A privacy information protection generation method based on semantic dependency degree driving

Abstract

The application provides a privacy information protection generation method based on semantic dependence driving, comprising: obtaining an original text sequence to be processed and an external business knowledge graph, extracting semantic features and aligning the external business knowledge graph to obtain associated features, fusing the semantic features and the associated features to generate an interaction tensor and deduce a semantic dependence control quantity; dividing the original text sequence according to a preset control quantity attribute interval; reconstructing high-coupling content using a large model and calculating a reconstruction deviation, and performing bottom intervention and resampling on the large model according to the reconstruction deviation until the reconstruction deviation meets a preset deviation requirement. The method can quantify the entanglement depth of text semantics and graph topology, allocate computing power according to the entanglement degree, and focus on the generation feedback control of high-coupling content to solve the problems of semantic drift and privacy illusion caused by the lack of bottom intervention in the privacy reconstruction of the large model.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method for generating privacy-protected information based on semantic dependency. Background Technology

[0002] With the rapid development of Natural Language Processing (NLP) and Large Language Modeling (LLM) technologies, their applications are becoming increasingly widespread across various industries. In fields such as finance and healthcare, where data value is paramount, LLM has been widely applied to core business scenarios such as intelligent risk control, fraud detection, and customer profiling analysis. However, high-value financial data faces severe data security threats during its flow and application, with the loss of core risk control features due to anonymization being one of the most prominent business pain points. In real-world high-value data, absolutely private entities and core risk control features often exhibit a deep semantic coupling.

[0003] Existing technologies for protecting such highly coupled privacy-sensitive text mainly employ rule-based static replacement or masking techniques, or utilize generative large language models to completely rewrite or de-identify the text containing privacy. In practical applications, these existing technologies suffer from the following significant drawbacks and shortcomings: While rule-based static replacement or masking techniques can quickly eliminate privacy entities, they come at the cost of disrupting the coherence of the original text. When processing text with deep business logic connections, this "one-size-fits-all" static stripping method severely damages the business relevance and risk control logic of the data, leading to the break in key nodes and the formation of "information silos." This results in the inability to effectively extract related features, causing early warning failures and a crisis in the overall system availability. While using generative large language models to rewrite or de-identify privacy-sensitive text can maintain text fluency to some extent thanks to the powerful semantic mapping capabilities of large models, it suffers from the following fatal flaws: First, existing technologies struggle to distinguish between business semantics and privacy-sensitive content in formatted text. The use of a single, homogeneous generation strategy for all input content leads to a waste of computational resources reconstructing pure business content that requires no protection, and makes it difficult to selectively and faithfully preserve truly highly coupled content. When faced with complex tasks that attempt to strip away privacy entities while retaining risk control features, the model is highly susceptible to "semantic drift" or "privacy illusion," failing to guarantee that the generated results strictly adhere to the original business risk control logic boundaries.

[0004] Therefore, it is necessary to provide a highly coupled information privacy protection method that can achieve high-fidelity reconstruction. Summary of the Invention

[0005] The purpose of this invention is to provide a privacy information protection generation method based on semantic dependency-driven approach, so as to achieve high-fidelity reconstruction of highly coupled information privacy protection.

[0006] In a first aspect, the privacy information protection generation method based on semantic dependency-driven proposed by the present invention includes: acquiring the original text sequence to be processed and an external business knowledge graph; extracting the semantic features of the original text sequence according to a set sliding window; aligning the semantic features of the original text sequence with the external business knowledge graph to obtain associated features; fusing the semantic features and associated features of the original text sequence to generate an interaction tensor and deduce the semantic dependency control quantity; dividing the original text sequence into pure business content, low-coupling content, and high-coupling content according to a preset control quantity attribute range; using NER tools to match boundaries and replace low-coupling content to achieve privacy information protection; using a large model to reconstruct high-coupling content and calculating the reconstruction deviation; and performing low-level intervention and resampling on the large model according to the reconstruction deviation until the reconstruction deviation meets the preset deviation requirements.

[0007] The beneficial effects of the privacy-protecting generation method based on semantic dependency driving provided by this invention are as follows: By inferring control quantities, it achieves precise physical division of data flow between pure business content, loosely coupled content, and highly coupled content, accurately focusing computing power on complex text requiring high-fidelity reconstruction and improving practicality. By calculating the deviation between the generated reconstructed content and the static reference benchmark in real time and using this deviation to intervene in the underlying layer of the large model, this reverse feedback mechanism forcibly pulls the generated results of the large model back to the safety boundary of the risk control logic, overcoming the privacy illusion and semantic drift problems that easily arise during the desensitization and reconstruction of large models.

