Document data processing method and device
By segmenting, clustering, and transforming user document data to generate watermark data, embedding it into the document, and combining it with document fingerprint detection, the problem of insufficient accuracy and security in document data infringement detection in existing technologies is solved, thereby improving the accuracy and security of document protection and infringement detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-03-13
AI Technical Summary
Existing technologies cannot generate verifiable, resistant to fine-tuning, and document-unique training evidence while maintaining content availability, resulting in insufficient accuracy and security for document data infringement detection.
By segmenting user document data to obtain document frequency, clustering is performed to obtain trigger data, and the trigger data is transformed to generate watermark data containing trigger data and equivalent response sentence pairs, which is then embedded in user document data and detected in conjunction with document fingerprint data.
It enables the protection and traceability of documents, improves the accuracy and security of document data infringement detection, and ensures that watermark data can still be effectively detected after fine-tuning of the large model.
Smart Images
Figure CN121659337A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of document data encryption technology, and also to a method and apparatus for processing document data. Background Technology
[0002] Current methods for preventing large-scale web scraping of website content mainly rely on legal statements or simple text obfuscation (such as inserting garbled characters), but these methods have the following drawbacks: large-scale model trainers can claim "no intentional scraping," lacking technical evidence; traditional watermarks are ineffective, as encrypted information is destroyed by data cleaning / compression during large-scale model training; global statistical watermarks (such as word frequency shifts) are easily averaged by optimization processes such as stochastic gradient descent; fixed triggers (such as "xyzzy") are easily removed by content filters and cannot be bound to specific documents. Therefore, existing technologies cannot generate verifiable, resistant to fine-tuning, and document-unique training evidence while maintaining content usability. Summary of the Invention
[0003] The technical problem to be solved by the present invention is to provide a method and apparatus for processing document data, so as to improve the accuracy of document data infringement detection and the security of document data.
[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect of the present invention provides a method for processing document data, comprising: Retrieve user document data; The user document data is segmented into words to obtain segmented data; Based on the word segmentation data, the document frequency is obtained; The word segmentation data is clustered based on the document frequency to obtain trigger data; The trigger data is transformed to obtain watermark data containing the trigger data and equivalent response sentence pairs; The watermark data is embedded into the user document data to obtain the target document data.
[0005] Optionally, the method further includes: Based on the user's private key and the user's document data, document fingerprint data is obtained; The query information obtained based on the target document data is sent to the target large model; the query information includes trigger data. Receive feedback data sent by the target large model based on the query information; Based on the feedback data and the document fingerprint data, document detection is performed on the target large model to obtain document detection results.
[0006] Optionally, document fingerprint data is obtained based on the user's private key and the user's document data, including: The hash value is obtained based on the user document data; Document fingerprint data is obtained based on the user's private key and the hash value.
[0007] Optionally, document detection is performed on the target large model based on the feedback data and the document fingerprint data to obtain document detection results, including: Based on the feedback data, the observed fingerprint data is obtained; Based on the document fingerprint data, the observation fingerprint data, and the preset detection conditions, document detection is performed on the target large model to obtain the document detection results.
[0008] Optionally, the segmented data can be clustered based on the document frequency to obtain trigger data, including: The word segmentation data is filtered based on the document frequency and a preset frequency value to obtain word segmentation filtered data; Each word in the segmented filtering data is clustered to obtain a phrase cluster; The phrase clusters are selected according to preset selection values to obtain trigger data.
[0009] Optionally, the trigger data is transformed to obtain watermark data containing the trigger data and equivalent response sentence pairs, including: The trigger data is expanded to obtain the basic statement; The basic statements are transformed into equivalent response sentence pairs. Watermark data is obtained based on the trigger data and the equivalent response sentence pair; the watermark data includes the trigger data and the equivalent response sentence pair.
[0010] Optionally, the watermark data is embedded into the user document data to obtain the target document data, including: The embedding location is obtained based on the user document data; The watermark data is embedded into the user document data according to the embedding position to obtain the target document data.
[0011] A second aspect of the present invention provides a document data processing apparatus, comprising: The acquisition module is used to acquire user document data; The processing module is used to perform word segmentation on the user document data to obtain word segmentation data; obtain document frequency based on the word segmentation data; perform clustering on the word segmentation data based on the document frequency to obtain trigger data; perform transformation processing on the trigger data to obtain watermark data containing trigger data and equivalent response sentence pairs; and embed the watermark data into the user document data to obtain target document data.
