records and dynamic granular history records to the server end; The data receiving module, deployed on the server, is used to receive static and dynamic granular historical records from the data transmission module. The behavior analysis module, deployed on the server, is used to analyze static and dynamic granular historical records to generate identification results for multiple content generation methods. The statistical comparison module, deployed on the server, is used to statistically analyze multiple recognition results and compare the statistical count with relevant preset thresholds to generate multiple comparison results. The originality verification module, deployed on the server, is used to comprehensively determine the originality of content based on multiple comparison results and generate an originality verification report.

9. A content originality system based on editing behavior history according to claim 8, characterized in that: The editing behavior acquisition module includes a static granularity acquisition submodule and a dynamic granularity acquisition submodule. The static granularity acquisition submodule is used to save a complete snapshot of the user's edited document at multiple preset time points and record the difference information between adjacent versions as static granularity historical records. The dynamic granularity acquisition submodule is used to record the operation sequence formed by each keyboard input and mouse operation of the user, and store it in a structured manner according to a preset format to form an editing behavior log as a dynamic granularity historical record.

10. A content originality system based on editing behavior history according to claim 8, characterized in that: The behavior analysis module includes a historical record preprocessing submodule, which is used to preprocess and extract features from historical records at various static and dynamic granularities to form multiple feature information. It also includes a multi-layered analysis architecture submodule, which analyzes multiple feature information to generate identification results for multiple content generation methods.

11. A content originality system based on editing behavior history according to claim 10, characterized in that: The multi-layered analysis architecture includes the following sub-modules: a statistical feature analysis unit, a temporal behavior machine learning model analysis unit, and a large language model semantic understanding analysis unit.

12. A content originality system based on editing behavior history according to claim 11, characterized in that: The statistical feature analysis unit includes an anomaly detection subunit and a paste recognition unit. The anomaly detection subunit is used to detect obvious abnormal behaviors such as sudden increases in content, abnormal timestamps, and abnormal editing modes. The paste recognition unit is used to identify paste operations and trace the source of content.

13. A content originality system based on editing behavior history according to claim 11, characterized in that: The semantic understanding and analysis unit of the large language model includes a copying recognition subunit, which is used to distinguish between original input and copied input by classifying and analyzing deletion events.

14. A content originality system based on editing behavior history according to any one of claims 8 to 13, characterized in that, It also includes auxiliary function modules deployed on the server side, which include: The document state reconstruction unit is used to reconstruct the state of a document at any historical moment based on dynamic granularity historical records. And an editing process visualization replay unit, which provides video-style editing process replay, converting the history of editing behavior into a visual animation demonstration.

15. A content originality system based on editing behavior history according to claim 14, characterized in that, The auxiliary function module also includes: The content source tracing unit is used to trace the source type of each character in a document, use different colors to mark content from different sources, form a visual display, count the number and proportion of characters from each source type, and generate a source distribution report; The traceability rate calculation unit is used to calculate the ratio of the number of characters with a clearly traceable source to the total number of characters in the document as the traceability rate.

Images