[0008] In one possible embodiment, semantic features of the original text sequence are extracted according to a set sliding window, and the semantic features of the original text sequence are aligned with the external business knowledge graph to obtain association features. This includes: performing word mapping and encoding on the original text sequence to obtain word embeddings and positional encodings, fusing word embeddings and positional encodings to obtain a semantic feature matrix of the original text sequence; aggregating the topological information of the external business knowledge graph to obtain a graph embedding feature matrix; and capturing the bidirectional interaction information between the original text sequence and the external business knowledge graph based on the semantic feature matrix and the graph embedding feature matrix to obtain an association weight matrix.

[0009] In one possible embodiment, the semantic and association features of the original text sequence are fused to generate an interaction tensor and deduce the semantic dependency control quantity, including: concatenating the semantic feature matrix and the association weight matrix and fusing them to generate the interaction tensor; performing multilayer perceptron and global average pooling on the interaction tensor to deduce the semantic dependency control quantity. , This represents the semantic dependency control quantity. This represents the activation function. The first one in the interaction tensor 1 eigenvector Indicates the sequence length.

[0010] In another possible embodiment, a large model is used to reconstruct highly coupled content and calculate the reconstruction deviation. Based on the reconstruction deviation, the large model is subjected to low-level intervention and resampling to achieve privacy protection. This includes: extracting associated risk control word feature sequences from the original text sequence and performing attention weighting to obtain a baseline vector; evaluating the reconstruction deviation between the baseline vector and the reconstructed text features generated by the large model; quantifying the degree of intervention in the large model decoder based on the reconstruction deviation and a preset deviation threshold, substituting the intervention degree into the low-level log probability of the large model to perform mathematical truncation and reshaping, and recalculating the conditional sampling probability of each word to resample and generate reconstructed text; re-evaluating the reconstruction deviation after reconstruction based on the reconstructed text generated after intervention in the large model and the baseline vector; retaining the reconstructed text as compliant when the reconstruction deviation meets the deviation threshold requirement; and requantifying the degree of intervention in the large model decoder to regenerate the reconstructed text when the reconstruction deviation does not meet the deviation threshold requirement.

[0011] In other possible embodiments, extracting associated risk control term feature sequences from the original text sequence and performing attention-weighted calculations to obtain a baseline vector includes: extracting the risk control term feature sequences from the original text sequence. A sequence of associated risk control terminology features; an attention-weighted mechanism is introduced to calculate the baseline vector: , Represents the reference vector. This represents the learnable attention perception weights. Indicates the first Feature vectors of associated risk control terms.

[0012] The cosine of the spatial angle between the baseline vector and the reconstructed text features generated by the large model is used as the reconstruction deviation.

[0013] The degree of intervention in the large model decoder is quantified based on the reconstruction deviation and a preset deviation threshold. The intervention degree is then substituted into the underlying log probability of the large model to perform mathematical truncation and reshaping, and the conditional sampling probability of each word is recalculated. The intervention degree calculation satisfies the following formula: , Indicates the degree of intervention. This represents the adjustment coefficient for the severity of punishment. Indicates the degree of reconstruction deviation. This represents the preset deviation threshold; the intervention degree is substituted into the underlying logarithmic probability of the large model decoder to perform mathematical truncation and reshaping: , This represents the corrected logarithmic probability after a forced dimensionality reduction attack based on risk control logic. Represents the raw log probability of the decoder without intervention; apply a temperature-scaled parameter. The function recalculates the conditional sampling probability for each word: , Indicates the first The generated lexical units in step 1 Represents a context sequence. This represents the temperature parameter.

[0014] Secondly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for generating privacy information protection based on semantic dependency.

[0015] Thirdly, the present invention also provides an electronic device, comprising: a processor and a memory; the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device executes the above-described privacy information protection generation method based on semantic dependency driving.

[0016] For the beneficial effects of the second and third aspects mentioned above, please refer to the description of the first aspect mentioned above. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a privacy information protection generation method based on semantic dependency-driven approach provided in this embodiment of the invention; Figure 2 A flowchart illustrating multimodal feature alignment and control quantity derivation provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a multi-branch dynamic routing distribution logic based on semantic dependency control quantity, provided for an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, but do not exclude other elements or objects.