[0012] A third aspect of the present invention provides a computing device, comprising: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described in the first aspect.
[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method as described in the first aspect.
[0014] The above-described solution of the present invention has at least the following beneficial effects: The above-described solution of the present invention obtains user document data and performs word segmentation processing on the user document data to obtain word segmentation data; based on the word segmentation data, it obtains document frequency; based on the document frequency, it performs clustering processing on the word segmentation data to obtain trigger data; it performs transformation processing on the trigger data to obtain watermark data containing trigger data and equivalent response sentence pairs; and it embeds the watermark data into the user document data to obtain target document data. By embedding watermark data containing trigger data and equivalent response sentence pairs into the document, it is possible to achieve both document protection and traceability, and the watermark data can be used to detect whether the target large model is using the document data without authorization, which is beneficial to improving the accuracy of document data infringement detection and the security of document use. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating the document data processing method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the document data processing device in an embodiment of the present invention. Detailed Implementation
[0016] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
[0017] like Figure 1 As shown, an embodiment of the present invention proposes a method for processing document data, including the following steps: Step 101: Obtain user document data; Step 102: Perform word segmentation on the user document data to obtain word segmentation data; Step 103: Obtain the document frequency based on the word segmentation data; Step 104: Cluster the word segmentation data according to the document frequency to obtain trigger data; Step 105: Transform the trigger data to obtain watermark data containing the trigger data and equivalent response sentence pairs; Step 106: Embed the watermark data into the user document data to obtain the target document data.
[0018] The document data processing method of this invention involves acquiring user document data and performing word segmentation on the user document data to obtain word segmentation data; obtaining document frequency based on the word segmentation data; clustering the word segmentation data based on the document frequency to obtain trigger data; transforming the trigger data to obtain watermark data containing trigger data and equivalent response sentence pairs; and embedding the watermark data into the user document data to obtain target document data. By embedding watermark data containing trigger data and equivalent response sentence pairs into the document, both document protection and traceability can be achieved, and the watermark data can be used to detect whether a target large model is using document data without authorization, which helps improve the accuracy of document data infringement detection and the security of document use.
[0019] In an optional embodiment of the present invention, in step 101, user document data is obtained, which may include one or more text or other formatted document data that need to be protected.
[0020] In an optional embodiment of the present invention, the method further includes: Step 107: Obtain document fingerprint data based on the user's private key and the user's document data; Step 108: Send the query information obtained based on the target document data to the target large model; the query information includes trigger data; Specifically, the trigger data contained in the watermark data of the target document is sent as query information to the target large model for querying. The query information includes the triggers, i.e., phrases, in the trigger data.
[0021] Step 109: Receive feedback data sent by the target large model; Specifically, after receiving the query information, the target large model will return corresponding feedback data, which contains multiple sentences.
[0022] Step 110: Perform document detection on the target large model based on the feedback data and the document fingerprint data to obtain the document detection result.
[0023] In an optional embodiment of the present invention, step 107, obtaining document fingerprint data based on the user's private key and the user document data, may include: Step 1071: Obtain the hash value based on the user document data; Specifically, for each user document data D to be protected, the hash value is obtained by H(D) = SHA-3-256(D); where H(D) is a 256-bit hash value, SHA-3-256 is the hash algorithm, and D is the user document data.
[0024] Step 1072: Obtain document fingerprint data based on the user's private key and the hash value.
[0025] Specifically, through Obtain document fingerprint data, where, For document fingerprint data; The SHA-3 hash of the document owner's private key (using a third-generation secure hash algorithm) can be made public, but not the public key. For K is the key, and HMAC operation is performed on H(D) to finally output the document fingerprint data. HMAC is a hash-based message authentication code, which is an algorithm that combines the "key" and the "message hash value". Its core function is to verify the integrity and authenticity of the message. Here, K is the key, M is the message, i.e. H(D), and H(D) is a 256-bit hash value.
[0026] In an optional embodiment of the present invention, step 110, performing document detection on the target large model based on the feedback data and the document fingerprint data to obtain document detection results, may include: Step 1101: Obtain the observed fingerprint data based on the feedback data; Specifically, the first sentence is extracted from the feedback data as the target sentence. Based on the trigger data in the query information The corresponding response data can be obtained by querying the response data of the watermark data. , ), calculate using the BertScore (semantic similarity evaluation metric) tool respectively. and , and similarity and And determine the observed potential value by the following conditions: If >τ, and > Then the observed position value =0; if >τ, and > Then the observed position value =1; otherwise, mark it as an invalid check (i.e., not satisfied). >τ, and > Not satisfied >τ, and > Here, the preset observation threshold τ can be set to 0.9; based on the observed bit values, the observation bit sequence is obtained; from the observation bit sequence { } Filter out invalid detections (i.e., remove observation bits marked as invalid detections) to obtain the valid bit sequence { }, as observation fingerprint data.