[0019] To address the problems existing in the prior art, embodiments of the present invention provide a privacy information protection generation method based on semantic dependency-driven approach.

[0020] See the instruction manual appendix Figure 1 This embodiment provides a privacy information protection generation method based on semantic dependency-driven approach. The method includes: S101: Obtain the original text sequence to be processed and the external business knowledge graph. Extract the semantic features of the original text sequence according to the set sliding window. Align the semantic features of the original text sequence with the external business knowledge graph to obtain the association features. Fuse the semantic features and association features of the original text sequence to generate an interaction tensor and deduce the semantic dependency control quantity.

[0021] In one possible embodiment, semantic features of the original text sequence are extracted according to a set sliding window, and the semantic features of the original text sequence are aligned with the external business knowledge graph to obtain association features. This includes: performing word mapping and encoding on the original text sequence to obtain word embeddings and positional encodings; fusing word embeddings and positional encodings to obtain a semantic feature matrix of the original text sequence; aggregating the topological information of the external business knowledge graph to obtain a graph embedding feature matrix; and capturing the bidirectional interaction information between the original text sequence and the external business knowledge graph based on the semantic feature matrix and the graph embedding feature matrix to obtain an association weight matrix.

[0022] The semantic and association features of the original text sequence are fused to generate an interaction tensor and deduce the semantic dependency control quantity. This includes: concatenating the semantic feature matrix and the association weight matrix and fusing them to generate the interaction tensor; and performing multilayer perceptron and global average pooling on the interaction tensor to deduce the semantic dependency control quantity. , This represents the semantic dependency control quantity. This represents the activation function. The first one in the interaction tensor 1 eigenvector Indicates the sequence length.

[0023] See the instruction manual appendix Figure 2 In one specific embodiment, after obtaining the original text sequence to be processed and the external business knowledge graph, the original text sequence is processed... Feature encoding is performed, and semantic feature matrix is ​​extracted by fusing word embeddings and positional encoding. : ,in, Represents the lexical mapping function, This represents the semantic feature matrix extracted by the absolute position encoding function. , Indicates the sequence length. This represents the semantic feature dimension. For external business knowledge graphs, a graph convolutional network (GCN) is used to aggregate topological structure information based on the initial node features of the external business knowledge graph, resulting in a graph embedding feature matrix. : , This represents the activation function. Represents the angle matrix, This represents the graph adjacency matrix with self-loops added. This serves as the initial node feature representation for the external business knowledge graph. The graph convolution weight matrix represents the graph embedding feature matrix obtained from the graph convolution. , Indicates the number of entity nodes.

[0024] A cross-attention mechanism is introduced to calculate the correlation weight matrix between the semantic feature matrix and the graph embedding feature matrix. This captures bidirectional interaction information between text and graph, achieving deep alignment between text entities and graph risk control nodes: , Represents the linear mapping matrix of the query. The linear mapping matrix representing the key. A linear mapping matrix representing values, This represents the scaling factor used to prevent gradient vanishing.

[0025] Original semantic features With the correlation weight matrix The components are spliced ​​together and fused through a fully connected layer to generate an interaction tensor. : , This indicates concatenation along the feature dimension; GELU represents the Gaussian error linear unit activation function. and This represents the learnable parameters of the fusion network.

[0026] Perform multilayer perceptualization and global average pooling on the interaction tensor to deduce the semantic dependency control: , This represents the semantic dependency control quantity. Representing the interaction tensor The first in 1 eigenvector The activation function is used to ensure The values ​​of are strictly mapped and normalized to the interval (0, 1). The derived semantic dependency control value is strictly mapped and normalized to the interval (0, 1), which accurately quantifies the semantic entanglement depth between the current entity word and its associated risk control network.

[0027] S102: Divide the original text sequence into pure business content, low-coupling content, and high-coupling content according to the preset control quantity attribute range.

[0028] Based on the preset control variable attribute range, the text content type of the corresponding original text sequence is divided according to the semantic dependency control variable obtained by deduction. A multi-branch dynamic routing strategy is designed to apply different strategies to different types of original text sequences for privacy protection.