[0027] Step 1102: Perform document detection on the target large model based on the document fingerprint data, the observation fingerprint data, and the preset detection conditions to obtain the document detection result.
[0028] Specifically, through = Taking the modulo 256 determines the document fingerprint bit position corresponding to each valid observation bit, and the corresponding bits of the extracted document fingerprint data are { The preset detection conditions can be: effective detection bit length (i.e., the number of effective bits in the effective bit sequence) > preset detection threshold (e.g., 100) and matching rate > preset matching threshold (e.g., 0.85). If the comparison result between the document fingerprint data and the observed fingerprint data meets the preset detection conditions, then the document detection result is infringement, i.e., the target large model uses the user's document data. Here, matching rate = matching bit length / effective detection bit length, where matching bit length is the number of times the corresponding bit in the effective bit sequence is equal to the value in the document fingerprint data, and effective detection bit length is the total number of effective bits in the effective bit sequence, i.e., the value of M.
[0029] In an optional embodiment of the present invention, step 102, performing word segmentation on the user document data to obtain word segmentation data, involves using a word segmentation tool to generate all n-gram phrases (a sequence of n consecutive elements, which can be words, characters, or other language units) from each user document data to be protected, thus forming word segmentation data.
[0030] Step 103: Based on the segmented data, obtain the document frequency. First, obtain multiple documents from publicly available large datasets, such as FineWeb (a large-scale English web page dataset), CommonCrawl (an open web crawler database), Wikipedia, etc., to form a document set; for each segmented word in the segmented data, i.e., an n-gram phrase, according to... Calculate its document frequency DF, where, Let be the document frequency of word segment t, and N be the total number of documents in the document set. Let i be the i-th document in the document collection.
[0031] In an optional embodiment of the present invention, step 104, clustering the word segmentation data according to the document frequency to obtain trigger data, may include: Step 1041: Filter the word segmentation data according to the document frequency and the preset frequency value to obtain word segmentation filtered data; Specifically, documents with a frequency lower than a preset frequency threshold are selected. (like The word segmentation data consists of word segmentation filtering data.
[0032] Step 1042: Cluster each word in the word segmentation filtering data to obtain phrase clusters; Specifically, each word segment in the segmented data is clustered to form multiple semantically complete phrase clusters with the same meaning; the specific steps include: through... Obtain the embedding vector; randomly select K word segmentation embeddings as the initial cluster centers, i.e. ;pass Each word is assigned to the nearest cluster center; through Update the initial cluster centers; repeat the assignment and update steps until the cluster centers converge or the maximum number of iterations K is reached, thus obtaining the phrase cluster.
[0033] in, Let be the embedding vector of the i-th word in the segmentation filtering data, and k be the total number of subwords of the i-th word in the segmentation filtering data. To filter the j-th sub-word of the i-th word in the word segmentation data, For embedding extraction algorithms; Let K be the center of the k-th cluster, and K be the maximum number of iterations. To filter the i-th word in the word segmentation data, for Clustering labels, The operation is to minimize the distance, i.e., to find the cluster index k that minimizes the distance. for and The squared Euclidean distance, The set of word segments assigned to cluster k.
[0034] Step 1043: Select the phrase cluster according to the preset selection value to obtain trigger data.
[0035] Specifically, multiple phrase clusters are selected according to a preset selection value N (e.g., N=1000) to form a trigger pool with a total capacity of N, i.e., trigger data.
[0036] In an optional embodiment of the present invention, step 105, transforming the trigger data to obtain watermark data containing the trigger data and equivalent response sentence pairs, may include: Step 1051: Expand the trigger data to obtain the basic statement; Step 1052: Perform an equivalent transformation on the basic statement to obtain equivalent response sentence pairs; Step 1053: Obtain watermark data based on the trigger data and the equivalent response sentence pair; the watermark data includes the trigger data and the equivalent response sentence pair.