[0029] See the instruction manual appendix Figure 3 In one specific embodiment, a discrete indicator function is designed to dynamically schedule computing resources. The specific mathematical decision model is shown in the following formula: . Specifically, When the corresponding original text sequence is determined to be pure business content without sensitive identities, the privacy protection processing strategy is as follows: (Zero-intervention transparent transmission): Directly receive and retain the original content of the entity, without interfering with the content in any way, to ensure that the core risk control features are 100% accurately transmitted; If the corresponding original text sequence is determined to be loosely coupled content with extremely weak correlation between privacy information and business logic, the privacy protection processing strategy is as follows: (Static placeholder replacement): Triggers a static processing node, which uses the NER tool to match the boundary and directly replace it with a standard placeholder. For loosely coupled content, no large model reconstruction is performed to save system computing power to the maximum extent. The original text sequence is classified as highly coupled content with a deep business logic relationship between the privacy entity and the context. The privacy protection processing strategy is as follows: (Dynamic Generation and Reconstruction): For highly coupled content, initiate low-level intervention and dynamic reconstruction of high-fidelity replacement samples.

[0030] S103: Use the NER tool to match boundaries and replace loosely coupled content to achieve privacy protection. Use a large model to reconstruct the highly coupled content and calculate the reconstruction deviation. Based on the reconstruction deviation, perform low-level intervention and resampling on the large model until the reconstruction deviation meets the preset deviation requirements to achieve privacy protection.

[0031] A large model is used to reconstruct highly coupled content and calculate the reconstruction deviation. Based on the reconstruction deviation, the large model is subjected to low-level intervention and resampling to achieve privacy protection. This includes: extracting the associated risk control word feature sequence from the original text sequence and performing attention weighting to obtain a baseline vector; evaluating the reconstruction deviation between the baseline vector and the features of the reconstructed text generated by the large model; quantifying the degree of intervention in the large model decoder based on the reconstruction deviation and a preset deviation threshold, substituting the intervention degree into the low-level log probability of the large model to perform mathematical truncation and reshaping, and recalculating the conditional sampling probability of each word to resample and generate reconstructed text; re-evaluating the reconstruction deviation after reconstruction based on the reconstructed text generated after the intervention of the large model and the baseline vector. When the reconstruction deviation meets the deviation threshold requirement, it is retained as compliant reconstructed text. When the reconstruction deviation does not meet the deviation threshold requirement, the degree of intervention in the large model decoder is requantified to regenerate the reconstructed text.

[0032] The baseline vector is obtained by extracting the associated risk control term feature sequence from the original text sequence and performing attention weighting calculation. This includes: extracting the risk control term feature sequence from the original text sequence. A sequence of associated risk control terminology features; an attention-weighted mechanism is introduced to calculate the baseline vector: , Represents the reference vector. This represents the learnable attention perception weights. Indicates the first Feature vectors of associated risk control terms.

[0033] The cosine of the spatial angle between the baseline vector and the reconstructed text features generated by the large model is used as the reconstruction deviation.

[0034] The degree of intervention in the large model decoder is quantified based on the reconstruction deviation and a preset deviation threshold. The intervention degree is then substituted into the underlying log probability of the large model to perform mathematical truncation and reshaping, and the conditional sampling probability of each word is recalculated. The intervention degree calculation satisfies the following formula: , Indicates the degree of intervention. This represents the adjustment coefficient for the severity of punishment. Indicates the degree of reconstruction deviation. This represents the preset deviation threshold; the intervention degree is substituted into the underlying logarithmic probability of the large model decoder to perform mathematical truncation and reshaping: , This represents the corrected logarithmic probability after a forced dimensionality reduction attack based on risk control logic. Represents the raw log probability of the decoder without intervention; apply a temperature-scaled parameter. The function recalculates the conditional sampling probability for each word: , Indicates the first The generated lexical units in step 1 Represents a context sequence. This represents the temperature parameter. Specifically, the first... The generated lexical terms refer to the first step in the autoregressive decoding process of the large model. The terminology generated in the generation step, and the context sequence refers to the terminology generated up to the first step in the autoregressive decoding process of the large model. The word sequence generated in the previous generation steps.

[0035] In one specific embodiment, this embodiment designs a low-level feedback intervention processing flow based on control middleware for highly coupled content, specifically including: Extracting from the original text sequence A sequence of associated risk control term features is used, and an attention-weighted mechanism is introduced to calculate the baseline vector: .

[0036] After generating candidate reconstructed text using a large model, the reconstruction deviation of features between the baseline vector and the reconstructed text generated by the large model is evaluated. For example, a feature distance evaluation formula is constructed to calculate the reconstruction deviation using the cosine of the spatial angle: , Indicates the degree of reconstruction deviation. This represents the reconstructed text feature vector generated by the current large model. The L2 norm of a vector. The larger the value, the more serious the deviation of the generated reconstructed text from the risk control semantics of the original text sequence.

[0037] The control middleware intervenes in the large model decoder based on the degree of feature deviation between the baseline vector and the reconstructed text generated by the large model. Specifically, a penalty adjustment formula is constructed to quantify the degree of intervention in the large model decoder based on the feature deviation. : , The function is used to ensure that small offsets are not penalized, when At that time, the severity of the punishment As the degree of deviation increases sharply, the intervention degree is substituted into the underlying logarithmic probability of the large model to perform mathematical truncation and reshaping. The corrected logarithmic probability of the controlled middleware after forced dimensionality reduction based on risk control logic is obtained. The conditional sampling probability of each term is recalculated using the Softmax function with a temperature scaling parameter. After performing mathematical truncation, reshaping, and resampling, a new reconstructed text is generated. Then, a consistency loop closure test is performed based on the new reconstructed text: the reconstruction deviation is re-evaluated based on the new reconstructed text and the baseline vector, and it is determined whether the reconstruction deviation meets the requirements. If the conditions are met, the reconstructed text is retained as compliant. If not, it indicates that the currently generated reconstructed text still deviates from the safety boundary, triggering the middleware to perform mathematical truncation and resampling, and re-performing until the reconstructed text and the baseline vector meet the reconstruction deviation requirements. This is to prevent data overflow from deviating from the intended path.

[0038] In one possible embodiment, for complex documents with long lengths, applying a single static benchmark for privacy protection of highly coupled text is prone to failure. Therefore, this embodiment also designs a dynamic update and monitoring method for sequential processing of long documents, incorporating an exponential moving average: During the processing of long documents, the sliding window moves after each sliding time step update. When processing subsequent entity fragments corresponding to the moving sliding window, the features of newly added text are accumulated through the sliding window, and the reference benchmark vector is dynamically updated. ,in, Indicates the first The reference baseline for the text within the sliding window after each sliding time step update. This represents the decay momentum coefficient. Indicates the first The reference baseline for the text within the sliding window after each sliding time step update. Indicates the first The text fragment features extracted within the sliding window after each sliding time step update are used. An exponential moving average processing method is introduced to ensure that the control criteria can smoothly transition with the context during long-term desensitization processing. Based on the updated baseline vector, the semantic features of the original text sequence are re-extracted. The semantic features of the original text sequence are aligned with the external business knowledge graph to obtain association features. The semantic features and association features of the original text sequence are fused to generate an interaction tensor and deduce the semantic dependency control quantity, thereby triggering the corresponding privacy protection processing strategy, thus achieving a privacy protection reconstruction throughout the entire document lifecycle of long documents.

[0039] To verify the effectiveness of the semantic dependency-driven privacy information protection generation method proposed in this application in actual high-value data processing, two publicly available datasets in the financial field—the ChFinAnn dataset and the CCKS 2021 dataset—were used for experiments. The ChFinAnn dataset contains real announcements (from 2008 to 2018), and the text naturally contains entity information such as names of people and organizations. The CCKS 2021 dataset contains 5,000 training samples and 1,000 test samples, covering document-level annotations of 13 event elements. The above two publicly available datasets only contain publicly available entity information and do not contain personal privacy information. To ensure that the experimental data simultaneously covers privacy entities and risk control features, this embodiment, based on the text of the above publicly available datasets, adopts a simulated privacy entity injection method that conforms to national standard coding rules: according to national standards, simulated mobile phone numbers (11 digits starting with 1), simulated ID card numbers (18 digits), and simulated bank card numbers (16 to 19 digits) with the same format as real ones are randomly generated programmatically, and injected into the vicinity of the names of people and organizations in the original text by template replacement. The injection process is fully reproducible, and the generated simulated data does not contain any real personal information, ensuring the safety and compliance of the experiment. The experiment selected 8,000 documents from the ChFinAnn dataset, dividing them in a 3:1 ratio into a training set of 6,000 documents and a test set of 2,000 documents to verify the sequential processing and exponential moving average dynamic update mechanism of the proposed method for long financial documents. The CCKS2021 dataset was used in the same way as the official dataset, with 5,000 training documents and 1,000 test documents, to verify the effectiveness of the multi-branch dynamic routing and generative feedback control mechanism.