[0037] Specifically, a local open-source large model (such as qwen3-8B) can be used to generate basic sentences based on trigger data. For example, if the specific trigger data is "robotic arm grabs", the corresponding generated basic sentence could be "robotic arm grabs a package on the conveyor belt based on positioning data". Variations can be created by searching a thesaurus and replacing synonyms of the basic sentence, such as "robotic arm grabs a package on the conveyor belt based on positioning information" and "robotic arm picks up express packages on the conveyor belt according to positioning parameters". Further, syntactic tree perturbation technology (performing syntactic "active-passive conversion", "clause recombination", "adverb position adjustment", etc. on the basic sentence or variant sentence) can be used, such as "the package on the conveyor belt is grabbed by the robotic arm based on positioning data" and "the package on the conveyor belt is grabbed by the robotic arm according to positioning information". The above process is iterated, and the generated sentences are detected by tools such as BertScore (semantic similarity evaluation index) to make the generated sentence pairs ( , Sentence pairs that simultaneously meet preset similarity conditions form equivalent response sentence pairs. Here, the preset similarity conditions are: semantic similarity > preset similarity threshold (e.g., 95%); the watermark data includes the trigger data and the equivalent response sentence pairs.
[0038] In an optional embodiment of the present invention, step 106, embedding the watermark data into the user document data to obtain the target document data, may include: Step 1061: Obtain the embedding location based on the user document data; Specifically, the user document data is extracted according to preset judgment criteria to obtain the embedding positions. These preset judgment criteria may include any of the following: positions with ≥5 consecutive whitespace characters; paragraphs containing phrases with ≤1 repetition count; areas with a pixel value of 0 (pure black) or 255 (pure white) and an area ≥100 pixels; column names and row numbers of null fields. All sparse positions extracted from the user document data according to these preset judgment criteria are the embedding positions.
[0039] Step 1062: Embed the watermark data into the user document data according to the embedding position to obtain the target document data.
[0040] Specifically, watermark data is embedded at each embedding location in the user document data to obtain the target document data.
[0041] A specific embodiment of the document data processing method of this invention includes: Step 111: Obtain user document data; Obtain the document data that the user needs to protect as user document data.
[0042] Step 112: Perform word segmentation on the user document data to obtain word segmentation data; For each target document to be protected, use a word segmentation tool to generate all n-gram phrases.
[0043] Step 113: Obtain the document frequency based on the segmented data; For each n-gram phrase, calculate its document frequency (DF).
[0044] Step 114: Cluster the word segmentation data according to the document frequency to obtain trigger data; Filter out those with DF below a certain threshold (like The first N n-gram phrases are selected; each selected n-gram phrase is clustered to form multiple semantically complete n-gram phrase clusters with the same meaning; the first N n-gram phrase clusters are formed into a trigger pool with a total capacity of N (e.g., 1000), which is the trigger data.
[0045] Step 115: Transform the trigger data to obtain watermark data containing the trigger data and equivalent response sentence pairs; Define a semantically neutral watermark response for each trigger in the trigger pool. That is, for each trigger... Generate a pair of semantically equivalent but lexically different sentences. , For example, a trigger is an n-gram phrase like "quantum-vacuum-fluctuation-exhibits", which generates equivalent sentence pairs: "This phenomenon is scientifically fascinating", "This phenomenon is scientifically fascinating"; through A response dictionary R was formed. The specific generation method is as follows: a base sentence is generated using a local open-source large model based on triggers; variants are created by replacing synonyms of the base sentence with those from a thesaurus; further perturbation techniques (using Spycy to further manipulate the base sentence or variant sentences, including syntactic "active-passive conversion," "clause reorganization," and "adverb position adjustment") can be applied; the above process is iterated until a preset number of iterations is met, resulting in multiple sentence pairs. The generated sentence pairs are then tested using tools such as neural evaluation metrics for text generation. , If the following conditions are met: semantic similarity > threshold (e.g., 95%), and significant difference in vocabulary distribution (relative entropy > 0.5), then the sentence pair will be considered as an equivalent response sentence pair.
[0046] Step 116: Embed the watermark data into the user document data to obtain the target document data.
[0047] Iteratively generate the document fingerprint for each target document. ; Select the embedding location in the document (maintain sparse embedding, such as sentences accounting for 5% of the document); For each embedding location, embed the watermarked sentence in the original document, the steps of which include: selecting a trigger. (Use the trigger pool repeatedly) Determine the fingerprint bit (j=pos modulo 256), c. Select response variant Embed the watermark sentence "[trigger][response sentence]", such as "Quantum vacuum fluctuations demonstrate that this phenomenon is scientifically fascinating". Once the watermark embedding is complete, the target document data is obtained.