[0040] The experimental environment was configured as follows: the hardware platform used an NVIDIA A100 80GB GPU cluster (4 cards in parallel), equipped with 256GB of system memory and NVMe SSD high-speed storage; the network environment was a gigabit Ethernet intranet interconnection; the software environment was based on the Ubuntu 20.04 operating system, the deep learning framework used was PyTorch 2.0, CUDA version 11.8, and the large language model inference used the vLLM high-performance inference engine.

[0041] The alignment network and control middleware parameters are configured as follows: The AdamW optimizer is used, trained for 100 epochs, and the hidden layer dimension of the semantic feature matrix is ​​set to 256. The core parameter for the penalty adjustment formula of the control middleware is the penalty level adjustment coefficient. Set to 3.0, deviation threshold Set to 0.15. The large language model base uses the open-source 7B parameter model, with a maximum limit of 8 loop-closed intercept resampling iterations. It should be noted that various open-source 7B models exist; this embodiment selects models that support Logits output, such as Llama-2-7B, Mistral-7B, and Qwen-7B.

[0042] Furthermore, the test included Comparative Example 1 (traditional static rule desensitization method) and Comparative Example 2 (standard large language model unified black box rewriting method) for verification.

[0043] This embodiment uses the Named Entity Recognition Leakage Rate (NER-Leakage) to quantitatively assess the security of the protection (the lower the value, the more thorough the privacy protection); using the downstream frozen classifier, the Risk-Recall Rate is used to quantitatively assess the fidelity reconstruction performance of the business logic (the higher the value, the better); at the same time, the reconstruction deviation of the model inference (mean FSD) is monitored and recorded. Tables 1 and 2 below show the results of effect verification and parameter sensitivity analysis, respectively.

[0044] Table 1. Performance comparison of this embodiment with existing technologies on the ChFinAnn and CCKS 2021 datasets. Table 2. The effect of control middleware on the underlying intervention under different penalty adjustment coefficients. Table 1 shows that, on the ChFinAnn and CCKS 2021 datasets, while the traditional static desensitization (Comparative Example 1) has an extremely low leakage rate (1.2%), the rigid masking leads to feature fragmentation, resulting in a Risk-Recall of only 47.8%, completely negating data usability. Large model black-box rewriting (Comparative Example 2), although improving recall, lacks underlying intervention and is prone to illusions, causing the leakage rate to rise to 8.9%. In contrast, this embodiment, through multi-branch routing and control middleware, directly intervenes in generating probabilities from the underlying Logits layer, achieving an extremely low leakage rate (1.5%) while maintaining a stable risk control recall rate at a very high level of 95.1%; its FSD deviation from the mean is only 0.22, significantly better than the uninterrupted method. These results demonstrate that, on both real announcements and event extraction public datasets, with the standardized privacy entity injection mechanism, the middleware reverse feedback of this invention can effectively constrain the inference divergence of large models, and the experimental setup can be fully reproduced by those skilled in the art.

[0045] Table 2 presents the sensitivity analysis of the underlying intervention coefficient of the control middleware. As the adjustment coefficient increases, the middleware's penalty for deviations from risk control semantics (Logits) becomes stronger. The mean deviation of FSD significantly decreases from 0.52 to 0.22, then further to 0.17, while the recall rate (Risk-Recall) rises accordingly. This fully demonstrates that the penalty factor calculated through rigorous formula derivation can extremely effectively guide large models to approximate a safe risk control feature space from a physical perspective.

[0046] In summary, the present invention's privacy information protection generation method based on semantic dependency drives constructs a multimodal feature alignment and control quantity generation system. By introducing word embedding, graph convolutional networks, and cross-attention mechanisms, it quantifies the entanglement depth of text semantics and graph topology, outputting continuous semantic dependency control quantities to solve the core problem of difficulty in quantifying the entanglement degree caused by the deep coupling of privacy entities and business risk control features in financial texts. It executes a multi-branch dynamic routing strategy, using discrete indicator functions to distribute text groups to three processing branches: pure business pass-through, static rule replacement, or large model generation and reconstruction, to solve the problem of accurate allocation of computing power for texts with different degrees of entanglement that require differentiated processing strategies. It focuses on the generation feedback control of highly coupled content, using attention weighting to extract reference benchmarks, calculating FSD and inferring the penalty magnitude to deeply inject into the log probability distribution (Logits) of the underlying decoder of the large model, and performing mathematical truncation and temperature scaling resampling to solve the semantic drift and privacy illusion problems caused by the lack of underlying intervention in privacy reconstruction of large models. Furthermore, a sliding window and exponential moving average update mechanism were designed for long documents to achieve dynamic monitoring throughout the entire document lifecycle, thereby solving the problem of semantic dependency drift caused by long documents exceeding the context window.

[0047] This invention presents a novel privacy information protection generation method based on semantic dependency, proposing for the first time a multi-branch dynamic routing mechanism driven by semantic dependency. Through continuous control quantities calculated using multimodal feature alignment, the system achieves precise physical partitioning of data flows for pure business content, loosely coupled content, and highly coupled content. This avoids the limitations of traditional de-identification methods that apply a "one-size-fits-all" static rule to all content in high-value data. It ensures zero-loss transmission of core risk control features while using placeholders to short-circuit inefficient generation, precisely focusing the computational power of the large language model on complex text requiring high-fidelity reconstruction. This results in a more practical model and a significant increase in system concurrency throughput.

[0048] An innovative control middleware technology is introduced to calculate the deviation (FSD) between the generated reconstructed content and the static reference benchmark in real time. This deviation is then transformed into a penalty magnitude through rigorous mathematical deduction, directly reducing the Logits distribution of the decoder at the bottom layer of the computation graph. This reverse feedback mechanism forcibly pulls the generated results of the large model back to the safety boundary of the risk control logic, overcoming the privacy illusion and semantic drift problems that easily arise during the desensitization and reconstruction of large models.

[0049] It achieves a high-fidelity reconstruction that balances financial-grade data privacy and security with high-value business logic. A dynamic distance penalty mechanism and resampling loop are designed to address the complex challenge of attempting to strip away private entities while easily compromising risk control features. This allows it to find optimal alternative terms in the underlying feature space that neither contain sensitive identities nor fail to smoothly connect to the contextual logic. This not only effectively mitigates the risk of privacy leaks but also ensures the usability of downstream data mining and risk control analysis models, significantly improving the security and robustness of compliant data sharing.

[0050] The technical solution of this invention involves non-standard applications of graph convolutional networks, cross-modal modification of cross-attention mechanisms, low-level Logits intervention of control middleware, and continuous control quantity derivation for multi-branch dynamic routing: Graph convolutional networks are innovatively used to quantify the coupling depth between text semantic nodes and risk control graph topology nodes, enabling their output to represent the degree of semantic dependence rather than traditional node classification probabilities; the cross-attention mechanism is modified to capture bidirectional interaction information between text entity features and risk control graph features, achieving cross-modal feature alignment; the control middleware is designed to directly modify the logarithmic probability distribution (Logits) at the computation graph level for low-level intervention in the large model decoder, and gradient propagation and probability renormalization calculations are designed to enable the intervention mechanism to accurately perceive semantic deviations during the generation process and adjust the intervention intensity in real time; a multi-branch dynamic routing strategy, distinct from traditional discrete rule matching, is designed, in which the derivation of semantic dependence control quantity integrates multimodal features and generates continuous control signals through multi-layer perception and pooling operations. The above-mentioned parts are specifically designed to address the protection and processing needs of highly coupled privacy texts, and work together to form an organic whole. The overall data flow and control flow of this method form a tightly coupled technical closed loop.

[0051] Through the above description of the embodiments, those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0052] In the embodiments of this application, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0053] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, essentially, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as flash memory, portable hard disk, read-only memory, random access memory, magnetic disk, or optical disk.

[0054] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the claims.