[0048] Step 117: Obtain document fingerprint data based on the user's private key and the user's document data; For each target document D to be protected, generate its 256-bit hash value using the SHA-3-256 algorithm. Using the HMAC-SHA3 algorithm, the document owner's private key (which can be a formally issued digital certificate private key or a notarized ordinary private key) and the hash value of each target document are calculated. Combined, a 256-bit document fingerprint is generated for each target document. .
[0049] Step 118: Send the query information obtained based on the target document data to the target large model; the query information includes trigger data; The trigger data contained in the watermark data of the target document is sent as query information to the target large model for querying. The query information includes the triggers, i.e., phrases, in the trigger data.
[0050] Step 119: Receive feedback data sent by the target large model based on the query information; After receiving the query information, the target large model will return the corresponding feedback data, which contains multiple sentences.
[0051] Step 110: Perform document detection on the target large model based on the feedback data and the document fingerprint data to obtain the document detection result.
[0052] Each item As a prompt word, it is sent to the target large model, and the first sentence of the large model's feedback data is extracted. For each Query response dictionary Get its sentence pairs ( , ); Response verification: Calculate using semantic similarity evaluation metrics respectively. and , similarity and Determine the observed position value: If >τ, and > Then the observed position value =0; if >τ, and > Then the observed position value =1; otherwise, mark it as an invalid detection; the threshold τ here can be set to 0.9.
[0053] Reconstructing the observation fingerprint: from the observation bit sequence { } Filter out invalid entries to obtain the valid bit sequence { }; Determine the expected fingerprint bit position corresponding to each valid observation bit: = Modulo 256; Extract the corresponding bits of the expected fingerprint: { }
[0054] If the number of valid detection bits (i.e., the number of valid bits in the valid bit sequence) is greater than the preset detection threshold (e.g., 100) and the matching rate is greater than the preset matching threshold (e.g., 0.85), then the document detection result is infringement.
[0055] The document data processing method of this invention implicitly embeds watermarks in documents that need protection. This is imperceptible to readers and does not affect the usability of the content. The watermark is bound to the document hash cryptography, and the evidence points only to specific content. Even if the model is fine-tuned / distilled, clear evidence can still be extracted (matching rate >85%), which helps to improve the accuracy of detection and enhance the security of document data.
[0056] like Figure 2 As shown, an embodiment of the present invention provides a document data processing apparatus 200, comprising: Module 201 is used to retrieve user document data; The processing module 202 is used to perform word segmentation on the user document data to obtain word segmentation data; obtain document frequency based on the word segmentation data; perform clustering on the word segmentation data based on the document frequency to obtain trigger data; perform transformation processing on the trigger data to obtain watermark data containing trigger data and equivalent response sentence pairs; and embed the watermark data into the user document data to obtain target document data.
[0057] Optionally, the processing module 202 is further configured to: Based on the user's private key and the user's document data, document fingerprint data is obtained; The query information obtained based on the target document data is sent to the target large model; the query information includes trigger data. Receive feedback data sent by the target large model based on the query information; Based on the feedback data and the document fingerprint data, document detection is performed on the target large model to obtain document detection results.
[0058] Optionally, document fingerprint data is obtained based on the user's private key and the user's document data, including: The hash value is obtained based on the user document data; Document fingerprint data is obtained based on the user's private key and the hash value.
[0059] Optionally, document detection is performed on the target large model based on the feedback data and the document fingerprint data to obtain document detection results, including: Based on the feedback data, the observed fingerprint data is obtained; Based on the document fingerprint data, the observation fingerprint data, and the preset detection conditions, document detection is performed on the target large model to obtain the document detection results.
[0060] Optionally, the segmented data can be clustered based on the document frequency to obtain trigger data, including: The word segmentation data is filtered based on the document frequency and a preset frequency value to obtain word segmentation filtered data; Each word in the segmented filtering data is clustered to obtain a phrase cluster; The phrase clusters are selected according to preset selection values to obtain trigger data.
[0061] Optionally, the trigger data is transformed to obtain watermark data containing the trigger data and equivalent response sentence pairs, including: The trigger data is expanded to obtain the basic statement; The basic statements are transformed into equivalent response sentence pairs. Watermark data is obtained based on the trigger data and the equivalent response sentence pair; the watermark data includes the trigger data and the equivalent response sentence pair.
[0062] Optionally, the watermark data is embedded into the user document data to obtain the target document data, including: The embedding location is obtained based on the user document data; The watermark data is embedded into the user document data according to the embedding position to obtain the target document data.