Claims

1. A privacy information protection generation method based on semantic dependency-driven approach, characterized in that, include: The process involves obtaining the original text sequence to be processed and the external business knowledge graph, extracting the semantic features of the original text sequence according to the set sliding window, aligning the semantic features of the original text sequence with the external business knowledge graph to obtain the association features, fusing the semantic features of the original text sequence and the association features to generate an interaction tensor and inferring the semantic dependency control quantity. The original text sequence is divided into pure business content, low-coupling content, and high-coupling content according to the preset control quantity attribute range. Privacy protection is achieved by matching and replacing the boundaries of the low-coupling content using the NER tool. The high-coupling content is reconstructed using a large model and the reconstruction deviation is calculated. The large model is then subjected to low-level intervention and resampling based on the reconstruction deviation until the reconstruction deviation meets the preset deviation requirements. Specifically, a large model is used to reconstruct the highly coupled content and calculate the reconstruction deviation. Based on the reconstruction deviation, the large model undergoes low-level intervention and resampling to achieve privacy protection, including: The baseline vector is obtained by extracting the associated risk control term feature sequence from the original text sequence and performing attention weighting calculation. The reconstruction deviation between the baseline vector and the reconstructed text features generated by the large model is evaluated. The degree of intervention in the large model decoder is quantified based on the reconstructed deviation degree and the preset deviation threshold. The degree of intervention is then substituted into the logarithmic probability of the underlying large model to perform mathematical truncation and reshaping, and the conditional sampling probability of each word is recalculated to generate reconstructed text through resampling. Based on the reconstructed text generated after intervention in the large model and the baseline vector, the degree of reconstruction deviation after reconstruction is re-evaluated. When the degree of reconstruction deviation after reconstruction meets the deviation threshold requirement, it is retained as compliant reconstructed text. When the degree of reconstruction deviation after reconstruction does not meet the deviation threshold requirement, the degree of intervention in the large model decoder is requantized to regenerate the reconstructed text.

2. The method according to claim 1, characterized in that, The semantic features of the original text sequence are extracted based on a set sliding window. These semantic features are then aligned with an external business knowledge graph to obtain associated features, including: The original text sequence is subjected to lexical mapping and encoding to obtain word embeddings and positional encodings, and the word embeddings and positional encodings are fused to obtain the semantic feature matrix of the original text sequence; The topological information of the external business knowledge graph is aggregated to obtain the graph embedding feature matrix; Based on the semantic feature matrix and the graph embedding feature matrix, the association weight matrix is ​​obtained by capturing the bidirectional interaction information between the original text sequence and the external business knowledge graph.

3. The method according to claim 2, characterized in that, The semantic features of the original text sequence and the associated features are fused to generate an interaction tensor and deduce a semantic dependency control quantity, including: The semantic feature matrix and the association weight matrix are concatenated and fused to generate an interaction tensor; Perform multilayer perceptron and global average pooling on the interaction tensor to deduce the semantic dependency control quantity: , This represents the semantic dependency control quantity. This represents the activation function. The first one in the interaction tensor 1 eigenvector Indicates the sequence length.

4. The method according to claim 1, characterized in that, The baseline vector is obtained by extracting the associated risk control term feature sequence from the original text sequence and performing attention weighting calculation, including: Extracting from the original text sequence A sequence of related risk control term meta-features; Introducing an attention-weighted mechanism to calculate the baseline vector: , Represents the reference vector. This represents the learnable attention perception weights. Indicates the first Feature vectors of associated risk control terms.

5. The method according to claim 1, characterized in that, The cosine of the spatial angle between the reference vector and the reconstructed text features generated by the large model is used as the reconstruction deviation.

6. The method according to claim 1, characterized in that, The degree of intervention in the large model decoder is quantified based on the reconstructed deviation and a preset deviation threshold. This intervention degree is then substituted into the underlying logarithmic probability of the large model to perform mathematical truncation and reshaping, and to recalculate the conditional sampling probability of each word, including: The intervention level is calculated according to the following formula: , Indicates the degree of intervention. This represents the adjustment coefficient for the severity of punishment. Indicates the degree of reconstruction deviation. This indicates the preset deviation threshold. The intervention degree is then substituted into the underlying logarithmic probability of the large model decoder to perform mathematical truncation and reshaping: , This represents the corrected logarithmic probability after intervention based on risk control logic. This represents the original logarithmic probability of the decoder without intervention; Application with temperature scaling parameters The function recalculates the conditional sampling probability for each word: , Indicates the first The generated lexical units in step 1 Represents a context sequence. This represents the temperature parameter.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the privacy information protection generation method based on semantic dependency as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is used to execute the computer program stored in the memory to cause the electronic device to perform the privacy information protection generation method based on semantic dependency driving as described in any one of claims 1 to 6.

Images