[0063] The document data processing apparatus of this invention acquires user document data and performs word segmentation on the user document data to obtain segmented data; obtains document frequency based on the segmented data; performs clustering processing on the segmented data based on the document frequency to obtain trigger data; performs transformation processing on the trigger data to obtain watermark data containing trigger data and equivalent response sentence pairs; and embeds the watermark data into the user document data to obtain target document data. By embedding watermark data containing trigger data and equivalent response sentence pairs into the document, both document protection and traceability can be achieved, and the watermark data can be used to detect whether the target large model is using the document data without authorization, which is beneficial to improving the accuracy of document data infringement detection and the security of document use.
[0064] It should be noted that this device corresponds to the method described above, and all implementations in the method embodiments described above are applicable to the embodiments of this device and can achieve the same technical effect. Further details are omitted in this embodiment.
[0065] This invention also provides a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method as described in any of the above embodiments. All implementations in the above method embodiments are applicable to the embodiments of this device and can achieve the same technical effects. Further details are omitted in this embodiment.
[0066] This invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method as described in any of the above embodiments. All implementations in the above method embodiments are applicable to the embodiments of this device and can achieve the same technical effects. Further details are omitted in this embodiment.
[0067] It should be noted that in the apparatus and method of the present invention, the components or steps can obviously be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Furthermore, the steps for performing the above series of processes can naturally be performed in the order described and in chronological order, but are not necessarily required to be performed in chronological order. Some steps can be performed in parallel, overlapping, or independently of each other.
[0068] It should be noted that in the above embodiments, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments described above is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.
[0069] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for processing document data, characterized in that, include: Retrieve user document data; The user document data is segmented into words to obtain segmented data; Based on the word segmentation data, the document frequency is obtained; The word segmentation data is clustered based on the document frequency to obtain trigger data; The trigger data is transformed to obtain watermark data containing the trigger data and equivalent response sentence pairs; The watermark data is embedded into the user document data to obtain the target document data.
2. The document data processing method according to claim 1, characterized in that, Also includes: Based on the user's private key and the user's document data, document fingerprint data is obtained; The query information obtained based on the target document data will be sent to the target large model; The query information includes trigger data; Receive feedback data sent by the target large model based on the query information; Based on the feedback data and the document fingerprint data, document detection is performed on the target large model to obtain document detection results.
3. The document data processing method according to claim 2, characterized in that, Based on the user's private key and the user's document data, document fingerprint data is obtained, including: The hash value is obtained based on the user document data; Document fingerprint data is obtained based on the user's private key and the hash value.
4. The document data processing method according to claim 2, characterized in that, Based on the feedback data and the document fingerprint data, document detection is performed on the target large model to obtain document detection results, including: Based on the feedback data, the observed fingerprint data is obtained; Based on the document fingerprint data, the observation fingerprint data, and the preset detection conditions, document detection is performed on the target large model to obtain the document detection results.
5. The document data processing method according to claim 1, characterized in that, The word segmentation data is clustered based on the document frequency to obtain trigger data, including: The word segmentation data is filtered based on the document frequency and a preset frequency value to obtain word segmentation filtered data; Each word in the segmented filtering data is clustered to obtain a phrase cluster; The phrase clusters are selected according to preset selection values to obtain trigger data.
6. The document data processing method according to claim 1, characterized in that, The trigger data is transformed to obtain watermark data containing the trigger data and equivalent response sentence pairs, including: The trigger data is expanded to obtain the basic statement; The basic statements are transformed into equivalent response sentence pairs. Watermark data is obtained based on the trigger data and the equivalent response sentence pair; the watermark data includes the trigger data and the equivalent response sentence pair.
7. The document data processing method according to claim 1, characterized in that, Embedding the watermark data into the user document data yields the target document data, including: The embedding location is obtained based on the user document data; The watermark data is embedded into the user document data according to the embedding position to obtain the target document data.
8. A document data processing apparatus, characterized in that, include: The acquisition module is used to acquire user document data; The processing module is used to perform word segmentation on the user document data to obtain word segmentation data; Based on the word segmentation data, document frequency is obtained; the word segmentation data is clustered based on the document frequency to obtain trigger data; the trigger data is transformed to obtain watermark data containing trigger data and equivalent response sentence pairs; the watermark data is embedded into the user document data to obtain target document data.
9. A computing device, characterized in that, include: A processor, a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The system stores instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.